They built a chip the size of a dinner plate. The hard part was never the silicon.

You are buying speed, from one customer, at a price that assumes 2028.

Cerebras built the fastest inference chip in the world by putting the whole model on a single wafer. That part works — OpenAI runs its fastest model on it in production, and the order book stands at $25.4B. The technology risk is largely retired.

What replaced it is customer risk and price risk. Two customers are 66% of revenue, the backlog is overwhelmingly one counterparty, and at 60x sales this is the most expensive name in AI silicon by nearly three times. Meanwhile Nvidia paid roughly $20B in December to buy the competing version of this architecture and now sells it as a tier inside its own platform.

What you actually own. The whole filing cabinet on the desk.

Every other chip company slices a 300mm wafer into hundreds of small chips. Cerebras doesn't slice — the entire wafer is one processor.

Why that matters: when an AI writes you an answer, the bottleneck isn't thinking, it's fetching. A normal GPU is a fast worker who keeps walking across the room to a filing cabinet. Cerebras built a worker with the whole cabinet on the desk. Less room for files, but no walking.

The result is speed. A token is about three quarters of a word, so 750 tokens a second is 560 words every second — against the 4 words a second you read at. It writes 142x faster than you can read.

On small jobs that gap is invisible. Scale the job up and it becomes the whole story.

They stopped selling machines and started renting them.

Two lines: hardware, where they sell you a box, and cloud, where they own the box and bill you for time on it. Hardware fell 23% last quarter while cloud grew 281%. That shift is deliberate — recurring revenue at better margins — and it is why capex is running above $500M a quarter.

But read the pivot the other way too. One semiconductor analyst put it bluntly: running your own token factory with your own hardware is not where a chip company wants to be, and you only do it if people aren't buying the hardware directly. Groq walked the same path into GroqCloud before Nvidia absorbed it.

The SeaMicro band, back together.

Founded around 2015 in Sunnyvale by five people who worked together at SeaMicro. CEO Andrew Feldman co-founded that company and sold it to AMD in 2012 for roughly $334M. CTO Sean Lie is a co-founder. CFO Bob Komin joined in April 2024 — he took Sunrun public and sold three companies to Yahoo, Pandora and Microsoft. An exit CFO, hired two years before the listing.

Feldman holds about 14.1M shares and Lie about 8.4M, almost all Class B carrying 20 votes each. Founder control is intact.

The scary number is mostly an accounting artifact.

GAAP gross margin fell from 31.1% to 14.2%. Everyone read that as the business breaking. Here is the actual bridge.

A warrant granted to OpenAI is booked as a reduction to revenue rather than as a cost. That single item is $44.3M in the quarter and explains essentially the whole 17-point drop. On the core basis, margin actually improved 940 basis points.

It is also structurally perverse: the faster OpenAI buys, the bigger the charge and the worse GAAP looks. The asset is $1.1B and runs to October 2031 — roughly 25 more quarters of this.

The risk moved from Abu Dhabi to San Francisco.

The top customer fell from roughly 87% of revenue in early 2024 to 34%. Genuine progress. But two names are still two thirds of the business, and filings show OpenAI revenue of $56.8M in the quarter — 31.5% of the total.

And OpenAI is not just a customer. It is simultaneously the largest customer, the holder of most of the $25.4B backlog, a $1B lender whose loan is forgiven if repaid in compute rather than cash, and a warrant holder for ~33M shares at $0.00001. No cash has been repaid on that loan — it is being worked off in billing credits.

TSMC holds every card.

Cerebras depends on one foundry, on 5nm. The real lock-in isn't the relationship — it's the reticle-stitching process, co-developed with TSMC over roughly a decade, which independent analysts say does not port to another fab. There is no second source at any price.

Worse, the filings state manufacturing is bought on a purchase-order basis with no capacity or volume commitments. Management says they've secured the wafers they need. That is an assertion, not a contract.

