Two alumni from Jump Trading and Coinbase have built what might be the most audacious—or perhaps most befuddling—pitch in the current AI infrastructure gold rush: a platform where you trade weekly blocks of GPU compute the way commodities desks trade futures contracts.
Computable, a New York venture that emerged from Y Combinator, claims to be the first platform for trading GPU hours with instant liquidity. The startup launched its debut auction last month on Hacker News, offering H100 nodes spanning August 2026 through January 2027, sold in tidy calendar-week increments. Bidding closed at the end of July.
Whether this represents genuine financial innovation or an overengineered solution hunting for a problem depends largely on how you read the gap between the company's marketing language and its legal disclosures. That gap is not small.
The Mechanics: An Auction With a Twist
Here's how the system works, at least in theory. You bid on whole eight-GPU nodes—contiguous weeks only—specifying your price per GPU-hour down to the cent. Put down a 10% deposit and you're in. Win the auction and you pay exactly what you bid; no uniform clearing price, just straight pay-as-bid settlement. Computable clips a 2% fee off the top.
The unusual wrinkle is what the company calls a "fragmentation charge." Lower bids can beat higher ones if they pack the calendar more efficiently, leaving fewer orphaned weeks. The selection algorithm, laid out in terms of service updated in late July, essentially penalizes anyone who would leave gaps in the inventory grid. Clearing prices get published after the fact.
Winners receive root SSH access to a virtual machine sitting on an entire node for whatever weeks they secured. Use it or lose it—unused capacity expires without refund.
Instant Liquidity, With Asterisks in 8-Point Font
The marketing pitch leans hard on liquidity. "We keep a standing buyback quote posted on every position," the website promises, "so unused capacity turns back into cash the moment you hit sell."
But Section 7.4 of the terms of service paints a different picture. The company maintains "no obligation to repurchase" accepted bids. Any buyback happens "at a price determined by the Company in its sole discretion," and crucially, "no buyback price...guaranteed." You also can't resell your position to anyone else.
So the "instant liquidity" on the landing page morphs into discretionary buyback offers—with no guaranteed price and no obligation to repurchase—buried in the legal copy. For anyone thinking of this as a hedging instrument—rather than a straightforward compute purchase with extra steps—that discrepancy matters quite a bit.
Compute as an Asset Class (Maybe)
Computable isn't alone in trying to financialize GPU access, which tells you something about the state of the market. Kalshi rolled out "Compute Forward Curves" in mid-July, framing them as regulated event contracts for hedging rental costs. A company called Pluto is pursuing CFTC approval for GPU futures. Compute Exchange launched a secondary market for reserve contracts just weeks ago.
The underlying reality is messy pricing and unpredictable availability. On-demand H100 rates swing wildly depending on provider: around $3.85 per GPU-hour at the low end (Nebius), north of $12 on Azure, according to aggregator data from earlier this year. CoreWeave was listing on-demand HGX H100s at $6.50 hourly in late July, with volume discounts available. Pluto's whitepaper claimed prices had dropped from roughly $8 per hour to near $2 in recent years, though marketplace realities vary.
For a startup planning a large training run months out, locking specific weeks at a known price carries obvious appeal. The trickier question is whether a two-person team—however pedigreed—can build and maintain the market-making infrastructure to deliver on the liquidity promises, especially the secondary market.
What We Don't Know
Computable has disclosed its founders' previous employers but not their names. The team numbers just two people, according to YC materials. There's no public information on node count, geographic region, interconnect specifications, or how frequently those buyback quotes actually refresh—if they do at all.
A commenter on Hacker News raised questions about potential CFTC entanglements, citing earlier attempts at compute trading. The company hasn't publicly addressed its regulatory posture.
The first auction results will clarify some things: clearing prices, bid volume, whether the liquidity mechanism works as advertised or remains theoretical. Until then, Computable is essentially a hypothesis—that compute scheduling is painful enough, and price volatility real enough, that weekly-granularity trading makes sense even without proven secondary liquidity.
For now, what exists is this: an eight-GPU node sold by the calendar week, accessed via SSH, carrying a 2% platform fee and a liquidity promise that occupies an uneasy space between the marketing copy and the terms of service. Whether that evolves into a genuine marketplace or a footnote in the AI infrastructure boom remains to be seen.
