← All Posts
· 39 min read

Cerebras Systems Inc. (CBRS)

Cerebras Systems (CBRS) built a computer chip the size of a dinner plate, and in May 2026 it turned that bet into the largest semiconductor IPO in history — $6.4 billion raised, a 68% first-day pop, and then a 58% slide from the high in six weeks. This deep dive unpacks what actually happened: a $25 billion contracted backlog dominated by a single OpenAI agreement that also makes OpenAI a lender and a prospective 10% shareholder, an 86% revenue concentration in two Abu Dhabi entities, and a deliberate margin collapse as the company rents its own systems back to serve demand faster. We work through the fundamentals, the wafer-scale architecture that lets Cerebras skip the HBM shortage entirely, the DOE and national-lab relationships, retail and institutional positioning including a rapidly building short base, and how CBRS compares to NVIDIA, Groq, and CoreWeave.

Deep Analysis of Cerebras Systems Inc. (CBRS)#

Sector: Electronic Technology

Industry: Semiconductors / AI Accelerators & Inference Cloud Infrastructure

This article is for informational purposes only and is not investment advice. Figures were gathered from public sources listed at the end.

Introduction#

Cerebras Systems Inc. is a Sunnyvale, California–based artificial intelligence infrastructure company built around a single contrarian engineering bet: instead of cutting hundreds of small chips from a silicon wafer and stitching them together with networking, Cerebras uses the entire wafer as one chip. Its Wafer-Scale Engine 3 (WSE-3) is roughly 46,225 mm² — about 57 times larger than the largest competing GPU die — carrying roughly 4 trillion transistors, ~900,000 AI cores, and 44 GB of on-chip SRAM. That architecture makes Cerebras extraordinarily fast at inference decode, the sequential, memory-bandwidth-bound half of running an AI model, and it lets the company sidestep the high-bandwidth memory (HBM) shortage that is currently the binding constraint for nearly every GPU maker on earth. Founded in 2015 by Andrew Feldman, Sean Lie, Gary Lauterbach, Michael James, and Jean-Philippe Fricker, the company sells two things: CS-3 systems for on-premises deployment, and time on those systems through Cerebras Cloud and partner clouds. Cerebras went public on May 14, 2026 at $185 per share after raising $5.55 billion (later $6.4 billion gross including the greenshoe) in what the company calls the largest semiconductor IPO of all time. The stock opened at $350, closed its first day at $311.07, and has spent the intervening three months giving most of that back. With roughly 700–780 employees, a $25 billion contracted backlog dominated by a single OpenAI agreement, and a customer list that until very recently was ~86% Abu Dhabi, CBRS is one of the most polarizing names in public markets — and its second-ever earnings report lands on August 12, 2026.

Fundamental Analysis#

Cerebras is financially stronger than most recently-IPO’d AI hardware companies and weaker than its headline numbers first suggest. The good news is real: revenue nearly doubled year-over-year in Q1 2026, gross margins expanded, the core (non-GAAP) operating loss narrowed to near breakeven, operating cash flow was positive, and the balance sheet is now flush with roughly $9–10 billion of combined pre- and post-quarter capital. The complications are equally real: the widely-cited 2025 GAAP net income of $237.8 million was driven almost entirely by a one-time, non-cash gain of roughly $363 million — on an operating basis 2025 was a $146 million loss — and management has guided to a sharp, deliberate margin contraction through the rest of 2026 as it rents capacity back from an existing customer to serve demand faster. Valuation metrics are also unusually noisy because of the dual-class structure and the pre-IPO stockholders’ deficit, so published P/E, P/B, and market-cap figures disagree wildly across data providers.

  • Q1 2026 revenue (quarter ended March 31, 2026): GAAP $193.4M, up 94% YoY from $99.5M and up 13% sequentially. Hardware $110.6M (+59%); cloud and other services $82.8M (+178%).
  • Core (non-GAAP) revenue: $191.3M, up 92% YoY — core hardware $111.6M (+60%), core cloud and services $79.8M (+167%).
  • Margins: GAAP gross margin 44.6% (hardware 41.3%, cloud 48.9%). Core gross margin 46.5%, up from 42.1% a year earlier and 41% in Q4 2025.
  • Profitability: GAAP operating loss $15.0M; GAAP net loss $14.0M (–$0.22/share, improved from –$0.46). Core operating loss just $3.5M (–2% margin); core net loss $2.5M. Adjusted EBITDA was positive $12.7M versus –$15.4M a year earlier.
  • Cash flow: Operating cash flow was positive $12.3M (vs. –$54.9M a year earlier), but capex of $132.0M and net investment purchases drove investing outflows of $236.6M. Financing brought in $2.04B (Series H plus the OpenAI loan).
  • Balance sheet (3/31/26): $2.75B cash and restricted cash plus $515.6M investments (~$3.3B total); total assets $4.95B; total liabilities $2.20B, including a $983M “loan from customer” (OpenAI) and $368M of customer deposits. Redeemable convertible preferred stock of $2.95B sat outside equity, leaving a stockholders’ deficit of $194.7M.
  • Post-quarter capital: ~$6.4B gross / ~$6.2B net from the May IPO, plus an $850M revolving credit facility closed in April. Liquidity is not the near-term problem.
  • Backlog: Remaining performance obligations of $25.0 billion at quarter-end (up from $24.6B at year-end 2025), the large majority tied to the OpenAI Master Relationship Agreement. Only a low-teens percentage is expected to convert to revenue across 2026–2027.
  • Commitments: Roughly $2.3 billion of non-cancelable data center leases at quarter-end, with about $1.6 billion more signed after the quarter — large fixed obligations against revenue that has not yet arrived.
  • Guidance: Q2 2026 core revenue ~$194M (+88% YoY) with core gross margin of just 36–38% and core operating margin of –30% to –32%. Full-year 2026 core revenue $855–865M (+69% at midpoint), core gross margin 38–41%, core operating margin –28% to –32%.
  • Valuation: Market capitalization is reported anywhere from roughly $45B to $64B depending on the share count used — approximately $51B on the ~226M share count most data providers publish, or about $64B counting all classes (~281M shares). Trailing P/E of roughly 190x is flattered by the one-time 2025 gain; price-to-book screens negative because providers are still using the pre-IPO stockholders’ deficit. No dividend.
  • Verdict: Liquidity and top-line momentum are strong; unit economics are improving but not yet proven; GAAP profitability, customer concentration, and near-term margin trajectory are weak. Cerebras is best described as well-capitalized and structurally advantaged on supply, but financially unproven at the scale its backlog implies — a company whose entire investment case rests on converting contracts into deployed megawatts.

