scx.ai logo

Event Recap

I Visited Computex 2026 — Here's What I Got Right and Wrong

David Keane reflects on his pre-Computex predictions after walking the show floor in Taipei. The ASIC wave is real, sovereign AI is table stakes, and the era of GPU monoculture is over.

By David Keane7 min read

Well — how did I do?

I published six predictions before Computex 2026. Then I walked the show floor in Taipei for four days and tested them against what was actually there.

Here's the honest scorecard.

1. Inference is eating the data centre ✅

What I said: The training vs. inference debate is over. The dominant share of compute demand in production AI now lives in inference. The metric everyone should be talking about isn't FLOPs or GPU count — it's tokens-per-watt.

What happened: For the first time in 2026, over 55% of global AI infrastructure spending shifted from training to inference. The AI inference market is now valued at USD 20–117 billion depending on how you count, growing at 27–43% CAGR. Deloitte called it "the year of AI inference." Every major infrastructure announcement at Computex was inference-first.

An easy win — it was already happening.

2. The CPU is back in the conversation ✅

What I said: Intel's Lip-Bu Tan would make the case that CPUs are a serious inference engine — not just orchestration, but actual inference work. The GPU:CPU ratio was shifting from 8:1 to 4:1 — and could invert.

What happened: Lip-Bu Tan keynote'd on June 2 with exactly that message. Intel unveiled Xeon 6+ (Clearwater Forest) — 288 efficient cores on Intel 18A, purpose-built for agentic inference. Their executives spelled it out: agentic AI tasks require significant CPU orchestration for coordinating tasks, retrieving information, accessing files, and interacting with software tools.

I wasn't expecting that level of density — does that mean even more opportunities to use the CPU as the orchestrator at lower costs than before? Probably.

3. SambaNova SN50 changes the geometry of the problem ✅

What I said: The SN50 delivers 5x lower latency and 3x the throughput of NVIDIA Blackwell B200 on agentic inference patterns, at a fraction of the power draw.

What happened: Intel announced new rackscale AI infrastructure built with Xeon processors and SambaNova SN50 RDUs, with Foxconn providing system integration. They described a disaggregated inference blueprint — GPUs for prefill, SambaNova RDUs for decode, Xeon for orchestration — boosting per-user token output by 2–3x compared to GPU-only systems.

A new enterprise inference cloud called Vector Core Compute was launched by Vista Equity Partners and Cambium Capital, running the same architecture.

The SN50 is shipping later in 2026. And yeah — I had a feeling that would happen. 🙂 Good to get this one right.

4. Power efficiency is the real bottleneck ✅ (with a twist)

What I said: The limiting factor isn't silicon — it's thermal and electrical. The companies that win the next phase will deliver more tokens per watt.

What happened: Confirmed — but better silicon solves this by design, not by accident.

A GPU-based AI rack draws ~140kW and needs water cooling. Our ASIC-based SCX setup runs at ~10kW. Same inference throughput. One-tenth the power.

SambaNova's SN50 delivers 3x the efficiency of NVIDIA's B200 on agentic workloads — under 20kW per rack, air-cooled.

You don't need a 110kW power shelf if your rack draws 10kW. You don't need liquid cooling if your chips were built for air-cooled environments from day one.

This one has to be right for us to scale AI Factories in markets that don't have unlimited power and water — and for markets that care about the environmental impact of the AI transition.

5. The ASIC wave is real ✅

What I said: Custom silicon shipments would hit nearly 27.8% of the AI chip market in 2026, growing at 44.6% year-on-year — nearly triple the growth rate of merchant GPUs.

What happened: Custom silicon shipments hit 27.8% of the AI chip market in 2026, growing at 44.6% year-on-year. Broadcom has a $73B AI chip backlog targeting $100B in annual revenue by 2027. Google is retiring x86 for custom TPUs with integrated Axion CPUs. Meta's MTIA v3 enters production mid-2026. TSMC projects 60% CAGR for AI chip revenue through 2029.

Every hyperscaler is building their own. The ASIC wave is here.

SambaNova's SN50 is the concrete proof point. Sixteen SN50 RDUs per rack, under 20kW, air-cooled, with massive industry-leading performance. That's the architecture SCX runs on: inference-native, purpose-built silicon.

The wave wasn't predicted. We were already surfing it. So I don't really get a win there.

6. Sovereign AI is the use case that makes all of this concrete ✅

What I said: Sovereign AI would be the frame that brings everything together.

What happened: Every hyperscaler is now talking about sovereignty. NVIDIA unveiled sovereign AI infrastructure initiatives across Southeast Asia and Latin America. DDN highlighted policy-based control and secure multi-tenant isolation for sovereign deployments. Phison launched a dedicated sovereign AI program.

All this is now table stakes. SCX built inside this frame before it was fashionable. Australian data jurisdiction. No CLOUD Act exposure. Air-gapped inference. Purpose-built for regulated industries — banks, government, critical infrastructure — that can't send sensitive workloads offshore.

What surprised me

The heterogeneous compute narrative. I expected it to be a conversation amongst the "AI nerd crowd" — but every major vendor at Computex was talking about it: the future is multi-silicon, workload-specific, and disaggregated.

The era of GPU monoculture is over.

What this means for SCX

We are at the right place, with the right partners, at the right moment.

Sovereign AI infrastructure — inference-first, heterogeneous, power-efficient, and purpose-built for agentic workloads — is exactly what we've been building. Computex 2026 gave me a real feeling that we have this right.

Now it's all about execution.

The days of treating AI infrastructure as a GPU-only story are over. We said that before Computex. The industry said it at Computex.

Make of that what you will.

Try it now at SCX.ai — we are giving away 1M tokens a day to all free accounts to see the performance of our system in your real-world applications and agents.

Give it a go.

Related Topics

Computex 2026SCX.aisovereign AISambaNova SN50ASICinferenceAI infrastructureTaiwan
I Visited Computex 2026 — Here's What I Got Right and Wrong