Cerebras Systems, the Silicon Valley chipmaker that constructed the world's largest industrial AI processor, erupted onto the Nasdaq on Wednesday, opening at $350 per share — practically double its $185 IPO worth — and rocketing previous a $100 billion market capitalization in its first hours of buying and selling. The debut immediately topped Cerebras as one of the crucial helpful semiconductor firms on Earth and validated a decade-long wager that the AI {industry} would ultimately demand a basically completely different type of chip.
The corporate offered 30 million shares at $185 apiece, elevating $5.55 billion in what Bloomberg reported as the largest U.S. tech IPO since Uber went public in 2019. The ultimate pricing shattered expectations: Cerebras initially marketed shares at $115 to $125, then raised the vary to $150 to $160 as investor demand surged, earlier than in the end pricing above even that elevated band.
"It is a new starting," Julie Choi, Senior Vice President and Chief Advertising Officer at Cerebras, instructed VentureBeat in an unique interview on the morning of the IPO. The corporate, she mentioned, plans to pour its recent capital into increasing the cloud infrastructure that has turn out to be the centerpiece of its development technique. "With this new capital, we're going to fill extra knowledge halls with Cerebras programs to energy the world's quickest inference."
The IPO caps one of the crucial dramatic company turnarounds in current tech historical past. Cerebras first filed to go public in September 2024 but withdrew the hassle greater than a yr later amid intense scrutiny over its near-total income dependence on a single buyer within the United Arab Emirates. The corporate refiled in April 2026 with a radically completely different enterprise profile: new partnerships with OpenAI and Amazon Web Services, a fast-growing cloud inference service, and a income base that had climbed 76% to $510 million in 2025.
How a dinner-plate-sized chip turned the muse of a $100 billion firm
To know the frenzy, it’s a must to perceive the silicon.
Cerebras builds one thing referred to as the Wafer-Scale Engine, or WSE — a single processor that occupies a whole silicon wafer, the dinner-plate-sized disc from which peculiar chips are minimize. The third-generation WSE-3 comprises 4 trillion transistors, 900,000 compute cores, and 44 gigabytes of on-chip reminiscence. It’s 58 instances bigger than Nvidia's B200 "Blackwell" chip and delivers 2,625 instances extra reminiscence bandwidth than the B200 package deal, in accordance with the corporate's S-1 filing with the Securities and Alternate Fee.
That bandwidth benefit issues enormously for AI inference — the method of working a educated mannequin to generate solutions. When a big language mannequin produces textual content, it predicts one token at a time, and every token requires the mannequin's total set of weights to maneuver from reminiscence to compute. This work is inherently sequential and can’t be parallelized, making reminiscence bandwidth the binding constraint on pace. Cerebras claims its structure delivers inference responses as much as 15 instances sooner than main GPU-based options on open-source fashions, a determine corroborated by third-party benchmarker Artificial Analysis.
"One of many architectural ideas after we constructed the wafer was: let's hold compute nearer collectively, in order that compute components can discuss to one another at decrease latency," Andy Hock, VP of Product at Cerebras, instructed VentureBeat. "Low latency is necessary to AI compute. It's a cornerstone of quick inference."
The founding perception was contrarian and, for a lot of the firm's life, commercially untimely. Cerebras's founders acknowledged in 2015 that AI workloads had been communication-bound issues — pace trusted how briskly knowledge might transfer between reminiscence and compute — and that one of the best ways to speed up that motion was to maintain every little thing on a single huge chip.
Wafer-scale integration had been tried and deserted repeatedly over the semiconductor {industry}'s 75-year historical past. Each earlier effort had failed. Cerebras solved the issue via two key innovations detailed in its S-1: a proprietary multi-die interconnect that stitches in any other case impartial die collectively on the wafer degree throughout fabrication, and a fault-tolerant architecture that routes round manufacturing defects utilizing redundant constructing blocks, just like how hyperscale knowledge facilities deal with server failures.
