Cerebras at $100B mcap — the inference bet pays out
Cerebras closing day one at $100B market cap is the public market's verdict on a question that was contrarian eighteen months ago: would inference become the dominant compute workload, and would wafer-scale architectures take share from NVIDIA's training-rack model? The answer to both is now yes — and the second-order effects are starting to reshape every comp in the stack.
Two facts you have to hold simultaneously to understand the Cerebras IPO. First: NVIDIA's Q1 fiscal-2027 data-center revenue was $75.2 billion, accounting for 92% of total sales, doubled year-over-year. The absolute dominance is intact. Second: the stock slid on guidance. The market priced the dominance correctly and then sold the stock anyway, because the inference-mix question is now bigger than the training-rack story.
Cerebras closing day one at ~$100B mcap is the other half of the same signal. The wafer-scale-for-inference thesis was a contrarian position through 2023-2024. The OpenAI $20B / 750MW deal in April moved it from contrarian to credible. The 68% Day-1 pop moved it from credible to consensus. Three quarters from contrarian to consensus is fast even for AI infrastructure.
What the public market is pricing
The market is pricing two structural shifts. First: inference will be larger than training, by a wide margin, by 2028. The Goldman 24× token-consumption forecast is the model that supports this view; the agentic-AI workload is what drives the inference share up; the per-token cost economics are what determines which chips win. Second: the inference market has a different competitive shape than the training market. Training rewards raw FLOPS; inference rewards $/token at scale, which depends on memory bandwidth, inter-chip data movement, and architecture-specific power efficiency — all areas where wafer-scale and LPU-class designs structurally outperform GPU-cluster topologies.
NVIDIA understood this; that's why the $20B Groq acqui-hire happened in February. The company that owns the training market is buying option-value on the inference market. But owning every promising inference architecture is not a viable defense; the procurement market will pick winners, and the OpenAI-Cerebras deal is the public-facing example of how the picking happens.
The chips-that-aren't-NVIDIA cohort
Cerebras, Groq (now under NVIDIA), AMD Instinct, Tenstorrent, SambaNova, and the China-domestic Huawei Ascend / Biren BR lines all compete on the same inference-market thesis. The differentiation is architectural: wafer-scale (Cerebras), language-processing-unit racks (Groq), MI-300/350 chip-and-software stack (AMD), RISC-V matrix accelerators (Tenstorrent), reconfigurable-dataflow (SambaNova), domestic-China substitution (Huawei, Biren). Each is betting that its specific architectural choice wins the $/token contest at scale.
None of them needs to take more than 15-20% of the inference market to be worth tens of billions. The Cerebras Day-1 valuation prices that outcome as already plausible for one vendor; the same outcome is plausible for two or three. The chips-that-aren't-NVIDIA cohort will probably aggregate to 30-40% of inference compute by 2028, with NVIDIA retaining the training rack monopoly and a still-large but no-longer-dominant inference share.
What this changes for the lab IPOs
The OpenAI IPO that filed its draft prospectus this week is going to be priced against this backdrop. Investors will read the S-1 with a model that includes "inference-cost-per-token in 2028" as an explicit input. The Cerebras-OpenAI deal disclosed there will be one of the most-scrutinized line items. Frontier-lab IPO valuations have always depended on assumed future inference economics; now there's a public-market reference price for the inference-architecture half of that assumption.
CNBC — Nvidia competitor Cerebras after wild IPO → · CNBC — Nvidia data center revenue doubles Q1 2027 → · eWeek — Cerebras targets $33B IPO challenging Nvidia →