I did not have “Attend Computex in Taipei to go deeper on AI infrastructure” on my bingo card heading into 2026. And when my friends at Intel asked if I’d like to attend to see what’s going on and what’s new, I was a little hesitant, as I’ve always viewed Computex as very consumer-ish. But based on what I was hearing, I took the trip.

AI dominated. This is not news. We all expected to hear about how AI is influencing people’s lives through the devices they carry and wear. But this went beyond the consumer, with a big focus on the datacenter technologies that drive these user experiences.

And on this topic, my belief has been reinforced that the industry is moving away from a single architecture, deployment model, or silicon provider. Rather, as we see AI activated in our lives through ubiquitous inference we will see a divergence. From the silicon to the system, these platforms will be more specialized, fragmented, and workload-specific.

That fragmentation isn’t a sign of market confusion. It reflects how enterprises actually deploy technology.

For as long as I’ve been in the IT solutions game, infrastructure decisions have largely been centered on performance, consolidation, and cost. AI introduces a much broader set of variables. Latency, power, scalability, cost, data sovereignty, operational simplicity — they all matter. More importantly, every workload places a different weight on those variables.

That’s the context for what I was looking for at Computex, and how my expectations squared with what I saw and heard.

The Datacenter Becomes the AI Activation Point

The clearest theme at Computex was the continued shift toward viewing the datacenter as the activation point for AI. Yeah, models are cool and capture headlines, but infrastructure is increasingly where competitive advantage is created.

There is no better proof point than from the custom chip designer Marvell. The company wasn’t simply another keynote presenter. Its presence reflected the importance of infrastructure plumbing in the AI era. When a company known for building the foundational silicon that enables AI infrastructure takes center stage, it suggests the industry is moving beyond fascination with models and toward a focus on deployment.

And when NVIDIA CEO Jensen Huang joined Matt Murphy (Marvell CEO) on stage to proclaim it the next trillion-dollar company, it crystallized that this was not a conference about algorithms. It was a conference about the systems required to operationalize them.

AI Specialization Is Now Mainstream

Another notable theme was the growing understanding and acceptance that specialized compute architectures will coexist for the foreseeable future. Whereas the datacenter of just a couple of years ago was largely (x86) CPU-centric, with GPUs for technical computing, analytics, or VDI workloads, the current landscape is shifting. Arm and RISC-V are being used to deliver more finely tuned accelerators and CPUs that move, process, store, and secure data. Companies like Marvell, Broadcom, and Qualcomm are joining the discussion. And relative newcomers like Cerebras ($101 billion IPO), Tenstorrent, and Furiosa are also becoming part of the mainstream discussion, not being treated as niche deployment scenarios.

Besides partaking in the Marvell keynote, Jensen Huang joined Arm CEO Rene Haas on stage. This was pretty symbolic — not because NVIDIA and Arm suddenly became partners, but because it demonstrated how large the opportunity has become. AI infrastructure is no longer a winner-take-all market. The message to the market was that Vera (the NVIDIA datacenter CPU) and AGI CPU (the Arm datacenter CPU) can coexist and thrive in this AI market.

As I watch this dynamic unfolding, I can’t help but think back to the early days of x86, when Intel went from PC CPU maker to a dominant datacenter architecture player. When first introduced, x86 had limited play (file and print). Within a few years, it was running the most demanding workloads.

Similarly, Arm continues to build a compelling case that it can become a primary compute architecture rather than a complementary one. Its keynote consistently reinforced the message that Arm is becoming increasingly central to the broader compute equation.

For me, the client remains the open question. At NVIDIA’s GTC event (running in parallel with Computex), the company launched its N1 CPU — targeting mainstream clients and laptops. While Huang joked about “finally getting it right,” I am curious to track adoption of this processor over time.

I compare this with Qualcomm’s experience a couple years ago launching its Snapdragon X — a very capable client chip that checks all the boxes around performance, battery life, and user experience. And maybe more importantly, Qualcomm did the heavy lifting that hardened Arm for Windows. Its challenge has been SKU selection and launching a platform that moved in large volume. (It entered the market at the very high end, where the volumes are lower).

I think NVIDIA will be able to move the market further toward Arm-client Windows adoption — and Qualcomm would benefit from this. But it’s important to note the foundational work done by the folks at Qualcomm.

Side note: Keep an eye out for coverage of Computex from my colleague Anshel Sag. He always has the most interesting and fun coverage on client, handheld, and other device types.

Pressure Builds on Traditional x86

Computex also reinforced a broader industry trend that deserves attention. For the first time in decades, x86 faces credible pressure from multiple directions simultaneously. Arm continues to gain momentum. Custom silicon is becoming more prevalent, with RISC-V alongside Arm. AI accelerators are taking on workloads once handled by traditional processors. Specialized inference architectures continue to emerge.

None of this suggests the end of x86. And to say that it does would be a little … silly. Enterprise infrastructure changes far too slowly for that narrative to hold. And when it comes to supporting the diversity of workloads in a datacenter, x86 remains unmatched.

What it does suggest is that the industry is entering a period in which x86 must compete in more places than it has in the past. And that x86 will have to coexist with these specialized chips and accelerators.

Intel’s announcements reflected that reality. The company didn’t really deliver a headline-grabbing moment, but that seems to have been intentional. Xeon 6+, built on the Intel 18A process node with up to 288 cores, represents meaningful execution against the company’s roadmap. And the rack-scale architecture announcements were directionally positive, while the Intelligence Center positioning provided an interesting framework for thinking about future datacenter deployments.

