4 Startups Solving the GPU Shortage

The frenzy to construct data centers has been spurred by the increasing demand for AI capabilities. This surge is primarily driven by the need to train and deploy complex AI models like LLMs (Large Language Models). There’s a shortage of crucial components like cooling systems and Nvidia chips, which are in unprecedented demand. It is estimated that the original version of ChatGPT alone required 10,000 GPUs to train. A new data center pops up every three days. To scale and broaden access to compute power, four startups are taking innovative approaches.

1) SF Compute

SF Compute supplies short-term rentals of compute for burst model training. The company maintains clusters of Nvidia’s high-performance H100 GPUs that are used to train AI models.

“There isn’t really a GPU shortage,” says co-founder Evan Conrad, “Theres a shortage of short-term rentals.” Most GPUs are tied up in long-term contracts with large enterprises, leaving smaller AI startups struggling to access necessary compute power. These startups can’t sink millions of dollars into long-term contracts, and only require short bursts of compute power to train their models.

SF Compute fills this gap by providing short-term rentals. The company offers the ability to burst to thousands of GPUs to train models without requiring long-term contracts. SF Compute sells to customers like Princeton University, PlayHT, and phind.

2) FluidStack

FluidStack adopts a decentralized approach to compute supply. The company offers access to 50,000 GPUs from its network of H100 and A100 chips, selling to customers like Character.ai, Replika, and Stanford University.

The company aggregates underutilized capacity from 1,000+ data centers across 50+ countries. FluidStack offers this compute power to startups to train new models or support existing customers. Like SF Compute, FluidStack offers short-term compute rentals with pricing options both by the hour and month.

3) Nanotronics

Nanotronics goes deeper into the supply chain, developing next-generation chip manufacturing facilities. The company’s “Cubefabs” are modular chip manufacturing facilities that can be shipped anywhere and assembled on-site. Cubefabs can be up and running within a year, producing semiconductors anywhere in the world.

The 14-year-old company has raised $162 million from Founders Fund and Intel’s late co-founder Gordon Moore. Its facilities are not designed for making the advanced GPUs from the likes of TSMC, but specific power chips like those used to route power in advanced data centers. Nanotronics is selling its Cubefabs for around $20 million, and aiming to have the first fab functional by 2025.

4) Exowatt

As GPU manufacturing continues to ramp up, eyes are turning toward the next limiting factor of the AI boom: energy. Exowatt, founded in Miami in 2023, is using solar energy to power data centers. Fresh off a $20 million Seed from a16z, Sam Altman, and Atomic, Exowatt’s modular system captures heat energy and stores it in a thermal battery that can be used for up to 24 hours. This differs from traditional solar technology which converts sunlight into electricity directly.

"Unlike traditional solutions that require significant upfront costs and extended setup times, Exowatt's modular system can be deployed rapidly and cost-effectively – and it's available this year," says Exowatt CEO and Co-Founder Hannan Parvizian. Exowatt’s mission is to provide “extremely low-cost energy that advances the capabilities of global AI infrastructure while protecting our planet,” according to Jack Abraham, Managing Partner of Atomic and Exowatt co-founder.

Conclusion

The AI boom has catalyzed significant demand for compute. Silicon Valley is investing in picks and shovels to support the surge, backing innovations in data center infrastructure and energy access. These developments continue the larger trend of increased interest in physical startups and hard tech we’ve seen over the past few years. AI is not just affecting the world of software, it’s resulting in tangible industrial innovation throughout the physical supply chain.