Key Products or Services#

Cerebras sells compute in two forms — the machine and the minute — and increasingly prefers the latter. The saleable unit is not the wafer but the CS-3 system, which customers either buy for their own data centers or rent time on through Cerebras Cloud. The strategic shift underway is from a hardware vendor that sold systems to sovereign buyers into an infrastructure operator that runs its own inference cloud and plugs into other people’s clouds, which is why cloud revenue is growing more than twice as fast as hardware and why management has warned hardware revenue will decline sequentially for the next few quarters as production is redirected into Cerebras-owned capacity.

  • Wafer Scale Engine 3 (WSE-3) and the CS-3 system: The core product. Built on TSMC’s 5nm node, ~4 trillion transistors, ~900,000 AI-optimized cores, 44 GB of on-chip SRAM, ~125 petaflops peak AI performance, clusterable to 2,048 nodes. Because the memory is SRAM printed on the logic wafer rather than separately packaged HBM, Cerebras avoids the HBM shortage, TSMC’s CoWoS packaging bottleneck, and 3nm capacity contention entirely.
  • Cerebras Cloud (inference and training): On-demand and dedicated capacity, plus model services. This is the fastest-growing part of the business and the intended long-term margin engine, with management targeting ~60% gross margin over time.
  • Partner cloud distribution — AWS: Announced March 13, 2026 and converted to a definitive agreement in June, the AWS collaboration puts CS-3 systems inside AWS data centers behind Amazon Bedrock in a disaggregated architecture — Trainium3 handles prefill, Cerebras handles decode, connected by Elastic Fabric Adapter. AWS is the first cloud provider for this solution. Management expects revenue impact in 2027, not 2026.
  • Partner silicon — AMD: Announced July 23, 2026 at AMD’s Advancing AI event, pairing AMD Helios rackscale systems (prefill) with the Cerebras WSE (decode). The companies model up to 5x higher tokens per second per watt versus a Cerebras-only configuration, available first through Cerebras Cloud in the second half of 2026. Note the baseline: that 5x is modelled by the two vendors against a Cerebras-only setup, not benchmarked independently against NVIDIA.
  • Manufacturing: All CS-3 production is in the United States — Cerebras claims to be the only accelerator maker manufacturing exclusively domestically. Flex is expanding dedicated CS-3 lines in Milpitas, California targeting roughly a 7x capacity increase through 2026; Sanmina has been added as a second contract manufacturer.
  • Data center footprint: Live capacity including a Scale Datacenter facility in Oklahoma City hosting 300+ CS-3 systems; up to 160 MW as anchor tenant at Bell Canada’s 300 MW Sherwood, Saskatchewan campus under a US$2.2 billion, 10-year lease; a fully pre-leased 40 MW critical-IT campus in Minnesota via CleanCore Solutions ($800M over 10 years); and a planned 200 MW European buildout across France, Norway, and Finland by end-2027, with first European capacity targeted before the end of 2026. Early discussions are underway for Israel, the UAE, Australia, Singapore, India, and Indonesia.
  • Enterprise and model partnerships: CrowdStrike (Falcon AI Detection and Response running on Cerebras inference, announced July 22, 2026), Lovable (latency-sensitive workloads on dedicated capacity, August 5, 2026), plus enterprise trials of Kimi K2.6 — the first trillion-parameter model served on Cerebras, at close to 1,000 tokens per second — and Google DeepMind’s Gemma 4.

Moats, Strengths and Weaknesses#

Moats#

  • Wafer-scale architecture as a supply moat: In a year when HBM is the industry’s hard constraint and memory is absorbing roughly 30% of hyperscaler data center budgets, Cerebras uses on-wafer SRAM. It also skips CoWoS packaging and 3nm capacity. As CEO Andrew Feldman put it on the Q1 call, the binding constraints for everyone else simply do not apply — an unusual and genuinely durable advantage while the shortage lasts.
  • Measurable latency leadership: Independent measurement by Artificial Analysis put Cerebras near 1,000 tokens per second on a trillion-parameter open-weight model. Management’s own on-stage demo showed a 21-second completion versus 4 minutes 37 seconds on a leading GPU cloud running the identical model and prompt. Speed is the product, and the lead is currently large.
  • The OpenAI relationship: A definitive agreement signed December 24, 2025 for more than $20 billion of compute across 750 MW through 2028, with an option for an additional 1.25 GW to 2 GW by end-2030. Cerebras went from signature to production deployment in 35 days, and is one of only two hardware vendors currently serving OpenAI models. GPT-5.4 already runs on Cerebras for OpenAI engineers and select customers.
  • Decade-long national-laboratory relationships: DOE labs were among the earliest adopters of the WSE, producing Gordon Bell Prize–recognized work and a formal Memorandum of Understanding with the Department of Energy in December 2025.
  • US-only manufacturing: A real differentiator in the current policy environment, and one that is hard for competitors to replicate quickly.

