Cerebras Systems filed its S-1 registration statement with the SEC on April 17, 2026, setting the stage for what became the largest AI chip IPO in history. The Sunnyvale-based startup listed on the Nasdaq under the ticker CBRS, ultimately pricing its shares at $185 apiece in May 2026 and raising $5.55 billion – including a 4.5 million-share over-allotment option exercised on top of the base 30 million-share offering, according to CNBC. With $510 million in 2025 revenue, a $10 billion OpenAI contract in hand, and a wafer-scale chip that is 57 times larger than Nvidia’s H100, Cerebras made a bold bet that the AI infrastructure market has room for a second major player. But behind the headline numbers lies a complex story of customer concentration risk, regulatory scrutiny, and a grueling 18-month journey from a withdrawn filing to a refiled prospectus.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The S-1 Filing: What Cerebras Revealed on April 17
Cerebras Systems officially announced the filing of its registration statement for a proposed initial public offering on April 17, 2026, per CNBC’s coverage of the Form S-1 targeting the Nasdaq ticker CBRS. The company had initially filed a confidential S-1 with the SEC in late February 2026, but the public filing on April 17 made the details available for scrutiny. According to the filing, Cerebras generated $510 million in revenue in 2025, up from an annualized run rate of roughly $272 million in the first half of 2024. Bloomberg’s April 2026 review of the SEC filing reported that Cerebras swung to $87.9 million in net income for 2025, reversing a $484.8 million net loss the company had posted in 2024.
The IPO is being led by Morgan Stanley, Citigroup, Barclays, and UBS as joint lead underwriters, with Mizuho and TD Cowen as bookrunners, and Needham, Craig-Hallum, Wedbush Securities, Rosenblatt, and Academy Securities as co-managers. The company is targeting a mid-May 2026 listing window, though final pricing and share allocation remain subject to market conditions. The proposed valuation range of $22 to $25 billion would make it one of the most valuable AI chip companies to go public since Nvidia itself.
“Cerebras is entering the public markets at a moment when investor appetite for AI infrastructure is at an all-time high, but so is scrutiny of unit economics,” said Stacy Rasgon, semiconductor analyst at Bernstein Research. “The $510 million revenue figure is impressive for a company that was doing sub-$100 million just two years ago, but investors will want to see the path to diversified revenue streams beyond two anchor customers.”
From SeaMicro to Wafer-Scale: The Cerebras Origin Story
Cerebras was founded in 2016 by Andrew Feldman (CEO) and Michael James (Chief Architect), who previously co-founded SeaMicro, a low-power server company acquired by AMD for $334 million in 2012. The founding insight was deceptively simple: traditional GPUs are constrained by the reticle limit of semiconductor lithography, which restricts chip size to roughly 800 square millimeters. Cerebras asked what would happen if you used the entire silicon wafer – all 46,225 square millimeters – as a single chip.
The result was the Wafer-Scale Engine (WSE), first unveiled in 2019 as the WSE-1, followed by the WSE-2 in 2021, and the current WSE-3 in March 2024. Each generation has roughly doubled performance while maintaining the same power envelope. The company has raised over $4 billion in private funding across eight rounds, including a $1.1 billion Series G round in October 2025 that valued Cerebras at $8.1 billion, according to The Next Platform, followed roughly four months later by a $1 billion Series H in February 2026 that nearly tripled the valuation to $23 billion, per Sacra. Key investors include Tiger Global, AMD, Fidelity, Alpha Wave Global, and Abu Dhabi’s G42.
The company’s journey has not been without setbacks. Cerebras originally filed an S-1 in September 2024, but was forced to pull the offering after the U.S. Committee on Foreign Investment in the United States (CFIUS) launched a review of G42’s minority stake. The investigation centered on concerns that AI chip technology could be transferred to China through G42’s historical business ties. Cerebras formally withdrew its IPO plans in October 2025, according to IPOScoop, the same month the CFIUS review concluded after G42’s stake was restructured to non-voting shares – clearing the path for the company to refile for a 2026 listing.
The WSE-3: 4 Trillion Transistors on a Single Wafer
The Cerebras WSE-3, fabricated on TSMC’s 5nm process, is the technical centerpiece of the IPO story. At 46,225 square millimeters, it is 57 times larger than Nvidia’s H100 GPU. The chip packs 4 trillion transistors, 900,000 AI-optimized cores, and 44 GB of on-chip SRAM memory with a bandwidth of 21 petabytes per second. Peak performance reaches 125 petaFLOPS in FP16 with sparsity optimizations. By August 2026, The Register detailed an upgraded WSE-3T variant that pushes those specs further, with each chip delivering 250 petaFLOPS alongside 44 GB of on-chip SRAM at 43.2 petabytes per second of bandwidth.
The architecture eliminates the off-chip memory bottleneck that plagues GPU-based systems. In a traditional GPU cluster, data must travel between the GPU’s on-chip cache (roughly 50 MB for an H100), external HBM memory, and across network interconnects between servers. Each hop adds latency and consumes power. By keeping 44 GB of SRAM directly on the wafer, Cerebras achieves memory bandwidth that is 7,000 times higher than an H100’s 3 TB/s HBM3 bandwidth, according to the company’s published specifications.
