By CerebrasLearning Objectives
- Understand what Cerebras Inference is and how wafer-scale chip technology works
- Compare Cerebras speed benchmarks to GPU-based and competing inference platforms
- Evaluate Cerebras pricing tiers and the significance of its OpenAI and AWS partnerships
What Is Cerebras Inference?
Cerebras Inference is an AI inference platform built on the Wafer-Scale Engine 3 (WSE-3) — the largest chip ever made. While traditional GPUs use chips the size of a postage stamp, the WSE-3 is the size of an entire silicon wafer: 4 trillion transistors, 900,000 AI-optimized cores, and 44 gigabytes of on-chip SRAM memory.
The result is inference speed that consistently outperforms every alternative. On large models (70 billion+ parameters), Cerebras delivers 2 to 6 times faster token generation than both Groq's LPU and NVIDIA's Blackwell GPUs.
💡Key Concept
Wafer-Scale Engine (WSE): Instead of cutting a silicon wafer into hundreds of individual chips, Cerebras uses the entire wafer as a single processor. This eliminates the communication bottlenecks between separate chips, allowing data to flow across 900,000 cores without leaving the chip.
Major Partnerships (2025-2026)
Cerebras has secured partnerships with three of the biggest names in AI:
- OpenAI (January 2026): A multi-year deal to deploy 750 megawatts of Cerebras wafer-scale systems for OpenAI inference — described as the largest high-speed AI inference deployment in the world, rolling out 2026-2028
- AWS (March 2026): WSE-3 chips coming to Amazon Bedrock, combining AWS Trainium for prefill with Cerebras CS-3 for decode. General availability expected in the second half of 2026
- Meta (April 2025): Powers the Llama API with up to 18 times faster inference than GPU-based solutions
Speed Benchmarks
Cerebras consistently leads inference speed benchmarks, especially on larger models:
| Model | Cerebras Speed | Groq Speed | Speedup |
|---|---|---|---|
| GPT-OSS-120B | ~3,000 tokens/sec | ~493 tokens/sec | ~6x faster |
| Llama 3.1 8B | ~1,800 tokens/sec | ~1,345 tokens/sec | ~1.3x faster |
| Llama 3.1 70B | ~450 tokens/sec | ~275 tokens/sec | ~1.6x faster |
| Qwen3 480B Coder | ~2,000 tokens/sec | Not available | Largest model hosted |
| Llama 4 Maverick | ~2,500+ tokens/sec | Available | ~2.5x faster than NVIDIA flagship |
📝Note
Cerebras's advantage grows with model size. On smaller models (8 billion parameters), the gap narrows. On frontier models (100 billion+), Cerebras pulls significantly ahead because the entire model fits in on-chip SRAM, avoiding the memory bottleneck that slows down GPU-based systems.
Supported Models
As of March 2026, Cerebras hosts a focused selection of major open-source models:
| Tool | Best For |
|---|
Pricing
- 1 million tokens per day
- 8,192 context length
- Higher limits
- Production use
- Coding-focused with discounted per-token rates
- High-volume coding workloads
- Dedicated capacity
- Fine-tuned models
- SLAs
Per-token pricing (approximate):
| Model | Input (per 1 million tokens) | Output (per 1 million tokens) |
|---|---|---|
| Llama 3.1 8B | $0.10 | $0.10 |
| Llama 3.1 70B | $0.60 | $0.60 |
| Llama 3.1 405B | $6.00 | $12.00 |
The free tier offering 1 million tokens per day is one of the most generous in the industry — enough for meaningful experimentation without a credit card.
WSE-3 vs. NVIDIA Blackwell
| Spec | Cerebras WSE-3 | NVIDIA B200 |
|---|---|---|
| Transistors | 4 trillion | 208 billion |
| Cores | 900,000 AI cores | 18,432 CUDA + 576 Tensor |
| On-chip memory | 44 GB SRAM | 192 GB HBM3e |
| AI compute | 125 petaFLOPS | ~4.5 petaFLOPS |
| Best for | Inference (speed leader) | Training + inference (flexibility) |
⚠️Warning
Raw specs do not tell the full story. NVIDIA's ecosystem (CUDA, cuDNN, TensorRT) supports virtually any model and workload. Cerebras excels at inference speed but has a narrower model catalog and does not support custom fine-tuning through its API yet.