Strengths#

  • Revenue nearly doubling (+94% GAAP, +92% core) with gross margin expanding and core operating loss narrowing to –2% of revenue.
  • A fortress balance sheet by early-stage standards: ~$3.3B at quarter-end, plus ~$6.2B net IPO proceeds and an $850M revolver behind it.
  • $25 billion of contracted backlog — visibility that essentially no other AI silicon challenger can claim.
  • Distribution through the world’s largest cloud (AWS via Bedrock) and a co-engineering relationship with the world’s number-two merchant accelerator vendor (AMD).
  • Strong sell-side support: roughly 10–11 covering analysts with a consensus that reads “Strong Buy” and an average 12-month target near $291–300, versus a recent price of about $227.
  • A widening non-UAE customer list — OpenAI, AWS, CrowdStrike, Lovable, Bell Canada — which is precisely the diversification the bear case demanded.

Weaknesses#

  • Extreme customer concentration: In 2025, MBZUAI accounted for 62% of revenue and G42 for 24% — about 86% from two UAE-affiliated entities that Cerebras’ own prospectus identifies as related parties. MBZUAI alone represented 77.9% of accounts receivable at year-end 2025. The concentration is migrating rather than disappearing: OpenAI now dominates the forward book.
  • Circular financing: OpenAI is simultaneously Cerebras’ largest customer, its lender (a ~$1.0 billion working capital loan at 6%, with interest waived if repaid through capacity delivery), a prospective ~10% shareholder via warrants for up to 33.4 million non-voting Class N shares, and a counterparty holding exclusivity provisions. If the Master Relationship Agreement terminates, the loan becomes immediately repayable.
  • Deliberate near-term margin compression: To serve backlog faster, Cerebras is temporarily renting its own systems back from an existing customer. The CFO guided to a 10–15 point hit to core cloud and services margin before recovering toward a 60%+ target. This disclosure — not the revenue number — is what drove the 19.6% single-day drop on June 24 and is the basis of several plaintiff-firm investigations.
  • Data center capacity is the binding constraint. Management said plainly that neither demand nor chip supply limits growth — physical data center space and power do. That is the same dogfight every AI infrastructure company is losing sleep over, and it is largely outside Cerebras’ control.
  • Large fixed commitments against unrealized revenue: roughly $2.3 billion of non-cancelable data center leases at quarter-end plus about $1.6 billion signed afterward.
  • Governance and disclosure: A dual-class structure left insiders with roughly 99.2% of voting power immediately after the IPO, and the S-1/A disclosed two material weaknesses in internal control over financial reporting. As an emerging growth company, Cerebras can defer Sarbanes-Oxley 404(b) auditor attestation for up to five years.
  • Lock-up overhang: The IPO lock-up expires on the earlier of the second trading day after Q3 2026 earnings or November 9, 2026 — a structural supply event roughly one quarter away.
  • The software gap: Cerebras competes against CUDA. Its architectural advantage in decode does not extend to general-purpose training and development workflows, where NVIDIA’s ecosystem remains the default.

News, Events and Partnerships#

The last ~180 days compress a private financing, the largest semiconductor IPO on record, a 108% first-day pop, a 50%+ drawdown, and a partial recovery into a single stretch. The arc runs from the February Series H at a $23 billion valuation, through a wildly oversubscribed May IPO that priced far above range and opened at nearly double it, into a June earnings report that beat on revenue but shocked the market with a margin guide, and then a July–August recovery driven by a rapid-fire series of partnership announcements. Operationally almost every headline has been positive; the negatives cluster entirely around margins, concentration, and valuation.

  • Feb 3–4, 2026 (positive): Closed a $1 billion Series H at a ~$23 billion post-money valuation led by Tiger Global, with Benchmark, Fidelity, Atreides, Alpha Wave Global, Altimeter, Coatue, 1789 Capital — and, notably, AMD — participating.
  • Mar 13, 2026 (positive): Announced the AWS collaboration to deliver disaggregated inference through Amazon Bedrock, pairing Trainium3 prefill with CS-3 decode. AWS is the exclusive first cloud for the solution.
  • Mar 16, 2026 (positive): Named anchor tenant for Bell Canada’s C$1.7 billion, 300 MW Sherwood campus near Regina, Saskatchewan, taking up to 160 MW — disclosed in IPO filings as a ~US$2.2 billion, 10-year lease. CoreWeave took the balance.
  • April 2026 (positive): Closed an $850 million revolving credit facility from a bank syndicate specifically to accelerate data center acquisition.
  • Apr 17 / May 4, 2026 (neutral): Filed and then launched the IPO at a marketed $115–125 range for 28 million shares, implying roughly a $26.6 billion valuation.
  • May 11–13, 2026 (positive): With orders reportedly 20x oversubscribed, the range was lifted to $150–160 and the deal upsized to 30 million shares, then priced at $185 — well above the raised range — raising $5.55 billion.
  • May 14–15, 2026 (positive): Debut. Opened at $350, closed at $311.07 (+68% on the offer price), setting the 52-week high of $386.34. The offering closed at 34.5 million shares including full greenshoe exercise, for $6.4 billion gross and roughly $6.2 billion net.
  • May 25, 2026 (positive): Added to S&P indices under Fast Track IPO Entry rules.
  • June 9, 2026 (positive): Quiet period expired; roughly nine to eleven firms initiated coverage, overwhelmingly at Buy, with an average target near $294 and Citigroup at the high end at $340.
  • June 23–26, 2026 (negative): First earnings as a public company. Revenue +94% and a narrower loss, but Q2 core gross margin guidance of 36–38% against 47% delivered sent shares from $226.72 to $182.26 — a 19.6% one-day decline. The stock bottomed at its 52-week low of $160.81 on June 26.
  • Late June onward (negative): Multiple plaintiff firms — Kaplan Fox & Kilsheimer and Block & Leviton among them — announced investigations into potential securities law violations, centered on whether the capacity rent-back and its margin impact were adequately disclosed at the time of the IPO.
  • July 9, 2026 (positive): Two announcements in one day — an expanded Flex partnership targeting roughly 7x CS-3 production capacity through 2026 at Milpitas, and a plan for 200 MW of European AI infrastructure by end-2027 across France, Norway, and Finland. Shares rose 11–12%.
  • July 10, 2026 (positive, indirect): The US Bureau of Industry and Security moved the UAE into Country Group A:5, easing license requirements for advanced AI computing exports to approved entities including G42 — a direct regulatory tailwind for Cerebras’ largest historical customer base.
  • July 22, 2026 (positive): CrowdStrike partnership announced; CrowdStrike will run Falcon AI Detection and Response models on Cerebras inference while Cerebras standardizes on Falcon internally. Shares rose about 7%.
  • July 23, 2026 (positive): AMD and Cerebras announced their disaggregated inference partnership at Advancing AI 2026, with Cerebras planning to deploy AMD Helios in its own data centers. CBRS rose while AMD fell about 3%.
  • July 24, 2026 (negative): Shares fell 9.5% to $199.12 in a broad AI-complex pullback, illustrating the name’s beta.
  • July 27, 2026 (positive): Mizuho raised its price target to $310 from $300, maintaining Outperform.
  • July 29, 2026 (positive): CleanCore Solutions signed a 10-year colocation agreement for a Tier 3 Minnesota campus 100% pre-leased to Cerebras — approximately $800 million initial contract value, potentially exceeding $3 billion with two renewal options.
  • Aug 5, 2026 (positive): Lovable partnership announced, bringing Cerebras inference to one of the fastest-growing AI software-creation platforms.
  • Aug 12, 2026 (upcoming): Q2 2026 results after the close — the first quarter to reflect the guided margin trough and the first read on OpenAI ramp velocity.

