Created in 2022, the young Californian shoot Inflection AI announced fundraising of $ 1.3 billion from its historical investors, including Microsoft and a new entrant Nvidia at the end of June. A significant but necessary contribution to help it build its gigantic cluster, including 22,000 H100 accelerators.
Mustafa Suleyman, co-founder, and CEO of Inflection AI, previously worked at Google as vice president of AI products and co-founded DeepMind, which was acquired by the latter in 2014. (credit: Inflection AI)
Start-up and artificial intelligence are words that go very well together… at least for Inflection AI, a young shoot from Palo Alto. The start-up, which is developing, in particular, a generative artificial intelligence technology competing with ChatGPT called pi.ai, has caught the eye of investors. The company announced at the end of June fundraising of $1.3 billion in a round table led by Microsoft, Reid Hoffman (co-founder of LinkedIn), Bill Gates (co-founder of Microsoft), Eric Schmidt (ex-CEO of Google) in which participated an Nvidia hefty newcomer. Taking into account the group’s first round of $225 million from Greylock, Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt, Mike Schroepfer, Demis Hassabis, Will.i.am, Horizons Ventures, and Dragoneer, the overall amount of funds granted reached $1.525 billion.
This contribution will finance the development of pi.ai and a gigantic cluster of 22,000 H100 Nvidia accelerators; when we know the unit price of each of these cards – approximately $30,000 after discount – a quick calculation brings out a bill of $660 million. Knowing that to function, this type of environment also requires other classic bricks, but no less essential (servers, memory, storage, network…).
Expected computing power in the long term of 22 exaflops
With this infrastructure, Inflection AI intends to support the training and deployment of the next generation of large-scale AI models, reaching a theoretical computing power of 22 exaflops over time. This is much more than the two exaflops potentially reached by Aurora, the monstrous supercomputer from Intel and HPE. However, unlike the second, the Inflection AI system still needs to be produced, which does not prevent the company from dreaming a little.
“We estimate that if we included our cluster in the recent TOP500 list of supercomputers, it would be second and close to first place, although it is optimized for AI applications rather than science. The cluster deployment is in progress, and we have already confirmed its performance in the recent MLPerf benchmark”. For the latter, 3,500 Nvidia H100 accelerators were mobilized, achieving the best results for natural language processing (NLP) LLM and BERT tests on a system hosted by CoreWeave, a cloud service provider specializing in GPU-accelerated workloads at scale.