Cerebras Systems

1.1 Identity, Stage, and Business Model

Cerebras Systems Inc. is a Nasdaq-listed AI-infrastructure company headquartered in Sunnyvale, California. Its commercial model combines on-premises wafer-scale supercomputers with cloud inference and training services, so the company participates in both equipment sales and usage-based compute. The current product line includes systems based on its Wafer-Scale Engine architecture; third-party histories record the CS-4 introduction in August 2026. The founding year is not fully reconciled: Wikipedia and The CODEW report 2015, while Sacra and note.com report 2016. The five named founders are Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker. The discrepancy is immaterial to product identity but should be resolved against the issuer's charter or prospectus before treating either year as definitive.[CO001, CO002, CO003, CO004]

1.2 Leadership, Founder Fit, and Governance Gaps

Andrew Feldman remains the chief executive, with Robert Komin as chief financial officer, Sean Lie as chief technology officer, and Dhiraj Mallick as chief operating officer in the current third-party roster. Founder-market fit is unusually direct: the founding group had worked together at SeaMicro, whose microserver architecture addressed data-movement and efficiency constraints before AMD acquired the company for $334 million in 2012. That continuity supports execution in hardware-software co-design, but it also concentrates institutional knowledge and strategic relationships in a small group. The assigned evidence does not establish a complete current board roster, committee composition, succession plan, or founder voting control. Leadership continuity is therefore supportable; governance depth and independence are not.[CO004, CO005]

1.3 Capitalization, Liquidity, and Operating Scale

Cerebras entered the public market after two large private financings. Sources report a $1.1 billion Series G at an $8.1 billion post-money valuation in September or October 2025, followed by a $1.0 billion Series H at a $23 billion post-money valuation in January or February 2026. The month differences are retained rather than harmonized. Cerebras then reported $6.4 billion of gross IPO proceeds in Q2 2026 and an April revolving credit facility of up to $850 million. These amounts are not a single "total raised": equity proceeds, debt capacity, and cash available are different measures. At March 31, 2026, the company reported $3.3 billion of cash, cash equivalents, restricted cash, and short-term investments. Q1 GAAP revenue was $193.406 million, up 94% year over year, while Trefis showed a $25.5 billion market capitalization at a $168.86 share price on October 1, 2026.[CO006, CO007, CO009, CO010, CO011, CO012, CO013]

1.4 Milestones, Partnerships, and Adverse Signals

The 2026 expansion rests on a company-announced multi-year OpenAI agreement for 750 megawatts of compute valued at more than $20 billion, alongside an AWS distribution partnership described in the Q1 release. The same filing warns that Cerebras depends on a limited number of significant customers, including OpenAI, G42, MBZUAI, and AWS. Trefis reports that three customers represented 76% of Q2 revenue. It also reports a $450.5 million Q2 GAAP net loss, with $386.6 million attributed to non-cash stock-based compensation; that accounting loss should not be described as cash burn. Regulatory history remains relevant: independent analyses say CFIUS scrutiny of G42 delayed the first IPO process until G42's voting rights were removed. The resulting profile is scaled and well-capitalized, but still dependent on concentrated counterparties and capital-intensive delivery.[CO008, CO014, CO015, CO016]

1.5 Exhibits

2.1 Market Boundary and Substitutes

Cerebras competes in AI compute infrastructure rather than in the whole semiconductor or software market. The included spend is dedicated accelerator hardware, integrated supercomputing systems, and cloud-delivered training or inference capacity for large AI workloads. General-purpose CPUs, end-user devices, model-development software, data labeling, and ordinary cloud storage sit outside this boundary. The company sells both on-premises systems and cloud access, so the relevant buyer choice spans capital purchases, reserved infrastructure capacity, and usage-based services rather than a single chip budget. Status-quo substitutes include GPU clusters, hyperscaler-owned accelerators, and conventional cloud AI services. Cerebras is therefore not entitled to the whole AI hardware forecast: its practical opportunity is the subset of workloads for which latency, model scale, or programming simplicity offsets ecosystem familiarity, switching cost, power, and facility requirements. That boundary also separates a technically addressable workload from spend Cerebras can realistically capture.[CM001, CM002, CM011, CM013, CM014]

