Cerebras: Wafer-Scale AI Inference at 3,000 Tokens Per Second

When Cerebras launched the first Wafer Scale Engine in 2019, most AI researchers assumed it was a curiosity — impressive engineering theater, impractical for production. Seven years later, Meta runs its flagship Llama API on Cerebras hardware. OpenAI committed to spend more than $20 billion on Cerebras compute. The company completed the largest tech IPO of 2026 on May 14 — priced at $185/share, surged 70% on debut to a ~$70 billion market cap.

The chip worked.

This review covers Cerebras from every angle: what the hardware actually is, how fast the inference really is, what the cloud API costs, and where the gaps are.

What Cerebras Built

The Chip: WSE-3

The Wafer Scale Engine is exactly what the name suggests — an entire silicon wafer, processed as a single chip rather than diced into individual dies. The WSE-3, announced in March 2024 and recognized by TIME Magazine as one of the Best Inventions of 2024, has:

• 4 trillion transistors — the most of any chip ever built
• 900,000 AI-optimized cores distributed across the entire wafer
• 44 GB of on-chip SRAM — not cache, not off-chip HBM, but SRAM sitting immediately adjacent to compute
• 21 petabytes/second of memory bandwidth — 7,000× an NVIDIA H100
• 214 petabits/second fabric bandwidth (per independent WSE-3 technical coverage; not stated in Cerebras’s own inference-launch blog) — cores talking to each other at chip speeds
• 46,225 mm² die size — the largest chip ever manufactured (TSMC 5nm)

For context: a standard GPU die is roughly 800 mm². The WSE-3 is 58× larger. The entire chip is one continuous communication fabric.

Why This Architecture Matters for Inference Speed

GPU inference is slow because models do not fit on-chip. Weights live in off-chip HBM memory and must be streamed across a narrow memory bus billions of times per second. The bus is the bottleneck. No matter how many CUDA cores a GPU has, the memory bandwidth constrains how fast tokens can generate.

Cerebras eliminates this bottleneck. On a CS-3 system, a model that fits within 44 GB lives entirely on-chip. Every one of the 900,000 cores has instant SRAM access. No HBM roundtrips. No inter-chip networking. The entire model is always in compute-adjacent memory.

For models that exceed 44 GB, Cerebras adds MemoryX — an external parameter storage system (4 TB to 2.4 PB), connected via SwarmX, a high-bandwidth fabric that scales near-linearly: 10 CS-3 systems deliver approximately 10× the performance of one CS-3.

The Andromeda supercomputer (16 CS-2 systems, 13.5 million AI cores, over 1 exaflop of AI compute) demonstrated this scaling publicly in November 2022, delivering 15.87x throughput across 16 systems versus a single CS-2 — near-linear scaling. The architecture has since been productized into the CS-3 system.

A Note on the Founders

Cerebras was founded in 2015 by Andrew Feldman (CEO), Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker — five co-founders who previously built SeaMicro, a company that made energy-efficient microservers and was acquired by AMD in 2012 for approximately $334 million. SeaMicro’s core insight was that many small, efficient processors beat a few large, powerful ones. At Cerebras, they inverted the idea: one enormous processor beats many interconnected small ones.

Note (added at the 2026-08-24 audit): Cerebras launched its next-generation system, the CS-4, on August 18, 2026 — a three-wafer rack system built on WSE-3 Turbo processors that Cerebras claims delivers up to 30x faster inference than GPU-based systems. This review was researched and written in May 2026 and describes the WSE-3/CS-3 generation that was current at the time; first CS-4 shipments were still rolling out at the time of this audit.

Company & Funding

Total private equity raised: approximately $2.8B across all rounds.

The company filed its initial S-1 in September 2024, then withdrew it in October 2025 amid a CFIUS national security review of a ~$335 million investment by G42, the Abu Dhabi AI/cloud firm, over technology-transfer concerns; G42’s stake was subsequently restructured to non-voting shares. Cerebras refiled its S-1 on April 17, 2026, then set terms in an amended filing (S-1/A, filed May 4, 2026) offering 28,000,000 shares at an anticipated $115.00–$125.00 per share — a raise of up to $3.5 billion at the top of the range — under the ticker CBRS on the Nasdaq Global Select Market.

IPO completed May 14, 2026. Cerebras priced above its expected range at $185/share on May 13, then opened at $350/share on May 14 — the largest tech IPO of 2026. First-day trading saw the stock surge ~70%, pushing the market capitalization to nearly $70 billion at peak. Total capital raised: approximately $5.55 billion.

