Welcome to
the Forum

Software development has matured over the years but fundamentally still comes down to beating your head against the wall while figuring out the art of giving directions to a machine that, at its foundation, “speaks” in on and off switches. The computers powering NASA in the race to the moon required punch cards that were literal stacks of paper fed into a machine, where the presence or absence of a hole represented a code translated into an electrical signal. A deck of punch cards represented a program. Later, programmers became experts in binary code and firmware until the introduction of high-level language compilers. Ironically, compilers and increasingly higher levels of programming languages led some to predict the demise of the art of software programming itself.

Since December 2025, there has been a lot of marketing and PR from AI model companies that "software engineering is dead" because of their internal experience in consuming their own AI tools for writing software. While researching that statement, I confirmed that the same AI model companies continue to post healthy numbers of job openings on their career sites for software engineering roles. How can both of these things be true? I believe the answer can be found going back into the history of computing, where many dire predictions were made about the disappearance of engineering and management jobs - yet today those professions continue to evolve and grow.  

The User Interface – Open Claw

We were overdue for a major update in how humans interact with machines, given the maturity of web-based design. Punch cards gave way to compilers and command lines, then to the GUI and web interfaces, and later to mobile. Each progression was revolutionary compared to how interaction worked in the prior generation. For example, web design and search engine optimization were refactored once mobile phone interfaces and app stores began to dominate customer experiences.

The original GenAI chatbots were seamless because web interfaces and mobile computers were already ubiquitous. But they weren’t the breakthrough that brought us closer to the dream of harnessing computing to help with nagging, annoying tasks. Interfacing with and harnessing the computer continued to be the nagging, annoying task, requiring proficiency in rote learning and grinding through lines of code.

Apple has a long history of shaking up the computing industry through breakthroughs in how we interact with devices—from the first Apple computer with a graphical user interface to the iPod, iPhone, and beyond. Open Claw represents a breakthrough because operating it does not depend on the old modes of human–computer interaction that chatbots required. There are many downsides related to security and vulnerabilities, but that isn’t the point. For meaningful differentiation in software solutions, finding ways to interact with users the way they want—rather than the way software developers envision—will be sustainable differentiation.

There are opportunities for an Open Claw–like corporate offering that includes lifecycle management, auditing, and governance-as-code, with built-in FinOps to ensure there are no unpleasant budgeting surprises from token usage.

Mechanics vs. Meaningful Differentiation

For the last 15 years, discussion forums like Reddit and Blind have been packed with prospective developers looking for the recipe for the fastest path into one of the “FAANG” companies. How many Big Tech parents pushed their kids into coding and robotics clubs to set them up for a career in SaaS? The goal often seems to be: get in, suck it up, work at a relentless pace, watch your stock go up, cash out, and move to the next rung.

As long as companies needed armies of developers to work in a production line—one that LeetCode could prepare you for—the assembly line from code camp to university to long-term employment at a large software company ran smoothly. GenAI and coding tools have been rattling these markets, and I’ve lost count of the social media influencers who have preached, “You won’t be replaced by AI; you’ll be replaced by someone using it,” especially to software developers over the last three years.

To those who believe that learning the mechanics of AI coding tools faster than the other 90% of equally panicked peers is the path to safety—as if it’s another “ace interviews at Google” course—here’s the reality: your use of AI tools is not a sustainable differentiator between long-term employment and becoming a bartender. CEOs—many of whose admins still print emails for them to read—may tell you that using AI tools is the goal and that you should start working with them or else. Tool usage and leverage are not sustainable differentiators, even if they serve as a short-term sifting mechanism.

Speed in the mechanics or methods of producing software is now a baseline, but it does nothing to explain why that work is worth doing in the first place. Software development has a long history of ever-higher levels of abstraction that demand more compute power but deliver better scale, faster feature development, and improved debugging. The increased efficiency of how humans direct machines isn’t interesting to anyone outside the tech industry unless it materially improves business outcomes for customers.

Yes, learn AI coding tools—they are the next level of abstraction in a long history of attempts to relieve developer frustration from grinding out lines of code. But for your own sake, learn them while working on problems that matter, and do it in a maintainable, enterprise-ready way.

Think Like an Owner - Make/ Buy/ Differentiate

There are so many AI tools and applications competing for business and there are equal numbers of non-developers working with coding tools. For software engineers who have ideas but no experience doing market research, there are now ready made research support systems from multiple vendors to help vet and verify the market opportunity for your idea. The fact that you do not have to pitch the idea through management to ensure a 100 person coding team gets assigned to develop the first version is both freedom and frightening. You are no longer held back by business analysis - just your decision on what to make or buy and how you take it to market. For non technical business owners who have AI products pitched to you nearly round the clock, focus on where you want your business to be unique and where you want to offload mechanics that are for you a chore or cost that is a necessity.  

While there are many decision frameworks around when to make or buy AI, AI agents, AI management platforms and AI applications, very few of the frameworks look at how to balance ease of use with long term supplier oversight and management. The decision is not just a technology skill question but needs to focus on whether your supplier will change access to their technology or whether the supplier has to change pricing models because their board or shareholders demand margin improvements. Talking to small and large enterprises, I find that the focus is on technology sustainability for mid to large companies who have access to technical skills - but perhaps more focus is needed on FinOps for AI where autonomous agents are being scaled. At smaller companies I advise, I find that the decisions start from a financial lens but could use structured thinking around long term differentiation - where does the owner or leader need to own or retain a key capability to set their business apart in the minds of their customers.  

The next few articles in the series will walk through at a deeper level some of the decision tradeoffs, strategic questions and frameworks I have used to drive clarity for major decisions customers and partners have faced.

‍

Ishween Kaur, Generative AI Lead at Salesforce, shares practical lessons on building multilingual AI systems, closing language gaps, and scaling with trust and real-world feedback.

Enterprise AI conversations still revolve around models. Benchmarks, context windows, and release cycles dominate the discussion.

But inside production environments, the real shift is happening elsewhere.

The competitive advantage in enterprise AI is moving away from model selection and toward the runtime architecture that surrounds it.

Once AI leaves the demo environment and enters core workflows, it must be governed, monitored, cost-controlled, and made resilient. That is not a model problem. It is an operational one.

From Pipelines to Context Engines

Traditional enterprise systems were built around deterministic pipelines. Data moved from source to warehouse to dashboard. Outputs were reproducible. Monitoring focused on throughput and uptime.

AI systems depend on something different: dynamic context.

Large language models (LLMs) rely on embeddings, retrieval layers, policy documents, customer history, and transactional signals that are often refreshed on tight cycles. When that context degrades, output quality degrades.

In production deployments, many perceived model quality issues are actually context failures:

• Stale embeddings
• Incomplete retrieval scopes
• Delayed ingestion cycles
• Missing domain ownership

The architecture therefore shifts from static data pipelines to context engines. These systems are designed around freshness service level agreements (SLAs), versioned vector stores, and controlled retrieval boundaries.

The model generates the answer. The context determines its reliability.

From Access Control to Behavioral Guardrails

Earlier governance models focused on who could access a system. AI introduces a different challenge: how the system behaves once accessed.

Outputs are probabilistic. Prompts vary. Users experiment. Sensitive data can surface in unexpected ways.

Modern AI runtimes increasingly include:

• Prompt filtering layers
• Output moderation services
• Personally identifiable information (PII) redaction pipelines
• Policy-aware orchestration logic
• Token ceilings by business unit

Consider a financial services copilot assisting relationship managers. A user asks for a client summary. The model has access to customer relationship management (CRM) notes, transaction data, and compliance documentation.

Without behavioral guardrails, the system could:

• Surface internal risk ratings not meant for disclosure
• Generate speculative language about client suitability
• Exceed approved advisory language

The runtime must intercept, evaluate, and shape responses before delivery.

Governance is no longer static access enforcement.  It becomes active runtime mediation.

From Predictable Infrastructure Spend to Usage Volatility

Traditional infrastructure cost planning followed relatively stable patterns. Workloads were forecastable. Capacity was provisioned accordingly.

AI inference introduces volatility.

Token consumption fluctuates based on prompt size. Adoption spikes increase inference load. Copilots embedded into daily workflows can quietly multiply request volumes.

Enterprises are responding by embedding cost awareness into the runtime layer:

• Model routing, using lightweight models for simple queries and premium models for complex reasoning
• Response caching strategies
• Tiered inference placement, internal endpoints versus external application programming interfaces (APIs)
• Real-time token dashboards
• Cost attribution per workflow

Without runtime-level cost instrumentation, AI initiatives can scale faster than financial oversight.

In this environment, cost architecture is structural.

From System Health to Decision Health

Traditional observability focuses on infrastructure metrics such as central processing unit (CPU) utilization, memory pressure, and latency.

AI systems require something more nuanced. A model can respond quickly and consistently while producing degraded or risky decisions. Decision health expands observability into the quality and impact of AI outputs.

In practice, this includes monitoring:

• Output drift over time
• Confidence scoring thresholds
• Hallucination frequency patterns
• Fallback invocation rates
• User correction frequency
• Escalation triggers to human review

If an AI assistant’s recommendations are increasingly overridden by users, the system may be technically healthy but operationally degrading. A rise in fallback activations may signal retrieval gaps or tightening policy enforcement.

AI systems are not just infrastructure components. They are decision amplifiers. Observability must reflect that reality.

A Production Pattern Is Emerging

Across industries, a common architecture is taking shape beneath enterprise AI deployments. Organizations are building structured data backbones, versioned embedding layers, policy-aware orchestration engines, guardrail services that mediate outputs, integrated cost monitoring tied directly to request routing, and decision-level observability with full audit trails.

The visible model can change.  The runtime scaffolding persists.

Over time, the reliability, governance posture, and economic efficiency of that runtime determine whether AI systems become trusted infrastructure or remain isolated pilots.

The Real Differentiator

Models will continue to evolve. Benchmarks will continue to shift.

Inside enterprise environments, models are becoming modular components.

The durable advantage lies in the architecture that governs them. The runtime manages context freshness, enforces policy, instruments cost, and monitors decision integrity.

Enterprise AI is no longer just a capability layer. It is an operational layer.

As with every operational layer before it, organizations that engineer it with discipline, not enthusiasm, will outperform those that treat it as novelty.

The future of enterprise AI will not be defined by who selects the best model.

It will be defined by who builds the most resilient system around it.

At VAST Forward 2026, VAST Data expanded the view of its ambitions for its AI operating system. The company dropped five major announcements in a single day: new agentic AI capabilities, a fully accelerated hardware stack built with NVIDIA, a global infrastructure control plane, a video intelligence partnership, and a formalized partner ecosystem. Taken together, they represent a vision of what enterprise AI infrastructure looks like when it stops being a collection of components and starts behaving like a unified, intelligent system.

Building the Thinking Machine: PolicyEngine and TuningEngine

The marquee announcement was the unveiling of two new capabilities of the VAST AI Operating System, VAST Data PolicyEngine and VAST Data TuningEngine. These two new services, slated for release by end of 2026, are designed to work in tandem inside the VAST DataEngine to create what the company calls a “thinking machine” — a system that doesn’t just execute AI pipelines but governs them, evaluates them, and improves on them automatically.

PolicyEngine functions as an inline enforcement layer for agentic workflows, applying fine-grained, tamper-proof controls on what agents can access, what tools they can invoke, and how they communicate with other agents. TuningEngine captures outcomes from those workflows and feeds them into fine-tuning pipelines using methods like LoRA, supervised fine tuning, and reinforcement learning, automatically generating candidate models for evaluation and deployment.

The result is a closed operational loop: observe, act, evaluate, improve. VAST co-founder Jeff Denworth framed it plainly: “Just as people are always learning, so should tomorrow’s applications.” For enterprises trying to deploy AI in regulated or mission-critical environments, the combination of zero-trust governance and automated model improvement in a single platform is a major step toward trusted, autonomous systems.

NVIDIA Collaboration Puts Acceleration Inside the Stack

The software story gets hardware muscle through an expanding collaboration with NVIDIA with an end-to-end, fully CUDA-accelerated AI data stack. The companies introduced CNode-X, a new class of NVIDIA-Certified servers that run the VAST AI Operating System directly on GPU-powered infrastructure. The deeper integration embeds NVIDIA libraries—cuVS for vector search, cuDF-based SQL acceleration via an engine called Sirius, and support for NVIDIA’s Context Memory Storage (CMX) platform—directly into VAST’s core data services.

The goal is to eliminate the fragmented stack problem: separate storage, database, and AI compute tiers that slow enterprise AI pipelines from pilot to production and add operational complexity at every seam. In doing so, it clears the path for agentic AI-enabled workloads to fulfill their promised potential. “CNode-X is CUDA-accelerated at every layer to give AI agents persistent memory so they can work on complex problems over days or weeks, and eventually years, without forgetting—opening the world to the next frontier of AI,” said NVIDIA Founder and CEO Jensen Huang.

CNode-X will come to market through OEM partners, including Cisco and Supermicro, giving enterprises a path to GPU-accelerated VAST infrastructure through vendors they already buy from.

Polaris Solves the Multi-Cloud Management Mess

VAST also announced Polaris, a global control plane purpose-built to provision, operate, and orchestrate distributed AI infrastructure across public cloud, neocloud, and on-premises environments. As AI pipelines span regions and providers between data collection, training, and inference, Polaris offers a centralized service delivery that converts disparate infrastructure instances into one operational environment.

Polaris is built on a Kubernetes-based architecture with a lightweight agent on every VAST node and operates as an intent-driven management layer. Administrators define the desired state of infrastructure, and Polaris coordinates the cloud-native services and VAST software to get there and keep it there. It supports cloud service provider partners, sovereign deployments, and multi-site, multi-cluster configurations under centralized management. It is available as part of VAST cloud deployments.

Video Intelligence Gets a Sovereign Deployment Path with TwelveLabs

VAST’s announced partnership with TwelveLabs, which develops video foundation models, introduced a partnership to help organizations extend video intelligence beyond public cloud deployments. Through the collaboration, the companies will support demand for deploying advanced video intelligence closer to where the data originates and is governed. The pitch is squarely at enterprises and government agencies sitting on massive video archives that can’t or won’t push data to a hyperscaler: media companies, financial services firms running surveillance-based fraud detection, and public sector agencies where data sovereignty is non-negotiable. TwelveLabs gains a deployment path into on-prem and neocloud environments, and VAST gains a compelling vertical use case to anchor its platform story.