2.2 Sizing Lenses and Demand Evidence

Two third-party forecasts provide broad upper bounds, not a reconciled market model. KoalaGains projects the AI accelerator market at more than $200 billion by 2030 with growth above 30% CAGR, while Sacra projects the broader AI hardware market at nearly $250 billion by 2030. The category definitions, base years, geographies, and calculation methods are not disclosed in the fetched text, so the endpoints should not be averaged or relabeled as Cerebras TAM. Company-scale evidence is more concrete but answers a different question: Cerebras reported $193.4 million of Q1 2026 GAAP revenue, split between $110.6 million of hardware and $82.8 million of cloud and other services, and disclosed an OpenAI agreement for 750 MW valued above $20 billion. These figures demonstrate demand and route-to-market expansion, but a customer contract is not market size and recognized revenue is not SAM. No retained source isolates the spend pool addressable by wafer-scale systems or provides adoption probabilities needed for a defensible SOM.[CM003, CM004, CM005, CM008, CM009, CM010, CM016]

2.3 Buyers, Users, Payers, and Adoption Path

The evidence supports several buyer groups but does not disclose most internal budget owners. Frontier model developers and hyperscalers buy or distribute dedicated inference capacity; national laboratories, research institutions, and sovereign programs purchase integrated systems for large scientific or national-scale workloads; enterprises and application developers can access the architecture through cloud services without owning hardware. The likely economic buyer is therefore an infrastructure, research, or technology executive, while day-to-day users are model engineers, researchers, and application teams. The route to adoption differs by segment: a laboratory can procure an appliance, a frontier lab can reserve multi-year capacity, a hyperscaler can integrate Cerebras into a broader service, and an enterprise can begin with cloud trials. Exact signatories, approval thresholds, sales-cycle length, and budget-line ownership remain private. Those omissions matter because technical users may value speed while finance and facilities teams evaluate total cost, power, utilization, and vendor risk.[CM002, CM005, CM006, CM007, CM012, CM014]

2.4 Growth Drivers, Adoption Constraints, and Valuation Relevance

The central demand thesis is that interactive and agentic applications increase the value of tokens delivered with low latency. An investor letter carried by Yahoo Finance describes fast inference as a rapidly growing compute subsegment, and Cerebras's OpenAI and AWS arrangements provide customer and distribution signals consistent with that thesis. The counterweight is a demanding adoption equation. GPU software ecosystems and hyperscaler breadth remain the default; on-premises systems carry multi-million-dollar pricing and specialized power requirements; cloud expansion requires data-center capacity and capital; and customer concentration can turn a large commitment into counterparty and execution exposure. Wikipedia reports that MBZUAI and G42 represented 62% and 24% of 2025 revenue, respectively, while KoalaGains independently describes more than 85% concentration in two customers. For valuation, the broad market forecasts justify a growth-market framing but not a precise share assumption. Underwriting should instead test contract conversion, diversified utilization, switching economics, and gross profit after infrastructure costs.[CM005, CM006, CM007, CM011, CM013, CM014, CM015, CM016]

2.5 Exhibits

3.1 The buyer can choose an incumbent platform, a specialist accelerator, or no new vendor

Cerebras does not face one clean peer set. Hardware buyers can remain with Nvidia, evaluate AMD or Intel, test specialist architectures from Groq, SambaNova, or Graphcore, or procure captive silicon through AWS, Google, and Microsoft. Cloud buyers can also buy managed compute from AWS, Azure, Google Cloud, Oracle, or CoreWeave instead of owning an accelerator system. Internal GPU clusters and continued use of existing cloud capacity remain status-quo substitutes. Sacra estimates Nvidia holds approximately 85% of the AI accelerator market, so Cerebras is challenging an installed platform rather than entering an open field. Public evidence in the assigned pool does not provide comparable funding, installed-base, or contract data for every named competitor; those cells remain explicitly unknown.[CP001, CP002, CP015]

