Computable

Computable is a marketplace for GPU compute that lets you buy, sell, and trade capacity by the calendar week. Instead of committing to long-term contracts or scrambling for spot instances when you need them, you bid on specific weeks of H100 GPU time and get predictable pricing for predictable windows. If your needs change and you don't use what you bought, you can sell the hours back through standing buyback quotes. It's a different model from the typical cloud GPU rental, built around the reality that compute demand is spiky and unpredictable for most teams doing AI work. The market is designed for flexibility rather than commitment.

The mechanism is a sealed-bid auction. You specify how many GPU-hours you want, which weeks you need them, and the price per GPU-hour you're willing to pay. Bids stay private until the auction closes, so you're not competing in a transparent bidding war where every price increase triggers another. This pricing model means you set your own value rather than accepting whatever the market is charging that day. Sample data on the site shows rates around $2.10 to $2.32 per GPU-hour, though actual prices depend on supply and demand for each auction window. The sealed format is designed to find market-clearing prices without the chaos of open bidding. You're not trying to outbid someone in real time; you're stating what the capacity is worth to you.

The flexibility here is the real selling point for teams with variable workloads. Traditional GPU contracts lock you in for months at a time. You're paying whether you're using the capacity or not. Spot instances from major cloud providers give you flexibility but disappear when demand spikes, often exactly when you need them most. Computable sits in the middle, giving you the ability to secure capacity for specific future weeks, then resell if your plans change. For teams doing ML training, inference at scale, or any workload where compute needs vary by project phase, this kind of week-level granularity is more aligned with how the work actually happens. Project timelines slip, scope changes, and compute needs shift. This market structure accommodates that reality.

The resale capability is worth emphasizing because it's unusual. If you buy capacity for a week and then your project timeline shifts, you're not stuck with unused hours. Standing buyback quotes let you sell back into the market. This liquidity transforms GPU hours from a sunk cost into something closer to a tradeable asset. For organizations managing multiple projects with uncertain timelines, the ability to reallocate capacity across time without penalty changes the risk calculus around procurement. You can buy conservatively for your likely needs and sell if you overbought, rather than underbooking and hoping spot availability holds. The liquidity is what makes the market work, and the standing quotes are how it's maintained.

Forward contracts add another layer of planning capability. You can lock in capacity for weeks or months ahead at prices you determine today. If you know you'll need significant compute in January but prices might spike by then, you can bid now and secure that capacity at your price. This is futures trading for GPU hours, which sounds exotic but maps directly to how mature markets handle commodities with variable supply and demand. The sealed auction format keeps the bidding from turning into speculation games, but the forward contract structure lets you hedge against price volatility. For teams with predictable project timelines, this forward visibility is valuable.

The current offering focuses on H100 nodes, which are the workhorse for serious AI training and inference workloads. The first auction covers August through January, with bidding closing at the end of July. This is an early-stage market, backed by Y Combinator, Jump Trading, and Coinbase investors, so the liquidity and depth will depend on adoption. If you're currently paying rental marketplace premiums or overcommitting to long contracts because you can't predict your needs precisely, this is worth watching. The team has experience in market structure, which shows in the auction design. The investor mix includes firms that understand both compute and markets.

There's no explicit per-unit pricing published because the whole point is that you set your own price through the auction. You access the platform, review available weeks and current standing quotes, submit sealed bids, and get results when the auction closes. For organizations that can plan their compute needs a few weeks out rather than scrambling for capacity on demand, this model offers cost control and flexibility that traditional GPU procurement doesn't provide. The alternative is either paying for capacity you don't use or getting locked out when supply is tight. Computable is betting that the middle ground is underserved.