Cosmos Gets a Real Partner Program

Finally, VAST formalized its Cosmos partner ecosystem into a unified global partner program, consolidating resellers, system integrators, independent software vendors, cloud providers, and advisory partners under a single framework with structured onboarding, tiered benefits, deal registration, and a centralized partner portal. Cosmos offers a clear engagement model for each partner type, from hardware platform partners validating architectures to consulting firms running deployment practices. H2O.ai and WWT are among the early participants.

The TechArena Take

Today’s announcements represent the most concrete evidence yet that VAST is building out its “thinking machine” vision a coherent, layered way, and is doing so with the right partners. The strategic logic is sound: if AI pipelines are becoming continuous, always-on systems, then the infrastructure layer needs to behave like an operating system—governing, learning, and adapting in real time. PolicyEngine, TuningEngine, and Polaris are all designed to elevate the company’s AI OS position.  

Meanwhile, its work with TwelveLabs, the formalization of its partner ecosystem, and the CNode-X collaboration with NVIDIA demonstrate that VAST is assembling a coalition to maximize its reach and impact. Each partnership extends VAST’s reach into a different part of the enterprise buying process, including technical validation, vertical use cases, and channel distribution. Together, they suggest a company that understands the AI OS can’t succeed alone.

Brookfield Asset Management has officially launched Radiant, a vertically integrated AI infrastructure company, through its acquisition of Ori Industries, a UK-based distributed AI cloud provider. The announcement marks the transition of Radiant from concept to active operations and signals a significant bet in the AI infrastructure space.

Radiant enters the market as one of the first bets from Brookfield’s AI Infrastructure Fund (BAIIF), which itself serves as the anchor for a broader $100 billion investment program. Financial terms of the Ori acquisition were not disclosed.

What Radiant Actually Is

As it launches, Radiant is targeting the delivery of high-performance, purpose-built AI compute to sovereign governments, telecom providers, and select large enterprises under long-term contracts.

Radiant will expand on Ori’s existing integrated AI Cloud assets, which operate out of more than 20 data centers around the world. The new infrastructure will be built on the NVIDIA DSX reference design, NVIDIA’s blueprint for what it calls AI factories, making Radiant an NVIDIA Cloud Partner. That designation matters because NVIDIA’s DSX architecture, which is designed to be Vera Rubin–ready, provides a standardized, scalable foundation for high-throughput AI workloads.

Alongside the long-term contract business, Radiant will continue to operate the Ori Global AI Cloud, a GPU-as-a-Service platform Ori built over seven years, for customers that need on-demand capacity and rapid deployment.

Brookfield’s head of AI infrastructure, Sikander Rashid, described the model plainly: “I think of it as a leasing business.” Radiant structures contracts to lock in revenue across the estimated five-year useful life of a chip cluster, with investment-grade customers committed to pay regardless of utilization. Brookfield has been explicit that it will not be taking on technology obsolescence risk in this model.

Energy as a First Principle

One differentiator Radiant is emphasizing is vertical integration that extends all the way to energy. AI compute is an energy consumption problem, and Brookfield’s existing portfolio includes power utilities and renewable generation assets.

Radiant’s ability to pair data center operations directly with behind-the-meter power generation represents a structural cost and reliability advantage. In a blog published with Radiant’s launch, Head of Product João Coelho noted, “Our behind-the-meter model co-locates AI Factories directly with massive-scale hydro, wind, solar, or nuclear generation. This is not an optimization of the datacenter; it is a re-architecture of the entire supply chain.”

Sovereignty as a Business Model

A growing number of national governments require that AI workloads processed on their behalf remain within their borders, and Radiant is explicitly positioning itself to meet that demand. Its sovereign framework is designed to go beyond simply deploying compute in-country.

As Radiant CMO Jonathan Symonds said, “Compute sovereignty depends on ownership of the supply chain: land, power, and capital.” Radiant plans to address all of these elements.

In terms of the technology, that means air-gapped control planes, hardware-rooted security, and single-tenant bare metal configurations that ensure sensitive datasets and model weights remain invisible to external parties, including Radiant’s own engineers. Open-weight model architectures are supported to reduce dependence on any single vendor’s proprietary stack.

The capital structure reinforces the sovereign pitch. Radiant argues that most AI infrastructure has been financed with short-term, high-cost capital—venture equity or private credit carrying hurdle rates around 20%—which creates incentive structures poorly suited to the stable, long-duration assets that national AI programs require. By financing at infrastructure-grade rates of approximately 5%, Radiant contends it can offer sovereigns compute offtake contracts of 3, 5, or 10 years with predictable, hedgeable pricing.

The TechArena Take

The launch of Radiant is a meaningful development in AI infrastructure. Sovereign governments and large enterprises increasingly want AI compute that is predictable in cost, physically located in-country, and not delivered by a US hyperscaler carrying geopolitical and data residency complications.  

Radiant is designed precisely for that buyer. With long-term contract structures, NVIDIA-validated architecture, and the energy integration to underpin reliable operations, Brookfield has assembled a credible stack.

The risk, as with any infrastructure-scale bet, lies in execution timing. AI chip generations turn over quickly, demand patterns from sovereign customers are still maturing, and $100 billion programs are easier to announce than to deploy. Brookfield’s contractual approach—locking customers into full payment regardless of utilization—reduces its downside but will require winning and maintaining the confidence of investment-grade counterparties in a competitive market.

Still, the resource moat is significant. Very few organizations can bring Brookfield’s combination of infrastructure experience, energy assets, and long-duration capital to bear on a compute leasing business. If the sovereign AI infrastructure market develops the way Brookfield is betting it will, Radiant will be well-positioned. Technology leaders evaluating AI infrastructure partnerships should pay close attention to what Radiant builds in its first 18 months of operation. As of today, the proof of concept is underway.

Back in November 2024, I wrote a blog which discussed the implications of cybersecurity in automotive. In that blog, I outlined some of the efforts and frameworks that have been established to address this evolving field as it applied to automotive while highlighting the ISO 21434 cybersecurity framework.  

With the advent of quantum computing, which promises to bring to deliver far greater computing performance than previously imagined, existing cybersecurity solutions are expected to be at risk once Cryptographically Relevant Quantum Computer (CRQC) become available. While a CRQC is not widely expected before 2029, the importance of recognizing and addressing this new form of cybersecurity attack today cannot be overstated. As an update to my previous blog, it seemed appropriate to explore the potential impact of Post Quantum Cryptography (PQC) on next-generation vehicles and the approaches and considerations that must be taken to address this looming threat.

From Connected Cars to Cryptographic Risk

While automotive OEMs race to deploy Software Defined Vehicles (SDV) with sophisticated connectivity, artificial intelligence (AI), and Over-the-Air (OTA) update capabilities, quantum computing, once confined to research laboratories, is rapidly approaching commercial viability, and with it comes the harsh reality of cryptographic obsolescence. Post Quantum Cryptography (PQC) represents the industry's attempt to address this looming threat, yet its implementation presents challenges as complex as the vehicles it aims to protect. For an industry already grappling with ISO 21434 compliance and the expanding attack surface of connected cars, the transition to quantum resistant security cannot be an afterthought.

PQC refers to a new generation of cryptographic algorithms that were designed to withstand attacks from both classical and quantum computers. Public key systems currently in use such as RSA and ECC (Elliptic Curve Cryptography), which have been effective to date, are based upon the premise that the available computing resources were insufficient to crack these codes.  With the arrival of CRQC, today’s public key systems are expected to be readily cracked, fully compromising today’s security infrastructure.

Unlike current public key systems such as RSA and ECC which rely on mathematical problems that quantum computers can solve exponentially faster, PQC algorithms are built upon hard mathematical problems that are believed to remain difficult to solve even for quantum systems. These include lattice-based cryptography, hash-based signatures, and multivariate polynomial equations. The National Institute of Standards and Technology (NIST) has begun standardizing these algorithms, with CRYSTALS-Kyber for key encapsulation and CRYSTALS-Dilithium for digital signatures emerging as primary candidates. As a quick explanation, the key exchange mechanism can be considered a “secret encoder and decoder ring” that is used to encrypt and decrypt the message, whereas digital signature ensures that the encrypted messages are coming from a trusted, known source.  For automotive systems, the transition to PQC isn't merely an upgrade, it's a fundamental architectural transformation.

Harvest Now, Decrypt Later: Why Timing Matters

The urgency to address this looming problem stems from a phenomenon security professionals call “harvest now, decrypt later.” Adversaries with access to quantum capabilities in the future could capture encrypted vehicle communications today and store them for decryption once quantum supremacy arrives. Given that vehicles remain operational for 15 to 20 years, data transmitted through V2X communications, OTA update channels, and telematics systems in 2024 could be vulnerable to retrospective decryption in 2035. The automotive supply chain compounds this risk; cryptographic vulnerabilities in Tier 2 or Tier 3 components may not manifest until years after deployment, creating liability exposure that extends across decades and multiple ownership transfers.

The attack surface for quantum-enabled threats mirrors and magnifies existing automotive vulnerabilities. Consider the Common Exposure Library identified in current threat modeling: WiFi, cellular connections, Bluetooth, TPMS (tire pressure monitoring systems), OBD-II (on board diagnostic) ports, USB interfaces, EV charging infrastructure, and V2X communications. Each of these vectors currently relies on cryptographic protocols that quantum computers will eventually compromise. V2X communications, particularly, present an acute concern; these systems depend on low-latency cryptographic handshakes between vehicles and infrastructure to prevent collisions and coordinate traffic flow. The computational overhead of PQC algorithms, often requiring larger key sizes and more processing cycles, threatens to introduce latency that could degrade safety-critical response times.

OTA Updates and Signature Trust

Over-the-Air updates, already identified as high-risk vectors for malware injection, face compound quantum threats. The digital signatures that authenticate OTA packages historically have employed traditional ECDSA or RSA schemes which will be vulnerable to quantum attacks. A malicious actor with future quantum capabilities could forge signatures for malicious firmware updates, effectively weaponizing the vehicle’s own maintenance infrastructure. The Software-Defined Vehicle architecture, with its billion lines of code and continuous update cycles, requires cryptographic agility—the ability to rotate algorithms without hardware replacement. Yet current automotive ECUs were designed with static cryptographic implementations, often burned into hardware security modules with decade-long lifecycles.

The intersection of Functional Safety (ISO 26262) and cybersecurity (ISO 21434) will become particularly challenging in the PQC transition. Safety-critical systems such as steering control, braking, and ADAS depend on predictable, high-speed timing and deterministic behavior. Many PQC candidates, while mathematically robust, exhibit variable, and extended execution times or impose system level requirements that could violate existing safety elements. Lattice-based algorithms, for instance, because of their inherent extended computational needs, can require memory allocations that may trigger watchdog timers or interfere with real-time operating systems.  Additionally, the extended computational time associated with these calculations may lead to exceeding the FTTI (Fault Tolerant Time Interval) of the vehicle, which is effectively the deadline that the system must beat once a fault is detected to prevent an accident. Furthermore, the threat agent risk assessment (TARA) processes mandated by ISO 21434 must now incorporate quantum-capable adversaries—nation-state actors with access to cryptanalytic quantum resources or organized criminal groups leasing quantum computing time through cloud services. In short, a lot of complexity now must be added to address critical threats.

The Hardware Reality Inside the Vehicle

Hardware constraints in general present formidable barriers to PQC deployment. Automotive microcontrollers, selected for cost efficiency and environmental resilience rather than computational headroom, often lack the memory and processing capabilities to execute post-quantum algorithms efficiently. A typical vehicle contains dozens of ECUs ranging from 32-bit microcontrollers with kilobytes of RAM to sophisticated infotainment processors and highly complex ADAS SoCs. Retrofitting PQC across this heterogeneous landscape requires either hardware replacement, which is prohibitively expensive (not viable) for vehicles already in service, or careful algorithm selection that balances security margins against resource constraints. Hybrid approaches, combining classical and post-quantum algorithms during transition periods, effectively double the cryptographic overhead.

Supply chain complexity amplifies these challenges. Automotive components source semiconductors from global foundries, incorporate software from hundreds of vendors, and integrate cryptographic modules from specialized providers. Coordinating a PQC transition requires synchronization across this ecosystem; OEMs must specify quantum-resistant requirements, chip vendors must implement hardware acceleration for lattice operations, and software suppliers must refactor cryptographic libraries. The “should strongly consider” language of ISO 21434, while providing flexibility, may prove insufficient to drive the coordinated industry response that PQC demands. Unlike the Jeep Cherokee vulnerability, which prompted immediate patches, the quantum threat offers no dramatic demonstration—only mathematical certainty of future compromise.

Data Privacy: The Archive Problem

The data privacy implications extend beyond vehicle control into the personal information ecosystem now embedded in modern automobiles. Biometric authentication data, payment credentials for EV charging, location histories, and occupant behavior patterns encrypted with current standards may persist in vehicle storage, cloud backups, and third-party databases for decades. Quantum-enabled decryption of this archive would expose not just current owners but entire household networks connected through vehicle telematics. A Nov. 2024 podcast that I participated in provided some real insights into the EV charging infrastructure security that are also very relevant and perhaps overlooked. Charging network operators must simultaneously protect real-time transaction integrity and ensure that historical charging patterns remain confidential against future quantum analysis.

Designing for Crypto-Agility Now

Addressing these challenges requires immediate architectural decisions with long-term consequences. Automotive cybersecurity teams must begin crypto-agility engineering; designing systems where cryptographic algorithms can be updated without hardware replacement, where certificate chains support algorithm diversity, and where secure boot processes can accommodate evolving signature schemes.

Algorithm diversity, in my opinion, is an admission to the fact that there is a real concern that the lattice-based algorithms may be cracked down the road, so alternative algorithms based on different math, Hamming and Hashing, are available. That said, an algorithm that was proposed by NIST for digital signage, which was deemed uncrackable after many years of review, was cracked within weeks of introduction using a relatively low-end microprocessor. In short, because CRQC are currently not available, there is no guarantee that PQC algorithms cannot be cracked leading to architectures that would require extreme amounts of agility and flexibility.

In summary, the transition to PQC cannot follow the automotive industry’s traditional model of generational updates; it must occur as a continuous capability evolution. As vehicles become software-defined platforms with connectivity lifespans exceeding their mechanical longevity, post-quantum readiness becomes not merely a security feature but a fundamental requirement for market viability.

TechArena Founder and Principal Allyson Klein was named Female Founder of the Year today at the 2026 Global Business Tech Awards in London.

Judged by an independent panel of leading technology experts, the awards honor companies and individuals whose work has added measurable, tangible value across customer experience, business management, data intelligence, and emerging innovation.