3.2 Wafer scale removes interconnect work, but architectural uniqueness is narrowing

Cerebras concentrates compute and memory on a wafer-scale processor, reducing the chip-to-chip communication and distributed-software burden found in clustered GPU systems. Independent and user-generated sources consistently describe this as the core differentiation, while Cerebras-oriented commentary reports strong inference performance on selected models. The counterpoint is important: a former Cerebras technical staffer wrote that several original design properties are no longer unique and that chiplets trade some runtime communication cost for better yield and incremental upgrades. Nvidia's CUDA software support also makes the competitive comparison broader than raw throughput. The evidence therefore supports a workload-specific advantage, not a universal platform lead.[CP003, CP004, CP005, CP016]

3.3 Cerebras is using both systems and cloud distribution to lower adoption friction

Publicly reported Cerebras packaging spans expensive systems and metered inference. One independent report places a CS-3 at roughly $2 million to $3 million list price and quotes Llama 3.1 405B inference at $6 per million input tokens and $12 per million output tokens; realized enterprise terms are not disclosed. Wikipedia separately describes nodes costing as much as $3 million and drawing 25kW. AWS distribution can reduce procurement and integration friction: Cerebras announced a disaggregated design in which Trainium 3 performs prefill and CS-3 performs decode. That heterogeneous model may enable multi-homing rather than requiring a full replacement, but the pool contains no customer-level migration, retention, or switching cost evidence.[CP006, CP007, CP008, CP009, CP010]

3.4 Capacity, ecosystem control, and buyer concentration can erode benchmark-led differentiation

Distribution is improving, yet execution risk remains central. Simply Wall St reported a planned European footprint and a Flex-supported target to increase CS-3 production capacity sevenfold, while Cerebras disclosed an OpenAI agreement for 750 megawatts valued above $20 billion. Those commitments may validate demand but also increase delivery and counterparty leverage. A September 2026 Seeking Alpha author argued that WSE-3's memory bandwidth is offset by limited capacity and rapid Nvidia advances. Cerebras itself warns that dependence on a small number of significant customers, data-center capacity, and effective competition are material uncertainties. The moat is therefore conditional: it lasts only if performance converts into reliable capacity, competitive economics, and a broader software and customer base before incumbents narrow the gap.[CP011, CP012, CP013, CP014, CP016]

3.5 Exhibits

4.1 Revenue Mix and Recognition

Cerebras reports two operating revenue categories rather than ARR: hardware and cloud and other services. In Q2 2026, GAAP revenue was $180.1 million, comprising $54.1 million of hardware and $126.0 million of cloud and other services. Core revenue was $209.9 million because the company removes data-center pass-through revenue and adds back non-cash customer-warrant amortization; the resulting core measure is therefore not interchangeable with GAAP revenue. Cloud was the growth engine, but public disclosure does not separate recurring API usage, reserved capacity, professional services, or other services within that category. The company also reported $25.4 billion of remaining performance obligations. That contracted amount is evidence of demand, not recognized revenue, and conversion depends on customer deployment choices, build schedules, and the company's ability to deliver capacity.[CI001, CI002, CI003, CI011]

4.2 Monetization and Go-to-Market Evidence

The public evidence supports a mixed hardware-and-cloud model but not a complete pricing model. Hardware is recognized through system sales, while cloud activity reaches customers through Cerebras-operated capacity and large partner or customer commitments. The disclosed OpenAI arrangement and the 600 MW capacity program indicate an enterprise, negotiated-contract motion rather than a transparent self-serve price book. Historical concentration also matters: a secondary source reports that MBZUAI and G42 represented 62% and 24% of 2025 revenue, respectively. New contracts may diversify that base, but no allocated source quantifies current concentration, sales cycles, CAC, commissions, discounts, renewal behavior, or channel economics. The fetched Private Shares Fund profile and Seeking Alpha text add no usable current financial detail, so they cannot close those gaps.[CI008, CI015, CI016]

