top of page
Business & Enterprise
AI Technology
Cloud Infrastructure
Cerebras
Commercialized wafer-scale computing into a full-stack AI compute business, using extreme chip-level performance to win sovereign AI programs and frontier inference demand, despite extreme customer concentration and NVIDIA’s dominant accelerator ecosystem.
1
MODEL
BUSINESS MODEL
Product–Service Hybrid
- Founded in 2015 by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker, former SeaMicro executives, to commercialize wafer-scale computing.
- Built the Wafer-Scale Engine, culminating in WSE-3: a 5nm, 4-trillion-transistor processor with 900,000 AI cores, designed to reduce data movement across chips.
- Sold CS-3 systems and supercomputers to governments, research institutions, enterprises, and AI organizations needing on-premise control.
- Launched Cerebras Inference in August 2024, converting wafer-scale systems into a token-metered cloud for high-speed LLM inference.
- Expanded through sovereign partnerships, including G42 and MBZUAI, and secured OpenAI’s multi-year commitment for 750MW of inference capacity, supported by a $1 billion working-capital loan.
1. Target a constraint that conventional architectures cannot solve without major tradeoffs, such as data movement, memory bandwidth, latency, or energy use.
2. Build the deepest technical layer first—chip, system, software, and reliability—because a hardware advantage is meaningless without a usable production stack.
3. Sell the first systems to customers whose requirements justify a novel architecture and whose deployment creates reference credibility.
4. Convert proprietary capacity into a metered service only after utilization, reliability, and support costs are understood.
5. Use anchor contracts to finance capacity, but diversify demand before concentration becomes a balance-sheet and renewal risk.
6. Avoid selling hardware as the end state; recurring cloud, managed capacity, or software services are usually necessary to smooth hardware-cycle revenue.
Cerebras scales through CS-3 system sales, cloud inference consumption, managed capacity, and multi-year committed contracts. Revenue grew from $78.7 million in 2023 to $510.0 million in 2025, a 76% year-over-year increase, and it reached an agreement for OpenAI to deploy 750 megawatts of Cerebras inference capacity over several years. But in 2025, MBZUAI represented 62% of revenue and G42 24%, so scalability depends on converting sovereign and anchor-customer demand into a diversified, recurring inference and systems business.
HOW THEY BUILT IT
HOW TO ARCHITECT IT
DISTRIBUTION MODEL
Enterprise Sales, Direct Sales, Partnerships, Developer/API-Led
HOW THEY OPERATIONALIZED
HOW TO REPLICATE WHAT WORKED
dm
- Sold CS-3 systems and supercomputers directly to governments, research labs, enterprises, and AI organizations through technical and executive enterprise sales.
- Used sovereign partnerships with G42, MBZUAI, and UAE-linked AI programs to create large, strategic deployments and national-scale references.
- Distributed Cerebras Inference through its own developer cloud, free credits, pay-per-token access, and enterprise plans.
- Expanded reach through AWS Marketplace, Microsoft Azure Marketplace, IBM watsonx, Vercel, OpenRouter, and Hugging Face, giving developers multiple procurement and integration paths.
What worked: use a small number of demanding, strategically important customers to prove a radical architecture, then expose the same capability through cloud APIs and marketplaces.
Trap if copied blindly: anchor-customer distribution becomes a liability when one or two buyers dominate revenue; without broad developer adoption or diversified contracts, the vendor inherits the customer’s budget cycles and strategic dependence.
| PATTERNS OF THIS MODEL
PATTERNS IN FULL-STACK AI HARDWARE-AS-A-SERVICE:
1. RADICAL ARCHITECTURE NEEDS ANCHOR DEMAND. Novel silicon usually requires a customer with enough scale and risk tolerance to validate the technology before broad markets adopt it.
2. HARDWARE SALES CREATE LUMPY REVENUE. System sales depend on procurement cycles, while cloud usage and managed capacity smooth revenue only if utilization is high.