“The WSE-3 is an engineering marvel, but the real question is whether wafer-scale computing can compete on total cost of ownership against Nvidia’s rapidly improving GPU roadmap,” said Patrick Moorhead, CEO of Moor Insights and Strategy. “Nvidia’s Blackwell B200 and the upcoming Vera Rubin platform are closing the performance-per-dollar gap with each generation.”
| Specification | Cerebras WSE-3 | Nvidia H100 | Nvidia B200 |
|---|---|---|---|
| Chip Area | 46,225 mm² | 814 mm² | 1,040 mm² |
| Transistors | 4 trillion | 80 billion | 208 billion |
| AI Cores | 900,000 | 16,896 CUDA + 528 Tensor | 18,432 CUDA + 576 Tensor |
| On-Chip Memory | 44 GB SRAM | 50 MB L2 Cache | 60 MB L2 Cache |
| Memory Bandwidth | 21 PB/s | 3.35 TB/s (HBM3) | 8 TB/s (HBM3e) |
| Peak AI Performance (FP16) | 125 PFLOPS | 1.98 PFLOPS | 4.5 PFLOPS |
| Process Node | TSMC 5nm | TSMC 4nm | TSMC 4nm |
The $10 Billion OpenAI Contract That Changed Everything
The single most important development in Cerebras’s IPO story is the $10 billion multi-year compute contract with OpenAI, disclosed in the S-1. The deal represents a transformative shift for a company that previously depended almost entirely on a single customer. In the first half of 2024, G42 accounted for 87% of Cerebras’s revenue, with $1.43 billion in long-term commitments. The extreme customer concentration was one of the primary risk factors cited in the 2024 S-1 filing and was likely a contributing factor in the decision to withdraw that prospectus.
The OpenAI contract positions Cerebras as a key inference provider for the company’s growing fleet of AI models. While OpenAI trains its largest models on Nvidia GPU clusters, the company has been seeking alternatives for inference workloads, where Cerebras’s wafer-scale architecture can deliver significantly lower latency. According to Cerebras, its CS-3 systems can serve large language models at speeds up to 15 times faster than comparable GPU-based solutions.
“OpenAI’s decision to sign a $10 billion deal with Cerebras is a clear signal that the AI industry is moving beyond single-vendor GPU infrastructure,” said Dylan Patel, chief analyst at SemiAnalysis. “This is not just about performance – it is about supply security. With Nvidia’s allocation constraints and 12-month lead times, OpenAI needed a credible alternative supplier.”
Revenue Growth and the Path to Profitability
Cerebras’s revenue trajectory reflects the explosive growth of the AI infrastructure market. The company’s H1 2024 revenue of $136.4 million annualized to roughly $272 million – already a significant jump from prior years. Cerebras Systems disclosed that full-year 2025 revenue of $510 million represents a 76% year-over-year increase from a $290.3 million 2024 base. The $87.9 million in net income for 2025 – a sharp reversal from the $484.8 million net loss posted in 2024, per the SEC filing reported by Bloomberg in April 2026 – suggests that Cerebras has reached meaningful profitability just as it heads to market.
The key question for investors is whether this growth rate is sustainable. The $10 billion OpenAI contract provides visibility into future revenue, but the timeline for recognition remains unclear. If the contract spans five years, it could contribute roughly $2 billion annually – a transformative figure that would place Cerebras among the fastest-growing semiconductor companies in history. However, the company’s S-1 cautions that contract terms are subject to modification and that actual revenues may vary significantly from committed amounts.
The broader AI chip market provides a tailwind. According to industry estimates, the AI accelerator market reached approximately $65 billion in 2025 and is projected to exceed $120 billion by 2028, driven by hyperscaler demand for training and inference compute. Nvidia currently commands an estimated Nvidia holds over **80%** share of this market, making even small share gains enormously valuable.[1][3][4] If Cerebras captures just 3 to 5% of the market by 2028, it would justify a revenue base of $3.6 to $6 billion – well above the current run rate.
The G42 Risk Factor: 87% Customer Concentration
Despite the OpenAI diversification, the G42 relationship remains the most significant risk factor in the Cerebras S-1. Abu Dhabi-based Group 42 (G42), an AI and cloud computing company with reported ties to the UAE’s sovereign wealth apparatus, accounted for 87% of Cerebras’s H1 2024 revenue. G42 invested $335 million in Cerebras through a direct share purchase, creating a customer-investor dynamic that raised red flags with regulators.
The CFIUS investigation centered on concerns that Cerebras’s wafer-scale technology could be transferred to China through G42’s historical business relationships. In 2024, G42 agreed to divest its Chinese investments and signed a partnership with Microsoft as part of a broader restructuring designed to address U.S. national security concerns. Cerebras’s S-1 notes that G42’s stake was restructured to non-voting shares as a condition of regulatory clearance.
Even with regulatory clearance, the concentration risk is stark. The S-1 discloses $1.43 billion in long-term commitments from G42, but losing this customer – or having the relationship further restricted by U.S. export controls – could eliminate a majority of the company’s revenue base. Analysts note that the transition from “one dominant customer to two” (G42 and OpenAI) is progress, but hardly constitutes a diversified revenue stream.
The Full UAE Picture: MBZUAI and G42 Together at 86% of 2025 Revenue
A closer reading of the April 17, 2026 S-1 by the Futurum Group’s research team adds important detail to the customer-concentration story for the full 2025 fiscal year, a breakdown CNBC corroborated in its own April 2026 coverage of the filing. The filing discloses that 86% of Cerebras’s $510 million in 2025 revenue came from just two UAE-affiliated entities: the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) contributed 62%, while G42 itself accounted for 24%. This breakdown reframes the risk profile from “single customer concentration” to “single-country concentration,” with Abu Dhabi-linked institutions collectively representing the vast majority of revenue.
The MBZUAI relationship is particularly notable because it is structured around research-grade Condor Galaxy supercomputer deployments rather than the commercial inference workloads that anchor the OpenAI agreement. MBZUAI’s procurement is closely tied to UAE national AI strategy and operates on multi-year capital cycles, which lends durability to the revenue stream but also concentrates exposure to a single state-aligned buyer. The 76% year-over-year revenue growth from 2024 to 2025 disclosed in the S-1 was therefore overwhelmingly UAE-driven, not a function of broad commercial diversification across hyperscalers or enterprise customers.
For prospective public-market investors, the implication is straightforward: the OpenAI Master Relationship Agreement is not just a growth catalyst but the principal mechanism by which Cerebras can credibly claim revenue diversification through 2026 and 2027. Without OpenAI ramping on schedule, the UAE share of revenue would remain near 86%, leaving the company structurally exposed to any future tightening of U.S. export controls on AI compute exports to the Middle East. The MBZUAI line item is also worth tracking separately from G42 in future quarterly disclosures, since the two entities operate under different governance structures despite sharing geographic origin.