Government Integration#

Cerebras has one of the deepest US government relationships of any AI silicon company, but almost none of it takes the form of large prime federal contracts. The relationship is instead a decade of research collaboration with the Department of Energy’s national laboratories, formalized in a Memorandum of Understanding signed December 18, 2025 in support of the White House’s Genesis Mission — the national initiative launched in November 2025 to use AI to accelerate scientific research. Where government exposure becomes financially material is on the export control side, and there the news over the past 180 days has been unambiguously good.

  • DOE Memorandum of Understanding (Dec 18, 2025): A framework for information sharing, joint R&D, and future agreements spanning large-scale scientific datasets, converged AI+HPC hardware and software, novel memory and I/O technologies, and AI “co-scientist” capabilities. It is a framework, not a funded award — but it positions Cerebras for follow-on procurement.
  • NNSA Tri-Labs program: Sandia National Laboratories leads a multi-year collaboration with Lawrence Livermore and Los Alamos under the Advanced Simulation and Computing program’s Advanced Memory Technology effort, targeting memory systems that could increase wafer-scale capacity for scientific simulation by up to 100x. Sandia’s April 2026 research materials feature Cerebras as a small-business success story, with CS-3 systems used for trusted AI models on secure internal Tri-Lab data.
  • Other lab work: Gordon Bell Special Prize–winning genomics work with Argonne National Laboratory in 2022, three consecutive years of Gordon Bell finalist collaborations, and faster-than-exascale results on molecular dynamics (Tri-Labs) and computational fluid dynamics (NETL).
  • Export licensing: Cerebras holds export licenses covering CS-2, CS-3, and future CS-4 systems for G42 and MBZUAI in the UAE, subject to rigorous security and compliance obligations.
  • CFIUS history: The Committee on Foreign Investment in the United States opened a review of G42’s minority stake after the original September 2024 S-1, forcing Cerebras to withdraw that filing. Clearance came in March 2025 after G42’s holding was restructured into non-voting shares. The concentration itself was never resolved by that clearance — it migrated to MBZUAI, which the prospectus identifies as a G42 related party.
  • BIS rule change (effective July 10, 2026): Commerce removed the UAE from Country Groups D:3 and D:4 and added it to Country Group A:5, creating an entity-specific approval framework that allows license-free advanced-computing exports to approved entities including G42 and Core42. This materially de-risks Cerebras’ UAE revenue stream — though it has drawn bipartisan congressional scrutiny over diversion risk, which means it is a policy that could be revisited.
  • Domestic manufacturing alignment: Exclusively US-based CS-3 assembly is a direct fit with the current administration’s semiconductor onshoring priorities, and a differentiator Cerebras has begun to market explicitly.
  • No identified large federal prime contracts: Government revenue appears to flow through lab system sales and research agreements rather than headline federal awards, so investors should not model a government segment.

Social Sentiment#

Retail sentiment on Cerebras has been violently bipolar, which is exactly what you would expect from a stock that doubled on day one and then lost more than half of that within six weeks. On StockTwits, CBRS registered the maximum “extremely bullish” reading (100/100) on its Nasdaq debut, drifted to “bearish” during the post-IPO grind lower, flipped from “bearish” to “extremely bullish” within 24 hours of the June earnings drop as dip-buyers piled in, and has remained a frequent top-ten trending ticker throughout. The dominant bull framing is straightforward and emotionally satisfying — “the Nvidia killer,” a genuinely differentiated architecture with a $20 billion OpenAI contract behind it. The dominant bear framing is equally crisp: 86% of last year’s revenue came from two Abu Dhabi entities, the 2025 “profit” was an accounting artifact, and the stock is priced for flawless execution. A widely circulated short thesis explicitly flags the November 2026 lock-up expiry as the structural inflection point, while cautioning that the thin float and expensive borrow make it a dangerous trade to be early on. Notably, this is one of the rare cases where retail and institutional narratives have converged rather than diverged — the same two facts, the OpenAI backlog and the UAE concentration, drive both the enthusiasm and the skepticism. Sentiment also got a visible boost when Cathie Wood’s ARK funds bought 105,616 shares on the debut day, a roughly $33 million stake that retail traders treated as validation.