3. SOVEREIGN DEMAND CAN SUBSTITUTE FOR EARLY COMMERCIAL BREADTH. Governments and national AI programs can fund frontier infrastructure, but they often concentrate revenue and introduce geopolitical exposure.
THE WEAKNESS: a hardware-led business can become dependent on a few strategic customers whose purchasing decisions are driven by national strategy, research priorities, or political relationships rather than repeatable commercial demand.
What companies with this model reveal
| OPPORTUNITY INTELLIGENCE
GOLDMINE 1 — PERFORMANCE-CONTRACTED INFERENCE. Buyers need guaranteed tokens per second, latency, cost per completed task, or energy per token—not merely access to a novel chip. Performance-backed contracts can convert architectural advantage into recurring revenue.
GOLDMINE 2 — SOVEREIGN AI FACTORIES. Nations need localized compute, data residency, talent development, and national models. A vendor that bundles systems, operations, training, and model-enablement can capture more value than a hardware supplier.
GOLDMINE 3 — INFERENCE-ORIENTED SYSTEM RENTALS. Many enterprises cannot justify buying a CS-3 system but need dedicated high-speed capacity. Capacity subscriptions and managed private deployments can turn lumpy hardware sales into recurring revenue.
THE PIT — CUSTOMER CONCENTRATION DISGUISED AS SCALE. Cerebras’s 2025 revenue was 86% dependent on MBZUAI and G42; large contracts create growth headlines but not necessarily a resilient revenue base.
THE SECOND PIT — CAPACITY COMMITMENTS WITHOUT FLEXIBLE SUPPLY. A 750MW OpenAI commitment creates enormous demand visibility but also requires power, systems, capital, and delivery execution over multiple years.
MOVE WITH CAUTION — Do not treat a large future commitment as current revenue; distinguish contracted capacity, deployed capacity, recognized revenue, and customer concentration before extrapolating growth.
Untapped Business Model / Gaps / Goldmines / Pits
Patterns & Insights
2
MARKET
mkt mt es
Cerebras competed against NVIDIA, AMD, Intel, Groq, SambaNova, and hyperscaler AI infrastructure. Rather than competing across general-purpose GPUs, it attacked the data-movement bottleneck through wafer-scale integration: WSE-3 placed 900,000 cores on one wafer, reducing inter-chip communication for large AI workloads. Its attack surface was extreme training throughput and inference speed for customers willing to adopt a new architecture. NVIDIA’s ecosystem, CUDA software base, and distribution remained the incumbent advantage Cerebras had to overcome.
MARKET TYPE
Emerging Market
ENTRY STRATEGY
Direct Entry
WHY THEY WON
AI accelerators and inference infrastructure are an Emerging Market: demand is expanding rapidly, architectures are unsettled, and buyers are experimenting with GPUs, custom ASICs, wafer-scale systems, and inference clouds. Cerebras exploited a structural opening created by the limits of distributed GPU clusters: communication overhead, memory bandwidth, and latency. Its wafer-scale approach changed the physical architecture of AI compute, but adoption required customers with large budgets and tolerance for a new platform. The transferable lesson is that emerging infrastructure markets reward radical architectural bets only when the entrant can prove a complete system and secure anchor demand.
EXECUTION
Cerebras entered directly with its own wafer-scale processor, systems, software, and later cloud, rather than entering through an incumbent’s platform or acquiring an existing business. Founded in 2015 by the former SeaMicro team, it spent years developing wafer-scale technology before launching the WSE-1 in 2019 and WSE-3 in 2024. The critical bet was that a single wafer-scale processor could outperform distributed GPU clusters for large AI workloads. This was viable because the founders had prior experience building and selling high-density server systems through SeaMicro’s $334 million acquisition by AMD. The approach required extraordinary capital, chip-yield management, thermal engineering, software maturity, and enterprise sales; it would fail if the system could not deliver reliable performance or if customers would not adopt a non-CUDA-centric platform.
