GPUs generate tokens one at a time, and each token waits on weights moving out of HBM. Cerebras, Groq and SambaNova remove most of that wait by keeping weights on the chip, in SRAM. All three sell speed. In 2026 they also took three very different corporate paths: an IPO, a licence to Nvidia, and a partnership with Intel.
Key Takeaways
- Cerebras — Wafer-scale chip, 44 GB SRAM per CS-3, about 2,700 tokens per second on GPT-OSS-120B. Public since May 2026 and contracted for 750 MW by OpenAI. The speed and scale leader.
- Groq — Compiler-scheduled LPU behind GroqCloud, $0.15 per million input tokens on GPT-OSS-120B. Technology licensed to Nvidia in December 2025; the founders left for Nvidia.
- SambaNova — SN40L dataflow chip with SRAM, HBM and DDR tiers, so one rack holds many models. The on-premises pick, backed by a $350M round and an Intel partnership in 2026.
- Bottom line — Use Cerebras for the fastest API, Groq for the cheapest fast API, SambaNova when the model has to run on your own hardware.
Side by side
| Feature | Cerebras | Groq | SambaNova |
|---|---|---|---|
| Architecture | |||
| Chip | Wafer-Scale Engine 3 in the CS-3 | LPU; Nvidia Groq 3 LPX rack in 2H 2026 | SN40L reconfigurable dataflow unit |
| Where weights live | On-chip SRAM, 44 GB per CS-3 | On-chip SRAM across many LPUs | Three tiers: SRAM, HBM, DDR |
| Many models per system | Large models span several systems | Large models span many chips | Designed to hold and swap many models |
| Performance and price | |||
| Public speed figure | ~2,700 tok/s on GPT-OSS-120B (Artificial Analysis) | Among the fastest on public charts | Fast; fewer public head-to-head figures |
| GPT-OSS-120B list price (per 1M tokens, Jan 2026) | $0.25 input / $0.69 output | $0.15 input / $0.75 output | Per-token on SambaNova Cloud; enterprise quotes |
| Buying | |||
| Access | Cloud API, dedicated capacity, on premises | GroqCloud API; LPX racks via Nvidia | Cloud API, dedicated and on premises |
| API style | OpenAI-compatible | OpenAI-compatible | OpenAI-compatible |
| Company status (Oct 2026) | Nasdaq: CBRS since May 2026; $10B OpenAI deal | Independent under CEO Simon Edwards; founders at Nvidia | Private; $350M round Feb 2026 after Intel talks stalled |
Verified 2026-10-31. Prices are list prices reported by The Next Platform in January 2026.
Cerebras: the speed and scale leader
Cerebras builds one chip from a whole silicon wafer. A CS-3 holds 44 GB of SRAM and is rated at 12.5 FP16 petaflops dense. OpenAI signed a $10 billion contract in January 2026 for 750 MW of Cerebras capacity through 2028, and the company listed on Nasdaq in May 2026, raising about $6.38 billion. The risk has moved from survival to concentration: one customer now anchors the growth story.
Groq: the cheap fast API, with an Nvidia question
Groq's LPU schedules every operation at compile time, which gives predictable latency. GroqCloud lists GPT-OSS-120B at $0.15 per million input tokens. In December 2025 Nvidia licensed the technology and hired the founders; Nvidia now ships its own Groq 3 LPX rack. GroqCloud keeps running, but the long-term hardware roadmap belongs to Nvidia.
SambaNova: the on-premises option
SambaNova's SN40L adds HBM and DDR tiers behind its SRAM, so one system can hold many large models and switch between them. That suits enterprises that want several models on their own hardware. Intel's acquisition talks stalled; in February 2026 SambaNova raised $350 million led by Vista Equity Partners and signed an Intel partnership.
Verdict
- Fastest tokens for agents and interactive products: Cerebras.
- Cheapest fast API on popular open-weight models: Groq, with a contract that survives a roadmap change.
- Fast inference on your own premises, many models per rack: SambaNova.
- Training, or any model these providers do not host: none of them; use GPUs or TPUs.
Related
- AI chip companies 2026: all eight Nvidia competitors scored.
- Google TPU vs AWS Trainium vs Nvidia GPU: the training platforms.
- AI inference platforms 2026: the serving layer above the chips.
- TPU vs GPU on our partner site CTAIO.
Frequently asked questions
Is Cerebras faster than Groq?
On the most-cited public benchmark of an open-weight model, yes: Artificial Analysis measured about 2,700 output tokens per second for GPT-OSS-120B on Cerebras' CS-3. Groq is also among the fastest providers. Speed varies by model and changes as both update their deployments, so check current figures for the model you plan to use.
What is the difference between an LPU and a GPU?
Groq's LPU (language processing unit) is a chip whose every operation is scheduled by the compiler ahead of time, with model weights in on-chip SRAM. A GPU schedules work dynamically and keeps weights in external HBM. The LPU gives lower, more predictable latency per token; the GPU is far more flexible and runs training as well as inference.
Did Nvidia buy Groq?
Not outright. In December 2025 Nvidia took a non-exclusive licence to Groq's inference technology, in a deal widely reported at about $20 billion, and Groq's founder Jonathan Ross and president Sunny Madra joined Nvidia. Groq remains independent under CEO Simon Edwards and still runs GroqCloud.
Is Cerebras better than Nvidia?
For single-user output speed on the models it hosts, Cerebras is faster than GPU-based services. For training, broad model choice and cost per token at high batch sizes, Nvidia GPUs remain the default. Most buyers use Cerebras as an inference tier next to GPUs, not instead of them.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.