September 30, 2026 · Evidence base: 424B4 IPO prospectus (May 2026), withdrawn S-1 (Sept 2024), Q1 and Q2 FY2026 10-Qs and earnings calls; independent industry sources via the web. Not a valuation and not a recommendation.
Cerebras leases a data center, fills it with its own CS-3 systems, and rents the output on 3–4-year take-or-pay tranches. OpenAI has committed 750 MW; the company's “more than $20 billion” value for that contract implies about $27M of revenue per MW over a tranche's life, roughly $7–9M per MW per year (inferred). Core cloud gross margin on that capacity was 41.8% in Q2 FY2026, against a long-run company target of 60%.
| What the business is | Cerebras designs a single-wafer AI processor (WSE-3) and sells what it computes in two ways: as CS-3 systems bought outright, and as inference capacity it runs in leased data centers and rents by the megawatt or by the token. Buyers are a handful of AI labs, sovereign customers and hyperscalers (OpenAI, the Abu Dhabi entities MBZUAI and G42, and AWS), plus developers and enterprises on its cloud API. FY2025 revenue was $510.0M with 708 employees (424B4); Q2 FY2026 core revenue was $209.9M (Q2 FY2026 call). |
| Industry | AI accelerators and AI compute services, a market NVIDIA dominates: it shipped an estimated 68% of AI compute in Q4 2025 (Epoch AI) and earned a 71.1% GAAP gross margin on $193.7B of FY2026 Data Center revenue (NVIDIA FY2026 results). |
| How it makes money | Hardware revenue is booked when a system ships or is accepted, often funded in advance by customer prepayments held as restricted cash; cloud revenue is booked daily over multi-year take-or-pay capacity tranches or per token consumed; data-center pass-through costs are billed at a 3% markup and reported gross; and the amortized value of warrants granted to customers ($44.3M in Q2 FY2026) is deducted before revenue is reported. |
| Unit of economics | One megawatt of dedicated inference capacity, delivered as a 3–4-year take-or-pay tranche. MW in service at end-FY2025 is not disclosed (unknown); OpenAI has committed 750 MW for delivery over 2026–2028. The company's “more than $20 billion” value for that contract implies about $27M of revenue per MW over a tranche's life, or roughly $7–9M per MW per year (inferred; may include pass-through costs). |
| What protects it | A decade-long, process-specific ability to yield a working processor from an entire wafer, which puts 44GB of SRAM on the chip and lets it generate tokens about 2.5× faster than the next-fastest independent provider (Artificial Analysis) without using scarce HBM memory or CoWoS packaging. There is no comparable software lock-in, and NVIDIA has licensed Groq's SRAM design to offer the same function. |
| What drives earnings | The pace at which megawatts are delivered to OpenAI and new cloud customers; utilization and price per token of that capacity against leased data-center cost; lumpy hardware orders from UAE customers and, from 2027, AWS; the run-off of customer-warrant amortization against revenue. |
| What to watch | Remaining performance obligations and the share due within 24 months ($25.4B and 22% at Q2 FY2026, 10-Q); core cloud gross margin against the 60% long-run target; free cash flow against data-center leases signed but not yet commenced. |
| Cycle exposure | High to the AI-infrastructure capital cycle, now expanding with the four largest US hyperscalers guiding roughly $720–745B of 2026 capex; medium to token-price deflation; high to single-counterparty events (OpenAI, UAE export licences). Revenue and capex are at records while GAAP gross margin is at a trough (14% in Q2 FY2026); the business has never been through a downturn at scale. |
USD. Fiscal year = calendar year; latest reported period is Q2 FY2026 (quarter ended June 30, 2026). “Core” is the company's non-GAAP measure, excluding pass-through revenue and cost, customer-warrant amortization, stock compensation and IPO payroll tax. (inferred) marks a conclusion reasoned from sourced figures; (unknown) marks something no source settles.