FOOTHOLD STRATEGY
Lighthouse Customer Strategy
Cerebras’ first traction came from high-performance computing and AI research organizations willing to deploy novel wafer-scale systems. Argonne National Laboratory, GlaxoSmithKline, Mayo Clinic, and the U.S. Department of Energy provided early scientific and enterprise validation. The decisive lighthouse relationship became the UAE ecosystem: G42 and MBZUAI deployed Cerebras systems at scale and represented 24% and 62% of 2025 revenue respectively. These deployments established sovereign-AI credibility and created the platform for the OpenAI capacity agreement announced in 2026.
GROWTH CAMPAIGN
CAMPAIGNS THAT WORKED
SO WHAT: Cerebras’ campaigns make a radical hardware claim credible by attaching it to named scientific, sovereign, and frontier-AI deployments rather than relying on abstract benchmarks.
KEY LEARNING: for disruptive infrastructure, credibility is built through demanding deployments that prove the system works under real constraints.
HOW TO USE: pair each major architectural claim with a named production deployment, a measurable performance metric, and a buyer type that validates the claim for the next market segment.
KEY LEARNING
THE STANDARD: In an Emerging Market, entrants win by making a new technical approach credible before the market standardizes around an incumbent architecture.
RULE 1 — ARCHITECTURAL BETS NEED PROOF SYSTEMS. Buyers require complete systems, software, support, and reference deployments, not chip specifications alone.
RULE 2 — EARLY DEMAND IS CONCENTRATED. Frontier labs, governments, and research institutions are often the only buyers able to absorb novel infrastructure risk.
RULE 3 — PRICING MOVES FROM SYSTEMS TO CONSUMPTION. Hardware sales establish the platform; cloud and managed capacity create recurring revenue and broaden adoption.
RULE 4 — EMERGING MARKETS REWARD SPEED BUT PUNISH CONCENTRATION. Fast growth built on one or two anchor customers can reverse quickly if their priorities change.
MARKET TYPE: Emerging Market (category).
gc
Market Context
| MARKET INTELLIGENCE
PATTERNS IN ENTERPRISE-SALES LED, PARTNERSHIP-EXTENDED HARDWARE DISTRIBUTION:
1. TECHNICAL REFERENCE CUSTOMERS REDUCE ADOPTION RISK. Scientific institutions, sovereign programs, and marquee AI companies make radical systems credible to subsequent buyers.
2. MARKETPLACES EXTEND REACH WITHOUT REPLACING ENTERPRISE SALES. AWS, Azure, and Hugging Face distribution helps developers discover the product, while large deployments still require direct sales.
3. ARCHITECTURE LAUNCHES MUST BE PAIRED WITH OPERATIONAL PROOF. Speed claims become commercially meaningful only when supported by reliability, software maturity, and deployment references.
4. THE WEAKNESS: enterprise-led hardware distribution can produce high revenue concentration, making valuation and growth highly sensitive to a few customer relationships.
| MARKET ENTRY PLAYBOOK
THE STANDARD: Direct entry is appropriate when the entrant owns a fundamentally different technology and can sell a complete solution directly to early adopters.
RULE 1 — Enter directly when the innovation requires close control over hardware, software, and customer deployment.
RULE 2 — Build a complete system before pursuing broad distribution; component-level innovation rarely sells itself in infrastructure.
RULE 3 — Use early customers to prove reliability and economics, then expand through APIs, marketplaces, and channel partnerships.
RULE 4 — Reject direct entry if the target market requires an incumbent’s software ecosystem or certification before adoption; Platform Entry or a strategic alliance may be necessary.
How to enter
| FOOTHOLD STRATEGY PLAYBOOK
THE STANDARD: A lighthouse-customer strategy uses a demanding early buyer to prove a radical product, unlock credibility, and create a path to broader adoption.
RULE 1 — Select a customer whose needs force the product to become production-grade.
RULE 2 — Convert the relationship into referenceable outcomes, not merely a pilot.
RULE 3 — Use the lighthouse to enter adjacent segments, but diversify before dependence becomes structural.
