Cerebras IPO: S1 Breakdown

Most CFOs I know didn't get into finance to chase receipts.

But that's where the time goes. Reviewing expenses that should have been blocked before they happened. Closing books that take longer to close than the month took to live. Approving spend that should have been approved automatically three days ago.

That's not a people problem. That's a systems problem.

That's why I use Brex. Agentic Finance that automates the receipts, catches out-of-policy spend before it hits the books, and closes your month in minutes instead of weeks. So your finance team stops spending on maintenance and starts spending on momentum.

35,000+ companies, including Mostly Metrics, Anthropic, DoorDash, and Coinbase, have already figured this out. See why it’s time to get Brex AF.

Cerebras builds the largest computer chip ever commercialized. It is 58 times the size of NVIDIA's flagship and roughly the size of a dinner plate. Which is also like the most American thing to say: Bigger is BIGGER!

Big as they are, the chips are the easy part of this tale. The financials are where it gets funky.

Cerebras did $510M in revenue in 2025, growing 76% y/y, and is sitting on a $24.6 billion backlog that is essentially one customer. That customer also loaned them a billion dollars and was issued 33 million shares for fractions of a penny. Who do you think it is?

They are filing to IPO this week at a range that implies a $33 billion market cap.

If you LOVED Coreweave you’re going to potentially maybe LIKE Cerebras.

Strap in. There’s a big chip sale going down. And I’m not talking Tostitos.

Every chip you've ever interacted with is really small for a reason. Chipmakers start with a silicon wafer (a round disc about the size of a vinyl record), stamp hundreds of identical chips onto it, and then slice the vinyl apart so each chunk can be packaged and sold individually. Your laptop and your phone each have one, and NVIDIA's flagship B200 GPU has two, mounted together.

The reason chips are small is because defects happen. Silicon is finicky AF; some part of every wafer has a flaw, and the bigger you make the chip, the higher the chance a flaw lands inside it and kills the whole thing. So the industry settled into making chips small and stamping a lot of them per wafer. For 75 years, the largest commercial chip was about the size of a postage stamp.

Cerebras kept the wafer intact. They invented a way for the chip to recognize defects and route the work around them, figured out how to cool a chip the size of a dinner plate without melting the motherboard, and worked with TSMC to develop a manufacturing process that had never existed before. Nobody had done this commercially to date. They call it the Wafer-Scale Engine, or WSE, and it is 58 times the size of NVIDIA's B200.

Here's why size matters for AI:

AI workloads spend most of their time moving data between chips, not computing. Which is something I learned when I was today’s years old, writing this here S1 breakdown.

Picture a GPU cluster as a giant room of accountants who can each do math, but every time one finishes a calculation, they have to walk it down the hall to another accountant to continue the work. The math isn’t hard for them. But they walk really fucking slow, and the walking takes up a ton of their energy.

Another way to think about it is NVIDIA sells you the accountants and the cables that connect them. Cerebras puts 900,000 accountants in the same room on a single piece of silicon, so the data never has to walk anywhere (and gets no lunch brekas). They claim this makes their system up to 15 times faster than NVIDIA on inference at a fraction of the power.

That pitch did not sell in 2020. From their own S-1:

So they sold to a handful of life sciences customers… and did… $25M in revenue in 2022. Womp womp.

Then inference became the workload that mattered the most.

The way modern AI works has changed in a way that finance people should care about, even if they don't care about the technology.

So inference is no longer a lookup. It is the bottleneck, and the bottleneck is exactly the thing Cerebras built their chip to solve, four years before the bottleneck existed.

That’s why this IPO is happening now. Cerebras spent five years explaining wafer-scale to a market that didn't need it, and then in early 2025 the market arrived at their doorstep. Revenue went from $78M in 2023 to $290M in 2024 to $510M in 2025. And then OpenAI signed a $20 billion deal in December and AWS signed a term sheet in March. Business is dancin like DJ Esco.

OK I’ll bite. I'm pulling down 500 megabits a second via Xfinity in this bizatch. Let’s keep going.

A quick scorecard before we get into the weird stuff.

To call out the CoreWeave in the room - the comparison is too clear. It’s important ot note that CoreWeave works WITH Nvidia, souping up their chips. Cerebras has TSMC manufacture their chips and then soup them up. Same same, but different.