“I founded TechArena to find a voice for the bold innovation driven by creators across the tech landscape,” Allyson said. “Since its inception, our platform has fostered a community of leadership from the world’s tech titans to the next wave of visionary startups. Our collaborations with inventors and market makers from across the value chain have accelerated IT strategy for navigating the inception of the AI era. This recognition is about the fantastic team we’ve built at TechArena, the choice collectively to dedicate our career aspirations to the north star of collaborative innovation, and all of the brilliant people who have shared their stories.”

A Silicon Valley Kid Who Never Lost the Wonder

Allyson grew up in Silicon Valley during the birth of semiconductors, in a home where technology wasn't abstract; it was dinner table conversation. Her father was an international marketing executive; her mother worked as a nurse for chip plants, bringing home stories about the intricate chemical processes behind the magic of fabrication. By the time Allyson spotted a glowing green Apple computer screen at a friend's house, she was already primed to find it mesmerizing.

Allyson spent 22 years at Intel, where her work went far beyond marketing individual products. She helped build ecosystems, crafted foundational industry narratives, and created initiatives that brought companies together around shared tech visions. She built the foundations of industry engagement that ushered in data center virtualization, cloud computing, 5G networks, and artificial intelligence. Her marketing strategy helped grow a $20 billion dollar business for the company and unquestioned leadership in the industry.

In 2009, when her boss told her to "go figure out social media," one of Allyson's two resulting recommendations was to start a podcast. Chip Chat launched as a weekly show and ran for 754 episodes, reaching over 20 million listeners and winning numerous industry awards. The insight behind it was simple but powerful: the best conversations about technology were happening in tech cafeterias, not in board rooms. Engineers came alive when given permission to talk about what they’d invented, and how they felt when their visions came to life.  

After Intel, Allyson led global marketing and communications at Micron, overseeing everything from CHIPS Act messaging to COVID-19 communications. But by 2022, something was missing.

“I missed creating content. I missed telling stories,” she said. “Those things gave me unique joy that leading massive marketing organizations never could.”

Building TechArena

TechArena was founded on the premise that the industry’s pace of innovation had fundamentally shifted, and conversations on the sidelines weren’t as valuable as direct access to inventors. This drove a conviction that the most important voices in technology aren’t always the loudest ones, and that insider knowledge creates a different kind of journalism. As Allyson puts it, “Most tech journalists don't have the background of living inside tech companies. At TechArena, we understand the shorthand.” That perspective has attracted an impressive range of guests and clients: the platform has featured companies representing more than $9 trillion in market cap, alongside 84 founders and CEOs of emerging tech startups who’ve shared their stories with TechArena’s audience of IT and cloud architects and infrastructure operations teams.

Every piece of TechArena content includes what the team calls the “TechArena take,” an opinion grounded in genuine insider experience. It’s a deliberate editorial choice that sets the platform apart.

What This Recognition Means

The Female Founder of the Year designation carries particular weight in an industry that still has significant ground to cover in terms of representation at the founding and leadership level. For the technology community TechArena serves, this award affirms that building something substantive, durable, and editorially credible is work worth recognizing. The community TechArena has built can strive farther and move faster in part because of the connections that they build together in the arena.

Allyson’s outlook on where technology is headed is characteristically optimistic. When cloud computing arrived, she recalls, the industry feared it would collapse the server market. Instead, new applications proliferated, new businesses were born, and human ingenuity found new expression. She sees AI the same way.

“Humans are going to have a renaissance in terms of what they can do based on AI innovation,” she said, “and while we re-calibrate on where intelligence is created between humans and machines, human to human interaction becomes even more essential and valued.” And she intends for TechArena to be there to tell those stories.

Congratulations, Allyson

The entire TechArena team extends our heartfelt congratulations on this well-deserved honor. It is a reflection of every interview conducted, every story pursued, and every voice given space to be heard. We look forward to continuing to build something worthy of this recognition.

To learn more about Allyson Klein and explore her work, visit techarena.ai/innovator/allyson-klein.

In higher education, information technology infrastructure often operates behind the scenes, quietly enabling learning without drawing attention to itself. For Rose-Hulman Institute of Technology, that philosophy recently drove a significant infrastructure transformation. The goal was straightforward: remove barriers so faculty and students can focus on research, teaching, and learning rather than wrestling with technology limitations.

During my recent TechArena Data Insights episode with Solidigm’s Jeniece Wnorowski and Justin Baker, systems administrator lead at Rose-Hulman, Justin shared how the institution modernized their infrastructure. The results demonstrate how strategic infrastructure investments can dramatically improve operational efficiency while directly supporting educational outcomes.

The Challenge: When Infrastructure Becomes the Bottleneck

Before its latest upgrade, Rose-Hulman’s previous infrastructure challenged system administrators in a variety of ways. Older, disparate systems that were pieced together created slowdowns in trying to do any sort of maintenance, from bringing systems back up if they went down to meeting the demand to roll out new software.

For a small IT team managing everything from student information systems to enterprise resource planning platforms and Microsoft 365 administration, these delays were a serious hindrance. The team needed infrastructure that would let them respond rapidly to emerging needs rather than constantly fighting the limitations of aging hardware.

“Upgrading made the most sense in terms of being able to get that speed and that ease of use….and making fewer points of failure,” Justin explained.

Doing More with Less Through Strategic Partnerships and Modernization

Rose-Hulman’s decision to upgrade by partnering with DataON and incorporating Solidigm solid-state drives (SSDs) as the storage foundation centered on technical compatibility. As a Microsoft shop running primarily Windows servers, Rose-Hulman saw DataON’s close collaboration with Microsoft as a perfect fit. In addition, DataON’s hardware expertise ensured the new infrastructure would support Rose-Hulman’s critical administrative and educational systems.

The performance improvements following the infrastructure upgrade were substantial. Scheduled maintenance windows that previously consumed six to eight hours now are completed in under three hours. Server deployment timelines have been compressed from up to two hours to 10-to-15 minutes. The team no longer needs to wait for “after hours” time blocks to do maintenance or fine tuning, and has time to address critical institutional systems.

“We’re able to run more with less,” Justin explained. “So we can focus on the types of things that allow us to add reliability or backup or something like that to our environment versus having to front-load most of the infrastructure for it just to run everything.”

Piloting a New Model for Access to Engineering Applications

Beyond upgrading core infrastructure, Rose-Hulman is exploring how Azure Local paired with Azure Virtual Desktop (AVD) and NVIDIA L4 graphics processing units (GPUs) can transform software delivery for students. The pilot deployment runs demanding engineering applications through virtual desktop infrastructure, eliminating the traditional constraint of needing powerful local hardware.

This approach addresses a longstanding challenge in engineering education: ensuring every student can access resource-intensive applications regardless of the device they own. By centralizing compute resources and delivering applications virtually, Rose-Hulman can provide consistent performance and eliminate student concerns around having the right high-performance device, or needing to make time to get to a lab to complete coursework.

The TechArena Take

Rose-Hulman’s infrastructure transformation illustrates how strategic technology investments can directly support educational missions in higher education. By partnering with vendors who understand their technology ecosystem and deploying high-performance storage solutions, the institution is achieving measurable operational improvements that cascade into better student experiences. For educational institutions managing tight budgets and small IT teams, efficiency gains translate directly into capacity for innovation and improved service delivery.

As Rose-Hulman continues expanding their Azure Local deployment and virtual desktop capabilities, they’re positioned to offer students greater flexibility and access while maintaining the high-performance infrastructure that engineering education demands. This balance between operational efficiency and educational excellence reflects the thoughtful approach required when infrastructure decisions directly impact student success. Learn more about Rose-Hulman Institute of Technology at www.rose-hulman.edu.

Cloud security conversations have matured. We talk about identity, Zero Trust, workload isolation, posture management. But one layer still gets treated as background configuration: Network architecture. And that’s where quiet failures begin.

Many cloud security issues don’t stem from advanced exploits. They stem from routing assumptions, Network Address Translation (NAT) shortcuts, Classless Inter-Domain Routing (CIDR) reuse, and peering decisions that were never revisited as the environment grew.

Cloud networking is easy to deploy. That does not make it easy to design correctly.

Routing is Enforcement, Not Just Connectivity

In cloud environments, routing tables determine more than reachability. They determine inspection paths. If traffic does not pass through a firewall, it is not inspected, regardless of how strong that firewall is.

Architecturally, this means:

• Default routes must be deliberate, not inherited.

• Route propagation in transit architectures should be controlled, not automatic.

• Inspection layers must sit in unavoidable traffic paths.

• Asymmetric routing should be tested, not assumed away.

A useful design question is simple:
Can any workload reach sensitive resources without crossing an inspection boundary?

If the answer is yes, the network design needs refinement.

NAT Strategy Should Be Intentional

NAT design affects attribution, monitoring, and policy enforcement.

When architecting egress, consider:

• Should workloads share egress Internet Protocols (IP)s, or should they be segmented?

• Is Source Network Address Translation (SNAT) capacity engineered for scale events?

• Can you correlate outbound traffic to specific workloads?

• Are fraud detection or allowlist controls dependent on stable egress identity?

Egress architecture should align with security assumptions. If your security model assumes consistent source identity, your NAT model must support it.

Otherwise, policy becomes guesswork.

CIDR Planning Directly Impacts Segmentation

IP address allocation is often treated as an early-stage task. It defines long-term flexibility.

Intentional CIDR planning should consider:

• Future regional expansion

• Hybrid integration

• Environment isolation (dev, test, prod)

• Growth without overlap

• Clear summarization boundaries for routing

When address space overlaps or becomes fragmented, segmentation logic becomes complex. Complexity increases error rates.

Segmentation clarity starts with clean IP design.

Transit and Peering Require Guardrails

Centralized connectivity models like transit gateways, hub-and-spoke, virtual Wide Area Network (WAN) are powerful.

They also centralize blast radius of an attack.

Architecturally:

• Route propagation should be explicit.

• Peering should include route filtering where possible.

• Environment boundaries should be enforced at the routing layer, not assumed.

• “Temporary” connectivity should have expiration or review processes.

Connectivity should be intentional and constrained.

Flatness in cloud rarely happens by design. It happens by accumulation.

Designing for Containment

The ultimate test of network architecture is containment.

If a workload is compromised:

• How many subnets can it reach?

• Can it bypass inspection?

• Does segmentation enforce least privilege at the network layer?

• Is sensitive data reachable from general compute environments?

Network design is not just about uptime. It defines how far compromise can spread. That is a security decision.

What Mature Cloud Network Architecture Looks Like

Strong cloud network design typically includes:

• Clear environment isolation

• Inspection points that cannot be bypassed

• Controlled route propagation

• Deliberate egress identity strategy

• Non-overlapping, scalable CIDR allocation

• Documented traffic intent between environments

It is rarely accidental. It is intentional. Cloud platforms abstract hardware, not responsibility. The network remains one of the few layers that can enforce unavoidable boundaries. When it is designed casually, security becomes fragile. When it is designed deliberately, it becomes a containment mechanism.

Cloud network architecture is not just foundational. It is decisive.

In a move that sent ripples through the burgeoning AI ecosystem, cloud computing giant Nebius announced its acquisition of Tavily, an Israeli startup making waves with its “agentic search” technology.

While official figures remain under wraps, reports peg the all-cash deal at an estimated $275 million, potentially climbing to $400 million with performance incentives. This isn't just another tech acquisition; it's a strategic chess move that could fundamentally reshape how AI agents are built, deployed, and scaled.

Tavily, founded in late 2024, has been a darling of the developer community, racking up over 3 million monthly SDK downloads and attracting a million-strong user base in record time. Their tech, specializing in real-time web retrieval for AI agents, addresses a critical pain point: hallucinations and outdated information that plague even the most advanced large language models (LLMs). With early funding from heavy hitters like Insight Partners and Alpha Wave Global, Tavily’s rapid, high-value exit underscores the intense demand for solutions that can ground AI in reality.

The combined entity aims to offer a full-stack solution for developers looking to build sophisticated AI agents. Imagine an AI that not only reasons effectively but can also instantaneously access and synthesize the latest information from the web. This integrated approach promises to streamline development, reduce latency, and, crucially, enhance the reliability of AI agents across various applications, from enterprise automation to customer service and beyond.

The market certainly seems to be listening. Nebius pointed to analyst projections that forecast the agentic AI market to explode from $7 billion in 2025 to a staggering $200 billion by 2034. This isn’t just growth; it’s a gold rush, and Nebius just staked a significant claim. Tavily’s continued operation under its own brand and the retention of its 30-person team, including CEO Rotem Weiss, suggests a smart integration strategy, preserving the innovative spirit that made Tavily so attractive in the first place.

TechArena Take

This isn’t merely a strategic acquisition for Nebius; it’s a declarative statement. For too long, the narrative in AI cloud has been dominated by the hyperscalers – AWS, Google Cloud, Azure – with their vast, vertically integrated empires. Nebius, often seen as a formidable player in high-performance compute, has made a bold play to differentiate itself by becoming the go-to platform for autonomous AI agent development.

The integration of Tavily's agentic search is a stroke of genius because it tackles the “black box” problem of AI head-on. By providing real-time, verifiable data, Nebius is directly addressing the trust deficit that has plagued AI adoption. This move positions them as a champion of “grounded AI,” a concept that will only grow in importance as AI agents take on more critical roles in our lives and businesses.

Nebius isn’t just buying a company; they’re buying a crucial piece of the future. By offering a complete agentic stack, they’re competing on capability and, more importantly, trust. This acquisition is a clear signal that the AI agent arms race is heating up, and Nebius just fired a warning shot across the bows of every major cloud provider. Keep a close eye on this space; the game just changed.

The rise of AI is exposing a widening gap between what modern data centers were designed to do and what AI workloads now demand. Boards and executive teams expect faster time-to-value from AI investments. Quietly, the infrastructure has become the bottleneck.

At AI Infrastructure Field Day 4 (AIIFD4), the Cisco Data Center Networking team addressed this gap head-on. Cisco made it clear they are not walking away from Ethernet. Instead, they are rethinking what Ethernet needs to become to reliably support the unique demands of AI workloads.

Why AI Workloads Break Traditional Data Center Design

AI workloads behave very differently from traditional enterprise applications. Training and large-scale inference generate long-lived, east west, GPU-to-GPU flows that are extremely sensitive to latency, jitter, and packet loss. Even minor congestion can cascade into stalled jobs, underutilized GPUs, and missed business deadlines.

During the session, a critical business consequence became obvious: time-to-first-token (TTFT) now matters as much as raw performance. Delays caused by network misconfiguration, troubleshooting blind spots, or prolonged deployment cycles directly erode the return on multimillion dollar GPU investments. In many cases, organizations lose months of effective depreciation time before AI clusters deliver meaningful value.