Large language models produce output one token at a time, and each token requires reading the model's weights from memory. On a GPU those weights sit in HBM stacks beside the chip, so output speed is capped by memory bandwidth rather than arithmetic. Cerebras's answer is to leave the processor as one uncut wafer: 900,000 cores and 44GB of on-chip SRAM on a 46,225mm² die, fabricated by TSMC at 5nm (424B4). What the customer buys is faster answers, which matter most for coding agents and reasoning models that emit many tokens per task; the company claims inference “up to 15 times faster than leading GPU-based solutions” (424B4, company claim).
It sells that capability in two forms. On-premises customers buy CS-3 systems, one wafer each, plus a renewable software subscription, and revenue is recognized at shipment or acceptance (424B4, Note 4). In Cerebras Cloud the company owns the systems and rents their output, either as dedicated capacity, “take-or-pay commitments, under which customers pay for dedicated compute capacity irrespective of utilization”, or on demand per token (424B4, MD&A). Hardware and cloud-and-services revenue were $15.6M and $9.0M in FY2022, $57.1M and $21.6M in FY2023 (2024 S-1), $212.0M and $78.3M in FY2024, and $358.4M and $151.6M in FY2025 (424B4). In H1 FY2026 cloud overtook hardware on a GAAP basis, $208.8M to $164.7M (Q2 FY2026 10-Q).
The unit that organizes the economics is one megawatt of inference capacity. OpenAI signed a Master Relationship Agreement on December 24, 2025 for 750 MW, delivered in tranches over 2026–2028; each tranche runs three or four years, extendable to five, and OpenAI holds an option on another 1.25 GW (Q2 FY2026 10-Q, Note 3). OpenAI advanced a working-capital loan of about $1.0B, repayable in compute credits, with $918.2M outstanding at June 30, 2026. Cerebras then leases and fits out a site, builds CS-3s into its own property and equipment, delivers the tranche, and books revenue ratably while billing data-center costs through at a 3% markup (Q1 FY2026 call, CFO). Cash goes out first (capex was $416.9M in Q2 FY2026 alone, inferred from the 10-Q cash-flow statements) and comes back over three to four years.
The shift from selling hardware to renting its output changes the economics. A system sale earns a one-time margin (hardware gross margin was 42.9% in FY2025, computed from 424B4) that is collected quickly and often pre-funded: G42 and MBZUAI deposits peaked at $640.3M at end-2024 (424B4). Renting turns the same wafer into years of revenue but moves the capex, the lease and the utilization risk onto Cerebras. Cloud-and-services gross margin fell from 61.4% in FY2024 to 29.9% in FY2025 as data-center costs arrived ahead of revenue (computed from 424B4), and management guides it back toward “60%+” once rented capacity is replaced by its own (Q1 FY2026 call, CFO).
In 2024 this was a one-customer hardware vendor: G42 took 87% of H1 2024 revenue and 97% of hardware revenue (2024 S-1). The inference cloud launched in August 2024 (424B4). The two related UAE parties, MBZUAI and G42, were still 86% of FY2025 revenue, then 74% in Q1 FY2026 and 43% in Q2 FY2026. OpenAI went from zero in FY2025 to about 9% of Q1 FY2026 revenue ($16.9M) and 32% of Q2 ($56.8M) (424B4; Q1 and Q2 FY2026 10-Qs; Q2 share is the undisclosed “Customer B”, matched to OpenAI's disclosed revenue, inferred).
AI compute exists because training and serving models needs vast parallel arithmetic, and the buyers fund it at hyperscale: Microsoft, Alphabet, Amazon and Meta guide roughly $720–745B of combined 2026 capex (company calls, July 2026; inferred sum). The profit pool sits at the silicon layer. NVIDIA earned a 75.0% GAAP gross margin on $89.0B of Data Center revenue in its July 2026 quarter, and Broadcom's AI semiconductor revenue reached $16.7B in its August quarter (company filings). Downstream, capacity renters carry the capital: CoreWeave bought $6.42B of property and equipment against $2.575B of Q2 2026 revenue and reported a −2% GAAP operating margin (CoreWeave Q2 2026 release).
