Cerebras IPO AI chip NVIDIA comparison: the Q2 math

September 29, 2026

The verb in the brief is the first thing to retire. Cerebras did not outpace NVIDIA in the quarter that followed its IPO. Cerebras reported GAAP revenue of $180.1 million for the three months ended June 30, 2026. NVIDIA reported $96.2 billion for the quarter ended July 26, 2026. Those are not the same 90 days. They are the nearest filed quarters, and the revenue ratio is about 534 to 1.

What Cerebras did do is close a greenshoe IPO, put about $6.2 billion of net proceeds on the balance sheet, and post a cloud line that nearly quadrupled. The Wafer-Scale Engine is a real machine: 44 GB of on-chip SRAM and 21 PB/s of memory bandwidth on the company datasheet. The 125 petaFLOPS figure on the CS-3 spec sheet is footnoted as sparse. This Cerebras IPO AI chip NVIDIA comparison is a filing scorecard for people who buy tokens or a CS-3, not a stock call.

Four gates. Stop at the first no. A yes on all four still needs a rerun on your model and your batch.

What May 15 actually sold

The June 30, 2026 Form 10-Q states the offering closed on May 15, 2026. Cerebras issued and sold 34,500,000 shares of Class A common stock at $185.00, including the underwriters’ option for 4,500,000 shares, exercised in full. Net proceeds were approximately $6.2 billion after discounts, commissions, and estimated offering expenses. The August 12 earnings exhibit describes the same deal as $6.4 billion gross. Both figures can be true. Thirty million shares at $185.00 is $5.55 billion, which is the base deal Nasdaq’s May 15 note reported. The greenshoe adds 4.5 million shares, about $833 million, and the gross with the shoe is about $6.38 billion. Net of fees is the $6.2 billion on the 10-Q. Use net when you are talking about cash. Use the base deal when you are reading a listing-day headline.

Immediately before the close, redeemable convertible preferred converted into 124,652,775 shares of Class B, one for one. A Schedule 13D amendment describes Class B as entitled to twenty votes per share. On the August 5, 2026 cover-page count in the 10-Q — 112,247,109 Class A, 111,601,424 Class B, 13,715,508 Class N — Class B’s votes are 2.23 billion against 112 million Class A votes. That is about 95% of the A-plus-B vote, before whatever vote Class N carries. This draft did not re-read the charter for the Class N vote. Public buyers of CBRS bought an economic slice. They did not buy control.

Cash and cash equivalents were $6.74 billion at June 30. Restricted cash was $685 million. Investments were $1.18 billion. The earnings exhibit rounds cash, restricted cash, and short-term investments to $8.6 billion of liquidity, and separately notes a revolving credit facility of up to $850 million. Capacity is not a draw. The same balance sheet shows a working-capital loan of $736 million. Read the cash line next to the loan, not instead of it. Customer warrants sit on the asset side at $168 million current and $961 million non-current. That is equity granted to customers, not a receivable you can spend.

The quarter after, in two accounting languages

GAAP total revenue was $180.1 million, up 74% from $103.3 million a year earlier. The mix moved, and the mix is the story. Hardware was $54.1 million, down from $70.3 million, about 23%. Cloud and other services were $126.0 million, up from $33.0 million, which the company calls 281%. For the six months ended June 30, hardware was $164.7 million and cloud was $208.8 million of $373.5 million total. Cloud is already 56% of the half. A chip-company narrative that ignores that split is reading the prospectus era, not the quarter.

GAAP gross profit was $25.6 million. GAAP gross margin was 14%. In the reconciliation table attached to the earnings exhibit, hardware’s GAAP gross margin is 1.8%: $54.1 million of hardware revenue against $53.1 million of hardware cost. Cloud’s GAAP gross margin is 19.5%. GAAP operating margin was negative 265%. The hardware appliance, on GAAP, barely cleared its own cost in the quarter. The growth is in the cloud line, and that line is not a 75% gross-margin business yet.