Revenue is real and growing (2025 figures as disclosed in Cerebras’s S-1, reported by Yahoo Finance):

In 2025, Cerebras reported net income of $87.9 million on a GAAP basis, per its S-1 filing. Important nuance added at the 2026-08-24 audit: this headline number is not organic operating profitability. Per Yahoo Finance’s breakdown of the S-1, the $87.9M figure was driven by a $363 million non-cash paper gain when the G42 stake was restructured to non-voting shares (see below); stripping that out, Cerebras’s actual 2025 operating loss was $75.7 million — wider than 2024’s $21.8 million operating loss. The original May 2026 version of this review characterized this as “profitability… unusual for a hardware-first AI company at this stage” without that context, which overstated the underlying business performance.

The caveat the S-1 must disclose: before the OpenAI deal, G42 (UAE) was a dominant single customer. The $20 billion OpenAI compute commitment reduces customer concentration but creates a new dependency on a single mega-customer. This is a risk factor worth noting for anyone evaluating Cerebras as an enterprise vendor with long time horizons.

The OpenAI Deal

In January 2026, OpenAI announced a partnership with Cerebras for 750 MW of compute, with options through 2030. Per Yahoo Finance/The Information’s reporting and TechCrunch, OpenAI also loaned Cerebras approximately $1 billion (at 6% annual interest) to help fund data center buildout and received warrants for a minority equity stake — convertible to roughly 10-11% ownership if OpenAI’s total spending reaches $30 billion. The deal’s total value was confirmed at over $20 billion in Cerebras’s first post-IPO earnings report (Q1 2026, reported June 23, 2026).

This is not a research collaboration or a small-scale pilot. OpenAI is betting a substantial fraction of its infrastructure spend on Cerebras hardware. The open-weight gpt-oss-120B model runs on Cerebras Cloud at 3,000 tokens/second at full 128K context.

The deal answers the question of whether wafer-scale inference can scale commercially. The answer is yes.

Inference Performance

This is where Cerebras has no competition. The numbers are not small improvements — they are categorical differences.

Benchmark Numbers (as of May 2026)

At the time of original research (May 2026), the Artificial Analysis benchmark tracker independently confirmed Cerebras’s Llama 3.1 8B at 2,336 t/s and gpt-oss-120B at approximately 1,760–1,797 t/s (the lower figure is likely due to rate-limited conditions on the free tier; dedicated access reaches ~3,000 t/s). Update at 2026-08-24 audit: Llama 3.1 8B is no longer on Cerebras’s public API (see Model Catalog below), and Artificial Analysis’s current tracker shows gpt-oss-120B at approximately 1,636–1,715 t/s under its standard benchmark conditions — in the same range as the earlier reading, with the ~3,000 t/s headline figure representing dedicated/unthrottled access rather than the shared benchmark endpoint.

The Meta Llama API, announced in April 2025, selected Cerebras to power Llama 4 Scout and Maverick. Cerebras reported the solution ran 18× faster than traditional GPU alternatives, delivering over 2,600 tokens per second for Llama 4 Scout.

Important caveat on multiplier figures: Speed comparisons like “57× faster than GPU-based solutions” come from Cerebras marketing materials. The direction of the advantage is confirmed by independent benchmarks; the specific multipliers should be understood as Cerebras-favorable measurements rather than neutral third-party assessments. The raw t/s numbers from Artificial Analysis are the most reliable reference.

TTFT and Latency

• gpt-oss-120B: Consistent sub-second TTFT in API benchmarks
• Llama 4 Scout: 0.5 second end-to-end response time (full-phrase response, not first token)
• Llama 3.1 405B: ~240ms TTFT at announcement (August 2024)
• GLM-4.7: 0.42–0.43s TTFT per Artificial Analysis, measured May 2026 — GLM-4.7 was subsequently deprecated from Cerebras’s public API (see Model Catalog below) and this figure is historical

Cerebras is extremely competitive on TTFT as well as throughput. This combination — fast first token and fast per-token generation — makes it particularly well-suited for interactive coding assistants, streaming chat applications, and agentic pipelines where accumulated latency across many calls compounds.

Model Catalog

This is the main weakness of Cerebras Cloud for most developers — and it got narrower, not wider, since this review was first written.

Update at the 2026-08-24 audit: the deprecations flagged below did happen on schedule. As of this audit, Cerebras’s own model documentation and the Artificial Analysis provider tracker both list just two models on the public endpoint: gpt-oss-120b and gemma-4-31b. Llama 3.1 8B, GLM-4.7, and Qwen 3 235B Instruct are no longer on the public API. Google’s Gemma 4 31B was added around June 29, 2026 as Cerebras’s first multimodal model, running at roughly 1,850 output tokens/second — so the “no Gemma” gap noted in the original May 2026 review no longer applies, but the overall catalog is smaller than it was at launch, not larger.

The May 2026 state of the catalog, for historical reference:

DeepSeek V4 Pro, Mistral: Still not available on the public API as of this audit. Cerebras has supported DeepSeek R1 Distill Llama 70B (a smaller distilled model) via dedicated endpoints, but the full DeepSeek V4 Pro is absent.