Insider Activity#

Insider activity has been mechanical rather than informative, and there is a structural reason for that: the IPO lock-up remains in force. Every Form 4 filed since the offering falls into one of three non-discretionary categories. First, share-class reclassifications executed immediately before the IPO — director Paul R. Auvil III moved 203,750 shares from Class A to Class B, director Elena A. Donio moved 33,701, and CFO Bob Komin filed similar restructuring transactions. Second, sell-to-cover transactions to satisfy tax withholding on vesting restricted stock units: CEO Andrew Feldman converted and sold 17,990 shares on June 25, 2026 at prices up to $184.73, while Chief Accounting Officer Yagnesh Patel sold 6,079 shares in the $162.34–$163.08 range the same day. Both filings explicitly state these are not discretionary transactions and are permitted exemptions under the lock-up. Third, in-kind fund distributions — director Eric Vishria reported 99,651 shares received through a pro-rata distribution from Benchmark Capital Partners VIII to its limited partners, with no cash changing hands.

What matters more than any of these is what is absent and what is coming. There have been no open-market discretionary purchases by insiders — no one is signalling conviction with their own cash at these levels. And Feldman still holds 14,038,631 Class B shares directly plus indirect positions through two GRATs, while the broader insider group controls roughly 99.2% of voting power. The lock-up expires on the earlier of the second trading day after Q3 2026 earnings or November 9, 2026, and that date, not any individual Form 4, is the insider-supply event worth marking on a calendar.

Politician Activity#

No US congressional or Senate trades in CBRS appear in available disclosure trackers — Quiver Quantitative explicitly returns no congressional trading data for the ticker. This is entirely expected for a company that has been public for under three months with a small tradable float; the STOCK Act’s 45-day disclosure lag alone means any post-IPO purchase might only now be surfacing. There is, therefore, no political-trading signal attached to this name in either direction.

Worth noting as adjacent context rather than as a trading signal: the AI silicon sector is drawing active congressional attention. Senators Elizabeth Warren and Richard Blumenthal opened a formal inquiry on March 20, 2026 into NVIDIA’s roughly $20 billion Groq transaction and the broader pattern of “reverse acquihire” structures, and the July 2026 easing of UAE export controls has prompted bipartisan scrutiny over technology diversion risk. Both threads touch Cerebras’ competitive and regulatory environment without involving the company directly.

Institutional Activity#

Institutional ownership is still forming, and the first genuinely comparable snapshot will not arrive until Q2 13F filings land in mid-August 2026. What is visible so far comes from Schedule 13G and 13D filings, and it is a mix of very large conviction positions, one notable trim, and a rapidly building short base.

  • Bullish: FMR LLC (Fidelity) reported beneficial ownership of 28,878,217 Class A–equivalent shares as of June 30, 2026 — 25.7% of the Class A class, including 20,443,122 Class B shares treated as convertible one-for-one. Fidelity also participated in both the Series G and Series H private rounds, so this is a long-held, deeply informed position rather than an IPO flip.
  • Bullish: Eclipse Ventures funds and managing member Lior Susan filed a Schedule 13D disclosing 13,466,197 Class B shares — 6.1% of total common, but with twenty votes per share, giving them governance influence well beyond their economic stake. The position was built across financings from 2016 to 2022.
  • Bullish: Strategic capital has come in from unusual places. AMD invested in the Series H — a GPU vendor buying into a wafer-scale competitor, which reads as a hedge against architectural disruption. Amazon reportedly purchased roughly $270 million of Cerebras shares alongside the AWS agreement. ARK Invest bought 105,616 shares on debut day across ARKK and ARKW.
  • Mixed: JPMorgan-affiliated entities initially disclosed 3,486,503 Class A shares (10.1% of the class), then filed an amended 13G on July 22, 2026 showing 3,100,807 shares (3.3%) — a meaningful reduction, though the shifting percentage denominator makes the trim harder to size than it looks.
  • Bearish / cautionary: Short interest reached 13.53 million shares as of June 30, 2026 — up 66.3% in two weeks and up roughly 613% since May. Depending on which float definition a data provider uses, that is anywhere from 16.9% to 39.2% of float; by one comparison it made CBRS’s short float roughly 6.6x AMD’s, 11.3x Broadcom’s, and 12.7x NVIDIA’s. Days to cover is low at around 1.4, so the position is crowded but liquid — capable of fueling a violent squeeze in either direction.
  • Cautionary: Legacy holders Benchmark (~17.6M shares), Foundation Capital (~15.3M), Alpha Wave (~12.1M), and G42 (~3.5M) all sit behind the same November lock-up, creating a known supply cliff that institutional buyers are visibly pricing around.

Political & Economic Landscape#

The macro backdrop for Cerebras is close to ideal on demand and genuinely fraught on everything downstream of it. Hyperscaler capital expenditure for calendar 2026 is tracking toward roughly $720–800 billion on the arithmetic of Q2 2026 guidance — Amazon near $220 billion, Alphabet $195–205 billion, Microsoft around $175 billion, Meta $130–145 billion — with Deloitte projecting total hyperscaler spend above $1 trillion. Demand for inference compute is not in question. What is in question is whether the physical and financial plumbing can keep up, and whether the revenue currently being booked across the sector reflects real deployment or scarcity pricing.

