SemiAnalysis' Doug O'Laughlin: Cerebras has a narrow path, Groq fits NVIDIA's GV200 rack, and TSMC is the AI buildout kingmaker

Doug O'Laughlin on AI Chip Dynamics, TSMC Supply, and the AI Buildout

Doug O'Laughlin of SemiAnalysis sees a compute shortage that is reshaping the chip landscape in ways nobody fully predicted — and the winners are not always the best-engineered products.

Cerebras: a narrow path

Cerebras has a viable but constrained role: inference at very high speeds for large models but only at small context window sizes, or for smaller models overall. The fundamental problem is architectural. The chip is an island — exceptional at compute in the middle of the workload, but moving data off it is hard. That limits how it integrates into the broader inference stack.

The market is big enough that Cerebras benefits from overflow demand, but O'Laughlin frames it as a solution that found its problem somewhat by accident, given that LLMs scaled far faster than anyone anticipated.

Groq inside Nvidia's GV200 rack

Groq fits into a specific and valuable seam in transformer inference. Prefill — loading all the weights — is compute-constrained, so it doesn't need much memory bandwidth. Decode is the opposite: extremely memory-bandwidth-limited. The GV200 rack can pass activations over to Groq's LPU, which runs on fast on-chip SRAM. Because the data is streaming activations rather than requiring a large scale-up topology, the IO constraints that hurt Cerebras don't apply. O'Laughlin describes Nvidia's move here as fitting Groq into a disaggregated inference pipeline, with the GV200 as the host rack.

Whether this becomes a 50% or 80% deployment pattern is genuinely unclear to him — the optimal configuration depends on model architecture and parallelism strategy, and those vary.

“In a shortage, it's not the best company who wins — you can look at NVIDIA's stock chart and that tells you. It's the second, third, fourth best companies where demand overflows. TSMC is kind of a kingmaker in terms of supply. There's no reason for them to let the market go out over its skis. The biggest bottleneck for data centers is power, so you can just move the data center to power and then hook it up with fiber.”

AMD and Intel

AMD's near-term priority is getting rack-scale interconnect working on the MI350. O'Laughlin thinks Lisa Su will get it right. He also wouldn't be surprised to see AMD pursue a fast SRAM offload chip for the feed-forward network layer within the next twelve months, though the field of viable candidates is small.

Intel's stock price is running ahead of its technical turnaround, in O'Laughlin's view. Lip-Bu Tan has stabilized the company and brought in the right people. The government-mandated customer deals — he references Amazon and Elon Musk's xAI — amount to top-down demand creation that Pat Gelsinger couldn't achieve from the bottom up in three years. Intel 14A is good enough given how tight TSMC's N3 capacity is, but execution risk remains real given Intel's history.

TSMC as kingmaker

TSMC is the binding constraint on the AI buildout, and O'Laughlin doesn't expect it to change fast. Clean rooms take roughly three years to bring online. Two years ago, TSMC would have needed near-perfect foresight to match today's demand — and that foresight didn't exist. The result is a structural lag where supply persistently trails demand, wafer pricing moves up, and TSMC responds incrementally rather than in step-change fashion. It is growing CapEx around 40%, but in absolute dollars those are already large numbers, and Taiwan is running short of semiconductor engineers.

That supply ceiling is quietly good for Intel, whose fab capacity absorbs overflow. Every unused clean room globally — old power fabs, display fabs — is being snatched up and retrofitted.

On the Elon Musk-Intel partnership, O'Laughlin thinks Musk is capable of compressing timelines, but by the time a new fab is ready, the broader supply response will likely have already partially closed the gap. He gives Musk credit for execution speed while remaining skeptical the timing will be decisive.

Data center siting and space

Municipal pushback on data centers is real but ultimately loses to economics. Power is the dominant site-selection variable now, not proximity to population centers. Two fiber pairs are sufficient to put a data center anywhere near cheap power, and O'Laughlin expects more buildout in the middle of nowhere as the ROI math increasingly favors it over urban density.

Space data centers don't pencil out in the near term — a pound of hardware costs roughly ten times more to put in orbit. The long-run case only works if the AI compute market grows so large that even 1% of it justifies a specialized off-planet supply chain. The more probable overflow valve for US zoning friction is Latin America, particularly Brazil, which has significant available power.

AI bubble or not

O'Laughlin is mildly concerned about a bubble but doesn't see the demand as fabricated. GPU prices are up, tools like Claude and Cursor are genuinely useful to him daily, and revenue across the stack looks real. His framing on valuation is that lab revenue multiples are orders of magnitude below dot-com peak multiples, and hyperscaler price-to-earnings ratios remain reasonable even under heavy CapEx. The pushback he acknowledges is that earnings can be engineered while free cash flow cannot — and on a free cash flow basis the picture is messier.

His longer view is that AI is likely bigger than the internet, a conviction he says he didn't hold two years ago. The scale of institutional disruption, including how governments relate to AI systems, is the underlying force driving sovereign AI investment — not the narrow question of whether French users need a French LLM.

Robotics he treats as much closer to the self-driving car trajectory than to AI software: definitely coming, but slower and less linear than the hype implies. The key difference between AI software and robotics is distribution — AI runs on infrastructure that already exists and can spread instantly over the internet.