Why Cerebras is betting its future on cloud inference as an alternative of {hardware} gross sales
For many of its life, Cerebras offered {hardware} — huge, water-cooled AI supercomputers put in on-premises at buyer services. That mannequin generated $358 million in {hardware} income in 2025. However the IPO prospectus reveals a strategic pivot that can outline the corporate's subsequent chapter: the transition to cloud-based inference services.
Cerebras launched its inference cloud in August 2024. In lower than two years, cloud and different providers revenue reached $151.6 million in 2025, up 94% from $78.3 million in 2024. The corporate now expects this section to comprise a considerably bigger share of complete income going ahead, pushed primarily by its huge cope with OpenAI.
"Cloud and mannequin APIs are the popular and pure consumption technique for inference providers and software builders," Hock instructed VentureBeat. "In order that was the pure packaging and go-to-market technique for the inference functionality."
Choi framed the cloud as a democratization play. "Whether or not that be an entrepreneurial developer, a startup, or a large group like OpenAI — the cloud has actually made it straightforward for folks to deploy and really feel the quick inference, the worth of it," she mentioned.
The economics of the transition are capital-intensive. Cerebras should lease knowledge middle area, manufacture and deploy its programs, and construct software program to handle capability — all earlier than recognizing recurring income. The S-1 warns bluntly that gross margins will decline within the close to time period as the corporate absorbs startup prices for cloud infrastructure. The corporate's gross margin already dipped to 39% in 2025 from 42.3% in 2024, pushed by increased knowledge middle prices. However the demand image seems formidable. "Each cloud system that we've deployed thus far, every one will get devoured up in capability," Hock mentioned. "We've been thrilled to see the demand for quick inference from Cerebras. We need to go sooner to service that market."
Contained in the $20 billion OpenAI deal that reworked Cerebras in a single day
The only most consequential enterprise relationship for Cerebras is its December 2025 settlement with OpenAI, beneath which OpenAI committed to purchase 750 megawatts of Cerebras inference compute capability over the following a number of years. The deal is valued at greater than $20 billion and contains provisions for OpenAI to buy an extra 1.25 gigawatts of capability, probably bringing complete deployment to 2 gigawatts.
The association goes far past a normal vendor-customer relationship. OpenAI and Cerebras are co-designing future models for future Cerebras {hardware} — a good suggestions loop that offers Cerebras visibility into frontier mannequin architectures earlier than they ship and provides OpenAI inference programs optimized for its particular workloads. The partnership moved from contract to manufacturing with exceptional pace. "After we introduced the partnership, we had the primary mannequin working in like 35 days," Choi instructed VentureBeat. "That was Codex Spark, and the engineers over at OpenAI simply had been like, thoughts blown."
Codex Spark, OpenAI's mannequin designed for real-time coding, permits builders to show natural-language directions into working software program in seconds utilizing Cerebras infrastructure. Choi described a deep cultural alignment between the 2 firms. "Our groups actually vibe as engineers. We're on the identical wavelength," she mentioned. "There's simply no quantity of pace that’s sufficient for these guys."
To fund the infrastructure buildout, OpenAI superior Cerebras a $1 billion working capital loan in January 2026, secured by a promissory observe maturing no later than December 31, 2032, bearing 6% annual interest. The mortgage might be repaid in money or via supply of compute capability. Nevertheless, the S-1 discloses important danger: if the MRA is terminated for any motive apart from OpenAI's materials uncured breach, OpenAI can seize management of the mortgage funds and demand quick compensation. OpenAI additionally holds a warrant to purchase up to 33.4 million shares of Cerebras Class N frequent inventory at an train worth of $0.00001 per share — basically free shares that vest as Cerebras delivers dedicated capability. On the IPO opening worth, the totally vested warrant can be value roughly $11.7 billion.