4.3 Margin, Cost Structure, and Cash Economics

Q2 demonstrates why GAAP and cash measures must remain separate. GAAP gross margin fell to 14.2%, versus 40.6% on the company's core basis; the reconciliation removes pass-through effects and adds back customer-warrant amortization and stock-based compensation. GAAP operating loss was $477.2 million, while core operating loss was $33.6 million, with $377.0 million of stock-based compensation in the reconciliation. Neither loss measure is cash burn. For the six months ended June 30, operating cash outflow was $47.5 million, while purchases of property and equipment were $548.9 million and total investing cash outflow was $1.31 billion. These figures show that infrastructure investment, more than operating cash use alone, is the central near-term capital demand. Per-system cost, utilization, and contribution margin remain private.[CI003, CI004, CI005, CI006, CI009]

4.4 Capital Adequacy and Financial Verdict

Cerebras reported $8.6 billion of cash, restricted cash, and short-term investments at June 30, 2026, alongside $918.2 million of working-capital-loan liabilities and $850 million of stated debt capacity. Restricted cash and borrowing capacity are not equivalent to freely deployable cash. The historical financing chronology is addressed in Company Overview; for forward adequacy, the relevant facts are the current liquidity mix, the 600 MW buildout, and management's raised full-year core outlook of $880 million to $890 million of revenue with a negative 19% to negative 17% core operating margin. Sacra separately estimates roughly $3 billion of burn across 2026-2027, but that is an external forecast rather than company guidance. The financial verdict is therefore conditional: revenue growth and liquidity are strong, yet revenue quality, unit returns, customer concentration, and unrestricted capital after committed infrastructure spend remain material diligence blockers.[CI007, CI008, CI010, CI012, CI013, CI014]

4.5 Exhibits

5.1 Product Portfolio and Customer Workflow

Cerebras packages the same core computing approach into several delivery forms rather than a single appliance. SEC-filed company material and an independent company profile describe both on-premises systems and cloud APIs, allowing a buyer either to deploy hardware or consume hosted inference and training capacity. CS-3 is the installed system referenced in the AWS architecture, while Sacra and Venture Atlas describe CS-4 as an announced rack built from three WSE-3 Turbo processors. Sacra also identifies CSoft and associated services as parts of the offer. The practical diligence distinction is maturity: CS-3 and cloud services have stated operating use, whereas the assigned evidence records CS-4 shipment guidance rather than independently verified production acceptance.[CE001, CE002, CE003, CE004, CE013]

5.2 Architecture and Technical Differentiation

The architectural thesis is to reduce cross-chip movement by using an entire wafer as the processor. Wikipedia's synthesis describes WSE-3 as a 215 mm-square wafer-scale semiconductor using switched fabric and SRAM; it attributes lower latency and fewer interconnect bottlenecks to that layout. The same source reports material trade-offs, including a 25 kW power draw and a price as high as $3 million per node, so the integration advantage cannot be separated from facility and capital requirements. Cerebras engineering job material further indicates that product execution spans RTL, synthesis, design-for-test, silicon bring-up, third-party IP, networking, storage, and physical deployment. The resulting stack is vertically integrated, but public evidence does not expose wafer yield, defect rates, or field failure data.[CE005, CE006, CE007, CE008, CE015]