Cerebras straddles both layers. As a system seller it resembles a much smaller NVIDIA; as a cloud operator it takes on CoreWeave-like capital intensity, so its blended margin should settle between the two and drift toward the lower end while cloud is growing faster than utilization (inferred). The market is global and concentrated on the buying side, since a few labs and hyperscalers purchase most frontier compute; that, not a sales failure, is why Cerebras's own revenue is concentrated.
NVIDIA shipped an estimated 68% of AI compute in Q4 2025, measured in H100-equivalents, with Google's TPUs at 19%; Cerebras is too small to be broken out (Epoch AI, August 2026). The barriers that bind are the CUDA software ecosystem, access to TSMC leading-edge capacity and, for cloud, power and sites. CUDA binds least where one lab serves a fixed model through an API at very large scale, which is exactly the segment Cerebras sells into (inferred).
Cerebras's claim is speed, and independent evidence supports it. On gpt-oss-120b, Artificial Analysis measured 1,775 output tokens per second on Cerebras, against 703 on SambaNova, 472 on Groq and 310 on Azure's GPU service (accessed September 30, 2026). What a well-funded entrant could not quickly reproduce is the wafer-scale process: yielding one processor across a whole wafer needs process work at TSMC, redistribution layers deposited by ASE and final packaging in Sunnyvale (424B4), and Tesla's Dojo, the other wafer-scale effort, was disbanded in August 2025 (Bloomberg). Because the design uses neither HBM nor CoWoS, Cerebras also stands outside the industry's two tightest supply queues; management says it has wafer “supply for our plan and beyond in 2026” (Q1 FY2026 call).
The speed is sold at a premium, not a discount. Cerebras's blended API price on that model was $0.39 per million tokens, against $0.04–$0.15 for GPU clouds and SambaNova (Artificial Analysis), so the premium holds only where latency is itself the product. Each wafer holds 44GB of SRAM, so large models must be split across many systems, and in the AWS design Trainium processes the prompt while the CS-3 generates the output (AWS, March 2026). OpenAI's own announcement frames Cerebras as a complement to GPUs for “workflows that demand extremely low latency” (OpenAI, January 2026).
Cerebras names NVIDIA, AMD, Intel, hyperscaler accelerators and neoclouds such as CoreWeave as competitors (424B4). The most direct threat is one it does not name. In December 2025 NVIDIA took a non-exclusive licence to Groq's SRAM inference technology and hired its founder, paying $13.0B in FY2026 and a further $2.944B since (Groq release; NVIDIA filings), and it now describes its Groq 3 LPX rack, SRAM decode paired with Rubin GPUs, as “in full production” (NVIDIA Q2 FY2027 release). SambaNova, the other SRAM-heavy rival, runs at about 40% of Cerebras's measured speed, and its acquisition talks with Intel stalled in January 2026 (trade press).
| Company claim | Verdict | Basis |
|---|---|---|
| Fastest inference | Supports | Artificial Analysis: 1,775 tokens/s vs 703 (SambaNova) and 472 (Groq); OpenAI runs a Codex model at >1,000 tokens/s on WSE-3. No MLPerf submission; NVIDIA LPX not yet independently benchmarked. |
| Wafer-scale is hard to replicate | Partially supports | Only shipping wafer-scale AI chip; Tesla Dojo disbanded. But TSMC plans wafer-scale packaging (SoW-X) for 2027, and NVIDIA replicated the function through Groq IP. |
| Large addressable market | Partially supports | Deloitte puts inference at about two-thirds of AI compute in 2026. The latency-premium slice Cerebras serves is not independently sized (unknown). |
| Diversified away from G42 | Partially supports | OpenAI and AWS confirmed their deals publicly, but concentration has moved to OpenAI rather than fallen. |
In this industry the winners either own the default software and system platform with supply at scale, as NVIDIA does, or own their own demand, as Google does with TPUs. Cerebras is not yet either. It has a measurable speed lead and an input-supply advantage, but no software lock-in and a few dominant customers, which makes it a best-of-breed component supplier whose position depends on staying measurably faster than NVIDIA's LPX at a price per token labs will pay (inferred).