I see this approach as the company recognizing that the datacenter–AI datacenter landscape is shifting. And rather than trying to own every chip in every server, it will be part of a bigger silicon universe — an unfamiliar but real future state. And, of course, its goal is to gain as much relevance (and market share) as possible in this heterogeneous world.

Perhaps most importantly, Intel appears to be approaching AI infrastructure with greater discipline than it has in recent years. The company is making progress without overpromising. That approach feels familiar. AMD followed a similar path during the early EPYC resurgence, focusing on targeted execution rather than broad market claims.

NVIDIA Treats the CPU as Strategic

Although NVIDIA largely reserved its major announcements for its GTC show, its influence at Computex was impossible to miss. Objectively, the company remains the benchmark against which much of the industry measures itself. Yet the most interesting NVIDIA announcements were not about GPUs.

Vera continues to gain momentum as NVIDIA expands its CPU strategy. The company provided additional detail around performance and customer adoption (including by Oracle Cloud Infrastructure), reinforcing the view that CPUs remain strategically important even in AI-centric environments.

This is especially true in agentic AI, where we see the CPU taking on a far more significant role and CPU-to-GPU ratios beginning to normalize. During the first wave of generative AI, CPUs often functioned primarily as host processors for increasingly large accelerator deployments. Agentic AI introduces additional responsibilities around orchestration, retrieval, tool execution, and workflow coordination. While it is too early to predict where CPU-to-GPU ratios will ultimately settle, I believe CPUs will become more important than many expected as AI systems become more operationally complex.

On the client side, N1 highlighted NVIDIA’s ambitions. As previously mentioned, while Qualcomm deserves significant credit for establishing the Windows-on-Arm foundation, NVIDIA may ultimately be the company that accelerates mainstream adoption by connecting client experiences directly to the broader AI ecosystem it already dominates.

Taken together, these announcements suggest NVIDIA increasingly views CPU strategy, client strategy, and AI strategy as inseparable.

Which raises an interesting question for me: What exactly is a CPU now?

When Is a CPU Not a CPU?

Maybe it’s me, but the industry’s terminology increasingly feels disconnected from the hardware itself. The industry still uses CPU, GPU, accelerator, DPU, NPU, and XPU labels, but the functional boundaries between those categories are increasingly blurred.

When many people think of CPUs, they still think of traditional x86 processors serving as the central point of orchestration within a system. And they think of a processor that supports virtually every workload in the datacenter.

Modern AI systems challenge that definition. Arm and NVIDIA have demonstrated that processors can play far more substantial roles than simply feeding accelerators. And both companies made architectural changes to optimize support for AI and cloud computing in terms of orchestration, retrieval, and coordination.

Likewise, modern GPUs increasingly look unlike the graphics processors from which they inherited their names. Rather, these are chips that are becoming increasingly bespoke for AI workloads. In fact, we saw AMD go so far as to deliver the MI300A for HPC and AI, and the “X” model for AI-first environments.

Perhaps the labels still matter. Perhaps they don’t.

What does matter is the broader trend. Compute is becoming increasingly heterogeneous. The boundaries separating CPUs, GPUs, accelerators, and custom silicon continue to blur as vendors optimize for specific AI outcomes rather than traditional compute categories.

Fragmentation Is a Feature, Not a Problem

This shift helps explain why so many new silicon vendors continue to emerge.

Enterprises will evaluate AI infrastructure differently depending on workload requirements. Some deployments will prioritize performance. Others will prioritize power efficiency. Some will focus on scalability. Others will focus primarily on economics. But no single architecture will optimize for every combination of requirements. And this reality creates space for multiple winners.

Companies like Furiosa put a spotlight on this trend. Its partnership with Broadcom is noteworthy. Not necessarily because of the technology involved, but because Broadcom rarely invests without conviction. And by working with this Korean startup, it is validating another player for the hyperscaler markets (and beyond).

In any case, the partnership clearly suggests that meaningful opportunities continue to exist outside the dominant architectures. And the implication is larger than any single company. Fragmentation is creating room for an expanding ecosystem of specialized providers.

The Distributed, On-Prem Future

I think the most interesting demonstration I saw at Computex was not a massive AI factory. It was Intel’s SuperClaw demonstration.

In a workstation-class environment, the system used four Intel Arc B70 GPUs and supported agentic workflows through a distributed operating model spanning client devices, local infrastructure, and cloud resources.

The economics were compelling. Four GPUs costing roughly $1,000 each delivered capabilities that many organizations could realistically deploy at the departmental level.

To me, this may ultimately be closer to the future of enterprise AI than many of the hyperscale-focused narratives dominating the industry today. Most organizations are not building AI factories. They’re looking for practical ways to deploy AI where work actually happens. That means smaller models. Distributed architectures. Local execution. Department-level deployments. Hybrid operating models.

The technology required to make that happen already exists. What is still missing is the packaging. And I think the company that successfully turns these components into an easily consumable enterprise solution may unlock one of the largest opportunities in AI infrastructure.

The Success of AI Requires Diversity

The takeaway from Computex is not that NVIDIA wins, AMD wins, Arm wins, or x86 loses. And it’s not that there is some clear predictable path forward.

No. To me, the big Computex takeaway is that AI infrastructure is fragmenting faster than many anticipated.

Different workloads require different architectures. Different enterprises require different deployment models. Different economic constraints produce different optimization decisions.

The next phase of AI infrastructure will not be defined by convergence around a single answer. It will be defined by the growing number of viable answers available.