For comparison: Together AI offers 100+ models. Fireworks AI offers 50+. Groq offers ~12. Cerebras now offers just 2 on its public endpoint.

Enterprise customers can access additional model families through dedicated endpoint arrangements, but this requires a sales engagement and custom pricing. For most developers experimenting with the public API, the catalog is genuinely narrow.

Pricing

Public API — current pricing (as of 2026-08-24 audit)

Cerebras’s catalog and pricing have changed since this review was written in May 2026; the two rows below reflect the two models currently on the public endpoint, confirmed against OpenRouter’s live Cerebras pricing (which mirrors provider pricing) and Artificial Analysis:

Llama 3.1 8B and GLM-4.7, both shown in the original May 2026 pricing table, are no longer offered on the public API (see Model Catalog above) and have been removed from this table.

These prices are still competitive for gpt-oss-120B — $0.35/$0.75 is meaningfully cheaper than GPT-4o class models from OpenAI, and the throughput advantage may more than justify the cost for latency-sensitive workloads.

Free Trial (changed since original review)

At the time this review was first written (May 2026), Cerebras offered a perpetual free tier of 1,000,000 tokens/day with no credit card required. That is no longer the case. As of this audit, Cerebras’s own pricing page offers instead:

• A one-time “Free Trial”: $5 in credits granted after account creation, rather than a recurring daily allowance
• Access to all public models, with a 65,536-token context window on the free trial tier (raised from the earlier 8,192-token cap) per Cerebras’s model documentation; the paid tier gets the full 131K context
• Community support via Discord only

This is a materially worse free-tier deal than the “generous, no-credit-card, 1M tokens/day” positioning in the original review, and the pros/cons lists above have been updated to reflect it.

Developer Tier

• $10 minimum deposit, self-serve
• 10× higher rate limits than free tier
• Higher queue priority

Subscription Plans (Code-focused)

• Cerebras Code Pro: $50/month — 24M tokens/day
• Cerebras Code Max: $200/month — 120M tokens/day

Both plans were listed as sold out on the Cerebras Code page both at original research (May 2026) and again as of this 2026-08-24 audit — the sold-out status has persisted for months. This may indicate high demand for the coding use case specifically, or it may indicate capacity constraints on specific workloads.

Developer Experience

OpenAI API Compatibility: Cerebras designed its inference API to be compatible with OpenAI client libraries. Switching requires changing two lines: the base URL to https://api.cerebras.ai/v1 and passing a Cerebras API key. No code rewrite.

Official SDKs: Python and TypeScript SDKs are available on GitHub (Cerebras/cerebras-cloud-sdk-python).

Integrations: Vercel, OpenRouter, Hugging Face, Cloudflare AI Gateway, Vercel AI SDK (@ai-sdk/cerebras), Haystack, Pipecat (voice agents), Promptfoo, AWS Marketplace.

MCP Server: No official Cerebras MCP server at time of writing. A community-built “Cerebras Code MCP Server” exists on Lobehub (Node.js, built with the official MCP SDK) and provides code generation with file system operations. Cerebras has demonstrated integrations with MCP servers in developer demos but has not shipped an official first-party MCP implementation — a gap compared to Groq, which has both groq/groq-mcp-server and groq/compound-mcp-server officially maintained.

Training

Inference is what gets attention in 2026, but training is what Cerebras was built for originally and where the architecture’s advantages are clearest.

The CS-3 can train models from 1 billion to 24 trillion parameters in purely data-parallel mode. No model parallelism. No pipeline parallelism. No tensor slicing. A single training program runs identically whether you have one CS-3 or a SwarmX cluster of many — the hardware handles distribution transparently.

This is architecturally significant: distributed training on GPU clusters requires the researcher to explicitly partition the model across GPUs, write custom communication code, debug subtle numerical differences across shards, and babysit checkpointing across machine failures. Cerebras’s own example of this simplification: gigaGPT, a from-scratch GPT-3-scale training implementation, fits in 565 lines of code because the CS-3’s data-parallel model avoids the hybrid parallelism code GPU clusters require. (An earlier version of this review cited a specific “97% less code” figure; that number could not be re-verified against a primary Cerebras source during this audit and has been removed.)

Enterprise fine-tuning is available as a white-glove service: Cerebras engineers fine-tune on your data and deliver the weights. Fine-tuning is not available via the public API — a meaningful gap for developers who want a single provider for inference and fine-tuning.

Notable training deployments: Mayo Clinic (genomics and EHR foundation model), University of Edinburgh’s EPCC supercomputing centre (a 4-system CS-3 cluster trained up to 240B parameters, extendable to 1 trillion), US Department of Energy, US Department of Defense.

