Cerebras vs Nvidia
Jan-Erik Asplund
TL;DR: After UAE regulatory scrutiny stalled its IPO, Cerebras raised $1B and is diversifying revenue by selling cloud inference alongside hardware—now powering AI coding tools like Windsurf and Cognition's Devin with 950+ tokens/second speeds that are 14x faster than frontier models. For more, check out our full report and dataset on Cerebras.
We last covered Cerebras in September 2024 as the company was retooling away from scientific deep learning and looking to challenge Nvidia for LLM and AI model training use cases.
Since our last coverage, Cerebras postponed its IPO following regulatory questions about its UAE revenue concentration (G42 accounted for ~83% of 2023 revenue) and instead raised $1B from Fidelity and Tiger Global.
Key points from our 2025 update:
- Cerebras's single-chip architecture loads entire AI models onto one chip with ~44GB of onboard memory, delivering inference ~10x faster than traditional multi-GPU setups—the company historically sold these as $2M hardware units to national labs, but launched a cloud inference API in summer 2024 that has since become the primary growth driver, serving customers like Perplexity, Notion, Windsurf, and Cognition on a pay-per-token basis.
- Where selling $2M chips to national labs like Argonne and Livermore limited Cerebras to a small pool of state-backed research customers with long sales cycles and high revenue concentration, selling inference via API creates higher velocity and opens up the much larger market of AI startups and enterprises that want speed without buying and operating specialized hardware.
- AI coding tools like Cognition and Windsurf are using Cerebras to run fine-tuned open-source models at 950+ tokens per second (13x faster than Sonnet 4.5 while achieving parity on coding tasks) letting them improve gross margins by routing specific problems to cheaper, faster models instead of paying frontier model prices for every inference call.
For more, check out this other research from our platform:
- Cerebras
- Groq
- Will Bryk, CEO of Exa, on building search for AI agents
- Kyle Corbitt, CEO of OpenPipe, on the future of fine-tuning LLMs
- Together AI: the $44M/year Vercel of generative AI
- Anthropic (dataset)
- OpenAI (dataset)
- Scale (dataset)
- Databricks (dataset)
- Hugging Face (dataset)
- OpenAI vs. Anthropic vs. Cohere