Cerebras vs. Nvidia: Can a Startup Challenge the GPU Monopoly?
The central competitive question for Cerebras is whether wafer-scale computing can carve out a sustainable niche against Nvidia’s GPU empire. Nvidia generated $130.5 billion in data center revenue in fiscal year 2026 (ending January 2026), giving it roughly Nvidia holds over **80%** market share in AI accelerators.[1][3][4] The company’s software ecosystem – CUDA, cuDNN, TensorRT – has created a developer lock-in that makes switching costs prohibitively high for most organizations.
Cerebras’s counter-argument rests on three pillars. First, the WSE-3’s on-chip SRAM architecture eliminates the memory wall that constrains GPU inference performance, delivering up to 15x lower latency on large language model serving. Second, the company’s CS-3 system simplifies deployment by replacing clusters of hundreds of GPUs with a single wafer-scale chip, reducing operational complexity. Third, Cerebras offers an inference-as-a-service model that removes the capital expenditure burden entirely, allowing customers to pay per token rather than purchasing hardware.
However, Nvidia is not standing still. The Vera Rubin platform, announced at GTC 2026 with 336 billion transistors, promises a 5x performance leap over the current Blackwell generation. AMD’s MI400 series, with 320 billion transistors, is also closing the gap. And custom silicon from hyperscalers – Google’s TPU v6, Amazon’s Trainium 3, Microsoft’s Maia 2 – is creating additional competitive pressure.
“The AI chip market is large enough to support multiple winners, but Cerebras needs to prove that wafer-scale economics work at hyperscaler volumes,” said Ben Bajarin, CEO of Creative Strategies. “The $10 billion OpenAI deal is proof of concept. Now they need three or four more deals like it to justify a $25 billion valuation.”
Valuation Analysis: Is $23 Billion Justified?
The proposed IPO valuation of $22 to $25 billion implies a forward revenue multiple of roughly 43 to 49x based on 2025 revenue of $510 million. This is aggressive by historical semiconductor standards – Nvidia trades at roughly 25x forward revenue, and AMD at approximately 10x – but not unprecedented for high-growth AI infrastructure plays. Cerebras’s February 2026 Series H round at $23 billion was led by Tiger Global, with participation from AMD and Fidelity, lending institutional credibility to the valuation.
The secondary market provides additional data points. Pre-IPO shares were trading at $102 to $107 per share on platforms like Forge and EquityZen, implying a fully diluted valuation of $26 to $28 billion – a premium to the proposed IPO range. This secondary market activity suggests that sophisticated investors believe the IPO is priced conservatively, with potential for a first-day pop similar to Arm’s September 2023 listing (which surged 25% on its first trading day).
The bull case for the valuation rests on the $10 billion OpenAI contract. If recognized over five years, this contract alone represents roughly $2 billion in annual revenue, nearly 4x the 2025 figure. Combined with G42’s $1.43 billion in commitments and organic growth from new customers, Cerebras could plausibly reach $3 to $4 billion in annual revenue by 2028, compressing the valuation multiple to a more reasonable 6 to 8x.
The May 4 Amended S-1: Pricing the Deal at $115 to $125 Per Share
On May 4, 2026, Cerebras filed an amended S-1 with the SEC that converted the April filing’s broad valuation guidance into concrete pricing terms. The amendment disclosed that the company will sell 28 million Class A shares at a price range of $115 to $125 per share, raising up to $3.5 billion at the high end of the range – substantially more than the $2 billion initially signaled in the April 17 filing. The mid-point of the range implies gross proceeds of approximately $3.36 billion before underwriting discounts, making this one of the largest AI infrastructure IPOs on the Nasdaq since Arm’s September 2023 debut. The ticker remains CBRS, with listing targeted for mid-May 2026.
The expanded raise reflects strong reverse-inquiry demand from institutional investors during the early stages of the roadshow. Initial price discovery on the secondary market had implied valuations of $26 to $28 billion, so the $115 to $125 range – which on a fully diluted basis aligns the deal with the upper end of those private prints – falls closer to the secondary signal than to the conservative anchor of the April S-1. The amendment also clarified the over-allotment option granted to the underwriting syndicate, with Morgan Stanley, Citigroup, Barclays, and UBS retaining a customary green shoe that could push total proceeds higher if exercised in full.
For comparison, the $115 to $125 range represents a meaningful premium to the $102 to $107 secondary market prints that circulated on Forge and EquityZen earlier in 2026. That gap suggests two dynamics at work: first, that the institutional order book is comfortably oversubscribed despite the larger share count; second, that pre-IPO secondary investors who bought at the lower band stand to realize an immediate paper gain of roughly 10 to 22% if the deal prices at the middle to upper end of the range. The pricing trajectory mirrors the Arm IPO playbook, where a measured initial range gave way to a stronger final print following a strong reception during the roadshow.
The amended filing also gives investors a sharper baseline for thinking about supply-and-demand on day one of trading. A 28 million share offering at $120 per share – the mid-point – translates to a free-float of roughly $3.36 billion at issue, which is a meaningful improvement over the thin float that characterized many recent AI-adjacent debuts. Greater float generally translates to lower opening-day volatility and a smoother path for index inclusion later in 2026, both factors that growth-mandate institutional investors weigh heavily when sizing positions.
Price Range Bumped to $150-$160 After 20x Oversubscription
Demand for the Cerebras book overwhelmed the initial marketing range. According to TECHi.com’s tracking of the deal in May 2026, the order book closed roughly 20x oversubscribed, prompting the underwriters to revise the price range upward from the initial $115 to $125 band – which Reuters had pegged at a $26.6 billion implied valuation at the high end – to a new range of $150 to $160 per share. At the upper bound of the revised range, the 28 million-share offering raises gross proceeds of approximately $4.8 billion, well above the $3.5 billion implied by the May 4 amendment and more than double the $2 billion telegraphed in the April 17 S-1.