How the Amazon Net Providers partnership might convey Cerebras chips to hundreds of thousands of builders
In March 2026, Cerebras signed a binding term sheet with Amazon Net Providers to turn out to be the first hyperscaler to deploy Cerebras systems inside its personal knowledge facilities. The partnership introduces a novel architectural idea referred to as disaggregated inference, which splits the 2 levels of AI inference — prefill (processing the consumer's immediate) and decode (producing the response) — throughout completely different {hardware} optimized for every job. Below this association, AWS Trainium chips deal with prefill, whereas Cerebras CS-3 programs deal with decode, linked by way of Amazon's Elastic Cloth Adapter networking.
In keeping with the AWS press announcement in March, the method goals to ship an order of magnitude sooner inference than what’s at the moment accessible. Hock offered technical element on why this works. "The interconnect necessities between prefill and decode programs really aren't that top, so we will use a standard interconnect between, say, Trainium and the wafer-scale engine and nonetheless ship that quick time to first token and that ultra-low latency token technology," he defined. "What the Trainium wafer-scale engine mixture actually provides us in that disaggregated or heterogeneous inference setup is all of the pace and vastly extra effectivity, so we will successfully serve extra tokens per unit rack area or kilowatt."
The partnership offers Cerebras one thing it has lengthy lacked: huge distribution. AWS serves hundreds of thousands of enterprise clients worldwide, and Cerebras programs deployed via Amazon Bedrock will turn out to be accessible to any developer inside their present AWS surroundings. "AWS has unbelievable attain," Hock mentioned. "The partnership is de facto about bringing that quick inference functionality — that form of best-in-industry, quick inference functionality delivered by wafer-scale engine and Trainium — to that broader market." The time period sheet additionally grants AWS a warrant to purchase up to approximately 2.7 million shares of Cerebras Class N frequent inventory at a $100 train worth, with vesting tied to product purchases past the preliminary lease.
The UAE buyer focus drawback that just about derailed the IPO — and whether or not it's actually solved
For all the thrill, Cerebras carries a danger that has haunted it since its first IPO try: buyer focus. In 2024, G42 — an Abu Dhabi–based mostly know-how conglomerate — accounted for 85% of Cerebras's total revenue. The corporate's September 2024 S-1 submitting drew heavy scrutiny over this dependence, compounded by questions on export controls for superior AI chips shipped to the UAE. Cerebras withdrew that filing.
The 2025 numbers present progress however not decision. G42's share of income declined to 24%, however Mohamed bin Zayed College of Synthetic Intelligence (MBZUAI), an Abu Dhabi establishment that may be a associated social gathering to G42, accounted for 62% of total revenue.
Collectively, the 2 UAE-linked entities nonetheless represented 86% of Cerebras's 2025 gross sales. The S-1 is candid about this risk, noting that MBZUAI accounted for 77.9% of accounts receivable as of December 31, 2025, and that U.S. export licenses for Cerebras programs shipped to G42 and MBZUAI require "rigorous safety and compliance obligations to stop diversion and abuse of our know-how."
Choi addressed the difficulty immediately, pointing to the OpenAI and AWS offers as proof of a broadening buyer base. "Now with OpenAI and Amazon, these are the identical sort of deep partnerships," she instructed VentureBeat. "We're a deep know-how firm. Our know-how has taken a decade to construct. We go deep in how we construct, and now we're going deep with two of the most important gamers — the most important AI lab, OpenAI, and the most important cloud, AWS."
Hock framed the client evolution as a development in market notion. "G42 precipitated the market to be intrigued and impressed," he mentioned. "No one within the enterprise is smarter, extra credible, or has higher attain than OpenAI and AWS. And so I believe OpenAI and AWS precipitated the market to shift from intrigued and impressed to — I'll name it curious and satisfied." Nonetheless, the S-1 warns that the OpenAI MRA itself "represents a considerable portion of our projected revenues over the following a number of years." Cerebras's enterprise will stay depending on a small variety of very giant clients for the foreseeable future — a structural function of the AI infrastructure market the place buildouts are measured in tons of of megawatts and billions of {dollars}.
Can Cerebras construct knowledge facilities quick sufficient to maintain up with runaway demand?
