We started GPU.ai because the most expensive part of building an AI company in 2026 should not be talking to procurement.
It is, though. Talk to any founder shipping a model right now and they'll tell you the same story: their team is small, their idea is good, and somewhere between week two and week six of any serious experiment, the work stops being about the model and starts being about the GPUs. Which provider has H200 SXM availability this week. Which one is cheaper for the next three months but locks you in for a year. Whose API takes seventeen minutes to spin a node and whose takes forty. Whose support pager actually wakes someone up at 3 AM in your timezone.
This is a procurement problem masquerading as an infrastructure problem. And procurement problems compound on the smallest, fastest teams the hardest — the exact teams that produce most of the interesting work.
What we believe
Three things, plainly:
Compute is the new electricity. It should be sold like electricity — by the second, at the lowest available price, with a meter you can read and a switch you can flip.
That sentence is the entire product. Everything else is consequences.
The aggregation layer is necessary. No single supplier has the right answer for every workload. A frontier-lab-grade training run wants InfiniBand-dense bare metal in a region with cheap power. A latency-sensitive inference workload wants memory bandwidth and proximity to users. A spiky fine-tuning job wants whatever GPU is cheapest in the next ten minutes. These are three different physical answers, and pretending one supplier can be best at all three is how you end up overpaying for the GPU and underdelivering on the product.
Transparency is a feature. You should always be able to see the actual market price for the GPU you're renting. You should always be able to see who you're renting it from. You should always be able to take your workload elsewhere on a Tuesday. We built our pricing engine on the assumption that customers will use it to negotiate against us — and we think that's good, because it forces us to actually be the best price.
Speed is the only moat that matters. Sixty seconds from CLI to SSH is not a vanity benchmark. It's the difference between a team that runs ten experiments a day and a team that runs one. Compounded over a year, that's the difference between a product and an autopsy.
What we're building toward
A version of the world where:
- A two-person team in São Paulo can rent the same Blackwell hours as a research lab in Mountain View, at the same price, with the same deploy time, with no negotiated contract.
- "Multi-region by default" is the boring choice, not the heroic one.
- The phrase "GPU shortage" stops meaning "my supplier doesn't have any" and starts meaning "the global aggregate market is short," which is a much rarer event.
- A startup's first AWS bill is no longer their cap table's biggest risk.
Who's building this
A small team, intentionally. Engineers from Amazon, NovaCore, BitSync. People who have run their own clusters, paid their own AWS bills, and been on the wrong side of a "your reserved capacity is now unavailable" email.
We came out of operating GPU clouds because operating one is the only way to understand what's actually broken about them. We started an aggregator because the answer to "what's broken" was: no single one of these is going to fix it on their own.
Why we're writing this
A few reasons. We've been asked the "why GPU.ai" question more times this quarter than the last three combined, and we owe people a real answer in writing instead of in pitch decks. We want the team we hire next year to know exactly what they're signing up for. And we want the customers who bet on us — especially the ones who bet early — to know that the thing they bought into is the thing we're still building.
If you're reading this and you're building something that needs serious compute, tell us about it. If you're reading this and you want to build the layer with us, we are hiring.
The compute layer is the most important piece of infrastructure of this decade. We intend to build it.
— Ranbir Badwal, Aditya Reddy