Core numbers add back non-cash amortization of customer warrants and strip pass-through data-center revenue and cost. On that basis the company reported core revenue of $209.9 million, up 103%, core gross margin of 41%, and core operating margin of negative 16%. The warrant add-back in the quarter’s reconciliation is $44.3 million. That is not cash collected. It is contra-revenue put back so the non-GAAP line looks like volume. The company’s own discussion of the adjustment says the amount and timing can move with customer deployment schedules, not with pricing. Use core to see the operating trend management wants you to see. Use GAAP when you are asking whether the quarter made money. It did not, on the operating line, in either language.

Guidance, as raised on August 12 and fetched with the exhibit: third-quarter core revenue of about $214 million to $216 million, core gross margin 38% to 40%, core operating margin negative 25% to negative 23%. Full-year 2026 core revenue was raised to $880 million to $890 million, from the $855 million to $865 million issued on June 23, with core gross margin 41% to 43% and core operating margin negative 19% to negative 17%. The CFO said the company plans to more than triple revenue in 2027, with no range attached. Against an $885 million midpoint, “more than triple” is arithmetic above about $2.7 billion. That is a verb in a press release, not a guidance range. Do not paste it into a model as if it were one.

The Cerebras IPO AI chip NVIDIA comparison on revenue

NVIDIA’s quarter ended July 26, 2026, not June 30. Revenue was $96.2 billion, up 106% from a year ago and 18% from the prior quarter. Data Center revenue was $89.0 billion, up 117%. GAAP and non-GAAP gross margin were both 75.0%. GAAP operating income was $63.7 billion. The next-quarter outlook was $108.0 billion, plus or minus 2%, at a 74% gross margin, plus or minus 50 basis points, with no Data Center compute revenue from China assumed. Those sentences are the company’s August 26 release, not a third-party estimate.

Nearest filed quarter	Cerebras, ended June 30	NVIDIA, ended July 26
GAAP revenue	$180.1 million	$96.2 billion
GAAP gross margin	14% (core 41%)	75%
Cloud or Data Center	Cloud $126.0 million	Data Center $89.0 billion
Operating result	GAAP margin (265%); core (16%)	GAAP operating income $63.7 billion

Not a matched 90-day window. Sources: Cerebras August 12, 2026 exhibit; NVIDIA August 26, 2026 release.

Set the two revenue lines next to each other and the horse race ends. $180.1 million is 0.19% of $96.2 billion. NVIDIA’s Data Center line alone is about 706 times Cerebras’s cloud line. Gross margin is 75% against 14% GAAP, or against 41% if you grant the core adjustments. None of that makes the wafer fake. It makes “outpace” the wrong verb for revenue, profit, and ecosystem. The GTC piece on Groq, Rubin, and the hardware bet is the NVIDIA-side document from this site. This one is the Cerebras filing. If the procurement question is who ships the default training cluster, the 10-Q is not the document that changes the answer.

The wafer claim, with the sparse-FLOP correction

The WSE-3 datasheet, fetched for this piece, lists 4 trillion transistors, 46,225 mm² of silicon, 900,000 cores, 44 GB of on-chip SRAM, 21 PB/s of memory bandwidth, and 214 Pb/s of fabric bandwidth, on a TSMC 5 nm process. It also says the die is 57 times larger than the largest GPU, without naming that GPU in the lines fetched. The CS-3 spec sheet lists 125 petaFLOPS and footnotes the figure as sparse. System power is 27 kW maximum, in 16U, on a liquid loop. Those are appliance facts. They are not a benchmark against your prompt.

SemiAnalysis, on May 13, 2026, put the dense FP16 or INT8 figure for one WSE-3 at 15.625 PFLOPS, and set that next to a B300 at about 13.5 PFLOPS of native FP4 and a Rubin GPU at 35 PFLOPS. Those are different numeric formats. A FLOPs table that ignores the format is marketing, including when the marketer is a critic. The architectural fact that survives the correction is the memory system: weights that fit in 44 GB of SRAM do not take a trip to HBM, and the bandwidth number on the datasheet is 21 PB/s. Peak sparse FLOPs are the number the booth print wants. Bandwidth is the number that can show up in decode latency, and only for a model that fits.