The revised pricing lifts Cerebras’s fully diluted valuation to roughly $48.8 billion, nearly double the $26.6 billion implied at the high end of the original $115 to $125 range. The jump pushes the IPO valuation well past the secondary-market signal of $26 to $28 billion seen on Forge and EquityZen earlier in 2026, and overshoots Cerebras’s own February 2026 Series H valuation of $23 billion by more than 2x. The acceleration mirrors the trajectory of recent AI infrastructure debuts where reverse-inquiry demand has consistently pushed final pricing toward – or past – the high end of the marketed ranges.
For pre-IPO holders who acquired shares on secondary platforms at $102 to $107 earlier this year, the $150 to $160 print represents an immediate paper gain of roughly 40 to 57% at pricing, well before any first-day trading pop. The dynamic underscores how decisively the AI infrastructure narrative has shifted since the original 2024 filing was withdrawn under CFIUS pressure: investor appetite for differentiated AI silicon is now strong enough to absorb a near-doubling of the deal’s headline economics during the marketing window. It also raises the bar Cerebras must clear post-listing – at $48.8 billion, the company is being priced as a top-tier AI infrastructure platform rather than as an unproven wafer-scale upstart.
The May 11 Upsize: 30 Million Shares and a $4.65 Billion Raise
The most consequential pricing development of the Cerebras roadshow came on Monday, May 11, 2026, when the company filed a second amended S-1/A that increased both the share count and the marketed price range. According to IPO Scoop, which tracks pricing amendments across the U.S. IPO calendar, the May 11 filing lifted the offering from 28 million shares to 30 million shares and revised the price range from $115 to $125 per share to $150 to $160 per share. At the top of the revised range, the deal is now expected to raise approximately $4.65 billion in gross proceeds – a meaningful step up from the roughly $3.5 billion implied by the May 4 amendment.
The pricing dynamics unfolded in two distinct stages. The May 4, 2026 amendment had set the initial marketed terms at 28 million shares at $115 to $125 per share, implying roughly $3.36 billion in gross proceeds at the midpoint, per IPO Scoop and Techi.com tracking of the deal – both of which referenced the original Cerebras press release dated April 17, 2026 announcing the initial filing. The cadence – initial filing on April 17, concrete pricing terms on May 4, upsized terms on May 11 – compresses the entire price-discovery process into a four-week window, an unusually fast progression that underscores the depth of institutional demand for AI infrastructure paper in May 2026.
The IPO Scoop entry for CBRS also lists Cerebras’s pricing-window market capitalization at $33,360.5 million – roughly $33.4 billion – with the listing scheduled for Nasdaq under the CBRS ticker. That market-cap figure, while sensitive to the final pricing point within the $150 to $160 range, anchors the deal at a level that exceeds the company’s February 2026 Series H valuation of $23 billion by approximately 45%. For pre-IPO secondary-market buyers who acquired shares at $102 to $107 on platforms like Forge and EquityZen earlier in 2026, the May 11 range implies an immediate paper gain in the range of 40 to 57% at pricing, depending on where within the band the deal ultimately clears.
Underpinning the upsize are the trailing financial metrics disclosed in the prospectus and tracked on IPO Scoop’s CBRS data page. The filing cites $510.0 million in revenue and $237.8 million in net income for the trailing 12 months ended December 31, 2025, figures sourced directly from the SEC filing. At the May 11 implied market capitalization of $33.36 billion, those trailing metrics imply a price-to-sales multiple of roughly 65x and a price-to-earnings multiple – on the trailing net income figure – of approximately 140x. Multiples at that level are aggressive even within the AI infrastructure category, and they place the burden squarely on the OpenAI Master Relationship Agreement to convert from contracted to recognized revenue on a timeline that supports the implied growth assumptions.
The decision to add 2 million additional shares alongside the price-range increase is also a meaningful structural change in the deal. Expanding the share count lifts the free float at issue from roughly $3.36 billion (28 million × $120 midpoint) to approximately $4.65 billion (30 million × $155 midpoint) – a 38% increase in dollar terms. Larger free floats typically translate into lower opening-day volatility, broader institutional eligibility for index inclusion, and a smoother absorption profile for the secondary-market shares that will become tradable once standard lock-up provisions expire. For an issuer like Cerebras, where institutional demand was reported to be roughly 20x oversubscribed during the roadshow, the choice to absorb some of that demand by selling more shares – rather than only by raising the price – signals a preference for distribution breadth over maximizing dilution-adjusted proceeds.
For investors sizing positions against the May 11 terms, the relevant question is no longer whether the IPO will price at the high end of its initial range, but whether the post-listing trading pattern can sustain the $33.36 billion implied market cap once lock-up expirations and follow-on supply enter the market later in 2026. The April 17 S-1’s disclosure of $510.0 million in trailing revenue remains the operating baseline, and the May 11 pricing range builds in a substantial premium for the conversion of the OpenAI Master Relationship Agreement into recognized revenue during the back half of 2026 and into 2027. The May 11 upsize, sourced from IPO Scoop and corroborated by the original April 17 Cerebras press release announcing the initial filing, is therefore the single most important pricing data point in the prospectus for any investor underwriting CBRS at the top of the revised range.
May 14, 2026 Final Pricing: Cerebras Prints at $185, Raises $5.5 Billion at a $56.4 Billion Valuation
On May 14, 2026, Cerebras priced its IPO at $185 per share, well above both the original $115 to $125 marketed range and the revised $150 to $160 band that the underwriters had set the previous week. According to TechCrunch’s reporting on the final print, the offering grew once again to 30 million shares, raising $5.5 billion in gross proceeds and implying a $56.43 billion fully diluted valuation at the IPO price, Reuters reported. CNBC subsequently reported that underwriters exercised a 4.5 million-share over-allotment option on top of the base offering, pushing total proceeds to roughly $5.55 billion. The final figure eclipses the roughly $4.65 billion that the May 11 upsize had pointed to and is approximately 57% above the midpoint of the original May 4 marketing range.