But the rest of the story is kinda eerie.

CoreWeave priced at $40 in March 2025, down from an initial $55 target if you recall, which was a dark cloud over the IPO. The stock opened flat, ran up 300% by June, then gave half of it back. It’s currently trading at roughly 19x current sales after $5.1B of 2025 revenue and bonkers growth. They are doing well. But also, the market does not know what to do with this category. And with Cerebras, it’s growing.

Cerebras has two revenue lines.

Hardware: $358M (70% of revenue)

They sell the Wafer-Scale Engine inside a fully integrated system called the CS-3, and they sell racks of those CS-3s wired together into what they call an AI supercomputer. Also, it sounds like a mid sized Audi with a twin turbo.

Customers buy these and put them in their own data centers. This is the original Cerebras business and it's where most of the revenue still comes from.

A miss on one customer is a miss on the quarter.

Cloud and other services: $152M (30% of revenue)

They launched the Cerebras inference cloud in August 2024. Customers can rent compute by the token, by the month, or for dedicated long-term capacity, and they can buy it directly from Cerebras or through partner marketplaces (AWS, Microsoft, IBM, Vercel, OpenRouter, Hugging Face).

Sit with that for a sec, bc it's the inverse of what the IPO narrative craves.

The standard SaaS-pivot story is that you start as a lumpy hardware or services business and migrate to recurring software revenue with higher margins, sticky retention, and a clean multiple.

Cerebras is doing the opposite. The hardware sale is a one-time invoice with reasonable margin. The cloud product is the recurring revenue line, and it's costing them more to deliver than the hardware they're trying to migrate customers away from.

Quarterly cloud gross margin in 2025 went: 68% → 26% → 16% → 21%. The Q2 crater is what happens when you stand up a bunch of new cloud capacity faster than you can fill it. The S-1 attributes the gross margin decline to

Two things to know about where this goes.

So terrible now, but maybe not terrible forever.

Either way, the company that's IPOing is not a software company. It's a hardware company with a margin-dilutive cloud bolt-on.

Two customers account for 86% of Cerebras's 2025 revenue.

Those customers are MBZUAI and G42, both based in Abu Dhabi. They are also related parties to each other, as defined by Accounting Standards Codification 850. Meaning, there’s some stuff going on here.

G42 is an Abu Dhabi-based technology holding company with subsidiaries spanning energy, finance, cloud, security, and healthcare. MBZUAI is the Mohamed bin Zayed University of Artificial Intelligence, also in Abu Dhabi, also operating in the orbit of the same sovereign-backed AI strategy. The S-1's own description:

Customers, vendors, partners, and research collaborators. Roommates?

Here's how the concentration has moved:

The accounts receivable is even more concentrated than the revenue.

Working capital concentration is a risk. So good thing they are investors!

Oh, then there's how the money moves between them.

G42 prepaid Cerebras $640.3M in 2024 to fund the build-out of the hardware they wanted Cerebras to deliver. That prepayment is what makes Cerebras's 2024 operating cash flow look healthy. Without it, OCF was deeply negative.

G42 also holds warrants. In December 2025, Cerebras issued G42 a warrant to buy 1,857,516 shares of Class N common stock at $0.01 per share. That warrant was fully vested and exercised in January 2026. In April 2026, Cerebras issued G42 another warrant, this time for 1,655,975 shares, also at $0.01 per share, also fully vested and exercised the same month.

So G42 is:

Doing the math… so like roughly half a billion in equity for the cost of a slightly used Nissan Altima.

There's also AWS.

In March 2026, Cerebras signed a binding term sheet with AWS to be the first hyperscaler to deploy Cerebras systems in its own data centers. The term sheet is binding on pricing, exclusivity, minimum capacity, and protections in favor of AWS, but the definitive agreement hasn't been negotiated yet.

AWS also got a warrant: up to 2,696,678 shares of Class N common stock at $100 per share, vesting tied to product purchases beyond the initial lease.

The AWS warrant is structurally different from G42's. The $100 exercise price means AWS pays real money (~$270M if fully exercised) and only vests if they actually buy a lot of compute. That's a more normal commercial arrangement.