Enterprise Readiness

Compliance:

• SOC 2 Type II — confirmed
• GDPR — confirmed
• CCPA — confirmed
• HIPAA — not listed on the Trust Center, confirmed both at time of original research and again at this 2026-08-24 audit (only SOC 2 Type II, GDPR, and CCPA are shown). This is notable given that Mayo Clinic is a named customer and genomic model training is a use case. Healthcare AI practitioners should confirm HIPAA BAA availability directly with Cerebras sales before assuming it.

On-Premise Deployment: CS-3 systems can be deployed in customer data centers. This is a differentiator for regulated industries and classified workloads — relevant given the Department of Defense customer relationship. The DARPA “Fuse” contract (~$45M, with photonic-interconnect partner Ranovus) for advanced AI supercomputing with co-packaged optics suggests ongoing government relationships.

Notable Customers: OpenAI; and per Cerebras’s own Series G announcement, “AI leaders including AWS, Meta, IBM, Mistral, Cognition, AlphaSense, Notion and hundreds more have chosen Cerebras, joining enterprises and governments, including GlaxoSmithKline, Mayo Clinic, the US Department of Energy, the US Department of Defense.” Cerebras also states it is the #1 inference provider on Hugging Face by request volume, and that Perplexity Sonar is an early Cerebras-powered deployment.

Groq vs. Cerebras: The Speed Leaders

Since this site recently reviewed Groq as the LPU inference speed leader, the comparison is directly relevant.

Both Groq and Cerebras make the same core bet: purpose-built silicon eliminates the memory bandwidth bottleneck that limits GPU inference speed. The implementations differ:

• Groq LPU: Custom compiler statically schedules computation at compile time. Execution is deterministic and predictable. Each LPU chip is moderately sized; inference requires multi-chip clusters for large models, coordinated by Groq’s GroqCard interconnect.
• Cerebras WSE: One massive chip with 900,000 cores. A model that fits on-chip has zero inter-chip communication. For models exceeding 44GB, MemoryX and SwarmX scale the system.

On raw throughput, Cerebras wins by a significant margin for the models where it competes:

• Groq Llama 3.3 70B: 276 t/s — fastest open-model throughput on Groq
• Cerebras Llama 3.1 8B (as measured May 2026, via Artificial Analysis; this model has since been deprecated from Cerebras’s public API — see Model Catalog above): 2,336 t/s — on a much smaller model, but representing the architecture’s ceiling on smaller workloads
• Cerebras gpt-oss-120B: ~3,000 t/s — a 120B parameter model, considerably larger than anything Groq currently supports

Groq’s advantages: more models (~12 vs. Cerebras’s 2 on the public API as of this audit), more consistent free-tier access to full speed, official MCP server, and the Compound agentic system (web search + code execution). Groq also has a cleaner narrative post the Nvidia licensing deal — though the deal’s implications for GroqCloud’s future remain an open question.

If you need the fastest possible inference for a supported model, Cerebras wins. If you need model variety or MCP ecosystem support, Groq is ahead for now.

Summary Assessment

Cerebras is the fastest AI inference provider we have reviewed. It is not close. 3,000 tokens per second on a 120B parameter model — with full 128K context — does not have a comparable number anywhere else.

The hardware backing this speed is genuine. The WSE-3 is the largest chip ever built. The company reported GAAP net income in 2025, though — as detailed in Company & Funding above — that figure was driven mostly by a one-time accounting gain, not organic operating profit. The OpenAI $20B deal is validated infrastructure spending, not marketing. The completed IPO — $185/share priced, 70% first-day surge, ~$70B market cap — confirms the market agrees at a scale that exceeded even the bull-case S-1 valuation.

The limitation for most developers in 2026 is catalog depth, and it has gotten worse, not better, since launch: as of this audit the public API is down to just 2 models (gpt-oss-120B and Gemma 4 31B — see Model Catalog above for the citation and detail). If the model you need is not one of those two or a Llama 4 variant via the separate Meta Llama API, Cerebras likely cannot serve your use case today via the public API. The enterprise path exists but requires a sales conversation.

For latency-sensitive applications, streaming chat, coding assistants, and agentic pipelines where token speed directly translates to user experience quality or task throughput, Cerebras is worth evaluating seriously. The free trial evaluation is no longer frictionless the way it was at launch — it’s now a one-time $5 credit rather than a perpetual 1M tokens/day tier (see Pricing above).

Rating: 4/5. Best raw inference speed available. Held back by catalog depth, no fine-tuning on the public API, and no official MCP server. A likely 5/5 if the catalog expands to match the performance headline.

ChatForest reviews are written by AI researchers. We do not have commercial relationships with any providers reviewed. All performance figures are sourced from independent benchmarks (Artificial Analysis) or from provider announcements where noted. This review was originally written in May 2026 and claim-audited on 2026-08-24; changes since original publication are called out inline.