Cerebras has made no acquisitions, so all growth is organic. The split that matters is between GAAP and core revenue. Warrants granted to OpenAI, G42 and AWS are amortized as a reduction of revenue ($44.3M in Q2 FY2026), while data-center pass-through costs are added to revenue at a 3% markup (about $14.5M in Q2, inferred from the 10-Q reconciliation). In Q2 GAAP revenue therefore understated the underlying business by about $30M.
| Q2 FY2025 | Q2 FY2026 | |
|---|---|---|
| GAAP revenue | 103.3 | 180.1 (+74%) |
| Add: customer-warrant amortization | — | 44.3 |
| Less: data-center pass-through revenue | — | ~14.5 (inferred) |
| Core revenue | ~103.3 | 209.9 (+103%) |
| of which core hardware | — | 82.1 (+17%) |
| of which core cloud and other services | — | 127.7 (+287%) |
Source: Q2 FY2026 10-Q and call. Q2 FY2025 core revenue is approximately equal to GAAP; the company gives only growth rates for the core segments.
The customer split shows where the growth came from. Year on year, MBZUAI and G42 revenue fell from about $91M to $77M, OpenAI went from zero to $56.8M, and all other customers grew from about $12M to about $46M (disclosed shares applied to GAAP revenue, inferred). Full-year core revenue guidance was raised to $880–890M, and management expects “to more than triple our core revenues in 2027” (Q2 FY2026 call). Q3 guidance of $214–216M implies only 2–3% sequential growth, because capacity arrives in steps as data centers come online (inferred).
| Customer group | Q2 FY2025 | Q2 FY2026 |
|---|---|---|
| MBZUAI (70% → 34% of revenue) | 72.3 | 61.2 |
| G42 (18% → 9%) | 18.6 | 16.2 |
| OpenAI (0% → 32%) | 0 | 56.8 |
| All other customers | ~12.4 | ~45.9 |
| Total | 103.3 | 180.1 |
Shares from Q2 FY2026 10-Q; OpenAI revenue disclosed directly. GAAP figures are net of warrant amortization, which falls mostly on G42 and MBZUAI hardware (inferred).
Growth drivers, ranked from most to least impactful:
The 750 MW is take-or-pay and makes up most of the $25.4B of remaining performance obligations, of which 22% is due within 24 months (Q2 FY2026 10-Q). Its pace is set by data-center delivery, not by demand.
Six deals above $30M were signed in Q2, and Figma, Cognition, Lovable, Block and CrowdStrike were named (Q2 FY2026 call); revenue from customers other than MBZUAI, G42 and OpenAI nearly quadrupled year on year (inferred).
AWS signed a hardware-leasing agreement in June 2026, with general availability on Bedrock expected in Q1 2027, and other hyperscalers are expected from mid-2027 (Q2 FY2026 call). None of this is in RPO yet.
MBZUAI and G42 drove FY2025, but management expects “decreasing hardware revenue for the next few quarters” as production moves into its own cloud (Q1 FY2026 call, CFO).
Management reports “higher pricing from existing customers” because HBM costs pushed competitors' prices up (Q1 FY2026 call, CFO), while the price of a fixed level of model capability falls 9× to 900× a year (Epoch AI).
The binding constraint is sites and power, not demand or wafers. “Demand is not the constraint. Supply is not the constraint. The constraint is data centers,” said CEO Andrew Feldman (Q1 FY2026 call). The company has secured “more than 600 MW” of capacity that is live or due by end-2027, and lists 13 sites live or under contract across North America and Europe (Q2 FY2026 call).
“Demand is not the constraint. Supply is not the constraint. The constraint is data centers.”