With OpenAI consuming 750 megawatts of dedicated capability and AWS preparing to deploy Cerebras systems in its knowledge facilities, the query is whether or not Cerebras can scale its bodily infrastructure shortly sufficient to serve everybody else. Hock acknowledged the stress. "It's a very good drawback to have when demand begins to outstrip provide. It doesn't imply it's a simple drawback to deal with," he instructed VentureBeat. "We've obtained to construct these extraordinary programs. We've obtained to obtain knowledge middle area. We've obtained to deploy programs there. Acquired to face up software program to fulfill our clients the place they’re."
The corporate is being deliberate about capability allocation. "We're attempting to be actually deliberate about how we allocate capability because it's constructed," Hock mentioned. "We're working in deep partnership to service the highest-priority clients and highest-priority markets."
Choi argued that the constraint really sharpens focus. "Typically when you might have much less of one thing, it forces you to be very deliberate," she mentioned. Past OpenAI, she named Cognition — the AI coding startup — and Block, led by Jack Dorsey, as important clients. "Jack participated in our roadshow as properly," Choi famous. "We're dashing up that total money-bot AI expertise inside Money App."
The S-1 discloses that Cerebras at the moment operates data centers in California, Oklahoma, and Canada, with plans to broaden internationally. The corporate executed non-cancelable knowledge middle leases in late 2025 with combination undiscounted future minimal funds of approximately $344 million, and in March 2026 signed a Canadian knowledge middle lease with anticipated minimal funds of roughly $2.2 billion over a 10-year time period.
The IPO proceeds — mixed with $1 billion from a January 2026 Sequence H most popular inventory spherical and the $1 billion OpenAI mortgage — give Cerebras a struggle chest exceeding $8 billion to fund the buildout. Whether or not that is sufficient to fulfill a market the place main clients are ordering capability measured in gigawatts stays an open query.
The Nvidia shadow: what Cerebras is de facto up towards within the AI chip wars
Cerebras enters public markets into the tooth of the most competitive semiconductor surroundings in a long time. Nvidia stays the dominant force in AI compute, controlling the overwhelming majority of the coaching and inference infrastructure market. Its GPU structure advantages from a deeply entrenched software program ecosystem constructed round CUDA, the programming framework that has turn out to be the de facto customary for AI improvement. Cerebras's S-1 explicitly acknowledges this, noting that "lots of our opponents profit from aggressive benefits over us, comparable to outstanding and cutting-edge know-how and software program stacks designed to maintain out new market entrants."
However Cerebras argues the inference market is structurally different from training — and that its structure has a basic benefit within the workload that issues most going ahead. As AI fashions have shifted towards reasoning, the place fashions carry out multi-step computation throughout inference to assume via issues, the number of tokens generated per request has exploded. Every token requires shifting full mannequin weights from reminiscence to compute, making reminiscence bandwidth the bottleneck. The S-1 cites Bloomberg Intelligence knowledge projecting that Cerebras's addressable portion of the AI inference market will develop from roughly $66 billion in 2025 to $292 billion by 2029, a forty five% compound annual development charge — considerably outpacing the 20% CAGR projected for AI coaching infrastructure.
Nvidia has clearly taken discover of the fast-inference menace. In December 2025, Nvidia acquired Groq — a startup whose tensor streaming processor structure extra carefully resembles Cerebras's method — for $20 billion.
Months later, Nvidia announced plans for Groq-based products, signaling that even the {industry}'s dominant participant acknowledges the restrictions of GPU structure for latency-sensitive inference. Cerebras additionally competes with customized silicon developed by hyperscalers — together with Google's TPUs and Amazon's Trainium chips — and a rising roster of AI cloud suppliers. Requested about Nvidia and Groq, Choi declined to interact. "We're feeling fairly good proper now," she instructed VentureBeat with a smile.