  • The memory shortage is Cerebras’ single biggest structural tailwind. HBM is sold out, each bit of HBM consumes roughly three times the wafer capacity of standard DRAM, memory is absorbing about 30% of 2026 hyperscaler data center budgets, and relief from new fabs is not expected before 2027–2028. Cerebras uses on-wafer SRAM and buys none of it. In a shortage, the company that does not need the scarce input wins share almost by default.
  • Power and data center capacity have replaced silicon as the constraint — and here Cerebras is on the wrong side of the trade. Management named data center capacity, not demand or chip supply, as the binding limit on 2026 growth, which is precisely why it is renting capacity back at a margin cost and signing leases across four continents.
  • The AI bubble debate is intensifying, and circular financing is at its center. Analysts increasingly point out that revenue curves are rising faster than physical deployment curves, and that huge remaining-performance-obligation numbers convert to cash slowly — Oracle’s $638 billion RPO with roughly 12% expected to convert within twelve months is the cautionary comparison. Cerebras’ own $25 billion backlog, with only a low-teens percentage expected across 2026–2027, sits squarely in that critique. The OpenAI structure — customer, lender, warrant holder, and exclusivity counterparty simultaneously — is the textbook example of the interlocking arrangements skeptics are worried about.
  • Export policy has turned decisively favorable. The July 10, 2026 BIS rule moving the UAE to Country Group A:5 removes license friction for Cerebras’ largest historical customer base and follows the November 2025 authorizations that let G42 and Saudi Arabia’s HUMAIN purchase advanced US silicon. The offsetting risk is that this is discretionary policy under active congressional scrutiny, and it can be reversed.
  • Domestic AI policy is supportive. The White House Genesis Mission, DOE lab partnerships, and onshoring incentives all favor a US-only manufacturer with a decade of national-security-adjacent compute work.
  • Consolidation pressure is rising in specialty AI silicon. NVIDIA’s roughly $20 billion Groq transaction — a non-exclusive license that also absorbed Groq’s founder and most of its senior team — reset expectations for what challengers are worth and how they exit. It also triggered a Senate inquiry, and it removed Cerebras’ most comparable independent competitor as a standalone engineering organization.
  • Rates and the broader tape are a live variable. With the US 10-year near 4.66%, an unexpected 23,000-job decline in July payrolls, and the Nasdaq at record levels around 26,690, the AI complex is priced for both continued easing and continued execution. A disappointment on either front hits high-multiple, unprofitable names like CBRS hardest.

The Competition#

Companies compared: NVIDIA Corporation (NVDA), Groq Inc. (private), CoreWeave, Inc. (CRWV)

NVIDIA Corporation (NVDA)#

NVIDIA is not merely Cerebras’ largest competitor — it is the entire frame of reference for the investment case, since the bull thesis on CBRS is essentially a claim that a meaningful share of inference will migrate off GPUs. NVIDIA’s advantage is not any single chip but the CUDA software ecosystem, the developer base built on it, and a supply chain that absorbs the overwhelming majority of the world’s HBM and advanced packaging capacity. NVIDIA has also publicly disputed the premise: Jensen Huang has argued that roughly a quarter of the inference market is genuinely latency-sensitive, against Feldman’s claim that the entire inference market is addressable for fast AI. Both positions are self-interested, and the honest answer is that nobody knows yet.

  • Market cap ~$5.5 trillion; trades around $224; CEO Jen-Hsun Huang; ~42,000 employees.
  • Highly profitable with a trailing P/E near 32 and price-to-book around 25.6 — a fraction of Cerebras’ multiple on every metric.
  • Q1 FY2027 data center revenue of $75.2 billion, against combined Big-Five quarterly capex of roughly $131 billion, implying NVIDIA captures well over half of every hyperscaler infrastructure dollar.
  • Also now a partial owner of the fast-inference thesis via its ~$20 billion Groq licensing deal, which brought Groq’s founder and key architects in-house.
  • Where Cerebras wins: single-stream decode latency and freedom from HBM. Where NVIDIA wins: everything else, including training, general-purpose workloads, software, and the ability to supply at scale today.

Groq Inc. (private)#

Groq is Cerebras’ closest philosophical competitor: a purpose-built inference chip (the Language Processing Unit) paired with a first-party inference cloud, sold on deterministic low latency, and — like Cerebras — heavily anchored by a Gulf sovereign relationship. It is included here despite being private because it is the single best like-for-like comparison for the fast-inference thesis, and because what happened to it in the last eight months is directly instructive for CBRS shareholders. In December 2025, NVIDIA entered a non-exclusive technology licensing agreement reportedly worth around $20 billion, under which founder Jonathan Ross and most of Groq’s senior technical staff joined NVIDIA while Groq remained independent, retaining its IP and GroqCloud business.

  • Last disclosed independent valuation of $6.9 billion (September 2025, $750M Series E); raised a further $650 million in June 2026 led by Infinitum and Disruptive.
  • GroqCloud serves more than 2 million developers, up from roughly 356,000 in 2024, across roughly 13 data centers in North America, Europe, and the Middle East.
  • Anchored by a $1.5 billion Saudi commitment tied to HUMAIN, the kingdom’s state AI champion — a near-mirror of Cerebras’ UAE dependence.
  • Also a Bell Canada tenant, anchoring a 7 MW facility in Kamloops, British Columbia.
  • The read-through for Cerebras is two-sided: it validates that incumbents will pay enormous sums for specialized inference IP, and it demonstrates how quickly a challenger’s engineering organization can be hollowed out without a formal acquisition.

CoreWeave, Inc. (CRWV)#

CoreWeave competes with Cerebras for the same inference dollars from the same customers, but with the opposite business model: it buys NVIDIA GPUs at scale and rents them, rather than designing its own silicon. It is the relevant public comparison for Cerebras Cloud specifically, and the two companies are literally co-tenants at Bell Canada’s Sherwood campus. CoreWeave’s advantage is speed of deployment and hyperscaler-grade contracts; its vulnerability is that it owns no differentiated technology and carries substantial debt against depreciating GPUs.

  • Market cap ~$49 billion; trades around $91; CEO Michael Intrator; ~2,189 employees.
  • Unprofitable (trailing P/E of –28.8; price-to-book ~9.6), like Cerebras but with a much heavier balance sheet.
  • Shares have fallen from a 52-week high of $153.20 to a low of $60.55 in July 2026 — a useful reminder of how brutally the market re-rates AI infrastructure names when the deployment story slips.
  • Positioned to deploy NVIDIA’s Vera Rubin NVL72 systems among the first providers in the second half of 2026, keeping it at the front of the GPU roadmap.
  • Cerebras’ counterargument is architectural: CoreWeave’s economics are downstream of NVIDIA’s pricing and HBM availability. Cerebras controls its own silicon cost curve.