In other words, long TTFT times mean expensive GPUs are sitting idle while the teams troubleshoot the network.

This is where the gap emerges. Traditional Ethernet is optimized for best-effort, north-south traffic. It was never designed for sustained, lossless, ultra-dense GPU communication. At the same time, many enterprises lack the operational appetite to introduce entirely separate fabrics just to support AI.  

Can Standard Ethernet Support AI at Scale?

Surprisingly, one theme that came through clearly was that plain Ethernet is not enough for modern AI clusters.

Standard Ethernet assumes packet loss is acceptable and recoverable. AI training does not. When one GPU waits on another due to congestion or dropped packets, the entire job slows down. No amount of compute spend can compensate for unpredictable network behavior.

Beyond performance, there is an operational issue. AI environments introduce unprecedented complexity across compute, storage, optics, and networking. Without deep visibility, network teams are often blamed first. But they usually don’t have the telemetry needed to prove where problems actually originate.

It’s hard to understand the challenge when you look at the complexity of a “small” 96 GPU network topology:

This is an executive level risk. AI failure modes are no longer isolated to IT, they impact product timelines, research velocity, and competitive advantage.

InfiniBand for AI: Strengths and Enterprise Tradeoffs

InfiniBand has long been the gold standard for HPC and AI training. It delivers native losslessness and extremely low latency, and it performs exceptionally well in controlled environments.

However, Cisco drew a clear contrast at AIIFD4. While InfiniBand works technically, it introduces business and operational challenges for enterprises:

It creates a separate fabric with specialized tooling and skills.

It limits multitenancy and segmentation, which are essential for shared enterprise AI platforms.

It offers limited end-to-end observability, particularly outside the fabric itself.

It complicates convergence with storage and front-end networks.

InfiniBand excels as a purpose-built backend fabric. But most enterprises aren’t building isolated AI factories. They are trying to operationalize AI alongside everything else.

How Cisco Is Making Ethernet Deterministic for AI

Cisco’s AIIFD4 appearance was not about replacing Ethernet, it was about evolving it.

Their approach combines Ethernet’s universality with AI-specific enhancements that deliver predictability and control. This transforms Ethernet from a best effort transport into a deterministic system fabric, capable of supporting AI training and inference without introducing separate operational silos.

Why Isolation Alone Isn’t Enough for AI Security

One of the most important themes from Cisco’s sessions was that security in AI data centers is about insight and control. It can’t be just about isolation.

Cisco’s AI-optimized Ethernet emphasizes:

Logical segmentation using EVPN-VXLAN, enabling strong multitenant isolation

Secure, TLS-based control plane communication in cloud managed environments like Nexus Hyperfabric

Proactive detection of physical layer issues, such as optic degradation, before they impact workloads

Job-level analytics that tie performance anomalies directly to infrastructure causes

The common thread is control: seeing problems early, understanding their impact, and fixing them before GPUs go idle.

This level of visibility simply does not exist in traditional InfiniBand environments. Cisco’s argument is that what you can see, you can secure, and what you cannot see becomes a business risk.  

The Business Case for AI-Optimized Ethernet  

Cisco’s appearance at AIIFD4 reframed the Ethernet versus InfiniBand debate as a business decision, not just a technical one.

For hyperscalers building single purpose AI factories, InfiniBand may remain the right choice. But for enterprises building multiple AI clusters, often incrementally, across teams and use cases, Cisco’s AI optimized Ethernet offers a compelling alternative: one fabric, one operating model, and one security posture.

The takeaway for executives is simple: the question is no longer whether Ethernet can support AI. The question is whether your Ethernet is engineered for determinism, visibility, and AI scale operations.

Cisco’s answer at AIIFD4 was clear. Enterprises don’t need a second fabric to keep up with AI. They need Ethernet that has been deliberately engineered for determinism, visibility, and scale.

Frequently Asked Questions: AI Ethernet and Deterministic Networking

Q: What is AI Ethernet?

A: AI Ethernet is Ethernet that has been deliberately engineered for AI workloads, with deterministic performance, lossless behavior, and end-to-end observability to support large GPU clusters at scale.

Q: Why isn’t standard Ethernet sufficient for AI workloads?

A: Standard Ethernet assumes packet loss is acceptable. AI training workloads are tightly synchronized, so even small amounts of loss or congestion can stall jobs and leave expensive GPUs underutilized.

Q: How does deterministic networking improve AI performance?

A: Deterministic networking delivers predictable latency and controlled congestion, which leads to faster job completion, higher GPU utilization, and more reliable AI production timelines.

Q: When does InfiniBand make sense for AI?

A: InfiniBand can be a good fit for hyperscalers or single purpose AI factories. Enterprises running shared, multitenant AI platforms often find its operational complexity and lack of convergence limiting.

Q: Why is observability critical for enterprise AI networking?

A: AI environments span GPUs, NICs, switches, and optics, making issues hard to diagnose without end-to-end visibility. Observability enables faster root cause analysis and reduces the risk of idle GPUs and lost value.

Q: Is AI Ethernet only about performance?

A: No. AI Ethernet also addresses operational simplicity, security, and risk by combining visibility, segmentation, and policy driven control as AI platforms scale. driven control as AI platforms scale.

Subscribe to our newsletter

It’s fair to ask whether AI in 2026 is a bubble. The echoes of the early 2000s are real: valuations running ahead of revenues, plenty of compelling tech, and plenty of fuzzy business models. We’ve seen this movie before.

But here’s what feels different this time. We’ve now seen AI deliver real, tangible value, from agentic systems like self-driving cars to generative models like ChatGPT, Gemini, and Claude. New workflows are already reshaping engineering and productivity. The value is real, even if the business models are still forming. What’s no longer speculative is what AI demands in practice: massive compute, running continuously, coordinated across thousands, and soon millions of processing elements.

And where compute goes, networking must follow.

Training isn’t just about FLOPS; it’s about keeping GPUs fed and synchronized—moving data between accelerators, memory tiers, and storage with tight timing. Inference at scale isn’t “lightweight” either. Agentic systems add constant coordination, state exchange, and feedback loops. This is persistent, symmetric traffic, less like consumer internet burstiness, more like an industrial control system that hates latency and variance.

So, while the top of the stack is still sorting itself out, the bottom of the stack is converging. Those infrastructure requirements are driving real decisions: AI-first data centers, power secured years out, liquid cooling systems designed in from day one, and campuses planned as a single distributed computer.

In the dot-com era, Alan Greenspan famously cautioned against “irrational exuberance.” What’s unfolding now feels more deliberate and methodical, albeit no less exuberant. It manifests not in pitch decks, but in data centers, power contracts, and miles of fiber.

From AI Debate to Infrastructure Bottleneck

Early in any technology cycle, progress is driven by ideas. Better algorithms. Smarter software. More elegant abstractions. Over time, however, the limiting factor shifts from what we can imagine to what we can physically deploy.

That shift is now unmistakable in AI.

Regardless of which hyperscaler wins, which model architecture dominates, or which application becomes the killer use case, the requirements inside the data center are converging quickly. AI systems must be dramatically faster, far denser, and far more tightly coupled than anything the industry has operated before—not just larger clusters, but clusters that behave as a single, synchronized system.

For years, optics and networking evolved as predictable plumbing. Bandwidth increased incrementally. Power budgets were manageable. Traffic patterns were relatively well behaved. That trajectory worked for cloud computing and the consumer internet.

AI introduces a discontinuity.

When that linear roadmap is mapped against the demands of large-scale training, generative inference, and agentic workloads, the gap becomes obvious. East–west traffic explodes. Latency consistency matters as much as raw throughput. GPUs grow intolerant of waiting. At scale, the cost, and energy, of moving data begins to rival the cost of computing on it.

This is how industries respond to step changes: they build the substrate first.

Hyperscalers and vendors are investing ahead of certainty—not betting on a single application or winner, but on the belief that AI will require fundamentally different physical systems. In doing so, they are running into a new reality: scaling AI is no longer gated by software ambition alone. It is increasingly constrained by three intertwined limits—speed, thermals, and power delivery.

Those constraints now define the AI infrastructure roadmap.

The Three Walls—and Why Optics Sits at the Center of All of Them

As AI systems scale, the industry is no longer debating abstract limits. It is colliding with three very concrete ones. They arrive together, reinforce each other, and cannot be solved independently.

These are the three walls now shaping AI infrastructure: speed, thermal envelope, and power delivery.

Wall #1: The Speed Wall (and the Copper Limit)

AI workloads demand orders of magnitude more data movement than previous generations of compute. Training large models requires constant synchronization across thousands of accelerators, while emerging agentic systems add persistent coordination and state exchange across distributed components.

To meet that demand, signaling speeds have been pushed relentlessly higher — and this is where physics intrudes.

At the frequencies required for modern AI interconnects, copper becomes a fundamental constraint. Signal integrity degrades rapidly with distance. Loss rises. Reach collapses dramatically from meters to centimeters. At scale, this creates a hard architectural ceiling.

This is not simply a matter of faster PHYs. As AI clusters expand beyond a single rack or building into “scale-across” systems, bandwidth and latency become inseparable. Propagation delay matters as much as throughput, and copper simply cannot preserve both over distance.

Optics relaxes this constraint by delivering far higher bandwidth while maintaining reach and latency as systems scale across racks, buildings, and campuses.

Wall #2: The Thermal Wall (Why “Just Go Faster” Fails)

Even where copper can deliver sufficient speed, it increasingly fails on heat.

As electrical signaling rates rise, resistive losses convert a growing share of energy directly into heat. In high-density AI racks, this creates a feedback loop: higher speed drives more heat, which demands more cooling, which consumes more power and constrains further scaling.

This is why liquid cooling has moved from an optimization to a requirement in modern AI infrastructure. At rack densities well beyond 100 kW, thermals increasingly shift from an operational concern to an architectural one.

Optics changes this equation by reducing resistive loss at the source. Moving data as light — and shortening or eliminating electrical paths through approaches like co-packaged optics — lowers heat generation and expands the thermal envelope available for compute.

At AI scale, optics isn’t about going faster. It’s about not melting the system while doing so.

Wall #3: The Power Delivery Wall (The Grid Becomes the Limit)

The final wall is the most unforgiving: power delivery.

In practice, many data centers are now constrained less by space or fiber availability than by access to electricity itself. New facilities are increasingly sited where power is available, near hydroelectric, nuclear, or renewable sources rather than where latency is most convenient.

In the cloud era, we measured success in Gigabits per second. In the Agentic era, one of the defining metrics increasingly becomes Joules per Inference. We are moving from a performance-constrained world to an energy-constrained one. Power must be budgeted hierarchically: per server, per rack, per row, per facility. One of the largest and fastest-growing consumers of that power is data movement, particularly the repeated conversion between electrical and optical domains.

The math is sobering. At scale, the energy spent moving bits can rival the energy spent computing on them.

Optics is central here not just because it is efficient, but because it enables efficiency everywhere. By doing more in light and less in copper — and by pushing optical interfaces closer to compute — operators can reduce energy per bit, per port, and per rack, freeing scarce power for actual computation.

This is what allows power-constrained data centers to continue scaling, and what makes it feasible to couple multiple facilities into much larger virtual systems.

Why Optics Solves All Three Walls at Once

These three walls are tightly coupled. Solving one in isolation makes the others worse. Faster electrical signaling increases heat. More cooling increases power draw. Greater power demand stresses both facilities and the grid, capping further scale.

This coupling is what makes AI infrastructure different from previous compute cycles.

Optics is unique because it relaxes all three constraints simultaneously. It delivers the bandwidth and reach required for scale-across architectures, reduces thermal load by minimizing resistive loss, and lowers energy consumption per bit, freeing scarce power for computation rather than transport.

That combination is why optics has moved from predictable plumbing to a first-order architectural consideration. Across components, systems, and emerging approaches like optical switching and co-packaged optics, the industry is increasingly using light to break limits that electrons can no longer navigate efficiently.

This shift applies not only to new builds. Existing data centers are being retrofitted to accommodate AI workloads, driving additional optical demand as legacy, copper-heavy designs are reworked to survive higher speeds, tighter thermal envelopes, and stricter power budgets.

Optics doesn’t eliminate tradeoffs, but at AI scale, it expands the feasible design space in ways no other approach can.

Is AI a Bubble?

We still don’t know which applications will dominate, which business models will endure, or which hyperscalers will capture the most value. Those questions remain open.

But one thing is no longer in doubt.

Whatever form AI ultimately takes, it will require a fundamentally new physical substrate — one that is faster, more deterministic, and dramatically more power-efficient than what came before. That substrate is being built now, and it is being driven by optics.

This is not speculation. It is infrastructure.

And infrastructure, once committed to at this scale, has a way of shaping the future regardless of who wins the race at the top of the stack.

History offers a useful parallel. After World War II, the United States embarked on an enormous infrastructure project: the interstate highway system. It was built without knowing exactly where people would live, which cities would boom, or which industries would dominate. It was built on a conviction that mobility would matter, and that the country would be better off prepared for wherever it led.

The AI infrastructure build-out has the same shape.

Data centers, power delivery, cooling systems, and optical interconnects are being constructed not because the industry has perfect clarity on applications or economics, but because it has conviction that AI will be foundational. Once that conviction takes hold, infrastructure becomes destiny.

This is why this moment feels different from past bubbles. Software cycles can inflate and deflate. Markets can overshoot and correct. But when an industry runs into hard physical limits, the response is not debate. It is construction.

Many AI companies will fail. Some valuations will reset. Entire categories will consolidate or disappear. That is how every major cycle unfolds.

But the infrastructure being built now will not vanish with the noise. Like the highways of the last century, it will outlive the narratives that justified its construction and quietly shape everything that comes next.

After World War II, we paved the country with concrete and asphalt. Today, we are doing it again, this time with photons, lasers, and fiber.

We are building massive highways of light.

The applications will change. The winners will shift. The economics will evolve.

But the highways will remain.

Subscribe to our newsletter

The first sign of global disruption is rarely a system outage. It is a quiet rise in alerts, a spike in phishing volume, or subtle misuse of valid credentials that look ordinary until it is not.  

During periods of instability, cyber risk does not suddenly appear. It compounds. Conflict acts as a force multiplier by exposing existing weaknesses, straining critical services, and pushing security teams into sustained high-alert mode. Recognizing this dynamic is essential for organizations that want resilience rather than reaction.