The earnings exhibit says Cerebras does not use HBM, CoWoS packaging, or 3 nm, and treats that as a supply-chain advantage while those parts are short. That is a real constraint story in 2026. It is also a ceiling. A model that does not fit in 44 GB does not get the SRAM trick unless you shard it, and sharding is the problem the wafer was built to avoid. The same exhibit says the company enabled OpenAI GPT-5.6 Sol at 750 tokens per second. That sentence is in an 8-K exhibit. It is not an independent benchmark, and it does not state context length, batch size, or precision. Treat 750 the way you would treat an NVIDIA MLPerf submission: directionally useful, not a substitute for your SLA.

The May 14, 2026 Form 424B4 is the prospectus. This draft did not re-parse that 95 MB file for the B200 comparison charts. Do not cite a 15x or a 2,625x multiplier from a secondary recap as if this page re-extracted it. Speed is an inference-cost input, not a trophy. If a coding agent is waiting on decode, tokens per second change the user’s clock. If the job is a batch of embeddings overnight, cluster throughput and dollars per million tokens matter more, which is the bill in the inference-cost piece. A decode chart does not price that job. The 2026 pricing guide is the API list. This section is only the machine.

Backlog, customers, and the lockup

Remaining performance obligations were $25.4 billion as of June 30, 2026. That is contracted future work, not revenue, and not cash. The same release says data-center capacity, live and under contract for delivery by the end of 2027, is more than 600 MW, and that manufacturing capacity is set to scale more than 10 times in 2026 at Flex, Sanmina, and Rocket EMS. The company says it secured the TSMC wafer supply it needs for the growth it is forecasting. “Secured” in a press release is not a disclosed volume or a price. Do not promote it to a purchase order you have not seen. Cash from the IPO can fund the build. It does not convert the backlog on a schedule the 10-Q has not printed.

Named customers in the 10-Q’s forward-looking section, the ones the company says it must retain, are OpenAI, G42, MBZUAI, and AWS. The August 12 exhibit adds Cognition, Lovable, Block, Figma, AlphaSense, GSK, and CrowdStrike. It also says an AMD disaggregated-inference setup, which the company claims can raise throughput up to 5 times, will be in production in the fourth quarter of 2026, with the same pattern expected on Amazon Bedrock in the first quarter of 2027. “Expected” and “up to” are the words in the sentence. Bedrock in the first quarter of 2027 is a date on a release until it is a region you can call.

The pre-IPO concentration story — heavy 2025 dependence on two Abu Dhabi-related buyers, beside the $510 million revenue figure Nasdaq attributed to the filing — is not restated as a percentage in the June 30 excerpts used here. If you need the 2025 customer table, open the 424B4. What the new quarter does show is a cloud line growing faster than hardware, and a logo list that now includes a frontier lab, a hyperscaler, and a security vendor. That is diversification in names. It is not proof the book is diversified until a filing prints a concentration percentage. The sections fetched on September 29, 2026 do not.

OpenAI is on the retain-this-customer list, and OpenAI is also having its own inference silicon built. The Jalapeño note on the brand hub is the other side of the same buyer. A lab that is both a named capacity customer and a future competitor for the same workload is not a five-year annuity. Contract length and termination rights matter more than the logo slide. The four-model comparison is the layer that has to fit on whatever chip wins that argument. A 44 GB SRAM ceiling is a model-choice constraint, not only a hardware spec.

The lockup is not a single 180-day cliff. The Motley Fool, on September 21, 2026, reading what it describes as the company’s published schedule, reported waves already eligible, further waves of about 19.4 million shares on September 30, October 14, and October 28, and a final release on the earlier of the second trading day after third-quarter results or November 9, 2026. That is journalism about a schedule, not a re-extraction of the prospectus lockup section. If you are marking a calendar, mark November 9 and the third-quarter print, and treat the intermediate dates as reported. This page is not a trading note.

The price, as fetched September 29, 2026 from a quote page: last close on September 28 was $196.74, against a $185.00 offer and a $311.07 first close. Since the first close, that page shows a return of about negative 37%. The stock is above the IPO price and well below day one. A market-cap figure of $46.74 billion on that page matches the August 5 share count times that close, to the quoted billion. The share count is six weeks older than the print. Do not treat either number as a valuation opinion. This piece does not have one. The May hardware-race piece on this site is the pre-IPO baseline. It is not a substitute for the 10-Q.