5.3 Deployment, Integration, Reliability, and Roadmap

The clearest integration design in the retained evidence is the AWS partnership announced in Cerebras's SEC-filed Q1 2026 release: Trainium 3 performs prompt prefill and CS-3 performs decode. This separates stages with different compute characteristics, but it also creates a cross-vendor orchestration and deployment dependency that is not resolved by chip specifications alone. A Cerebras-authored infrastructure job listing describes rack layouts, bills of materials, power allocations, cable maps, deployment-readiness reviews, validation checks, and root-cause analysis, which is evidence of intended operating discipline rather than proof of achieved reliability. CS-4 adds roadmap upside, but independent acceptance tests, service-level history, and named shipment verification were not available in this pool.[CE002, CE003, CE008, CE009, CE013]

5.4 Performance, Software Adoption, and Trust Controls

Cerebras reports more than 1,000 tokens per second for Codex-Spark and says Kimi K2.6 approached 1,000 tokens per second in an Artificial Analysis measurement, while Gemma 4 31B ran an order of magnitude faster based on that index. These are named results, but the company's own website cautions that performance comparisons may rely on third-party benchmarking or internal testing and vary by workload, configuration, date, and model. Sacra separately identifies developer dependence on Nvidia's mature CUDA ecosystem as a software-adoption risk for Cerebras's bespoke stack. The assigned pool does not establish security certifications, privacy controls, a public status history, support SLAs, or formal quality attestations; that is a diligence gap, not evidence that such controls are absent.[CE010, CE011, CE012, CE014, CE016]

5.5 Exhibits

6.1 Customer Segments and Buying Motions

Cerebras reaches buyers through several distinct motions rather than one homogeneous customer funnel. Its historical base includes national laboratories, sovereign AI organizations, and life-sciences or industrial enterprises using on-premises systems for large scientific and model workloads. The newer inference motion addresses frontier-model developers and hyperscalers: OpenAI has a large direct capacity agreement, while AWS is intended to distribute Cerebras-backed inference to startups and enterprises. A third surface consists of enterprise and developer integrations such as Notion and Cline. Contract form differs by surface: the allocated evidence describes outright system purchases with renewable software subscriptions, reserved cloud capacity under take-or-pay terms, and on-demand usage. Those channels identify buyer and user groups, but they do not establish how many active paying customers Cerebras has in each segment.[CU001, CU002, CU003, CU004, CU009, CU014]

6.2 Adoption Signals and Named-Customer Proof

The clearest current adoption signals are contracted capacity, services revenue, announced trials, and specific integrations rather than a disclosed customer count. Cerebras reported $82.8 million of GAAP cloud and other services revenue for the quarter ended March 31, 2026, up 178% year over year. OpenAI agreed to deploy 750 megawatts of high-speed inference compute under a multi-year deal valued above $20 billion, and Cerebras announced a multi-year AWS partnership. The status labels matter: Kimi K2.6 and Gemma 4 were described as enterprise customer trials, while the AWS announcement describes a distribution and deployment plan rather than evidence of realized end-customer usage. Sacra separately reports deployed enterprise surfaces at Notion and Cline, but the allocated pool contains no primary customer documents for those outcomes.[CU003, CU004, CU005, CU006, CU010, CU011, CU012]

6.3 Durability, Expansion, and Concentration

Counterparty concentration remains the central customer-quality risk. BBAE and Wikipedia report that MBZUAI and G42 represented approximately 62% and 24% of 2025 revenue, respectively, or 86% in aggregate. The 2026 OpenAI and AWS relationships broaden the named-customer set, but Cerebras itself still lists OpenAI, G42, MBZUAI, and AWS among the significant customers on which it depends. This is not the same as demonstrated diversification: the allocated sources do not disclose recognized revenue by those newer relationships, deployment milestones, cancellation protections, or the share of capacity already utilized. Simply Wall St further frames under-filled capacity or changed OpenAI and AWS deployment choices as risks to utilization and margins. Renewable subscriptions and take-or-pay cloud terms can improve contractual durability, but no public NRR, GRR, churn, renewal, or cohort data in the assigned pool tests realized retention.[CU007, CU008, CU013, CU014, CU015, CU016]