The May 14 print puts CBRS roughly 2.45x above the company’s February 2026 Series H valuation of $23 billion and well past the $26 to $28 billion secondary-market signals from Forge and EquityZen earlier in 2026. For pre-IPO secondary buyers who acquired shares at $102 to $107, the $185 print represents an immediate paper gain of roughly 73 to 81% at pricing, before any first-day trading pop. The trajectory – initial $22 to $25 billion guidance in the April 17 S-1, revised to $26.6 billion at the high end of the May 4 amendment, lifted to roughly $48.8 billion at the May 11 upsize, and finally cleared at $56.43 billion on May 14, per Reuters’s reporting on the sold 30 million-share offering – compresses one of the largest price-discovery rerates of any AI infrastructure IPO on record.
The final pricing also recasts how investors should read the 2025 financial disclosures, which Bloomberg’s April 2026 review of the SEC filing pinned more precisely than earlier press coverage. The headline figures – $510 million in 2025 revenue, up 76% year over year, alongside an $87.9 million net profit – now anchor the valuation. That profit figure is particularly striking when set against 2024, when the company posted a net loss of $484.8 million; the year-over-year swing from a nearly half-billion-dollar loss to a profitable year was one of the principal factors that allowed the deal to reprice from $115 to $185 over a three-week marketing window. At $185 per share on $510 million in trailing revenue, CBRS clears at roughly 111x trailing sales, a multiple that prices in continued conversion of the OpenAI Master Relationship Agreement into recognized revenue through the back half of 2026 and into 2027.
Inside the 750 Megawatt OpenAI Compute Deal
The May 2026 disclosures also sharpen the technical contours of the OpenAI contract. The agreement, announced in January 2026, commits Cerebras to supply 750 megawatts of low-latency AI compute power to OpenAI under a multi-year arrangement valued at over $10 billion. The 750MW figure is significant in its own right: it places Cerebras alongside the largest hyperscale data center deployments globally and signals that the contract is not a pilot or a hedging arrangement, but a primary inference platform commitment.
For context, 750MW of dedicated AI compute is roughly equivalent to the combined power draw of two large hyperscale data center campuses, or approximately the electrical capacity needed to serve hundreds of thousands of concurrent large-language-model users at sub-second latency. Given that the CS-3 system consumes approximately 25 kilowatts per node, 750MW translates to on the order of 30,000 CS-3 systems if dedicated entirely to compute load – a manufacturing volume that will stretch TSMC’s wafer-scale capacity and validate Cerebras’s claim that wafer-scale economics can scale to hyperscaler workloads.
The contract structure also clarifies why OpenAI specifically pursued a non-GPU supplier. OpenAI’s inference traffic for ChatGPT and its API products has scaled into the billions of tokens per day, and latency has emerged as the binding constraint on user experience for real-time applications like Advanced Voice and code generation. Cerebras’s claim of up to 15x lower latency than GPU clusters becomes a strategic asset when measured against OpenAI’s user growth trajectory rather than a marketing line. The 750MW commitment, combined with the 28 million share offering at $115 to $125 disclosed in the May 4 amendment, gives investors a clearer line of sight from contracted compute capacity to recognizable revenue than was visible in the April filing.
The $20B+ Master Relationship Agreement and AWS Bedrock Term Sheet
The Futurum Group’s S-1 teardown also clarifies the legal architecture of the OpenAI engagement. The agreement is not a single purchase order but a binding Master Relationship Agreement (MRA) valued at over $20 billion – roughly twice the $10 billion headline figure that initially circulated in press coverage. The MRA covers the 750 megawatts of contracted AI inference capacity disclosed in the S-1, with explicit expansion provisions that allow OpenAI to scale the commitment up to 2 gigawatts by 2030. At the upper bound, the deployment would represent one of the largest single-customer compute commitments ever recorded for a non-GPU silicon supplier.
The “master relationship” framing is materially different from a fixed-price hardware purchase order. An MRA establishes ongoing commercial terms that can flex with capacity additions, technology generations, and pricing renegotiations. For Cerebras, that translates into a multi-year runway for both WSE-3 and the planned WSE-4 generation to flow through the same legal vehicle, smoothing revenue recognition and reducing the friction of repeated procurement cycles. For OpenAI, the structure secures inference capacity at scale without locking it into a single hardware generation as the workload mix between training and inference continues to shift toward serving.
Beyond OpenAI, the S-1 also discloses an AWS Bedrock integration term sheet that would make Cerebras-served models available through Amazon’s managed foundation model platform. While the term sheet is not yet a leading agreement and carries no contractual revenue commitment, the disclosure is strategically important: it signals that Cerebras is being evaluated for inclusion in the hyperscaler distribution layer rather than competing against it. If the AWS engagement converts into a production deployment, it would directly address one of the central risk factors flagged in the April S-1 – the absence of meaningful hyperscaler distribution beyond a single anchor customer – and provide a third leg of revenue diversification on top of MBZUAI, G42, and the OpenAI MRA.
Key Investors and the Capital Stack
Cerebras has raised over $4 billion in private funding across eight rounds. The most recent Series H round in February 2026 raised $1 billion at a $23 billion valuation, led by Tiger Global Management. Previous rounds included a $1.1 billion Series G in October 2025 that valued the company at $8.1 billion, according to The Next Platform, and a $250 million Series F. Notable investors include AMD, which has a strategic interest in supporting alternatives to Nvidia’s dominance, and Fidelity Management, which has been one of the most active crossover investors in late-stage AI companies.
The presence of G42 as both a customer and an investor adds complexity to the capital structure. G42’s $335 million share purchase gives it a meaningful equity position, but the restructuring to non-voting shares limits its governance influence. The S-1 filing will need to disclose the exact ownership breakdown, related-party transaction details, and any lock-up provisions that could affect post-IPO trading dynamics.