When the wafer is the right machine

Four questions. Stop at the first no. The diagram above the fold is the same test, shorter.

Does the model, at the precision you will actually serve, fit in 44 GB of on-chip SRAM? If it does not, you are back to sharding, HBM, and a GPU cluster, which is the job NVIDIA’s Data Center line is already being paid $89 billion a quarter to do. A 70B-class model in a low precision often fits. A long-context frontier model with a large KV cache often does not. Measure the footprint on the build you will serve, not on the demo checkpoint. A yes here is a measurement, not a brochure claim.

Is the scarce resource tokens per second on the decode path, rather than overnight throughput? Interactive agents, inline security checks of the kind the CrowdStrike sentence describes, and coding tools that stall on the next token are the jobs where the datasheet’s bandwidth can show up in a stopwatch. Batch scoring, embedding backfills, and training runs are usually a dollars-per-token problem. Price those against an API list, not against a booth chart. If your users do not feel decode latency, you are not buying the thing Cerebras is differentiated on.

Can you accept a single-vendor wafer, a GAAP operating margin that is still largely negative, and a core margin that depends on a warrant add-back? Cerebras is public and liquid. It is also, in its own 10-Q, dependent on a short list of customers it must retain, including a lab that is building a competing inference chip. That is a counterparty fact. Ask for the remaining term, the termination right, and what happens to your token price if that contract is restated. A logo is not a capacity reservation.

Is the alternative actually a CS-3 in your hall, or Cerebras cloud? On-prem is a power and water problem: 27 kW maximum and a liquid loop, per the spec sheet fetched September 29. The local-models piece is the decision frame for owning the box. Cloud is a token SLA with someone else’s power bill, and it is where the revenue mix already went — 70% of second-quarter GAAP revenue, 56% of the half. If you cannot answer the power question, you are buying cloud whether the white paper says appliance or not. Hardware revenue falling while cloud revenue quadruples is the company telling you the same thing.

If you clear all four, ask for a quote against your prompt mix and your batch-1 latency, and make the vendor rerun the 750 tokens-per-second claim on your model. If you fail any one, the default machine is still a GPU, rented or owned. NVIDIA’s quarter is the size of that default. Cerebras is a public inference specialist with a real memory-system advantage, a cloud mix that is already larger than its hardware mix, and an income statement that does not yet look like the company in the headline. Buy the axis you measured. Do not buy the verb.

FAQ

Did Cerebras outpace NVIDIA after the IPO?

No, on any revenue or profit line in the nearest filed quarters. Cerebras Q2 GAAP revenue was $180.1 million. NVIDIA’s quarter ended July 26 was $96.2 billion, at a 75% GAAP gross margin against Cerebras’s 14%. The wafer’s lead, if you grant the datasheet, is memory bandwidth for models that fit on the die.

What is core revenue, and why is it higher than GAAP?

Core revenue adds back non-cash amortization of customer warrants and removes pass-through data-center revenue. For Q2 2026 the company reported core revenue of $209.9 million against GAAP revenue of $180.1 million. The warrant add-back in that quarter’s reconciliation is $44.3 million. Core is a trend line. GAAP is the income statement.

Is 125 petaFLOPS the number to compare with a B300?

No. The CS-3 spec sheet footnotes 125 petaFLOPS as sparse. SemiAnalysis put dense FP16 or INT8 for one WSE-3 at 15.625 PFLOPS, in the same neighborhood as a B300’s FP4 peak and below a Rubin GPU’s, on different formats. Compare bandwidth and your measured tokens per second, not the sparse peak.

When do locked Cerebras shares come out?

Not on one date. As reported by The Motley Fool on September 21, 2026, from the company’s schedule, further waves land on September 30, October 14, and October 28, and the remainder on the earlier of the second trading day after third-quarter results or November 9, 2026. Confirm against the prospectus before you trade around it. This is not a trading note.

Is the $25.4 billion backlog cash?

No. Remaining performance obligations of $25.4 billion as of June 30, 2026 are contracted future performance, recognized as the work is delivered. Cash, restricted cash, and short-term investments were the $8.6 billion liquidity line. A working-capital loan of $736 million sits on the other side of the balance sheet.