Income is surging, however the monetary advantageous print reveals a extra sophisticated image
The monetary image that emerges from the S-1 is one in every of fast scaling with important underlying complexity. Income surged from $78.7 million in 2023 to $290.3 million in 2024 to $510 million in 2025 — a greater than tenfold improve over three years. The corporate reported GAAP internet earnings of $237.8 million in 2025, however this determine is closely influenced by a $363.3 million one-time achieve from the extinguishment of a ahead contract legal responsibility associated to a most popular inventory association. Stripping out that achieve and stock-based compensation, Cerebras's non-GAAP internet loss was $75.7 million in 2025, widening from a $21.8 million non-GAAP loss in 2024.
Working losses deepened as properly. Cerebras lost $145.9 million from operations in 2025, up from $101.4 million the prior yr, as the corporate invested closely in analysis and improvement ($243.3 million, up 54%) and gross sales and advertising and marketing ($70.6 million, up 237%).
The corporate burned $10 million in working money circulate in 2025, a pointy reversal from the $452 million of money generated in 2024 — a yr boosted by $640 million in buyer deposit inflows, primarily from G42 and MBZUAI. The S-1 warns that gross margins will face near-term stress from startup prices for cloud infrastructure, buyer warrant amortization, and pass-through knowledge middle bills.
The trail to this second was something however easy. Cerebras shipped its first programs in 2020 and 2021 — earlier than the market was prepared. Because the founders wrote within the prospectus: the corporate "had constructed one thing extraordinary, however the market wasn't prepared." The ChatGPT second in late 2022 modified every little thing.
By early 2025, Cerebras's pace benefit — lengthy an answer seeking an issue — turned urgently related as AI coding brokers, deep analysis instruments, and real-time voice functions demanded the type of low-latency inference that GPU clusters struggled to ship. The S-1 describes a market the place AI coding brokers "barely existed in 2023" however collectively generated "billions in ARR in 2025," and the place 42% {of professional} code is now AI-generated or assisted.
What Cerebras should show to justify a $100 billion valuation — and what occurs if it may't
Wanting ahead, Hock signaled that the present technology of {hardware} is only the start. "Wafer-scale engine three and CS-3 just isn’t the tip of the story. It's only the start," he instructed VentureBeat. "We have now a multi-year know-how roadmap that continues constructing on wafer-scale know-how, accelerating efficiency, growing effectivity, supporting bigger scale."
The S-1 confirms that Cerebras intends to expand on-chip memory and bandwidth, enhance interconnect density, and leverage future course of node advances — and discloses that the corporate has already obtained export licenses for future CS-4 programs destined for the UAE.
The corporate additionally faces an online of operational dangers that may check any group, not to mention one which has by no means operated as a public firm. It relies upon fully on TSMC for wafer fabrication, with no long-term provide dedication. Its knowledge middle leases stretch for years, whereas its inference buyer contracts are sometimes shorter-term or consumption-based, making a mismatch between mounted prices and variable income. It has recognized material weaknesses in its inside controls over monetary reporting. And its most necessary buyer relationship — with OpenAI — contains exclusivity provisions that prohibit Cerebras from working with sure named OpenAI opponents, probably limiting future diversification.
Whether or not Cerebras can maintain a $100 billion-plus valuation will rely upon its skill to execute towards all of those challenges concurrently: constructing knowledge facilities at unprecedented pace, manufacturing wafer-scale chips at scale via a single foundry, navigating export controls on its most profitable worldwide relationships, and competing towards an Nvidia that has proven it won’t cede the inference market with no struggle.
However Cerebras has at all times been constructed on a willingness to try what others mentioned was inconceivable. Wafer-scale integration had stumped the semiconductor {industry} for its total existence. Now a chip the scale of a dinner plate — as soon as dismissed as an engineering curiosity — powers the quickest AI inference on the planet, serves the world's main AI lab, and simply debuted on the Nasdaq to a valuation that dwarfs firms many instances its age. The world, it seems, was prepared. As Hock put it to VentureBeat, recalling the journey from the lab to the buying and selling flooring: "The IPO isn't the tip of the story. It's the start."