How Cerebras stacks up#

Cerebras occupies an unusual position: it is far smaller than NVIDIA, far more differentiated than CoreWeave, and — following the NVIDIA-Groq transaction — effectively the last well-capitalized independent challenger with a genuinely distinct architecture. The technical claim is measurable and holds up under independent benchmarking: on identical models and prompts, Cerebras generates tokens roughly an order of magnitude faster than leading GPU deployments. The commercial claim is where the argument lives. NVIDIA’s position is that fast inference is a slice of the market; Cerebras’ position is that no technology market has ever been won by the slower option. History broadly favors Feldman’s framing, but the counterexample matters: dial-up lost to broadband on cost-parity, not at a premium price, and Cerebras’ tokens currently cost more.

One structural note on the peer set: AMD does not appear above despite being the world’s number-two merchant AI accelerator vendor, because it now occupies three roles simultaneously — competitor, Series H investor, and co-development partner on disaggregated inference. It is treated in The Proxy section below.

MetricCerebras (CBRS)NVIDIA (NVDA)Groq (private)CoreWeave (CRWV)
Market cap$51B ($64B all classes)~$5.5T~$6.9B (last disclosed round, Sept 2025)~$49B
Recent price~$227~$224Not publicly traded~$91
Business modelWafer-scale systems + owned inference cloudMerchant GPU silicon + systems + CUDA softwareLPU silicon + GroqCloud inferenceRents NVIDIA GPU capacity as a cloud
ProfitabilityGAAP net loss; core operating near breakevenHighly profitableNot disclosedNet loss
Trailing P/E~190 (flattered by a 2025 one-time gain)~32Not applicableNegative
Employees~700–780~42,000Not disclosed post-NVIDIA deal~2,189
Key edgeFastest decode; no HBM, CoWoS, or 3nm dependencyCUDA ecosystem, scale, supply allocationDeterministic low latency; large developer baseDeployment speed; hyperscaler-grade contracts
Key riskCustomer concentration; data center capacity; circular dealsValuation; custom-silicon share lossFounder and team absorbed by NVIDIADebt load; NVIDIA dependence; concentration
  • CBRS: the highest-conviction, highest-variance expression of the fast-inference thesis, with real backlog and real concentration risk.
  • NVDA: the safe way to own AI compute, and the direct source of the risk to Cerebras’ addressable market.
  • Groq: the cautionary and validating comparison — proof the technology is valuable, and proof that independence is fragile.
  • CRWV: the pure capacity play, useful mainly for understanding how harshly the market punishes deployment slippage.

The Proxy#

Advanced Micro Devices, Inc. (AMD)#

There is no parent company or holding structure that offers indirect ownership of Cerebras, so the cleanest proxy is the one company that is simultaneously a shareholder, a partner, and a competitor: AMD. AMD participated in the $1 billion Series H at a $23 billion valuation in February 2026, meaning it holds actual Cerebras equity. Five months later, on July 23, 2026, the two companies announced a joint disaggregated inference architecture pairing AMD Helios rackscale systems with the Cerebras Wafer-Scale Engine, with Cerebras planning to deploy Helios inside its own data centers and the combined solution launching first through Cerebras Cloud in the second half of 2026. Owning AMD gives an investor exposure to the same core thesis — that inference is fragmenting away from monolithic GPU deployments toward workload-specialized silicon — at roughly a fifth of the multiple, with actual profitability, far deeper liquidity, and none of the UAE concentration or lock-up overhang. The obvious catch is dilution of exposure: AMD’s Cerebras stake is immaterial against a roughly $797 billion market capitalization, and AMD competes with Cerebras for the same inference budgets even as it partners on them.

  • Market cap ~$797 billion; trades around $484; CEO Lisa Su; ~31,000 employees; trailing P/E ~159 with a price-to-book near 12.3.
  • Holds Cerebras equity via the Series H round, alongside Tiger Global, Benchmark, Fidelity, Atreides, Alpha Wave Global, Altimeter, Coatue, and 1789 Capital.
  • Co-developing the disaggregated inference solution with Cerebras, modelled at up to 5x higher tokens per second per watt versus a Cerebras-only configuration — though that figure is vendor-modelled against a Cerebras baseline, not independently benchmarked against NVIDIA.
  • Broader AI momentum independent of Cerebras: a $5 billion investment in Anthropic and an up-to-2-gigawatt Instinct MI450 deployment agreement, both announced in the same July week.
  • Best for investors who want the specialized-inference thesis without single-customer risk; less suitable for anyone specifically underwriting wafer-scale architecture, since AMD’s fortunes turn overwhelmingly on its own GPU roadmap.
  • Narrower, higher-beta alternatives: Flex Ltd. (FLEX, ~$47B) manufactures every CS-3 and is scaling that line roughly 7x through 2026, making it a supply-chain-levered play on Cerebras volume. CleanCore Solutions (ZONE, ~$85M) has a Minnesota campus 100% pre-leased to Cerebras under an ~$800 million ten-year agreement — by far the most concentrated public exposure to Cerebras’ buildout, and correspondingly the most speculative, trading under $1 with fifteen employees and a 52-week range of $0.22 to $7.82.

The Big Picture for Cerebras#

Cerebras has done something genuinely difficult: it built a fundamentally different computer, proved it is an order of magnitude faster than the incumbent at the fastest-growing workload in technology, and then converted that into a $25 billion contracted backlog and distribution through the world’s largest cloud. The architectural advantage is not marketing. In a year when HBM is sold out through 2026, CoWoS packaging is exhausted, and 3nm capacity is spoken for, Cerebras needs none of it. That is a real, quantifiable edge over every GPU-based competitor, and it is the single most underappreciated fact in the story. The demand side of the equation — hyperscaler capex tracking toward $720–800 billion this year, inference displacing training as the dominant workload, sovereign AI programs proliferating from Ottawa to Abu Dhabi — is about as favorable as any company could ask for.