Weaknesses Become Visible Under Pressure

Periods of disruption do not create new classes of cyber risk. They reveal gaps that already exist but are often tolerated under normal conditions. Identity systems, access controls, and operational shortcuts become pressure points when speed and availability take priority. Data from IBM Security shows that compromised credentials and misuse of valid accounts remain among the most common initial access vectors in major breaches, and incidents involving valid credentials take longer to detect and cost more to remediate. When organizations rely heavily on cloud services and remote access, these weaknesses become easier to exploit, not harder.

Critical Services Have Low Tolerance for Disruption

The impact is most visible where failure carries immediate consequences. Energy, healthcare, transportation, and communications systems operate with little tolerance for disruption. Advisories from the Cybersecurity and Infrastructure Security Agency consistently warn that elevated risk environments increase attempted intrusions against critical services. Even short-lived outages or degraded performance can affect safety, continuity, and public confidence. In these environments, the perception of instability often causes as much damage as the technical event itself.

Ambiguity Complicates Response

Cyber activity also becomes harder to classify during periods of instability. Analysis from Europol highlights how financially motivated attacks, espionage, and disruptive activity increasingly overlap. For defenders, this ambiguity complicates response decisions, regulatory obligations, and communication strategies. Familiar technical indicators can suddenly carry unfamiliar consequences, forcing teams to operate with incomplete information.

The Human Cost of Sustained Alert

The strain is not limited to systems. Sustained high-alert conditions place continuous pressure on security teams, particularly those responsible for incident response. SOC surveys from the SANS Institute show rising fatigue and burnout across security operations roles. Prolonged stress reduces detection accuracy, slows response times, and increases the likelihood of error. In this context, burnout becomes a measurable security risk rather than a workforce concern.

Tools Alone Are Not Enough

It is tempting to assume that advanced tooling, automation, and threat intelligence can neutralize these challenges. While technology improves visibility and response speed, it does not eliminate structural weaknesses. Tools cannot replace clear decision-making, effective communication, or well-rested teams. Post-incident reviews repeatedly show that organizations fail not because of missing tools, but because coordination and judgment break down under pressure.

The World Economic Forum continues to rank cyber insecurity among the top global risks precisely because it compounds during uncertainty. Conflict does not pause cybersecurity. It accelerates it. Organizations that invest in identity protection, realistic incident planning, and sustainable operating models are better positioned to absorb prolonged instability.

Conclusion: Resilience Priorities During Instability

• Reduce credential risk: tighten MFA coverage, harden recovery flows, and monitor for suspicious sign-ins and privilege changes.
• Assume disruption in critical workflows: test continuity plans for outages, degraded performance, and vendor dependency failures.
• Simplify incident decision-making: pre-define escalation thresholds, comms owners, and legal/regulatory triggers for ambiguous events.
• Protect the responders: use rotations, on-call limits, and automation to prevent sustained fatigue from degrading detection and response.
• Practice cross-functional coordination: run tabletop exercises with IT, security, legal, comms, and operations so judgment holds under pressure.

The question for leaders is no longer whether disruption will occur; it is whether their systems, decisions, and people can sustain pressure when it does.

References

• Cost of a Data Breach Report 2025: https://www.ibm.com/reports/data-breach
• Internet Organised Crime Threat Assessment (IOCTA): https://www.europol.europa.eu/publications-events/main-reports/iocta-report
• SANS 2025 SOC Survey: https://www.sans.org/white-papers/sans-2025-soc-survey
• Global Risks Report 2024: https://www.weforum.org/publications/global-risks-report-2024/

As organizations deploy AI models at scale, a new set of challenges has emerged around operational efficiency, developer velocity, and infrastructure optimization. A recent conversation with Solidigm’s Jeniece Wnorowski and Brennen Smith, head of engineering at Runpod, revealed how cloud platforms are rethinking the entire AI stack to help developers move from concept to production in minutes rather than months.

The Economics of AI Inference

Runpod operates 32 data centers globally, providing graphics processing unit (GPU)-dense compute infrastructure for small companies and enterprises building and deploying AI systems. This service is crucial considering the economics of modern GPUs, where a single system with 8 GPUs can cost hundreds of thousands of dollars. Runpod understands that the compute hardware is only part of the equation. “Storage and networking…glue these systems together,” Brennen said. “By ensuring that there’s high quality storage paired up with these GPUs….we have been able to show that this results in a markedly better experience.”  

On top of this, the company provides a sophisticated software stack that allows developers to go from their idea to production in minutes, across training and inference use cases. The goal is to “Make it so developers and AI researchers can focus on what they do best, which is actually delivering value to their customers,” Brennen said.  

The ability to rely on optimized infrastructure is becoming even more important as organizations move from training to deployment. Smith likened training infrastructure to traditional business capital expenditures, noting that the high up-front costs see a return on investment over a long period of time. In inferencing, organizations deal with ongoing operational realities, grappling with scaling, efficiency, and delivering value to customers daily. As a result, Runpod has engineers specifically looking at inference optimization. With the rise of AI factories, “How well these systems are run from an operational excellence perspective will dictate the winners and losers,” Brennen said. “You run an inefficient factory, you’re out.”

Storage as the Innovation Accelerator

One of the most important insights from our conversation addressed storage, which is now seen as a hidden bottleneck in AI. Brennen recounted how his engineering team recently investigated Docker image loading times. While unrelated to a specific large language model (LLM) activity, developers flag issues like slow loading times as hurting their overall workflow. This gets in the way of things needing “to magically just work.”

For the solution, Brennen reiterated that storage is what glues the system together. “What we have found is every time, as long as we are optimizing our storage, we are able to make the data move faster,” he said. And when data movement is optimized, entire development cycles accelerate.

Runpod recently launched ModelStore, a feature in public beta that leverages NVMe storage and global distribution to make AI models seem to appear “like magic.” What previously took minutes or hours now happens seamlessly, compressing development iteration cycles. For organizations under pressure to deliver AI capabilities quickly, these time savings compound into significant competitive advantages.

Brennen emphasized that faster developer cycles enable teams to fail fast and iterate more effectively to deliver successful outcomes. When CTOs receive mandates to implement AI, their success depends on giving teams tools that accelerate innovation rather than creating additional friction.

The Convergence of Infrastructure and Code

Looking ahead, Brennen identified the convergence of infrastructure and software as a transformative trend. The goal is to enable code to self-declare and automatically establish the infrastructure required to run it, freeing developers from thinking about infrastructure so they can focus on their code and creating value aligned to business logic. “Anything we can do to make it even easier to get global distribution, that’s a hugely powerful paradigm,” he said.

The TechArena Take

Runpod’s emphasis on developer experience demonstrates that sustainable AI deployment requires thinking holistically about the entire infrastructure stack. The company’s focus on making complex infrastructure feel magical to developers reflects a broader industry recognition that reducing friction accelerates innovation.

As AI moves from experimentation to production deployment, organizations that optimize for developer velocity and operational efficiency will have a significant advantage from their ability to accelerate time to value. For organizations evaluating AI infrastructure partners, Runpod’s approach offers a model that balances performance, scalability, and ease of use.

Connect with Brennen Smith on LinkedIn to continue the conversation, or visit Runpod’s website and active Discord community to explore how their platform might support your AI initiatives.

Humans desperately want to find patterns, meaning in patterns, and to create and connect.

We anthropomorphize constellations, animals, elements, and now AI. From Pygmalion to Frankenstein to the internet and computer games like The Sims, our urge to be a creator as part of human Imago Dei is a thread throughout human history. Going back to Milton’s Paradise Lost, we desperately want to say, “Did I solicit you from darkness to life?” to a creation of ours.

In today’s viral moment of February 2026, we have Claudebot/Moltbot/OpenClaw/? patterned after a pinnacle of human achievement, said no one ever, Reddit.

Yes, it is consistent that what fascinated humans with Victor Frankenstein continues to fascinate us now. It is clear that humans can’t help but step close to mistaking the technical with labels that evoke transcendence. Of course, it is curious and telling of a deep human need that training LLMs on Reddit would result in a pseudo-religion. The extent to which Moltbot/OpenClaw is a mirror reflecting back ourselves will be a subject of ongoing study, just as much as cybersecurity professionals are studying the security implications.

The Positives: Building on the Fundamentals

In all of this, there are some positives that point to directions the next innovator can build on. Like the iPhone, the fundamentals that came together in a novel, breakthrough approach weren’t necessarily new. Crucially, AI breakthroughs aren’t about models and model benchmarks anymore; we have shifted to applications and services that neutralize model identity. A few trends that were improved or extended include:

Messaging Apps: Internally at my “day job” company, individuals were building assistants/agents that they had to schedule a Teams call with to continue training their assistant like a junior employee. For SaaS/FAANG, Slack and WhatsApp have been the natural communications channels. Moltbot/OpenClaw messages you back proactively, extending other chatbots that require a ‘check back” from the user.

Personalization and Memory: Most chatbots have improved saving state over the last few years. Even free versions can hold a conversation history so you don’t experience 50 First Dates with every new chat. Private GPTs and avatar chatbots trained on years of an individual’s writing have been around for almost two years. Thanks to how the internet and remote work have conditioned us, those interactions were starting to feel like we were collaborating with a team member rather than a program. Tying into point #1, if the channel is the same for a human team and an AI agent, who or what is on the other side can start to matter less than the task that is being completed. It can even feel like you’re really connecting because Moltbot remembers you.

A Cruise Director for Your Life: Years ago, a woman I was in a leadership cohort with caused the entire room to burst into laughter because she said, “I need a work wife!” There is a reason a faithful and patient personal assistant is a constant sidekick in movies about rock stars and the rich. Someone who knows you and who proactively directs you on what to focus on, where to go, and even arranging your day will make “adulting” easier for us all. Personalized assistants are now democratized.

The Downsides: Governance and Agency

There are also some downsides. As a certified AIGP (AI Governance Professional) in tech for years, I have seen that this technology has been unruly for its own creator. A technology that can be incredibly powerful only if given full system access can be incredibly powerful against you and your system.

Vulnerability: LLMs still fall prey to prompt injection, data poisoning, and model drift. They are probabilistic rather than deterministic. LLMs can’t always differ between a legitimate prompt and a prompt hidden in what should be benign information fields. Set limits upfront and mandate agent behaviors on specific tasks to check back before taking actions outside specific guardrails you set in advance.

Security: There was a joke years ago about how Gen X were raised with the fear of sharing personal information with strangers online or getting into a stranger’s car. The following generations pioneered social media and Uber… cybersecurity pros are rightly raising alarms about Moltbot. For now, you have to be prepared for securing your system and data, API keys, and tokens, setting limits and mandating agent behaviors on specific tasks. Over time, agentic security controls and governance will catch up and be more off-the-shelf for average users. Until then, assume a defensive driving posture like you’re riding a motorcycle in a third-world country without wearing a helmet.

You Own It: More than anything, open-source agentic AI means you have to have agency yourself. It sounds great to be your own billion-dollar, one-person company. it sounds amazing to have your own personal assistant. The quality of your ideas, your ability to reach farther, and your ability to refine faster with a critical eye will determine your success. Your technical ability to expand and secure your setup is something you own for yourself.

When I think back to the last OCP Global Summit 2025, one of the most memorable sights on the show floor wasn’t a chip or a server tray. It was the racks.

Meta’s Open Rack Wide (ORW) specification introduced adouble-width form factor that looked, at first glance, almost counterintuitive, especially in an industry moving toward disaggregation.

But ORW is a useful clue about where AI infrastructure actually is right now. We may be headed toward disaggregated systems, but today’s highest-performance AI deployments are still heavily constrained by short-reach, high-lane-count copper connections, plus the physical sprawl of power delivery, networking, and cooling that modern platforms demand. In other words, the rack is increasingly behaving less like furniture and more like the computer.

The Evolution of Open Rack

The Open Rack specification has been a cornerstone of hyperscale data center design for years. Unlike traditional 19-inch racks, Open Rack was designed from the ground up for large-scale cloud and AI deployments. Its signature21-inch width improves airflow and its powered busbar simplifies power delivery while reducing cable clutter.

Over time, Open Rack evolved to meet the growing demands of AI and high-performance computing. The original ORV1 specification introduced a 12V busbar, ORV2 improved scalability and cooling, and ORV3 moved to 48V—enabling higher power density and making liquid cooling easier to integrate (via rear-mounted manifolds). Then came ORV3 HPR (High Power Rack), which pushed further with added depth and more robust power management to support the most demanding AI servers while maintaining compatibility with the ORV3 standard.

For a while, ORV3 HPR seemed like the pinnacle of rack design. But as AI workloads continued to push the limits of power and cooling, even HPR began to show its constraints.

The Case for a Wider Rack

The industry is undeniably moving toward disaggregation—separating IT load, power, and cooling into distinct systems. Draft specifications and roadmaps for dissagregated power architectures targeting 100kW today and up to 1MW-class racks over time are already being shared through the OCP community, so a wider rack design might seem like a step backward. However, before we can fully embrace disaggregation at rack scale, we need to overcome the limitations of copper-based electrical connections. The sheer number of electrical and signaling leads—plus distance, loss, and power constraints—required to connect rack systems at scale presents significant challenges. Until those challenges are resolved, many AI deplyments favor a “scale-up” architecture over a “scale-out” approach.

There’s another factor at play: the physical layout of compute systems is expanding. As GPU die sizes grow, so do the memory, networking, and power delivery systems that support them. In short, while we know disaggregated systems are the future, we still need an intermediate solution to bridge the gap. That’s where Open Rack Wide (ORW) comes in.

ORW scales up the HPR’s feature set to accommodate much larger, heavier, and more power-intensive AI systems. With double the width of ORV3 racks and a slightly taller frame, ORW provides the space and structural integrity needed for next-generation AI platforms.

What Makes ORW Different?

ORW isn’t just a bigger rack—it’s a reimagined platform designed for the unique demands of AI. At 1200mm wide (compared to ORV3’s 600mm), ORW offers significantly more real estate for high-density compute trays, liquid cooling manifolds, and power distribution systems. It supports up to 3500 kg of IT gear—more than double the capacity of ORV3 HPR, and is engineered to handle the thermal and electrical loads of modern AI workloads. (Fun fact: ORW is also affectionately known as “BFR” — Big Freaking Racks.)

One of the most compelling aspects of ORW is its flexibility. The specification supports multiple power architecture options, including legacy ORV3 power shelves, side power racks for low- or high-voltage DC input, and even native high-voltage busbars that distribute power directly within the rack. This adaptability ensures that ORW can evolve alongside AI infrastructure, whether for training clusters, inference workloads, or hybrid deployments.