6.4 What the Evidence Does and Does Not Prove

The named relationships provide credible evidence that Cerebras can sell to sophisticated AI laboratories, cloud platforms, sovereign institutions, national laboratories, and enterprises. They do not yet prove a broad, durable revenue base. User-base figures attached to Notion or Cline describe the distribution surface of those products, not the number of paying Cerebras customers or the amount of Cerebras-attributable usage. Likewise, agreement value and megawatts are commitments, not recognized revenue, customer retention, or capacity utilization. Investment diligence should therefore reconcile contract schedules to recognized revenue, request segment-level active account and renewal cohorts, separate pilots from production traffic, and quantify top-one and top-five customer exposure after the OpenAI and AWS ramps.[CU003, CU004, CU005, CU010, CU011, CU016]

6.5 Exhibits

7.1 Risk ranking and investment implication

The risk profile is dominated by correlated dependencies rather than a single isolated failure mode. Cerebras itself identifies a limited set of significant customers, including OpenAI, G42, MBZUAI, and AWS, while an independent summary attributes 62% and 24% of 2025 revenue to MBZUAI and G42, respectively. The announced OpenAI commitment adds scale but does not by itself diversify counterparty exposure: it shifts more execution risk toward one large deployment, its financing arrangements, and the infrastructure required to serve it. Manufacturing risk is also concentrated because the available evidence describes TSMC as the only capable manufacturer and the process as unusually specialized. The heatmap therefore ranks customer, deployment-capital, and foundry/yield risks as the most consequential residual exposures. Ratings are analyst judgments based on the cited public evidence, not measured probabilities.[CR005, CR006, CR009, CR010, CR011, CR012, CR013, CR014, CR015]

7.2 Regulatory, legal, IP, and security posture

The clearest adverse regulatory precedent in the assigned evidence is the reported CFIUS scrutiny surrounding the earlier G42-linked IPO process. That event is historical, but it demonstrates that customer identity, governance rights, export controls, and national-security review can affect financing and market access. The 2026 underwriting agreement contains company representations that required legal proceedings were disclosed, that no security breach or notice had occurred, and that no material written IP challenge was pending or threatened. Those are useful disclosure controls, not independent proof that incidents, claims, or vulnerabilities are absent. Residual exposure remains material until counsel, security auditors, and management provide current jurisdiction-by-jurisdiction records, incident logs, export-control analyses, and an IP ownership/retention map.[CR001, CR002, CR003, CR004]

7.3 Operational, quality, and capital-execution risk

Wafer-scale differentiation creates a coupled operational risk: the architecture depends on a specialized manufacturing process, high-power systems, and adequate data-center capacity. Independent sources flag supply bottlenecks and technological obsolescence, while the public company filing records a $4.590 million product-warranty provision for the three months ended March 31, 2026. That accounting provision is not a failure-rate measure because the assigned evidence does not disclose units shipped, claims frequency, severity, or yield. Financial execution also remains exposed: Cerebras reported a Q1 GAAP operating loss and projected deeply negative core operating margins for Q2 and full-year 2026 while an analyst estimates substantial 2026-2027 cash consumption. Liquidity raised is not equivalent to cash available for every commitment, so deployment milestones and restricted uses require reconciliation.[CR005, CR006, CR007, CR008, CR013, CR014, CR015, CR016]

7.4 Partner dependencies, mitigations, and thesis-break triggers

Mitigation maturity is mixed. AWS distribution broadens the route to market, and IPO proceeds improve financing capacity, but neither removes OpenAI deployment concentration, Gulf-linked historical concentration, TSMC dependence, or the need to fund capacity before demand is converted into recognized revenue and cash. The diligence standard should therefore be contractual and measurable: obtain revenue concentration by customer, termination and minimum-purchase terms, capacity acceptance schedules, foundry allocation and yield history, warranty cohorts, security audit results, and downside liquidity forecasts. Thesis-break triggers are cancellation or material deferral of the OpenAI rollout, failure to reduce single-customer revenue dependence, repeated misses against capacity milestones, persistent negative operating margins without a funded path to breakeven, or an export-control, cybersecurity, IP, or quality event that restricts deployment.[CR001, CR003, CR005, CR008, CR009, CR010, CR011, CR012, CR013, CR014, CR015]

