Nodus Compute launches from YC with a plan to raise GPU use from five percent

The AI buildout is funded with debt, and the cards it pays for sit idle most of the time. Nodus Compute launched on Y Combinator this week with that claim as the opening of its post. The YC page says GPU clusters run at an average of five percent use. It says the sector has taken on more than 500 billion dollars of debt for new hardware. The post had 4 votes in its first hours.

What the product does

Nodus hosts many AI jobs on the same pool of GPUs. The YC post lists the jobs by name: training, reinforcement learning, evaluation, inference, and agent sandboxes. Nodus says a single card can hold several models at once. It says the card can switch between them in near real time. It says it fills the gaps with batch work. The post includes a demonstration of one H100. The card pretrains a language model while a game of Super Mario plays on the same hardware. The company says it uses AI agents to rewrite GPU kernel code so the jobs pack tighter.

Who runs it and who uses it

The founders are Viswa and Pranav Varshney. The YC post says Viswa was an engineer at Citadel Securities and at Snowflake. Pranav built infrastructure at AWS, Google and Uber. Both worked at Scale AI. The YC post says more than 200 teams have signed up since the company started taking customers last month. The post does not name any of them. It gives no revenue figure and no round size. It does not name a lead investor.

What stays unsaid

The five percent average use rate is a figure Nodus states on its own page. Reveneau did not check it against a third party. The 500 billion dollar debt figure is from the same post. The 200 team figure is the only traction number Nodus offers. The YC post does not say how many of those teams pay. It does not say how much compute they run. The Super Mario demonstration proves one card can run two jobs at once. It does not say what the throughput cost is when the two jobs compete for memory.

What this changes for a team building software: a buyer of GPU time now has a vendor with a different sales pitch. Most GPU vendors sell capacity by the card and by the hour. Nodus sells use rate. A buyer should ask Nodus for one number. It is the average use rate on a live cluster the vendor runs. If the average is still five percent on a Nodus cluster, the pitch does not survive the pilot. If the average is higher, the saving is the gap, measured in the buyer's own spend.

Source

• Nodus Compute launch page, Y Combinator
• Nodus Compute website