The S-1 names the dependency layer cake in its risk factors. They name the four customers who are essentially the whole company:

Cerebras also has supplier concentration on the back end. They use one foundry, TSMC, to manufacture every wafer, and the S-1 acknowledges they have no formalized long-term supply commitment with TSMC, who also fabricates wafers for Cerebras's competitors (including NVIDIA, who is many times larger and buys many more wafers). On top of that, the data centers hosting the cloud business are leased, not owned.

So the picture is:

Reading the cap table is like attending a Targaryen wedding. The customers are also the owners. The owners are also the lenders. The lenders are also your sister.

In December 2025, Cerebras signed a Master Relationship Agreement (sounds like a Uranium production deal) with OpenAI. OpenAI committed to purchase 750 megawatts of cloud compute capacity over multiple years, with options to expand to 2 gigawatts. The deal is valued at more than $20 billion and represents most of the $24.6B backlog the IPO pricing is built on.

Alongside the MRA, Cerebras issued OpenAI a warrant to purchase 33,445,026 shares of Class N common stock at an exercise price of $0.00001 per share. At the $155 IPO midpoint, that warrant is worth approximately $5.18 billion in equity. The exercise cost to OpenAI to take possession of all 33 million shares is $334.45. So, like, about what it costs to take my three kids to a character breakfast at the Contemporary Hotel in Disney World.

OpenAI also advanced Cerebras a $1.0 billion working capital loan. It accrues 6% interest, matures in December 2032, and is repayable in cash, compute capacity, or hardware. Dealer’s choice!

It is secured, and it comes with strings: if the MRA is terminated for any reason other than OpenAI's material uncured breach, OpenAI can direct the bank to freeze the cash and demand immediate repayment of principal plus interest.

The warrant vests in three tranches.

So Cerebras's largest customer is also the customer most economically incentivized to keep buying more.

Success be expensive (and dilutive) AF.

RPO recognition schedule:

That's ~$1.85B/year of recognized revenue over the next 24 months from backlog alone, growing to ~$5.3B/year in years 3-4. If it plays out, 2026-2027 combined revenue from the backlog is more than 3x what they did in 2025. That is the bull case.

The asterisks: it assumes

Any one of those slips and recognition shifts out out out.

1. They told the SEC they don't have a great accounting function.

In their own S-1:

I've spent enough time in audit meetings to know what "ineffective segregation of duties" means in practice. Basically, it means somebody is writing the journal entries and reviewing the journal entries, which is the kind of thing that gets you on the front page of the Wall Street Journal in a bad way. Remember: this is the company about to go public at a $33B implied market cap, with $24.6B in remaining performance obligations under a complex revenue recognition agreement (the OpenAI MRA), telling you in writing that their controls over revenue recognition suck.

2. One foundry to make chips. No long-term supply commitment.

Every Cerebras wafer is manufactured by TSMC, with no formal long-term supply agreement. TSMC also manufactures wafers for NVIDIA, AMD, and most of Cerebras's competitors, all of whom buy a lot more wafers than Cerebras does. If TSMC reduces allocation, raises prices, or prioritizes one of those competitors, Cerebras has no second source. It's not like you can call up Samsung Foundry on Monday and have wafers shipping by Friday.

Wafer-scale adds a wrinkle: Cerebras buys an entire wafer per unit on a specialized manufacturing process built in partnership with TSMC, which is not easily portable to another foundry. Bigger is bigger risk?

3. The UAE concentration is a geopolitical bet, as well as a customer bet.

86% of 2025 revenue came from two related Abu Dhabi entities. That's a customer concentration problem and a political concentration problem layered on top of each other. G42 has drawn scrutiny from the U.S. government over historical ties to China, which is why Microsoft put $1.5B into them in 2024 (basically buying G42 a U.S. passport). The Commerce Department continues to update its export framework for advanced semiconductors in jurisdictions where G42 and MBZUAI operate, and the framework has gotten tighter, not looser, over the last 18 months.

If the U.S.-UAE technology relationship shifts the wrong way, a meaningful portion of Cerebras's revenue base disappears with limited notice. That's not a hypothetical, by the way. They call it out as an explicit risk factor in the S-1.