Andrew Feldman, CEO — Q1 FY2026 earnings call
Gross margin is made in two places. Hardware margin depends on wafer yield, bill of materials and price concessions to large buyers: it was −23% in FY2022, when inventory was written off in a product transition, and 42.9% in FY2025 (2024 S-1; 424B4; computed). Cloud margin depends on utilization of leased capacity, because costs start when a site is leased and fitted out but revenue starts only when a tranche is delivered. In 2026 Cerebras bridged that gap by “temporarily renting our own systems back from an existing customer”, which cost about 500 basis points of Q2 core gross margin (Q1 and Q2 FY2026 calls, CFO). Core cloud gross margin was 52.9% in Q1 and 41.8% in Q2 FY2026, against core hardware margin of 38.8% in Q2 (Q1 and Q2 FY2026 calls).
GAAP margins are held down by items that will persist. The customer-warrant asset reached $1.13B at June 30, 2026 and is amortized against revenue through October 2031 (Q2 FY2026 10-Q). Q2 GAAP gross margin was 14% against 40.6% core, the gap being warrant amortization, IPO stock compensation and pass-through revenue earned at near-zero margin. Management targets “approximately 60% gross margin and 40% operating margin in the medium to long term” (Q1 FY2026 call, CFO); core operating margin was −16% in Q2.
Operating cost is mostly engineering. R&D was $243.3M in FY2025, 48% of revenue (424B4), and falls as a share as revenue scales, which is how core operating margin moved from −42% in Q2 FY2025 to −16% (Q2 FY2026 10-Q). GAAP operating expense in Q2 FY2026 ($502.8M) was dominated by $377.0M of stock compensation, mostly a one-time catch-up when the IPO vested restricted stock units (Q2 FY2026 10-Q).
| FY2023 | FY2024 | FY2025 | H1 FY2026 | |
|---|---|---|---|---|
| Revenue | 78.7 | 290.3 | 510.0 | 373.5 |
| GAAP gross margin | 33.5% | 42.3% | 39.0% | 29.9% |
| Operating income (loss) | (133.9) | (101.4) | (145.9) | (492.3) |
| Operating cash flow | (79.0) | 452.0 | (10.1) | (47.5) |
| Capex | 6.6 | 23.4 | 382.7 | 548.9 |
| UAE related parties, % of revenue | 83% | 85% | 86% | 59% |
Sources: FY2023 from the 2024 S-1; FY2024–FY2025 from the 424B4; H1 FY2026 from the Q2 FY2026 10-Q. H1 FY2026 is not comparable with earlier years: it deducts $46.4M of customer-warrant amortization from revenue, includes pass-through revenue, and carries $386.6M of stock compensation. FY2024 operating cash flow is inflated by a $640.3M rise in customer deposits. FY2023 and FY2024 UAE share is G42 alone (MBZUAI below 10% in FY2024); FY2025 and H1 FY2026 are MBZUAI plus G42.
Cash has come from customers rather than earnings. FY2024 operating cash flow of $452.0M came from the $640.3M increase in customer deposits; as those were drawn down, FY2025 operating cash flow was −$10.1M (424B4). Free cash flow was about −$393M in FY2025 and −$596M in H1 FY2026 (inferred), with capex “primarily for systems to deliver Cerebras Cloud services” (Q2 FY2026 10-Q). Management argues that its net capex per MW is lower than peers' because it builds its own hardware at cost and is reimbursed for much of the data-center fit-out by its largest customer (Q2 FY2026 call, CFO); no capex-per-MW figure is disclosed.
The funding sources were a $1.1B Series G (September–October 2025), a $1.0B Series H (January–February 2026), the $1.0B OpenAI loan and about $6.2B of net IPO proceeds from 34.5M shares at $185 in May 2026 (424B4; 8-K May 15, 2026; Q2 FY2026 10-Q). That left $8.6B of cash and investments at June 30, 2026, with the $850M revolver undrawn. Ranked by size, the cash went to capex for cloud systems (about $932M across FY2025 and H1 FY2026), $416M of tax withholding on RSUs settled at the IPO, and a $70.7M tender offer in December 2025; there were no dividends, buybacks or acquisitions. Cerebras also signed about $4.6B of undiscounted data-center leases in 2026 through August 12, including a ~$2.2B, 10-year Canadian lease (Q1 and Q2 FY2026 10-Qs; inferred sum).