The central question is not whether the technology works or whether demand exists. It is whether Cerebras can pour concrete fast enough. Management said it plainly on the Q1 call: neither demand nor chip supply constrains growth, data center capacity does. That is why the company is renting its own systems back from a customer at a 10–15 point margin cost, why it has committed to roughly $3.9 billion of data center leases, and why the second-half revenue ramp that underpins the full-year guide is entirely back-loaded. Execution on megawatts brought online — not tokens per second, not press releases — is the only KPI that matters over the next four quarters.

Around that operational question sit three risks that will not resolve quickly. Customer concentration has migrated rather than disappeared: 86% of 2025 revenue came from two related UAE entities, and the diversification story simply swaps that for dependence on OpenAI, which is also the lender, a prospective 10% shareholder, and the holder of exclusivity provisions. The July 2026 easing of UAE export controls helps materially, but it is discretionary policy under congressional scrutiny. And the lock-up expiring on the earlier of the second trading day after Q3 earnings or November 9, 2026 will release a very large supply of insider and venture shares into a float that is currently small enough to make short interest of 13.5 million shares look like anything from 17% to 39% of tradable stock, depending on how you count.

If Cerebras brings capacity online roughly on schedule, if the OpenAI deployment ramps through the back half of 2026, if AWS contributes in 2027 as guided, and if gross margin troughs in Q2 and recovers toward the 60% target, then the current price will look like it was set during a temporary crisis of confidence about an accounting line item. That is the bull case, and the analyst consensus near $291–300 is underwriting it. If instead data center timelines slip, the margin trough proves deeper or longer, or a single large counterparty renegotiates, the stock has already demonstrated it can lose 20% in a session and 58% from its high in six weeks.

For the next several years, three things deserve more attention than anything else in the news flow: contracted megawatts actually energized each quarter, the emergence of a material customer outside OpenAI and the UAE, and the trajectory of core cloud gross margin back toward management’s 60% target. The August 12 report is the first real data point on all three. Cerebras is not a company that will be left behind by where the AI market is going — it is arguably closer to the center of it than any company its size. The realistic risk is that the equity gets repriced hard on the gap between a $25 billion backlog and the far smaller amount of it that becomes revenue any time soon.

Sources#

Related

All Posts →
Moderna, Inc. (MRNA)

Moderna, Inc. (MRNA)

On August 19, 2026, Moderna and Merck announced that their personalized mRNA cancer therapy hit its endpoints in the Phase 3 INTerpath-001 melanoma trial — the first positive late-stage readout for an mRNA cancer therapy in history — and the stock closed up roughly 178% in a single session. This deep dive examines whether a company whose revenue has fallen from $19 billion to under $2 billion, whose federal contracts were terminated by an HHS openly hostile to mRNA, and which still loses roughly $780 million a quarter has genuinely earned a $70 billion valuation overnight. We break down the fundamentals behind the headline, the newly approved mFLUSIVA flu vaccine, the $2.25 billion patent settlement, the 50/50 economics Moderna shares with Merck, uniformly one-directional insider selling, the institutional split that saw Fidelity and T. Rowe Price buying while D.E. Shaw exited, and how Moderna compares to BioNTech, Pfizer, and Novavax. The catch: the announcement disclosed no hazard ratios, no effect sizes, and no survival data, and roughly a third of the move was short covering against a 13.5% short float. Read on for what the market has already priced in — and what it hasn't seen yet.
Corning Incorporated (GLW)

Corning Incorporated (GLW)

Once known mostly for Gorilla Glass and Pyrex, 175-year-old Corning (GLW) has become one of the most foundational physical-layer suppliers of the AI boom — the company that makes the optical fiber, cable, and connectivity wiring hyperscale data centers together. This deep dive unpacks the run of multibillion-dollar deals (Meta, Amazon, and an NVIDIA equity partnership) and the upgraded "Springboard" plan targeting $35–40 billion in sales by 2030, then asks whether a profitable, dividend-paying blue-chip that has risen roughly sixfold since 2023 still has room to run — or whether a triple-digit P/E, cyclical glass and solar exposure, and heavy insider selling are warning signs. We break down the fundamentals, the marquee partnerships, CHIPS Act funding and government ties, social and institutional sentiment, and how Corning stacks up against optical rivals Coherent, Lumentum, and Amphenol. The verdict: Corning is positioned to lead the optical layer of the AI buildout, but its stock has, for now, run ahead of even its improving fundamentals. Read on for the full picture.
Ouster, Inc. (OUST)

Ouster, Inc. (OUST)

Once a SPAC-era lidar hopeful, Ouster, Inc. (OUST) has reinvented itself as a "sensing and perception for Physical AI" platform — pairing its chip-based digital lidar with cameras, AI compute, and software — and has rocketed to multi-year highs on the back of its native-color Rev8 launch, defense and smart-infrastructure wins, and a powerful regulatory tailwind. This deep dive examines whether Ouster's clean balance sheet, NDAA-compliant "trusted-supplier" moat, and 700-plus BlueCity deployments justify a stock trading near $50 at roughly 9x sales — or whether persistent losses, lumpy margins, heavy insider selling, and a freshly doubled authorized-share count are warning signs. We break down the fundamentals, the Rev8 product cycle and marquee partnerships, social and institutional sentiment, the NDAA Section 164 ban on Chinese lidar, and how Ouster stacks up against rivals Aeva, Innoviz, and China's Hesai. The verdict: Ouster is squarely aimed at where 3D perception is heading and is the best-capitalized U.S. pure-play in a protected procurement lane, but it remains a high-beta bet that has already priced in years of flawless, dilution-light execution.