4. The cloud business is unproven and they're betting the company on it.

Cerebras launched the inference cloud in August 2024. As of the S-1, it is fewer than 18 months old. The OpenAI MRA, the AWS term sheet, and most of the $24.6B backlog are all structured as cloud capacity, not hardware.

The founders and the board still hold meaningful equity, but Cerebras has been a heavily funded private company for almost a decade and institutional names dominate the cap table.

Top holders (post-IPO, before any over-allotment):

Feldman and Lie have been a duo since SeaMicro in 2007, which AMD acquired in 2012. They both did time at AMD post-acquisition, then co-founded Cerebras in 2016. So the operating partnership has 18 years of continuity. Feldman is a founder-CEO running his second company, not a hired gun.

Combined founder ownership is 8.3%, which is on the lower end for an IPO-stage company. Cerebras founders got diluted harder than most, because the business needed a lot of capital before the inference moment showed up.

In January 2026, Cerebras closed a $1.0 billion Series H at $89.02 per share, with Alpha Wave, Benchmark, and Fidelity all participating. Four months later, they're pricing the IPO at a midpoint of $155, which is a 74% step-up from Series H to IPO.

For context, Series G priced at $36.23 per share in September 2025. The company was reportedly thinking of going public at that point but did the round instead. So inside 12 months: $36 → $89 → $155. 4.3x in a year, almost all of it riding on the OpenAI deal signing in December 2025.

The other thing to flag is that Fidelity put in $700 million at the Series G round and another $100M at Series H. That’s a fat IPO position.

Cerebras has three share classes. Class A (one vote per share, the public IPO shares), Class B (20 votes per share, held by insiders/founders), and Class N (non-voting, held by OpenAI and G42 via their warrants). Class B will represent 99.2% of voting power immediately after the offering. The new public shareholders own 14% of the economic upside and 0.8% of the voting power. LOL.

At the $155 midpoint, Cerebras prices at roughly $33B on $510M of 2025 revenue and ~$2.5-3B of expected 2026 revenue (using the RPO recognition schedule plus the base business). That's an EV/forward revenue multiple of ~10-13x.

Where the public peers trade right now

Cerebras is pricing at a premium to every peer except NVIDIA. The semiconductor peers (AMD) and the neocloud peers (CoreWeave, Nebius) both trade at half the multiple Cerebras is asking for. Only NVIDIA gets a higher multiple, and NVIDIA prints free cash flow at industrial scale, which Cerebras does not.

What's actually defensible:

The midpoint is the high end of what's defensible with the current information. The NVIDIA-style multiple is what the OpenAI deal narrative will be selling. Whether it sticks depends on whether the market reads the OpenAI backlog as "huge customer win" or as "huge customer dependency."

Andrew Feldman was a defendant in SEC v. Pereira in 2008 related to his time as VP of Corporate Marketing and Corporate Development at Riverstone Networks in 2001-2002, where the SEC alleged he was aware of and aided in sales transactions that were improperly accounted for. He settled with the SEC without admitting wrongdoing, agreed to a permanent injunction against future securities violations, and paid $289,507 plus interest. In a parallel DOJ action, he pled guilty to one count of circumventing accounting controls of an issuer. He was sentenced to three years probation and a $5,000 fine.

These are 18-year-old events from a different company, and Feldman wasn't barred from serving as an officer or director. He's gone on to run SeaMicro (sold to AMD) and now Cerebras. The S-1 discloses all of it in his bio. Standing next to the present-day material weakness disclosure on revenue recognition controls, it's the kind of thing the audit committee is presumably already on top of.

The S-1 says the three share classes "are identical, except with respect to voting and conversion rights."

The Tranche 2 OpenAI warrant (~$864M in equity to OpenAI) vests when Cerebras's market cap clears $40 billion on a 30-day rolling average. At the $155 midpoint, Cerebras opens at $33B. So the company is one good month of trading away from automatically transferring almost a billion dollars of equity to its largest customer.

The lead-left underwriter is Morgan Stanley. The bracket is Citi, Barclays, UBS, Mizuho, and TD. Then further down the cover: Needham, Wedbush, Rosenblatt, Academy, Credit Agricole, MUFG, First Citizens, and Craig-Hallum Capital Group LLC.

Disclosures: None of this is investment advice. Do your own homework.

Wishing you an IPO with ample float and no material accounting weaknesses,

CJ