Liquid cooling is another key feature. ORW’s design accommodates high-power liquid-cooled busbars, which are essential for managing the heat generated on the busbar by the power delivery of today’s AI chips. This focus on cooling efficiency aligns with the industry’s push toward sustainable, high-performance data centers.

Industry Collaboration and Early Adoption

ORW isn’t just a Meta project—it’s an open standard developed in collaboration with industry leaders. The base specification for ORW was announced by Meta at the OCP Global Summit 2025, and it quickly gained traction. Companies like AMD, Wiwinn, and Rittal debuted their own ORW-based designs at the summit, showcasing the specification’s potential. AMD’s "Helios" rack-scale reference system, for example, leverages ORW to deliver optimized performance for AI clusters, while Wiwinn unveiled its double-wide rack architecture for next-generation AI workloads. Rittal, meanwhile, is preparing ORW-compatible enclosures and accessories for mass production later in 2026. This collective effort underscores the importance of open standards in shaping the future of AI infrastructure.

It’s worth noting that not everyone is on board. NVIDIA, for instance, is advancing  vertically integrated rack-scale systems and architectures that don’t necessarily map cleanly to ORW. But for those committed to open standards, ORW offers a compelling path forward. The AMD design exemplifies this as it integrates GPU, CPU and networking into a single, cohesive rack system for large-scale AI and High-Performance Computing (HPC) workloads.

Challenges on the Road Ahead

Developing ORW wasn’t without its challenges. The increased size and weight of the rack required new manufacturing approaches, including automation and bolt-together assembly techniques to simplify production and shipping. Testing presented another hurdle: traditional test equipment couldn’t handle ORW’s 3500 kg payload, forcing the team to partner with automotive and aerospace testing facilities to validate the design.

Standardization is also critical. For ORW to succeed, the OCP community must continue to refine the specification and ensure interoperability across vendors. This collaborative approach is what makes open standards like ORW so powerful—they bring together hyperscalers, vendors, and researchers to solve shared challenges.

The Future of AI Infrastructure

ORW represents a foundational shift in data center design. It addresses today’s power, cooling, and space constraints while laying the groundwork for future advancements. As the industry works toward full disaggregation, ORW provides a scalable, open platform that can evolve with the needs of AI workloads.

By providing a bridge to the future, ORW enables the industry to innovate today while preparing for the next wave of data center evolution.

Fermilab’s Silvia Zorzetti explains how quantum computing and sensing are evolving, where they outperform classical systems, and what’s next for the field.

For the last several years, the media industry has framed its future as a codec war: free versus licensed, open versus proprietary, AV1 versus HEVC and its successors. On the surface, the debate feels rational. Compression efficiency has always mattered, and it still does. Without it, global streaming at scale would not exist.

But the codec fixation has become a distraction.

The market is no longer defined by how efficiently bits are compressed in isolation. It is being reshaped by whether entire systems can guarantee experience behavior end-to-end. By “system,” I mean the full chain: encoding, transport, wireless edge, client buffering/playout, and the control loops that coordinate them. Consumers do not churn because of subtle compression artifacts; they churn because experiences fail—buffering during a live touchdown, audio drifting out of sync, latency breaking immersion. These failures are not codec failures. They are system failures.

Efficient bits cannot compensate for fragile delivery.

Codec Wars Are a Component Debate

For two decades, the industry optimized the payload. Engineers worked relentlessly to represent more information per bit while preserving perceptual quality and creator intent. The results were extraordinary: lower bitrates, higher fidelity, and an explosion of global video delivery.

That work succeeded because the environment allowed it to succeed: media consumption was largely passive, buffers could mask uncertainty, and users tolerated occasional degradation when networks misbehaved.

That environment no longer exists.

In the agentic AI era, media consumption is no longer passive. It is increasingly mission-critical, and failures are no longer cosmetic—they can be catastrophic. Experiences now span real-time interaction, immersion, and safety-adjacent workloads where timing and continuity are non-negotiable.

Today’s dominant failure modes are not caused by insufficient compression. They are caused by path fragility, especially at the wireless edge. Interference, congestion, multipath fading, and contention are not engineering oversights — they are physical realities. Even the most deterministic core network cannot repeal the laws of radio physics.

If a media experience depends on a single path behaving perfectly, it does not matter how advanced the codec is or how efficient the compression may be. When that path degrades, the experience suffers—and too often, it breaks.

The codec debate keeps asking one component to solve problems that belong to the system.

“Free” Doesn’t Eliminate Complexity — It Relocates It

Much of today’s codec discourse centers on cost. Royalty-free codecs are often presented as the inevitable future, eliminating licensing friction and unlocking innovation. For hyperscalers with vast engineering budgets, this trade can be rational. Royalties are exchanged for compute and internal optimization. But for much of the ecosystem, the economics are more complicated.

As the old systems engineering adage goes, complexity is conserved.

In any large system, removing one form of complexity does not make it disappear — it displaces it. When standardized licensing frameworks are removed, complexity migrates into less visible, more variable domains. Encoding efficiency often requires more compute. Hardware acceleration becomes fragmented across silicon platforms. Integration, validation, and debugging burdens shift from the ecosystem to individual product teams. IP risk moves from a shared framework onto each adopter’s balance sheet.

“Free” codecs do not eliminate cost; they transform a known, predictable expense into a distributed operational tax that grows with scale.

The real decision is not between free and paid. It is a choice about where complexity lives, and whether it is managed once at the ecosystem level or repeatedly inside every organization.

The Market Shift: From Components to System Guarantees

As media evolves toward real-time, immersive, and safety-adjacent use cases, the competitive frontier is moving decisively upstream. Differentiation no longer comes from compression efficiency alone. It comes from whether the system can guarantee behavior under non-deterministic, hostile edge conditions.

This is the defining transition underway: media is no longer optimized as a signal, but engineered as a system.

Instead of asking codecs to compensate for unpredictable networks, systems must be designed to tolerate unpredictability by construction. Reliability can no longer depend on a single path behaving perfectly. It must emerge from coordination across multiple paths and layers.

Redundancy becomes the new reliability.

The Architectural Answer: Cooperative, Multi-Path Delivery

Today’s media delivery architecture is largely an act of faith. The cloud compresses content. The player buffers it. The network does its best. Each layer operates with limited awareness of the others’ constraints or priorities.

The codec does not know when the Wi-Fi link is about to degrade.

The network does not know the next frame carries a safety alert.

The player hopes the buffer is deep enough to hide the chaos.

This architecture was sufficient for a world of passive viewing. It is insufficient for a world of precision and mission-critical applications.

Coded Multisource Media Format (CMMF) represents a critical architectural pivot. Rather than treating delivery as a single fragile stream, it enables cooperative, multisource systems where media can be reconstructed from multiple paths simultaneously.

In plain terms, CMMF is an industry-standard container that enables robust, low-latency media streaming by allowing content to be delivered simultaneously from multiple network sources, like different CDNs or network paths. Instead of sending identical copies of the data, CMMF uses linear, network, or channel coding to split media into coded “symbols”. A client can then pull unique coded pieces from several locations and reassemble the original stream once enough pieces are collected. This approach increases reliability, improves throughput, and reduces rebuffering—without the inefficiency of storing full duplicate streams everywhere, making it ideal for modern multisource and multipath delivery architectures.

This is not about making one pipe bigger. It is about orchestrating multiple pipes intelligently.

Unlike basic connection bonding, multisource coding avoids redundant traffic while dramatically improving effective Network QoS. Wi-Fi and cellular links become a unified connectivity pool rather than mutually exclusive choices. The client assembles the experience from whichever paths are healthy at any given moment.

Physics remains hostile — but it is rarely hostile everywhere at once.

AI further amplifies this shift. Traditional streaming protocols are reactive by design. Quality drops after packets are lost. Buffers drain before adaptation begins. For real-time and immersive experiences, that response comes too late.

A cooperative system can observe conditions continuously, predict degradation, and adapt preemptively. Critical frames are rerouted before failure becomes visible. The experience does not stall or degrade — it simply continues.

The technology to do this exists today. The challenge now is not invention; it is adoption: moving cooperative delivery from standards and trials into repeatable, mass-market deployment.

Where Guarantees Matter — and Where They Don’t

Advocates of “good enough” media often argue that consumers will not pay for this level of precision. And for TikTok dance videos or Instagram streams watched on a bus, they are right.

But the growth engines of the next decade are not passive or disposable. They are high-consequence, real-time, and immersive experiences where failure is not a minor annoyance—it is a liability. These are the domains where guarantees become the product.

In the automotive cockpit, media becomes mixed-criticality. Entertainment and safety signals coexist on the same system. A collision warning cannot buffer behind a map update or a game download. Entertainment can degrade; safety cannot.

In live sports, latency is no longer a technical metric — it is a business metric. When fans learn about a touchdown from social media before seeing it on screen, value is destroyed. Determinism sells time.

In XR and spatial computing, the governing constraint is biological. Motion-to-photon latency and its variance determine whether an experience feels natural or induces nausea. There is no buffer in XR. Timing must be exact, every time.

Across these domains, the pattern is unmistakable. “Good enough” fails not because quality is too low, but because time is no longer negotiable. These are the markets where determinism moves from a technical aspiration to a commercial and experiential requirement—and where system-level cooperation becomes the only viable path forward.

Why Standards Decide Whether This Scales

Vertically integrated stacks can deliver exceptional experiences when one company controls the entire pipeline. That model works — but it does not scale across global ecosystems of creators, silicon vendors, OEMs, operators, and platforms.

History is clear: when industries hit a complexity wall, they standardize.

Wi-Fi did not achieve mass adoption through proprietary turbo modes. It scaled when interoperability became the baseline and innovation moved up the stack. Media delivery is approaching the same inflection point.

Deterministic, cooperative delivery cannot scale as a collection of proprietary silos. It requires shared assumptions, reference behavior, certification, and long-term stewardship. Standards turn fragile integrations into predictable markets. They allow creative intent and timing guarantees to survive the journey intact — regardless of who built each layer.

Without standards, cooperative delivery remains a premium feature. With standards, it becomes infrastructure.

Conclusion: The System Is the Product

The era of competing on cheaper bits is ending. The era of competing on guaranteed experience has begun.

Over the last decade, the industry rebuilt the nervous system — more deterministic networks, faster optics, better wireless. Now it must upgrade the signal itself.

The codec, the transport, and the player are no longer independent optimization problems. They are a single system, and they must be designed as one.

Value is migrating from components to architecture.

From efficiency to reliability.

From isolated optimization to cooperation.

The winners of the next era will not be those who compress bits most aggressively, but those who ensure experiences arrive intact, on time, and without compromise — even when the environment in between is hostile.

For the past two years, enterprise AI postmortems have sounded the same. A pilot stalls. Results look inconsistent. Trust erodes. The verdict follows quickly: the model is immature, the tools are unstable, the technology moved too fast.

That explanation is convenient. It is also wrong.

AI did not introduce fragility into enterprise data platforms. It exposed what was already there.Long before large models showed up, many platforms were held together by undocumented assumptions, fragile transformations, and ownership gaps everyone learned to work around. AI did not break those systems. It removed the ability to ignore their weaknesses.

What teams are facing is not an AI failure. It is a systems reckoning.

Data Debt Is Structural, Not Cosmetic

Data debt is often framed as bad quality or missing fields. That framing misses the point. The real debt is structural. It lives in pipelines no one fully owns, logic that exists only in people’s heads, and transformations that accumulated over years without a clear contract.

Traditional analytics could tolerate this. Dashboards aggregate. Reports smooth over inconsistencies. When something looks off, an analyst adjusts a filter or adds a footnote. Time absorbs the problem.

AI does not.

AI pipelines pull from multiple sources, assemble context, and produce outputs that appear authoritative. Every hidden assumption becomes an input. Every undocumented rule becomes a risk. Every unclear boundary becomes a debugging exercise with no obvious owner.

Consider a familiar enterprise pattern. A customer dimension evolves over a decade. Marketing owns part of it. Finance applies overrides. Operations enrich it downstream. No one owns it end to end. Queries reference it through layers of views. The system works because people know where it breaks.

Introduce an AI system that needs customer context in near real time. The cracks surface immediately. Conflicting attributes. Missing lineage. Output shifts no one can explain. The AI did not create the inconsistency. It forced it into the open.

This matters because AI compresses feedback loops. Issues that once took quarters to surface now appear in days. What used to be background noise becomes a blocking problem. Debt that was once survivable becomes operationally expensive.

This is a well-understood pattern in data platform maturity discussions: when assumptions aren’t explicit, systems fail under new latency, reliability, and trust requirements.

Lineage and Ownership = Operational Requirements

Trust is the currency of AI systems. Without it, outputs are questioned, bypassed, or quietly ignored. Trust does not come from model accuracy alone. It comes from traceability.

When an AI output is challenged, the first question is rarely about hyperparameters. It is about provenance. Where did this data come from. Why does it say this. What changed since yesterday.

Lineage answers those questions. Ownership makes the answers actionable.

This is not about governance theater or compliance checklists. It is about operational clarity. Who owns this dataset? What assumptions does it encode? Who signs off when it changes?

This is also where many enterprise AI efforts stall: trust breaks when teams can’t answer provenance questions consistently.

In practice, that means contracts, tests, and change management around critical datasets—not just documentation.

Dashboards could survive ambiguity because they were passive. AI systems are not. They summarize, recommend, and influence decisions in real time. That shift raises the bar.

A report could be wrong for weeks with limited impact. An AI recommendation can trigger action immediately. Confidence must extend beyond the output to the system that produced it.

Many platforms struggle here because clarity was deferred. Storage scaled. Compute scaled. Understanding did not. The result is a technically impressive platform that no one can fully explain. AI makes that state unsustainable.

Teams that treat lineage and ownership as first-class concerns move faster, not slower. They spend less time debating what the system is doing and more time improving it.

Cost Blowups Reveal Design Gaps

Another common complaint about AI is cost. Training runs are expensive. Inference adds up. Storage grows faster than planned. Budgets get burned.

The instinct is to blame the workload. The reality is less flattering.

AI workloads punish inefficiency. They amplify waste that already existed. Redundant datasets, unnecessary joins, over-retained history, and poorly scoped transformations were tolerable when they powered nightly reports. They become ruinous when they sit on the critical path of AI systems.

Poor data hygiene leads to runaway cost because the platform does more work than it needs to. It processes data that should have been archived. It enriches context that is never used. It recomputes logic that should have been materialized once.

Cost control is an architectural outcome, not a finance exercise. When engineers understand data flows end to end, they can design for efficiency. When they do not, cost becomes an external constraint imposed after the fact.

This is why cost governance has moved upstream into engineering practice: measure unit costs, instrument pipelines, and design to avoid waste.