The $2 billion in targeted IPO proceeds will likely be allocated to manufacturing expansion, R&D for the next-generation WSE-4, and working capital to support the OpenAI contract ramp. Cerebras relies on TSMC for fabrication, and securing wafer-scale production capacity at the volumes required by a $10 billion contract will require significant capital investment and supply chain coordination.
The CS-3 System and Inference-as-a-Service
Cerebras ships its chips as part of the CS-3 system, a complete inference and training platform built around a single WSE-3 wafer. The CS-3 consumes approximately 25 kW of power and replaces hundreds of GPUs for specific workloads. The company claims that a cluster of CS-3 systems can scale to 2,048 nodes delivering 256 exaFLOPS of aggregate compute, enough to train and serve the largest frontier models. By August 2026, Cerebras had followed the CS-3 with the CS-4 system, which packs three of the upgraded WSE-3T chips into a single unit to deliver 750 petaFLOPS of AI compute, according to Yahoo Finance.
In addition to hardware sales, Cerebras has built an inference cloud service that allows customers to access WSE-3 compute on demand. This service directly competes with Nvidia’s DGX Cloud and the growing ecosystem of GPU cloud providers like CoreWeave and Lambda. The inference-as-a-service model is particularly compelling for Cerebras because it converts hardware sales into recurring revenue, improving predictability and lifetime value per customer.
The company claims its inference service can deliver up to 3 million tokens per second for large language model workloads, with latency advantages that are particularly significant for real-time applications like conversational AI and code generation. OpenAI’s reported interest in Cerebras for inference – rather than training – aligns with this positioning, as inference now accounts for the majority of compute spending at large AI labs.
Market Context: The AI Chip IPO Wave
Cerebras is filing into one of the most active IPO markets for AI infrastructure companies in history. Databricks is pursuing a $134 billion listing that would be the largest enterprise software IPO ever. CoreWeave, the Nvidia-backed GPU cloud provider, went public in March 2026 amid debate over its $30 billion debt-fueled expansion. SiFive, the RISC-V chip startup, raised $400 million at a $3.65 billion valuation and is targeting an IPO. The broader semiconductor sector has seen its combined market capitalization exceed $5 trillion for the first time, driven by insatiable AI demand.
The Cerebras IPO arrives at a particularly favorable moment. The Nasdaq Composite has recovered from its tariff-driven sell-off in late 2025, and AI chip stocks are trading near all-time highs. Nvidia’s market capitalization exceeds $3.5 trillion, providing a valuation benchmark that makes Cerebras’s $23 billion target look modest by comparison. ASML’s strong Q1 2026 earnings – with revenue of €8.8 billion and 53% margins – have reinforced investor confidence in the semiconductor supply chain.
“The window for AI chip IPOs is wide open, but it won’t stay that way forever,” said Matt Kennedy, senior IPO strategist at Renaissance Capital. “Cerebras is smart to file now while investor enthusiasm for AI infrastructure remains high. If interest rates rise or the broader market corrects, this valuation could be much harder to achieve six months from now.”
Risk Factors Investors Should Watch
The Cerebras S-1 outlines several material risk factors beyond customer concentration. Manufacturing risk is paramount: the WSE-3 is fabricated on a single 300mm silicon wafer at TSMC, and defect rates at wafer scale are inherently higher than for traditional die-sized chips. Cerebras has developed proprietary redundancy and yield optimization techniques, but scaling production to meet the OpenAI contract will test these systems at unprecedented volumes.
Export control risk is another significant factor. The U.S. government’s expanding restrictions on AI chip exports to China and certain Middle Eastern countries could affect Cerebras’s ability to serve G42 and other international customers. The CFIUS resolution addressed G42’s specific investment, but broader export control changes – such as the October 2025 updates to the Entity List – could introduce new constraints on wafer-scale technology transfers.
Competitive risk from Nvidia’s roadmap is perhaps the most existential threat. Nvidia’s annual release cadence – Hopper in 2023, Blackwell in 2024, Vera Rubin planned for 2026 – means that the performance gap between GPUs and the WSE-3 narrows with each generation. If Nvidia’s next-generation chips close the inference latency gap, Cerebras’s primary competitive advantage could erode.
| Risk Factor | Severity | Mitigation |
|---|---|---|
| G42 customer concentration (87% H1 2024) | High | $10B OpenAI contract diversifies revenue |
| CFIUS and export control restrictions | Medium-High | G42 stake restructured to non-voting shares |
| Wafer-scale manufacturing yield | Medium | Proprietary redundancy techniques at TSMC |
| Nvidia competitive response (Vera Rubin) | High | WSE-4 development underway |
| GAAP profitability (net loss $75.7M) | Medium | Non-GAAP net income of $237.8M in 2025 |
| TSMC fabrication dependency | Medium | Long-term supply agreements in place |
What This Means for the Broader AI Chip Market
A successful Cerebras IPO would send a powerful signal to the AI chip market: there is room for architecture-level innovation beyond the GPU paradigm. For years, the semiconductor industry has watched Nvidia accumulate an estimated Nvidia holds over **80%** share of the AI accelerator market with growing unease.[1][3][4] Hyperscalers have responded by developing custom silicon – Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia – but these chips are captive to their respective cloud platforms and do not serve the broader market.
Cerebras represents a different path: a merchant silicon company that sells to anyone willing to buy, much as Nvidia does, but with a radically different architecture. If the IPO succeeds and the company executes on the OpenAI contract, it could encourage further investment in alternative AI chip architectures. Companies like Groq (with its Language Processing Unit), SambaNova (with its reconfigurable dataflow architecture), and Tenstorrent (with Tensix cores) are all pursuing non-GPU approaches to AI compute. A validated Cerebras IPO at $23 billion would raise the ceiling for all of them.
For Nvidia, a successful Cerebras listing is more of a validation than a threat – at least in the near term. Nvidia’s data center revenue of $130.5 billion dwarfs Cerebras’s $510 million, and the CUDA software ecosystem remains an unmatched competitive moat. But the mere existence of a credible, publicly traded alternative could give hyperscaler customers more negotiating clout, potentially pressuring Nvidia’s margins over time.