RULE 4 — In capital-intensive markets, a lighthouse should also improve financing access, procurement scale, or operating knowledge. Cerebras’ UAE relationships provided all three, but also created concentration risk.
How to get the first strong position
MARKET PATTERNS & PLAYBOOK
3
MONEY
money rev pri
Cerebras historically relied on large CS-3 system sales, which accounted for $358.4 million, or 70%, of its $510.0 million 2025 revenue. It added Cerebras Inference in 2024, generating $151.6 million, or 30%, of 2025 revenue through token consumption and compute rental. The OpenAI agreement for 750MW of committed inference capacity and a $1 billion working-capital loan represents a move toward multi-year predictable capacity revenue, but it also increases delivery and financing obligations.
REVENUE MODEL
Product Sales, Usage-Based, Contract Revenue
PRICING MODEL
Custom Enterprise Pricing, Usage-Based Pricing, Value-Based Pricing
WHY THEY WON
Customers buy CS-3 systems and related equipment for on-premise AI supercomputing, consume Cerebras Inference by token, or rent dedicated compute monthly or annually. Revenue expands through additional systems, larger clusters, more tokens, longer cloud commitments, and managed capacity. Hardware sales create large but lumpy revenue; inference and committed-capacity contracts create more recurring visibility. Margins depend on wafer yields, system cost, power, data-center operations, financing, and utilization of cloud capacity.
Cerebras prices CS-3 systems and supercomputers through custom enterprise quotes, while Cerebras Inference uses pay-per-token pricing. Its 2026 pricing page offers a free tier with $5 in credits, self-serve funded accounts starting at $10, and production-scale enterprise plans. The strategic consequence is a hybrid model: hardware captures large upfront deployments, while token pricing broadens developer adoption and creates recurring demand.
TARGET AUDIENCE
Governments and sovereign-AI programs (enterprise/custom; want national AI infrastructure, data control, and large-scale systems).
Research institutions and national labs (enterprise/custom; want high-throughput training and scientific AI compute).
Frontier AI labs such as OpenAI (committed capacity; want high-speed inference at scale).
AI-native developers and enterprises (free, self-serve, or enterprise inference; want fast token generation through API or marketplace access).
CUSTOMER BUYING BEHAVIOUR
Large buyers begin with a strategic AI or research requirement, evaluate wafer-scale performance against GPU alternatives, and conduct technical validation before purchasing systems or committing capacity. Procurement involves executive sponsorship, multi-year planning, financing, and deployment milestones. Developers more often begin with free credits or API access, compare token speed and price, then move to funded usage or enterprise plans as workloads scale.
| PRICING INTELLIGENCE
What makes this model effective & make customers pay
Customers are not buying chips; they are buying a way to train or serve AI workloads faster, with less data movement and more control over critical infrastructure.
RULE 1 — PRICE THE WORKLOAD OUTCOME, NOT THE TRANSISTOR COUNT. Buyers evaluate time-to-train, tokens per second, energy use, and operational reliability.
RULE 2 — USE HARDWARE SALES TO ESTABLISH THE PLATFORM. Large system deployments create credibility and capital, but they should lead to recurring services rather than remain one-time transactions.
RULE 3 — USE TOKEN PRICING TO BROADEN THE FUNNEL. Pay-per-token access lets developers test the architecture without buying a system.
RULE 4 — DO NOT CONFUSE COMMITTED CAPACITY WITH REALIZED REVENUE. A 750MW commitment creates future revenue potential but requires delivery, power, financing, and utilization execution.
The buyer is really buying accelerated AI throughput and infrastructure control, proven by system purchases, token consumption, and multi-year capacity commitments.
PRICE & REVENUE
| Revenue Risk - The biggest threat to revenue stability
Extreme customer concentration: MBZUAI and G42 together represented 86% of Cerebras’ 2025 revenue; changes in their funding, priorities, or deployment plans would materially affect revenue.
Lumpy hardware revenue: CS-3 system sales depend on large procurement decisions, creating revenue timing risk even when demand is strong.