Teams that scale AI treat efficiency as a design requirement. They ask hard questions early. What data is actually needed. What freshness is justified. What assumptions can be encoded once instead of recalculated repeatedly. That discipline pays off well beyond AI use cases.

The Counterargument, Taken Seriously

A common objection is that AI itself is too unstable for enterprise use. Models evolve. Outputs vary. The pace of change makes durable systems impossible.

There is truth here, but it is incomplete.

Teams with disciplined data foundations are scaling AI today. They are not chasing every new capability. They focus on reliability, clarity, and ownership. When models change, they adapt because their data layer is not a black box.

The difference is not talent or tooling. It is systems thinking. Organizations that treat data platforms as long-lived products rather than one-time projects have fewer surprises. They know what they own and where it breaks. AI becomes an extension of the platform, not a threat to it.

Blaming AI immaturity avoids a harder conversation. It is easier to say the technology is not ready than to admit the platform was never as solid as assumed.

Conclusion

AI did not break enterprise data platforms. It told the truth about them.

For years, many organizations optimized for output over understanding. They shipped faster than they documented. They scaled storage before ownership. They accepted ambiguity because it was convenient. AI removes that option.

This is not a failure story. It is an opportunity. AI acts as a forcing function that pushes data platforms toward maturity. It rewards clarity and penalizes shortcuts. It turns invisible debt into visible risk.

The path forward is not to pause AI adoption. It is to take data platforms seriously as long-term systems. Invest in ownership. Make lineage explicit. Design for efficiency. Treat context as infrastructure.

Teams that do this will find that AI does not destabilize their platforms. It strengthens them.

Hedgehog CEO Marc Austin joins Data Insights to break down open-source, automated networking for AI clusters—cutting cost, avoiding lock-in, and keeping GPUs fed from training to inference.

There was a time in the late ‘90s when the first dot-com boom was underway, mobile phones were going mainstream, and personal computers were finally becoming portable. But the internet was still a physical destination. It was a place you went to, a connected desktop in a home, office, or internet café, not a parallel universe you could access on the go.

I started my career in the trenches, developing some of the industry’s first 802.11 (Wi-Fi) transceivers. Looking back, Wi-Fi wasn’t so much a technology breakthrough as it was an inevitable response to a shift in human behavior. We wanted to communicate. We wanted access to the riches of the internet. We wanted our computers. And we wanted them with us, all the time.

It was a chaotic, fragmented, loud and wildly innovative time. We had proprietary “Turbo Modes,” one-upmanship, conflicting standards, and dozens of startups and few “grown up” companies, claiming they owned the future. But eventually, and rapidly, an ecosystem developed that transformed Wi-Fi from a novelty into an essential utility, sitting beside water and electricity, within 20 years. There was an explosive catalyst, a gold rush, and an eventual consolidation around standards.

Today, as I watch the optics industry in 2026, I can’t help but feel the same electric hum in the air. The same fragmentation. The same confusion. The same catalytic force. The same gold-rush energy. The same inevitability that something is about to happen, must happen, to enable the latest seismic shift in human behavior.

If you squint your eyes, it almost looks like 2003 all over again…

When the Wi-Fi Ecosystem Learned to Grow the Pie Together

What’s often forgotten about early Wi-Fi is that its success was not inevitable. Wireless is a shared medium. There is no such thing as a private RF universe. Interference, broken roaming, competing beam forming mechanisms, and failed interoperability didn’t just hurt competitors, it damaged customer trust in the entire category. If Wi-Fi was unreliable, it wouldn’t matter who had the fastest radio. The market itself would be dead on arrival.

That realization changed behavior. Through a few painful stumbles, the industry learned that there had to be a common baseline, a set of rules everyone followed, to keep the air clean and the experience predictable. Differentiation still mattered, but it had to be built on top of a shared foundation, not at its expense.

Competitors worked together. Standards bodies matured. Interoperability test beds emerged. Certification programs enforced compliance and guardrails. Vendors argued fiercely, but within boundaries that preserved the viability of the ecosystem. It wasn’t altruism. It was survival. Grow the pie first, then fight like hell for your share of it.

This ecosystem balance only worked because the cast of characters was diverse, and complementary. There were large, established players acting as the adults in the room, setting expectations around enterprise reliability, security, and scale. There were aggressive startups injecting energy, new ideas, and technical breakthroughs that pushed the state of the art forward. And there was Intel.

Intel wanted to make mobile computing inevitable. Creating a new, fast-growing category for higher-margin mobile processors was simply good business. But Intel did something unprecedented: it put its own balance sheet behind the ecosystem. A $300 million Centrino marketing campaign, unheard of at the time, made Wi-Fi synonymous with mobility, reliability, and interoperability. It was a spark that turned momentum into a conflagration.

Intel wasn’t alone. Cisco built enterprise-grade wireless networks that IT could trust. Microsoft pulled wireless deep into the operating system, normalizing it for developers and users alike. Dell and other OEMs made Wi-Fi table stakes in mobile computing. The ecosystem had champions. It had shepherds. And it had plenty of unruly sheep. Together, that unlikely combination produced one of the most successful infrastructure transitions in modern technology history.

Wi-Fi didn’t win because one company dominated early. It won because enough powerful players decided, independently, and selfishly, that growing the pie together mattered more than grabbing the biggest slice first.

When Optics Stopped Being Predictable Plumbing

For a long time, optics was boring, in the best possible way. Optical networking was reliable, predictable, and largely invisible. Bandwidth increased on a steady cadence. Power budgets were understood. Distances were fixed. Traffic patterns were well behaved. As long as you followed the playbook, the system worked.

10G became 25G became 40G became 100G became 400G. Roadmaps were clear. Margins were thin but stable. Optics was foundational, but rarely strategic.

Then AI broke the playbook.

The rise of large language models, agentic systems, and massive multi-modal workloads are driving an insatiable demand for compute, that simply does not fit inside traditional data center assumptions. Training and inference push processing density, east–west bandwidth, and latency sensitivity into entirely uncharted territory.

Clusters no longer scale in a single dimension. They scale up, packing more compute into a rack. They scale out, spreading workloads across rows and halls. And increasingly, they scale across, connecting multiple data centers into a single logical system.

At each step, the network had to keep pace. Exceptionally expensive GPUs cannot sit idle waiting for data. As clusters stretch across racks, buildings, and campuses, the network stops being a background transport and becomes a gating factor for utilization, determinism, and overall system efficiency.

Then the industry hit a power wall.

The constraint is no longer real estate or fiber, it is megawatts. New data centers are being built where power is available, not where latency is optimal and convenient. And that power constraint applies to everything: compute, switching, cooling, and optics alike.

The result is a mandate optics has never faced before, and must now satisfy simultaneously:

1. Dramatically increase compute scale.

2. Deliver higher speed and tighter determinism so GPUs never wait.

3. Reduce power consumption per bit, per port, per rack, per data center.

That combination changes optics from predictable plumbing into a first-order architectural constraint.

Copper interconnects, once “good enough,” are becoming a barrier at scale. Signal integrity, power loss, and reach limits are no longer theoretical, they are operational. Co-packaged optics, long discussed in labs and roadmaps, are now moving into real deployments, bringing optics closer to switch silicon and reducing copper distances and power consumption. Pluggable optics no longer monopolize the design space. Optical switching is re-emerging not as an experiment, but as a necessity.

In this world, optics stops being plumbing. It becomes both the limiting factor and the enabling force of AI infrastructure.

The Return of Chaos (and Opportunity)

Just like early Wi-Fi, this shift triggers a burst of simultaneous, multi-vector innovation.

Startups attack every layer at once: new modulation schemes, novel laser technologies, co-packaged optics, disaggregated control planes, fiber automation, thermal management, and power-aware networking. Incumbents are forced to re-architect product lines that were stable for a decade. Conferences fill with competing visions, overlapping claims, and incompatible approaches.

It feels chaotic. Because it is. But once again, chaos is not a failure mode, it is a signal that the industry is very much alive.

And once again, chaos requires gravity. In the AI era, NVIDIA plays the role Intel once did. Again, not out of altruism, but out of self-interest. NVIDIA’s GPUs, interconnect requirements, and system architectures now define the shape of modern AI clusters. Their need for scale, efficiency, and determinism forces the entire optical ecosystem to evolve faster than it otherwise would. Like Intel with Centrino, NVIDIA is pushing the levers that expand the market, because doing so directly expands its own opportunity.

The hyperscalers are doing the same. Meta, Google, Microsoft, Amazon, and others are committing tens of billions of dollars to build AI infrastructure capable of supporting agentic workloads at planetary scale. They are willing to fund new architectures, absorb early inefficiencies, and accept real risk to break through existing limits.

Why the Optics Landscape Feels So Familiar

If this all feels strangely familiar, it should.

Fragmentation? Check. The optics industry today looks a lot like Wi-Fi did in the early 2000s; fragmented, noisy, and bursting with parallel innovation. Dozens of companies are attacking adjacent problems simultaneously: DSPs, lasers, co-packaged optics, thermal management, fiber automation, disaggregated control planes. No single approach has emerged as “the” answer, and that uncertainty is driving experimentation in every direction at once.

Intellectual chaos? Check. The intellectual chaos is unmistakable. Conferences are filled with competing visions and overlapping claims, with multiple companies promising order-of-magnitude breakthroughs through fundamentally different architectures. Wi-Fi went through the same debates, MIMO, MU-MIMO, interference with incumbent RF systems, number of streams, multiple versions of beamforming, proprietary turbo modes vs standards. None of those questions had clean answers at the time, and optics is no different today.

Massive funding inflows? Check. A gravitational pull toward consolidation? Absolutely.

Capital is flowing freely, another familiar signal. Investors and operators alike sense that optics is no longer incremental plumbing; it’s a breakout category with strategic importance. That gravitational pull inevitably leads toward consolidation. We saw this clearly with Marvell’s recent acquisition of Celestial.ai, a move that signals the era of standalone components is ending. Just as Wi-Fi eventually centered around a small number of dominant silicon platforms, optics will likely converge around a handful of dominant players who can integrate those disjointed innovations into a platform.

And most importantly: a forcing function? YES! Wi-Fi needed to cut the wire, then work in dense deployments, then enable low-power IoT, and now act more deterministically to power our agentic future.

Optics needs to make AI scale physically possible; within switches, across racks, across data centers, without collapsing the grid that feeds it.

When a market has a forcing function, it must evolve. There is no choice.

What Wi-Fi Can Teach Optics About the Road Ahead

We’ve seen this movie before, and Wi-Fi left behind a few hard-earned lessons that optics would be wise to absorb.

Lesson 1: Standards and Interoperability always win — even when it's messy.

Never bet against Ethernet and never bet against Wi-Fi. Proprietary performance advantages are tempting early on, but shared infrastructure lives or dies by common language. Great compromises in the days of 802.11g and 802.11n brought the industry together and left proprietary turbo modes as window dressing for the retail market. Coopetition flourished and the winners were those who embraced it. Optics will face the same tradeoffs, and the ecosystems that prioritize interoperability early will ultimately outlast those that don’t.

Lesson 2: The market rewards companies that grow the pie.

Intel didn’t sell access point silicon. Microsoft didn’t sell radios. Dell didn’t care which chipset won; they bought from everyone. What they all cared about was expanding the market itself. Their success came from making Wi-Fi inevitable, not exclusive. Optics needs its own version of that mindset.

Lesson 3: Simplification beats elegance.

Wi-Fi became ubiquitous not because it solved the RF problem perfectly, but because it made the technology easy for millions of people to deploy. Optics is approaching a similar inflection point. Operators aren’t asking for more clever architectures; they’re asking how to deploy across dozens of data centers, manage thermal and power budgets, automate fiber paths with fewer humans in the loop. Elegance helps, but simplification wins.

Lesson 4: The winners are ecosystem players.

The most successful Wi-Fi companies didn’t just ship chips; they built platforms. Reference designs, SDKs, certification programs, developer ecosystems, and trusted brands mattered as much as raw performance. Optics now has the same opportunity, but only if the industry thinks beyond feeds, speeds, and component optimization.

Where Optics Goes From Here

If the analogy holds, and I believe it does, then optics is entering a decade defined by startup energy, vendor consolidation, architectural standardization, and deep vertical integration. Complexity will be abstracted away. New platforms will emerge. The conversation will shift from components to systems, and eventually to experiences.

The companies that win won’t just be the fastest or the most clever. They’ll be the ones that make optics predictable, operable, and trustworthy at scale. They’ll lean into interoperability before the market forces it. They’ll treat power, cooling, and fiber as software problems. They’ll partner with kingmakers rather than trying to outmuscle them.

And most importantly, they won’t try to own the whole pie. They’ll grow it.

Because every major networking revolution, Ethernet, Wi-Fi, cloud, and now AI fabrics, follows the same arc: breakthrough, fragmentation, chaos, consolidation, and ubiquity. Optics is squarely in the fragmentation and chaos phase.

That’s not a bug. It’s the signal that the industry is alive again.

Conclusion: The Déjà Vu Is Real

When I sit in modern AI datacenters and look at the optical racks, I feel the same thing I felt holding a pre-standard 802.11g PCMCIA card in 2002:

“We don’t fully know what we’re building yet, but when we do, it will reshape the entire industry.”

Wi-Fi unlocked mobility. Optics will unlock AI at scale. And just like Wi-Fi, the winners won’t be the ones who optimize a component in isolation. They’ll be the ones who understand that ecosystems, not components, determine the future.

In a move that signals a significant restructuring of the semiconductor IP landscape, Synopsys and GlobalFoundries (GF) today announced a definitive agreement for GF to acquire Synopsys’ Processor IP Solutions business. The deal, which includes the ARC processor family and related software development tools, marks a pivotal moment for both companies as they sharpen their focus on the burgeoning Physical AI opportunity.  

The transaction, expected to close in the second half of calendar year 2026, will see Synopsys’ Processor IP portfolio—ARC-V™ (RISC-V) and ARC® CPU IP, DSP IP, Neural Network Processing Unit (NPU) IP, and related software development tools including ARC MetaWare Development Toolkits—move into the GlobalFoundries ecosystem. The transaction also includes Synopsys’ ASIP Designer™ and ASIP Programmer™ tools for automating the design and implementation of application-specific instruction-set processors (ASIPs).  

GF’s announcement also calls out the included ARC product lines as ARC-V, ARC-Classic, ARC VPX-DSP, and ARC NPX NPU, and says that upon closing, these assets and expert teams will be integrated with MIPS, a GlobalFoundries company.  