Predictions: Five Scenarios for Cerebras Post-IPO
Based on the S-1 data, market conditions, and competitive dynamics, here are five scenarios for Cerebras in the 12 to 24 months following its IPO:
1. Revenue acceleration to $1.5 billion by 2027. If the OpenAI contract begins ramping in H2 2026 and G42 commitments are recognized on schedule, Cerebras could nearly triple its 2025 revenue within 18 months. This would validate the growth story and compress the valuation multiple to more sustainable levels.
2. A second hyperscaler deal by Q4 2026. The OpenAI contract demonstrates product-market fit for inference workloads. At least one additional hyperscaler – likely Meta or a major cloud provider – will evaluate Cerebras for latency-sensitive inference, potentially resulting in a multi-billion-dollar commitment.
3. WSE-4 announcement in early 2027. Cerebras has maintained a roughly two-year cadence between wafer-scale generations (WSE-1 in 2019, WSE-2 in 2021, WSE-3 in 2024). The WSE-4, likely on TSMC’s 3nm process, could double performance again and maintain the architecture’s competitive positioning against Nvidia’s Vera Rubin.
4. Acquisition interest from a hyperscaler. At $23 billion, Cerebras is expensive but not out of reach for companies like Microsoft (which has expressed interest in in-house AI silicon) or Oracle (which is restructuring around AI infrastructure). A strategic acquisition would provide instant access to wafer-scale technology without the development risk.
5. Export control headwinds by mid-2027. The U.S. government’s expanding AI chip export restrictions could affect Cerebras’s ability to serve international customers, particularly in the Middle East. Any tightening of controls on wafer-scale technology – a category that did not exist when current regulations were drafted – could introduce new compliance burdens.
The Underwriter Syndicate and IPO Mechanics
The choice of Morgan Stanley as lead left underwriter is significant. Morgan Stanley has led some of the most high-profile tech IPOs of the past decade, including Rivian’s $11.9 billion offering in 2021 and Arm’s $4.87 billion listing in 2023. The firm’s involvement signals confidence in investor demand and suggests that the $2 billion raise is well within the market’s absorption capacity.
The broad underwriter syndicate – with 10 banks named in the filing – reflects both the deal’s complexity and Cerebras’s desire to maximize retail and institutional distribution. The inclusion of mid-tier firms like Needham, Craig-Hallum, and Rosenblatt indicates that Cerebras is targeting semiconductor-specialist investors in addition to the generalist growth funds that typically dominate large tech IPOs.
Lock-up provisions, which typically restrict insider sales for 180 days post-IPO, will be a key factor in post-listing trading dynamics. With over $4 billion in private capital invested, early investors and employees will hold significant positions that could create selling pressure once the lock-up expires. The S-1 filing does not yet specify the exact lock-up terms, but standard 180-day restrictions are expected.
How Cerebras Compares to Recent AI IPOs
The Cerebras IPO enters a market that has seen mixed results for AI-adjacent listings. Arm Holdings went public in September 2023 at a $54.5 billion valuation, raising $4.87 billion in the largest IPO of that year. The stock has since more than doubled, driven by AI licensing revenue growth. CoreWeave, which listed in March 2026, has faced more skepticism, with analysts questioning whether its $30 billion in debt-financed GPU capacity is sustainable.
Cerebras sits between these two extremes. Unlike Arm, which earns royalties on virtually every smartphone chip sold, Cerebras is a hardware company with concentrated customer relationships and binary execution risk. Unlike CoreWeave, which resells Nvidia GPUs as a service, Cerebras designs and sells its own silicon, giving it higher gross margins but also higher R&D costs. The $510 million in revenue and $237.8 million in non-GAAP net income suggest a more mature business than CoreWeave’s at the time of its listing, but the 87% customer concentration in H1 2024 is a red flag that has no parallel among recent tech IPOs.
May 2026 Synthesis: Three Disclosures That Now Define the CBRS Investment Thesis
Three disclosures from the April 17 and May 4, 2026 S-1 filings – when read together – now anchor the Cerebras investment thesis far more concretely than the April prospectus alone did. The first is the May 4, 2026 amendment, which converted the original $22 to $25 billion valuation guidance into a marketed range of $115 to $125 per share for 28 million Class A shares on the Nasdaq under ticker CBRS, implying a fully diluted valuation near $26.6 billion at the top of that band. The second is the $20 billion Master Relationship Agreement (MRA) with OpenAI covering 750 MW of AI inference compute capacity with an option to expand to 2 GW by 2030. The third is the customer-concentration disclosure showing that two UAE-based entities – MBZUAI and G42 – accounted for 86% of 2025 revenue.
The pricing trajectory itself is the loudest signal in the May 2026 roadshow. Underwriters revised the marketed range upward to $150 to $160 per share after the order book closed more than 20x oversubscribed, a level of demand that historically only appears for AI infrastructure deals where the institutional bid is structurally undersupplied. The move from the initial $115 to $125 band to the revised $150 to $160 band is a roughly 28% increase in price during the marketing window, and it occurred without any new operational disclosure – meaning the entire repricing was driven by demand discovery rather than fresh fundamental data. That dynamic typically compresses first-day pop expectations but tightens the post-IPO support level, because the marginal buyer in the book has already been forced to underwrite the higher valuation.
The MRA’s 2 GW by 2030 expansion option is the disclosure most likely to be underappreciated by generalist investors reading the prospectus. A move from 750 MW to 2 GW would represent roughly a 2.7x increase in contracted capacity over a four-year window, and at the CS-3 system’s approximate 25 kW per node power draw, the 2 GW upper bound would translate to a fleet on the order of 80,000 wafer-scale systems dedicated to OpenAI inference alone. Whether or not that option is exercised, the existence of a contractual ceiling at 2 GW reframes the OpenAI relationship from a single $20 billion line item into an open-ended capacity reservation that competes directly with the largest GPU-cluster build-outs publicly disclosed by hyperscalers in 2025 and 2026.