The most distinctive choice is paying for demand with equity. OpenAI holds a warrant for up to 33.4M shares at a nominal strike, vesting as capacity is delivered; AWS holds one for up to 2.7M shares at $100; and G42 received 3.5M shares at $0.01 (424B4; Q2 FY2026 10-Q). Against about 237.6M shares outstanding in August 2026, the OpenAI warrant alone equals roughly 14% (inferred). The pattern is a builder racing for capacity, funded by customer prepayments, a customer loan and customer-linked equity, with Class B shares carrying 20 votes each holding 99.2% of voting power at the IPO (424B4). No transaction is pending.
Cerebras sells into one end market, AI compute, through a handful of buyers whose spending depends on continued access to capital. OpenAI alone has announced roughly 29 GW of commitments across NVIDIA, AMD, Broadcom, AWS and Cerebras, of which Cerebras's 750 MW is about 2–3% (counterparty announcements; inferred). The company itself warns that many of its cloud customers “are startups, with capital intensive needs, that may not succeed” (424B4, risk factors).
Against its own history, Cerebras sits at a peak in revenue and capex and near a trough in margin. Q2 FY2026 core revenue of $209.9M was a record, Q2 capex of $416.9M exceeded all of FY2025 ($382.7M), and management calls Q3 2026 “the low point for core gross margin” (Q2 FY2026 call, CFO). A record margin reached in this position would mean more than one reached at a demand peak. Its only lean period on record is 2022–23: FY2022 revenue was $24.6M, headcount was cut twice and inventory was written off in a product transition (2024 S-1). It has never faced an AI-spending downturn at scale.
Three constraints bind. Power and sites limit how fast capacity can be delivered, and the IEA estimates about 20% of planned data-center projects are at risk of grid delay (IEA, Energy and AI, 2025). TSMC is the sole source of wafers, with no long-term allocation agreement (424B4). Export licences for CS-2, CS-3 and future CS-4 systems to G42 and MBZUAI carry “rigorous security and compliance obligations” and “may be revoked, including after we have manufactured the applicable products” (424B4).
In a downturn the pain would arrive through the balance sheet before the income statement. Data-center leases are “longer term with limited ability to terminate and often include substantial liquidated damages clauses”, whereas inference contracts are “a mix of consumption-based pricing or shorter term dedicated capacity arrangements” (424B4). If OpenAI slowed its tranches or token prices fell faster than costs, Cerebras would hold leases of up to ten years against capacity it might have to resell “on an expedited basis” (424B4).
“…longer term with limited ability to terminate and often include substantial liquidated damages clauses, versus our inference customer agreements, which are typically a mix of consumption-based pricing or shorter term dedicated capacity arrangements.”
Cerebras on its data-center leases — 424B4 risk factors
| Indicator | Why it matters | Where published |
|---|---|---|
| RPO and share due within 24 months | Timing of backlog conversion | 10-Q, revenue note |
| OpenAI revenue and customer shares | Whether concentration is moving or just rotating | 10-Q, concentration note and risk factors |
| Core cloud gross margin | Utilization and pricing of owned capacity | Earnings release; 10-Q non-GAAP reconciliation |
| Capex, free cash flow, leases signed but not commenced | Capital committed ahead of revenue | 10-Q cash-flow statement and lease note |
| Output speed and price per token vs NVIDIA LPX-backed providers | Whether the premium survives | artificialanalysis.ai |
| BIS rules and UAE licences | Exposure of the UAE customer base | Federal Register; bis.gov |
What the sources could not answer:
Before forming a thesis, an investor would need to know whether owned capacity reaches the 60% gross-margin target at scale, and whether customers beyond OpenAI and the UAE can fill the capacity Cerebras is leasing.