Synopsys: Shedding the Cores to Focus on the “Plumbing” of AI

For Synopsys, the divestiture looks like disciplined portfolio management. By offloading its processor business, Synopsys is doubling down on its leadership in interface and foundation IP.

“We are focusing our IP resources and roadmap to further our leadership in essential interface and foundation IP while winning new, high-value opportunities that advance our position as the leading provider of engineering solutions from silicon to systems,” said Sassine Ghazi, president and CEO of Synopsys.  

This focus is more than just marketing speak. As AI chips become increasingly complex, the bottleneck is rarely the processor core alone; it’s the high-speed connectivity (PCIe, CXL, DDR) and the fundamental logic libraries that enable multi-die/chiplet architectures. Synopsys is positioning itself to be the indispensable provider of the “connective tissue” that powers AI from the cloud to the edge, while continuing to dominate the EDA software market where they optimize implementations for all processor ecosystems.  

GlobalFoundries: Moving Up the Value Chain

For GlobalFoundries, this acquisition is an aggressive step toward becoming a platform provider rather than a pure-play foundry. By acquiring ARC and integrating it with MIPS, GF is building a more complete “Physical AI” stack.

Physical AI refers to the deployment of AI in the tangible world—wearables, robotics, automotive, and industrial IoT—where power efficiency and custom silicon are paramount. By owning the processor IP, GF can offer its customers more tightly integrated, end-to-end solutions, lowering the barrier to entry for companies that want to move quickly from concept to high-volume manufacturing.

“This acquisition doubles down on our commitment to advancing our leadership in Physical AI,” noted Tim Breen, CEO of GlobalFoundries. “By combining Synopsys’ ARC IP and MIPS technologies with GF’s advanced manufacturing capabilities, we are lowering the barrier for customer adoption.”  

Transaction Details and Timeline

Assets transferred: The Synopsys Processor IP portfolio includes ARC-V™ (RISC-V) and ARC® CPU IP, DSP IP, NPU IP, related software development tools including ARC MetaWare Development Toolkits, plus ASIP Designer™ and ASIP Programmer™. GF additionally describes the included ARC product lines as ARC-V, ARC-Classic, ARC VPX-DSP, and ARC NPX NPU.  

• Impact on Synopsys: Synopsys says the transaction is not material to its business and that terms are not being disclosed.  

• Closure: Anticipated in the second half of calendar year 2026, subject to customary closing conditions including regulatory approvals.  

• Continuity: Synopsys will retain and continue to grow its design IP portfolio spanning logic libraries, embedded memories, interface IP, security IP, and subsystems.  

• Continuity for customers: Synopsys and GF say they will work together to ensure Processor IP customers are supported through the transition without disruption.  

TechArena Take

The divestiture of Synopsys’ Processor IP Solutions business fits the pattern of the “New Synopsys” story arc: a company increasingly defining itself as an engineering-solutions platform from silicon to systems, especially after Synopsys completed its acquisition of Ansys in July 2025.

Layer on the NVIDIA partnership news from December 1, 2025—where NVIDIA announced an expanded strategic partnership with Synopsys and disclosed a $2 billion investment in Synopsys common stock (at a stated purchase price of $414.79 per share)—and the strategic emphasis on simulation, digital twins, and AI-accelerated engineering workflows becomes even clearer.  

For GF, this is a “Foundry 2.0” play. In a world where specialized AI silicon is the new gold, being “just” a manufacturer isn’t enough. By owning the IP (ARC and MIPS) and packaging it with software tools, GF is positioning itself to deliver more “foundry-ready” platforms—particularly for physical AI use cases where power, latency, and tight integration matter.  

The industry is watching closely. This deal consolidates ARC and MIPS under one roof. If GF can successfully integrate these teams and maintain the neutrality required to keep ARC customers comfortable through the transition, it will have carved out a serious niche in the Physical AI era.

Over the past several weeks, escalating AI storage demand and lack of supply has begun to dominate tech headlines.

Industry coverage has pointed to enterprise HDD supply tightening sharply—Tom’s Hardware recently reported enterprise drives can be on backorder for up to two years, and it also noted HDD prices rose about 4% in Q4 2025, the biggest increase in eight quarters. Reuters reported in early December that AI-driven demand is contributing to a broader memory supply crunch, with manufacturers prioritizing higher-margin products and customers scrambling for allocation.

At the same time, the NAND market is flashing its own warning lights. TrendForce forecast that NAND Flash contract prices could rise 33–38% quarter-over-quarter in Q1 2026 as memory makers prioritize server and AI-related demand. And on the supplier side, Tom’s Hardware reported (citing Nomura) that SanDisk is expected to raise enterprise 3D NAND pricing for SSDs aggressively in Q1 2026—potentially more than doubling in some cases—tying the move to AI-driven storage demand and near-term supply pressure.

That backdrop matters for news out of VAST Data this week. In a briefing, the company framed the shortage as a market inflection point—and introduced a Flash Reclamation Program designed to repurpose NVMe SSDs already sitting inside customer environments, alongside a broader push around inference key-value (KV) cache persistence aligned with NVIDIA’s Inference Context Memory Storage (ICMS) platform direction.

In the briefing, VAST co-founder Jeff Denworth positioned the company as a meaningful consumer of enterprise flash via customer deployments, and framed the market as facing compounding constraints: HDD shortfalls pushing more demand into enterprise SSDs (especially QLC), plus a fresh wave of AI infrastructure requirements.

VAST’s Two Headline Moves

VAST says it will launch a Flash Reclamation Program designed to repurpose NVMe SSDs already sitting inside customer estates—including drives currently deployed behind other platforms—so customers can stretch existing media rather than wait on new allocations. In the Q&A, VAST was explicit that this can mean pulling SSDs from existing systems and redeploying them under VAST after rapid qualification.

Second, VAST argued that inference is about to generate a new class of storage demand as context moves from GPU memory into shared NVMe tiers, enabling faster reuse of prior context for long, multi-session workloads.

That second point maps closely to NVIDIA’s own platform messaging. NVIDIA has described ICMS as a BlueField-4-powered approach intended to extend inference context memory for multi-turn agentic AI and to support high-bandwidth sharing of KV cache across systems.

Meanwhile, the “HDD delays → more flash demand” narrative continues to circulate in the channel, with DigiTimes-linked reporting (and follow-on coverage) describing extended enterprise HDD lead times and increased interest in QLC alternatives.

TechArena Take

VAST’s messaging lands because it’s not trying to create a problem—it’s trying to name one that independent sources are already surfacing.

The most revealing part isn’t the performance claims. It’s the go-to-market posture. A “reclaim the flash you already own” program is a shortage-era motion: it assumes constrained allocation, long lead times, and customers willing to tolerate disruption to free up scarce media.

On the AI side, KV cache is quickly becoming the next battleground for storage architecture narratives. NVIDIA’s ICMS framing makes KV cache persistence feel inevitable for long-context, multi-turn agents, and it creates a new category of “storage that behaves like memory.” VAST is positioning itself as the software and data-services layer around that shift—where efficiency, protection, and lifecycle controls become part of the ICMS-era value prop, not an afterthought.

In other words: the shortage story is bigger than VAST, but VAST’s response is a useful signal. When infrastructure vendors start building programs around reuse and reclamation—not just new boxes—it’s a sign the market expects constraints to persist, not clear up in a quarter.

Subscribe to our newsletter

I remember the early days of Wi-Fi, developing some of the industry’s first 802.11a/b/g transceivers. Back then, the mission was singular and remarkably simple: cut the wire.

Wireless has always evolved around its biggest pain point. First speed, then density, then IoT. Every era shifts when a new problem becomes the one we can’t ignore.

In the early years, the entire industry was engaged in a breathless race to make the air look like Ethernet. We obsessed over modulation schemes and channel widths, fighting physics to push throughput from 2 Mbps to 11 Mbps to 54 Mbps, and eventually toward Gigabit performance. Companies stacked on proprietary “Turbo Modes” and pre-standard features to squeeze out every bit and position themselves competitively.

And we won. The speed gap closed. Wi-Fi didn’t just catch wired performance at the residential edge and the enterprise edge — in many places it surpassed it.

Once raw throughput was “good enough,” the priority shifted. We moved from chasing speed to chasing density:

Can we make this work in a packed stadium?

On a subway platform in Tokyo?

In a high-rise where 200 access points sit next to and on top of one another?

That era led us to borrow techniques from cellular: OFDMA, MU-MIMO, BSS Coloring — tools to solve the wireless “cocktail party problem,” the RF equivalent of a noisy room where many devices speak at once and the network must separate overlapping conversations.

Then came the third wave: the Internet of Things. Suddenly, the devices connecting to our networks weren’t just laptops and phones; they were sensors, cameras, thermostats, wearables, industrial controllers, and all kinds of headless endpoints no one wants to update until it’s too late. The number of “things” began to outpace the number of people.

We realized that hauling all that data back to the cloud was often wasteful, so we started pushing compute outward — toward gateways, access points, and edge nodes — processing data closer to where it was created. The mindset shifted from performance to outcomes. Sensor networks don’t require much bandwidth, and no one cares what protocol they are using; they care about how the data is being used to make their lives better.

Today, we are hitting a new inflection point — one that makes the previous shifts look incremental.

In many enterprise environments, human client growth is no longer the main scaling driver. The next explosion in networking isn’t coming from people watching Netflix or scrolling Instagram. It is coming from autonomous agents. And unlike people, AI agents do not forgive “best effort.”

To see why, imagine a modern fulfillment center. Not humans pushing carts, but a hive of hundreds of Autonomous Mobile Robots weaving past each other at speed. Each robot negotiates right-of-way with a central controller, with safety systems watching for conflicts — a single distributed organism connected by an invisible wireless tether.

If that tether stretches into a noticeable hiccup — tens of milliseconds in the wrong moment — the system doesn’t “buffer.” It stops. A momentary disruption becomes a full-aisle shutdown. This is where “best effort” becomes a business risk rather than a minor annoyance.

The New Consumer: Human vs. Machine

To understand why the network architecture must change, you have to understand the difference between a human user and an AI agent.

Humans are incredibly adaptive. If you are on a Teams call and the video freezes for 500 milliseconds, you might grimace and cry out to your deity of choice, but your brain fills in the gap. If a web page takes an extra second to load, you wait. We are built to tolerate variance. Our networks were designed around this tolerance; we built best-effort systems that prioritized maximum throughput over consistent timing.

AI agents (robots, autonomous logistics bots, digital twins, and XR interfaces) are not adaptive in the same way. They require precision.

If a warehouse robot loses reliable connectivity at the wrong moment, it doesn’t “buffer”; it performs a safety stop. If an XR experience slips into noticeable lag, the user gets disoriented, or nauseous (“clean up on aisle 3”). These “users” don’t care about peak speed. To an AI agent, performance isn’t measured in gigabits per second; it’s measured in bounded variance.

Determinism means engineering to strict upper bounds on latency, jitter, and packet loss, and then meeting those bounds every time. “Good” is no longer a high average throughput. “Good” is the mathematical guarantee that 99.9999% of packets will arrive within a fixed window (e.g., 10 ms), regardless of RF congestion, multipath, or compute/buffer delay.

We are moving from an era of bandwidth to an era of determinism.

Wireless as the AI Nervous System

If the modern data center — with its massive GPU clusters — is the brain of the AI revolution, the wireless edge is the nervous system.

A brain in a jar is useless. To function, intelligence needs sensory input from the physical world. It needs to know who is in the room, where the asset is, what the environmental context is, and what the expected action (intent) will be.

This is the new mandate for the wireless edge. We must pivot from building “dumb pipes” that simply move data to building a sensory fabric that feeds context and intent to the enterprise AI.

This shift requires three fundamental architectural changes.

1. Determinism over throughput

We need to stop marketing “fast” and start engineering “predictable.” The industry is acknowledging this reality, and Wi-Fi 8 is shaping up to emphasize ultra-high reliability in hostile RF environments, not just another massive jump in peak PHY rate.

This is a tacit admission that the race for raw speed is no longer the primary battle. The future of wireless lies in scheduling the air with the same seriousness we apply to wired switching: prioritization, admission control, traffic classification, roaming behavior that doesn’t spike tail latency, and continuous measurement of what the network is actually delivering.

Whether via private 5G or reliability-focused Wi-Fi evolution, the network must support SLA-like behavior for latency-sensitive machine traffic. For network designers, this flips the planning model: instead of asking “How fast can we make it?” we now ask “What is the worst-case delay this robot, vehicle, or agent can survive?” Determinism becomes the budget we engineer around.

2. Identity as the control plane

In a world of autonomous agents, the distinction between “Wi-Fi” and “cellular” is often a distraction. The agent doesn’t care about the protocol; it cares about the outcome. We need a unified identity layer that can abstract away the radio physics.

A security robot moving from the parking lot (5G) into a warehouse (Wi-Fi) shouldn’t experience a policy gap. The policy must follow the identity, not the port.

In practice, this means policies can no longer live primarily in VLANs or subnets. They must live with the identity itself — tied to a device, workload, or agent — and remain consistent as it roams across spectrum, transport, topology, and physical location.

3. The network as an immune system (distributed zero trust)

When humans click on phishing links, we train them to be better. You cannot “train” an infected thermostat or a compromised sensor. As we flood our networks with headless devices, the attack surface expands exponentially.

Security can no longer be a perimeter overlay; it must be intrinsic to the fabric. In this model, the chain of trust starts at the edge. The access point stops being a passive pipe and becomes an enforcement point: identity-based segmentation, continuous verification, and rapid containment at the first hop.

Architecturally, the edge is no longer a passive on-ramp; it is the first line of defense that can shrink blast radius immediately and feed high-fidelity telemetry into centralized policy and response.

The Opportunity

We spent the last 20 years building networks that were excellent at delivering content to people. The next 20 years will be about building networks that deliver context from the physical world to AI models.

This is not just an upgrade cycle. It is a fundamental reimagining of why we build networks in the first place. The edge is no longer just about connectivity. It is the sensory interface for the AI era.

If you’re a network or infrastructure leader looking at this shift, the key question isn’t “how fast can the wireless network go?” The question is: can we support real-time, deterministic applications? Can we make policy follow identity across domains? Can we contain threats where they originate, not after they spread?

The technology to build this exists today. The “things” are already here. The agents are waking up.

We are done designing for human patience. Now, we must build the nervous system for machine precision. The 'Best Effort' era is over. The Deterministic era has begun.

Rose-Hulman Institute of Technology shares how Azure Local, AVD, and GPU-powered infrastructure are transforming IT operations and enabling device-agnostic access to high-performance engineering software.