The 76% year-over-year revenue growth figure disclosed in the April 17 S-1 also lands differently when paired with the 86% UAE concentration. Growth of that magnitude – taking the company from a roughly $290 million 2024 base to $510 million in 2025 – would normally be presented as evidence of broad commercial pull. In Cerebras’s case, the same 76% growth was overwhelmingly funded by MBZUAI’s 62% revenue share and G42’s 24% share, meaning that the headline growth rate is, for now, a function of UAE national-AI procurement cycles rather than diversified demand. That reframing matters because the public-market valuation premium typically applied to 76% growth assumes that growth is distributed across customers and verticals; here, it is not.
Profitability sustainability is the third axis investors will discount during May 2026 pricing. The April 17 S-1 disclosed a widening non-GAAP net loss of $75.7 million against the $510 million revenue base, which raises a question the prospectus does not fully resolve: whether wafer-scale unit economics improve materially as the OpenAI MRA ramps from contracted to recognized revenue, or whether the gross-margin profile of the inference-as-a-service model structurally caps near-term operating clout. The widening loss against 76% top-line growth suggests that operating expenses are still scaling roughly in line with revenue, which is acceptable during a hyperscaler ramp but will need to inflect by the time the company reports its first two post-IPO quarters in 2026.
Read together, the May 2026 disclosures move the Cerebras thesis from “wafer-scale upstart with a marquee customer” to “single-architecture inference platform with a 2 GW contractual ceiling, priced at the upper bound of secondary-market signals, and structurally exposed to UAE concentration until the OpenAI MRA ramps.” That is a more bankable thesis than what was visible in April, but it is also a thesis that requires the OpenAI MRA to convert from contracted capacity into recognized revenue on schedule. Investors sizing positions ahead of the mid-May 2026 listing will be paying the revised $150 to $160 range explicitly for that conversion – and the post-IPO trading pattern will reflect whether the market believes the conversion timeline disclosed in the prospectus is realistic.
What Happens Next: Timeline and Catalysts
With the S-1 now public, the next steps follow a well-defined IPO process. Cerebras will enter a quiet period during which management cannot make public statements about the offering beyond what is in the prospectus. The SEC will review the filing and may request amendments, a process that typically takes four to six weeks. If the review proceeds smoothly, the company could begin its roadshow in early May, with pricing and listing targeted for mid-May 2026.
Key catalysts to watch include: the SEC’s response to the S-1, any additional customer announcements that could diversify the revenue base, Nvidia’s response at upcoming industry events, and broader market conditions. The federal funds rate, currently at 4.25 to 4.50%, remains elevated, but the Federal Reserve’s signaled rate cuts in H2 2026 could provide a tailwind for growth-stock valuations.
For investors, the Cerebras IPO represents a rare opportunity to invest in a differentiated AI chip architecture at a pivotal moment in the market’s development. The risks are real – customer concentration, regulatory uncertainty, and competitive pressure from Nvidia – but so is the upside if wafer-scale computing proves to be the right architecture for the inference-dominated future of AI compute.
Frequently Asked Questions
When is the Cerebras IPO date?
Cerebras filed its public S-1 registration statement on April 17, 2026, and filed an amended S-1 on May 4, 2026, with concrete pricing terms. The company is targeting a mid-May 2026 listing on the Nasdaq under the ticker symbol CBRS. Final pricing and the exact listing date depend on the SEC review process and market conditions.
What is the Cerebras IPO valuation?
The May 4, 2026 amended S-1 initially priced the offering at $115 to $125 per share for 28 million Class A shares. After roughly 20x oversubscription during the roadshow, the range was lifted to $150 to $160 per share, implying a fully diluted valuation of approximately $48.8 billion and gross proceeds of about $4.8 billion. The company’s most recent private round in February 2026 valued it at $23 billion.
How much revenue does Cerebras generate?
Cerebras reported $510 million in revenue for 2025, up 76% year over year, with 86% of that revenue coming from two UAE-affiliated entities – MBZUAI (62%) and G42 (24%). The company reported non-GAAP net income of $237.8 million for 2025.
What is the Cerebras WSE-3 chip?
The WSE-3 (Wafer-Scale Engine 3) is Cerebras’s flagship AI chip, fabricated on TSMC’s 5nm process. It covers an entire 300mm silicon wafer at 46,225 square millimeters, making it 57 times larger than Nvidia’s H100. It contains 4 trillion transistors, 900,000 cores, and 44 GB of on-chip SRAM.
Who are Cerebras’s biggest customers?
Cerebras’s largest customers in 2025 were MBZUAI (62% of revenue), G42 (24% of revenue), and OpenAI, which signed a binding $20 billion+ Master Relationship Agreement for 750 megawatts of low-latency AI inference capacity, expandable to 2 gigawatts by 2030. The S-1 also discloses an AWS Bedrock integration term sheet.
How does Cerebras compete with Nvidia?
Cerebras competes with Nvidia primarily in AI inference workloads, where its wafer-scale architecture delivers up to 15x lower latency than GPU-based solutions. However, Nvidia dominates the broader AI chip market with roughly 80% share and a deeply entrenched CUDA software ecosystem that creates significant switching costs.
What was the final Cerebras IPO price on May 14, 2026?
On May 14, 2026, Cerebras priced its IPO at $185 per share, above both the original $115 to $125 range and the later $150 to $160 range. The offering grew to 30 million shares and raised $5.5 billion in gross proceeds, implying a $56.4 billion fully diluted valuation at the IPO price – roughly 2.45x the company’s February 2026 Series H valuation of $23 billion.
How did Cerebras’s profitability change from 2024 to 2025?
Cerebras swung from a net loss of nearly $500 million in 2024 to a $237.8 million net profit in 2025 on $510 million in revenue, per TechCrunch’s coverage of the prospectus. The year-over-year swing of roughly three-quarters of a billion dollars in net income was a key factor that allowed the May 2026 IPO to reprice from a $115 to $125 marketed range to a final print of $185 per share.