Ecosystem dependency: NVIDIA’s CUDA ecosystem and broad GPU availability remain the default for many AI teams, limiting switching unless Cerebras delivers a decisive workflow advantage.
Capacity-commitment execution risk: The OpenAI agreement for 750MW of inference capacity requires multi-year capital, power, systems, and operational delivery; failure to deploy on schedule could affect revenue and customer trust.
Financing risk: The $1 billion OpenAI working-capital loan supports growth but ties Cerebras’ balance sheet to a major customer relationship and future capacity delivery.
Cerebras reported $510.0 million in 2025 revenue, up 76% year over year, and completed its Nasdaq listing in 2026; these are company-reported financials, while future committed-capacity revenue remains subject to deployment and contract conditions.
Where the model can break
4
MOTION
motion ge cs
Cerebras’ primary revenue motion is enterprise sales: CS-3 systems, sovereign deployments, supercomputers, and multi-year capacity agreements require technical validation, executive sponsorship, procurement, and delivery planning. Its 2025 revenue mix—70% hardware systems and 30% cloud—shows that enterprise system sales remain the commercial core. The OpenAI 750MW commitment and G42/MBZUAI relationships further demonstrate a structured enterprise and sovereign sales process rather than a purely self-serve motion.
GROWTH EXPANSION MODEL
Market Development, Platform Expansion
HOW THEY EXPAND
Cerebras expanded from wafer-scale chips into complete CS systems, sovereign AI supercomputers, and a public inference cloud. Market development occurred as it moved from research and HPC buyers to sovereign AI programs, enterprises, and frontier-model companies. Platform expansion occurred through Cerebras Inference, developer APIs, marketplace distribution, and ecosystem integrations with AWS, Azure, IBM watsonx, Vercel, OpenRouter, and Hugging Face. The OpenAI agreement represents a further move into committed inference capacity at national-infrastructure scale. The dependency is that each expansion requires capital, power, systems supply, and operational delivery.
COMPETITIVE STRATEGY
Differentiation
HOW THEY COMPETE
Cerebras faced NVIDIA, AMD, Intel, Groq, and SambaNova. Its battleground was large-scale AI training and high-speed inference. Its comparative advantage was wafer-scale integration, which reduced inter-chip data movement through a single wafer-scale processor. It differentiated by selling an alternative architecture and complete systems rather than competing solely on GPU price or CUDA compatibility. The constraint was that NVIDIA’s ecosystem, developer familiarity, and distribution remained difficult to displace, while Cerebras’ early revenue depended heavily on sovereign anchor customers.
GROWTH ENGINE
GTM
Content Flywheel
Cerebras’ growth engine begins with a strategic customer’s need for extreme AI compute—national AI capability, scientific research, or frontier inference. Enterprise sales and solution engineering translate that need into a system deployment, sovereign supercomputer, or committed cloud-capacity agreement. Successful deployment creates a reference, which supports additional government, enterprise, and AI-lab opportunities. The inference cloud then converts the same architecture into metered developer demand, broadening the funnel beyond large system buyers. The engine compounds when deployments prove reliability, cloud usage grows, and new capacity contracts finance further systems. It breaks if customer concentration rises, wafer yields or system delivery falter, or GPU alternatives deliver sufficient performance at lower cost and greater ecosystem compatibility.
Cerebras runs an enterprise-led GTM system with a developer-cloud layer for broader adoption. The primary motion is direct enterprise and sovereign sales: Cerebras sells CS-3 systems, supercomputers, and committed capacity to governments, research institutions, AI labs, and large enterprises. Technical proof, executive relationships, and strategic partnerships drive these large deployments. The secondary motion is product-led: developers access Cerebras Inference through free credits, token-based billing, and cloud marketplaces. The founder-level implication is that radical hardware requires a high-touch enterprise motion first; self-serve distribution can broaden demand only after the platform is proven at scale.