💡Key Concept
Cerebras's role in the three-way inference shift. Stratechery's Ben Thompson argued in his May 11, 2026 piece "The Inference Shift" that AI compute is bifurcating into three workload categories that need fundamentally different hardware: training (GPUs win on bandwidth + ecosystem), answer inference (where token speed for human-facing chat matters most), and agentic inference (where humans aren't in the loop and memory capacity + cost-per-token matter more than raw speed). Cerebras's WSE-3 is positioned by Thompson as the canonical "answer inference" play — 21 petabytes per second of on-chip SRAM bandwidth versus 3.35 terabytes per second of HBM on the NVIDIA H100. When the next response in a conversation is what the user is waiting on, that bandwidth gap becomes practical latency. For agentic workloads where the model is making many tool calls without a human watching, the framework predicts cost-optimized memory-heavy hardware — possibly using slower, cheaper DRAM — will out-economize either GPUs or Cerebras-class speed silicon.
Company Details
| Detail | Info |
|---|---|
| Founded | 2016 |
| CEO | Andrew Feldman |
| Headquarters | Sunnyvale, California |
| Employees | ~750-800 |
| Latest Funding | $1 billion Series H (February 2026) |
| Market Cap | $66 billion (post-IPO close) |
| Total Raised | ~$2.9 billion private + $5.5 billion IPO proceeds |
| Key Investors | Tiger Global (lead); Benchmark; Fidelity; AMD; Coatue |
| IPO | Debuted May 14, 2026 — 28 million shares priced at $185 (above $115 to $160 range), $5.5 billion raised, stock more than doubled on debut to close at $311 for $66 billion market cap |
| Notable Customers | OpenAI; AWS; Meta; Group 42; Saudi MBZUAI; Mistral; Perplexity; Mayo Clinic; US Department of Energy |
| Website | cerebras.ai |
Strengths
- Fastest inference on large models — 2 to 6 times faster than Groq and NVIDIA on 70 billion+ parameter models
- Wafer-scale architecture — 4 trillion transistors on a single chip eliminates inter-chip communication bottlenecks
- Major partnerships — OpenAI (750 megawatt deployment), AWS (Bedrock integration), Meta (Llama API)
- Generous free tier — 1 million tokens per day at no cost
- Frontier model support — runs models up to 480 billion parameters (Qwen3 480B Coder)
Limitations and Considerations
- Narrower model catalog — fewer models than Together AI or GPU cloud providers; focused on major open-source models
- No custom fine-tuning via API — you cannot upload or fine-tune your own models (unlike Together AI or AWS SageMaker)
- Inference only — Cerebras Cloud does not offer model training (though on-premise CS-3 systems support training)
- AWS integration not yet live — Bedrock availability announced for H2 2026 but not generally available yet
- Ecosystem maturity — NVIDIA's CUDA ecosystem is vastly more developed; Cerebras is still building out developer tooling
IPO and Public-Market Debut
Cerebras Systems completed the largest US tech IPO of 2026, pricing 28 million shares at $185 — above the $115 to $160 range — and raising roughly $5.5 billion in proceeds. The stock more than doubled on debut Thursday, opening at $385 and closing at $311 for a $66 billion market cap. The S-1 named OpenAI, Group 42, Saudi Arabia's MBZUAI (Mohamed bin Zayed University of Artificial Intelligence), and Amazon Web Services as top customers, and Cerebras swung to profitability on $510 million of 2025 revenue. OpenAI is one of Cerebras's largest customers under a multi-year contract worth more than $10 billion and holds a $1 billion secured loan plus warrants for over 33 million shares — making OpenAI a meaningful post-listing shareholder.
The successful debut validates the thesis that frontier labs will pay a premium for Nvidia alternatives when inference economics work, and gives Cerebras a public-market currency to chase Nvidia's data-center share more aggressively. It also positions Cerebras as the first AI-specialized silicon company to reach a frontier-scale public listing and serves as a practical test of Nvidia's GPU pricing power and frontier-lab compute lock-in. Ben Thompson's "Inference Shift" framework — see the InfoBox above — helps frame the answer-inference workload bet behind investor demand: when the next response in a conversation is what a user is waiting on, the on-chip SRAM bandwidth gap between WSE-3 and HBM-based GPUs becomes practical latency.
Key Takeaways
- Cerebras Inference delivers the fastest AI inference available, powered by the WSE-3 — the world's largest chip with 4 trillion transistors
- Speed advantage grows with model size: 6 times faster than Groq on 120 billion parameter models, with support for models up to 480 billion parameters
- Major 2026 partnerships with OpenAI (750 megawatt deployment) and AWS (Bedrock integration) validate the technology at massive scale
- Free tier offers 1 million tokens per day — ideal for experimentation; the company completed the largest US tech IPO of 2026 on May 14, pricing 28 million shares at $185 (above the $115 to $160 range) for $5.5 billion raised, with the stock more than doubling on debut to close at $311 for a $66 billion market cap and OpenAI a meaningful post-listing shareholder