7.5 Exhibits

8.1 Investment Thesis and Structural Anti-Thesis

The core investment thesis for Cerebras Systems rests on its proprietary Wafer Scale Engine architecture, which eliminates interconnect latency bottlenecks by integrating an entire silicon wafer into a single processor. With WSE-3 delivering 4 trillion transistors and 900,000 AI-optimized cores, Cerebras provides 10x to 20x speed advantages in AI inference workloads compared to legacy GPU clusters. This hardware differentiation enabled Cerebras to secure a transformative multi-year compute agreement with OpenAI for 750MW valued at more than $20 billion, expanding remaining performance obligations to $24.6 billion as of year-end 2025. Conversely, the structural anti-thesis emphasizes that headline profitability remains distorted by non-cash items, while operating losses persisted at $145.9 million in 2025 and $15.0 million in Q1 2026. Furthermore, extreme customer concentration across OpenAI, G42, MBZUAI, and AWS creates acute counterparty sensitivity, where any regulatory intervention, deployment delay, or commercial renegotiation would directly impair revenue run rates.[CV001, CV002, CV003, CV004, CV005, CV006, CV007]

8.2 Valuation Context, Multiple Realities, and Investment Stance

Cerebras' private-to-public trajectory represents one of the steepest multiple expansions in semiconductor history. The company raised a $1.1 billion Series G in September 2025 at an $8.1 billion post-money valuation, followed rapidly by a $1.0 billion Series H in February 2026 led by Tiger Global at a $23.0 billion post-money valuation. In May 2026, Cerebras priced 30 million Class A common shares at $185 each in its Nasdaq IPO, raising $5.55 billion in gross proceeds and opening at $350 per share before stabilizing around $242 per share, representing a $53 billion market capitalization. At this valuation level, CBRS trades at approximately 222x trailing 2025 earnings, though GAAP net income of $237.8 million was driven by a non-recurring $363.3 million non-cash gain from extinguishing forward contract liabilities rather than sustained core profitability. Operating margin guidance of negative 28% to 32% for full-year 2026 highlights ongoing infrastructure buildout costs. Given these valuation dynamics, we maintain a Research-More stance with medium confidence, a high risk rating, and a stretched valuation assessment.[CV008, CV009, CV010, CV011, CV012, CV013]

8.3 Scenario Analysis, Downside Triggers, and Thesis-Break Criteria

Evaluating Cerebras requires modeling contrasting scenarios across multi-year infrastructure delivery and commercial conversion. In our bull case, Cerebras executes smoothly on its 750MW OpenAI deployment, scales AWS Bedrock adoption for disaggregated prefill/decode architectures, successfully ramps the CS-4 system built on 5nm process technology, and converts its $24.6 billion backlog at healthy core gross margins approaching 50%, supporting an implied valuation of $75 billion. In our base case, infrastructure deliveries face typical datacenter power and cooling constraints, with full-year 2026 core revenue landing near guidance of $855 million to $865 million and 2028 revenue scaling toward $4.5 billion, supporting a fair equity value of $42 billion to $48 billion. In our bear case, customer concentration backfires through delayed OpenAI cluster rollouts, G42 regulatory headwinds resurface under CFIUS oversight, and capital expenditure burn consumes available cash, compressing multiples toward $22 billion. Crucial diligence priorities include auditing WSE-3 wafer yield economics with TSMC, tracking non-G42/non-OpenAI customer diversification, and monitoring cash burn against the $3.3 billion cash reserve.[CV003, CV005, CV014, CV015, CV016]

8.4 Exhibits