The offsetting advantage is real though. Cerebras uses no HBM, no CoWoS packaging and no leading-edge node — the three tightest bottlenecks in AI silicon. As Feldman puts it, the constraints facing the industry don't apply to them.

The order book is enormous and very slow.

Backlog is roughly 29x guided revenue. In any normal industry that's extraordinary. But only about a fifth lands inside two years, so it is contracted and a long way from the income statement.

The constraint is Cerebras's own — over 600MW live or contracted, and they're renting systems back to serve demand they can't host. Which means a new customer doesn't unlock revenue. It joins a queue. That is the flaw in the argument that one big deal re-rates this stock.

Where the demand leaks is straightforward: any latency-sensitive workload Cerebras can't serve on time gets served by Nvidia instead.

Nvidia already bought the other version of this idea.

In December 2025, Nvidia licensed Groq's inference technology and hired its team including founder Jonathan Ross — reported at roughly $20B. Groq survived, re-based at a $3.5B valuation down 49%, and now resells Nvidia GPUs as a cloud partner. It is no longer a chip company.

Nvidia now ships that IP as NVIDIA Groq 3 LPX, the seventh chip of the Vera Rubin platform, claiming 35x higher throughput per megawatt on very large models. It arrives 2H 2026.

Note what Nvidia changed. LPX is not memory-on-chip only — each rack pairs 128GB of fast memory with 12TB of conventional memory and runs alongside GPUs. Nvidia didn't accept the pure approach. It turned it into a tier inside its own platform, which is far harder to compete with than a rival chip.

Qualcomm is the 2028 threat, not the 2026 one — and note their Meta deal is a CPU agreement contributing nothing until 2028, which a lot of coverage gets wrong.

And the Meta question everyone is asking.

Meta already works with Cerebras — it powers fast inference inside the Llama API, announced at LlamaCon, hitting about 2,600 tokens/sec on Llama 4 Scout. But read what that is: Meta is reselling Cerebras speed to outside developers, not buying compute for itself. Meta is a distribution channel, not a customer.

The proof is in the filings — OpenAI, G42, MBZUAI and AWS are named. Meta is not. Mizuho thinks Meta could be announced as a full customer in 2H 2026, which would make it the third major cloud customer. That is a forecast, not a signed deal.

The complication: Meta is spending $115-135B in 2026 building four generations of its own inference silicon with Broadcom, and describes that program as inference-first. A large external contract would land in the same window it ramps its own chips.

Three reasons the technology may not be as strong as it looks.

Worth knowing the precedent. Gene Amdahl tried wafer-scale at Trilogy Systems in 1980, burned roughly $230M, and shipped nothing — killed by yield, lithography and layers delaminating under heat. Cerebras solved that with far finer redundancy: about a million physical cores of which 900,000 are exposed, so bad ones are simply mapped out.

The most expensive name in AI silicon, by nearly three times.

On the FY26 guide the multiple is ~46x. On FY27 consensus of $2.95B it's 13.9x. On Morgan Stanley's contracted path to ~$6B by 2028, roughly 6.8x. To sit at the silicon median today, Cerebras needs about $1.85B of revenue against an $885M guide.

You are paying today for 2027 to 2028 execution. This is not a 2026 story at any price. The fair offset: consensus three-year growth is 145% versus 48% for Nvidia, so growth-adjusted it's less absurd than it looks.

Fair value, on FY27 revenue.

The calendar, in order.

Sitting right on the shelf.

CBRS daily. Rounded base off the $160.81 low, price back above both moving averages, and heavy volume on the August advance.

The structure is a clean rounded bottom and the volume profile supports it — distribution into the June lows, a long quiet middle, then rising volume on the advance. What it hasn't done is reclaim $265. More urgent is the downside: at $202 price is directly on the $200 shelf. Hold it and the base is intact. Lose it and the next tested floor is 20% lower.

Expensive, and worth it.