SUSTAINING MOATS
High Customer Lock-In
moat
The moat is Cerebras’ wafer-scale architecture and the systems, software, and manufacturing knowledge required to make it commercially viable. By integrating massive compute and memory on one wafer, it attacks the data-movement constraint that limits conventional distributed AI systems. The moat strengthens as wafer yields, system reliability, software maturity, and deployed reference systems accumulate. It weakens if GPU-based systems achieve comparable performance more cheaply or if Cerebras cannot scale production and power capacity.
| MOAT INTELLIGENCE
THE STANDARD: A technology moat lasts when the architecture creates a performance advantage that requires deep, system-level expertise to replicate.
RULE 1 — WAFER-SCALE INTEGRATION IS A SYSTEMS BARRIER. Cerebras had to solve chip yield, thermal management, power delivery, packaging, and software—not merely design a larger processor.
RULE 2 — ANCHOR DEPLOYMENTS VALIDATE THE ARCHITECTURE. Deployments with G42, MBZUAI, national labs, and OpenAI demonstrate that the technology can operate in demanding production environments.
RULE 3 — CLOUD EXPOSURE CREATES A SECOND ADOPTION PATH. Cerebras Inference lets developers test wafer-scale performance without buying hardware, expanding the evidence base beyond large system customers.
THE SIGNAL: Cerebras reported revenue growth from $78.7 million in 2023 to $510.0 million in 2025, with WSE-3 offering 900,000 cores and 125 petaflops of peak AI performance. The moat does not protect against customer concentration, NVIDIA’s ecosystem gravity, or execution risk in delivering 750MW of committed inference capacity.
Why this company remains defensible
ARR & TAKEAWAY
ARR Journey - what to do at each stage
PRE-$1M ARR — PURSUE THE ARCHITECTURAL BET. Cerebras was founded in 2015 by the former SeaMicro team to commercialize wafer-scale computing, a problem widely viewed as technically prohibitive.
$1–5M ARR — PROVE THE FIRST SYSTEM. The company focused on chip and system engineering, building the technical credibility required to sell a radically new architecture.
$5–10M ARR — WIN EARLY ADOPTERS. It targeted research institutions and high-performance computing organizations that could validate wafer-scale systems in demanding workloads.
$10–50M ARR — EXPAND INTO ENTERPRISE AND SOVEREIGN AI. Cerebras moved from research systems toward strategic customers such as G42 and MBZUAI, creating larger deployments and sovereign references.
$50–100M ARR — PRODUCTIZE INFERENCE. In August 2024, it launched Cerebras Inference on WSE-3 and CS-3, creating a token-metered cloud channel beyond hardware sales.
$100M+ ARR — CONVERT ANCHOR DEMAND INTO COMMITTED CAPACITY. Cerebras reported $510.0 million revenue in 2025, then secured OpenAI’s agreement to purchase 750MW of inference capacity over several years, supported by a $1 billion working-capital loan. WATCH: concentration, capacity delivery, and financing now determine whether growth becomes durable.
Rule: In radical infrastructure, use demanding anchor customers to prove the architecture, then convert credibility into recurring capacity before concentration becomes a structural risk.
COPY PLAYBOOK : What Worked → What Failed → What to Replicate → What to Avoid
THE STANDARD: Build a complete alternative architecture, prove it with demanding anchor customers, then convert the validated platform into recurring cloud or capacity revenue.
SEQUENCE:
1. Solve a fundamental system constraint that incumbents address only indirectly, and build the full hardware-software-operational stack.
2. Win a small number of customers whose scale and requirements validate the architecture in production.
3. Convert proven capacity into metered services, managed deployments, or multi-year contracts—but diversify before anchor customers dominate revenue.
WORKED: Cerebras used wafer-scale engineering and sovereign AI deployments to establish credibility, then launched Cerebras Inference to broaden adoption beyond system buyers.
CAUTION: Do not treat anchor-customer contracts as diversified demand; Cerebras’ 2025 revenue was 86% dependent on MBZUAI and G42. Also avoid committing to massive future capacity without a proven ability to finance power, systems, and delivery at the required pace.
bottom of page