Show HN: Gpubook – An order book for GPU compute

jesse_portal1 pts0 comments

gpubook — a compute exchange

📖">

GPU rental is a listing service.<br>It should be an exchange.

You're paying $0.45/hr for a 5090 right now. Someone three listings down is paying $0.28 for the same card from the same region. Neither of you knows. There's no order book. No visible market. Just a wall of listings and a prayer.

If you're a provider, you're guessing what to charge. Set it too high and your GPU sits idle. Too low and you're leaving money on the table. There's no demand signal. No way to sell tomorrow's capacity today. No way to build a reputation that commands a premium.

Meanwhile, compute is becoming the most important commodity in the world. CME is launching compute futures. DRW and Jump Trading are placing bets. Everyone agrees GPU-hours will be traded like oil — but nobody is building the physical exchange where actual machines change hands.

We are.

A gpubook contract represents one GPU-day of compute: a specific GPU class, a specific date, a specific reliability grade. Contracts trade on an order book. The price reflects what the market actually thinks compute is worth — not what a provider guessed when they listed their rig.

Date<br>Bid<br>Ask<br>Spread<br>Vol

May 14 today<br>0.22<br>0.25<br>0.03<br>42

May 15<br>0.33<br>0.35<br>0.02<br>98

May 20<br>0.38<br>0.42<br>0.04<br>45

Jun 01<br>0.30<br>0.35<br>0.05<br>12

That's a 5090-standard-A order book. $/hr by date. Bid, ask, spread, volume. Today's contract is decaying — it's 6pm, there's 6 hours of compute left, and the price reflects that. May 20th is elevated because the market expects a demand spike. You can see all of this. Nobody can today.

Providers earn a reliability rating — like bond grades. AAA means you always deliver, specs always match, uptime is near-perfect. C means you're new and unproven. The market prices in the difference. A 5090-standard-AAA contract costs more than a 5090-standard-C . Nobody sets these prices. The market does.

This means a hobbyist with a 5090 in their basement can sell compute alongside a datacenter operator with a rack of H100s. They don't compete on brand. They compete on price and reliability. The rating is earned from telemetry — continuous, objective, automated. Both buyer and provider run agents. The numbers don't lie.

There are no disputes. No refunds. No customer support tickets. If a provider delivers garbage, their rating drops and their contracts trade at a discount. If they default entirely, their rating craters. The market enforces quality, not the platform.

Three types of people use the exchange.

Providers — you have GPUs sitting idle between jobs. Or you're sourcing cheap compute from Vast and reselling it. Either way, you're leaving money on the table without demand visibility. On gpubook, you list contracts for specific dates at your price. You see what buyers are bidding. You build a rating that lets you charge more than the guy who flakes every third rental. A AAA provider with consistent uptime charges 30-40% more than a C-rated newcomer — and the market sets that premium, not you.

Consumers — you need a 5090 for a training run next Tuesday. Right now you're scrolling Vast hoping the machine you pick isn't garbage. On gpubook, you buy a 5090-standard-A contract for Tuesday at the market rate. You know exactly what you're paying, you know the provider's reliability grade, and if you finish early you sell the remaining hours back. One command to buy, one to SSH in. No wasted compute, no mystery machines.

Traders — you see that H100 demand spikes every December (budget flush) and before every NeurIPS deadline. You buy contracts for those dates now while they're cheap, and sell when consumers need them. You never touch a GPU. You forecast supply and demand and profit from being right. The volatility is real — H100 prices dropped 57% in 18 months and spiked 10% in a single month. This is a tradeable market with no exchange. Until now.

This is what it looks like:

# what's trading?<br>$ gpubook 5090-standard-A

DATE BID ASK SPREAD TIME LEFT<br>May 14 0.22 0.25 0.03 6h 12m<br>May 15 0.33 0.35 0.02 30h 12m<br>May 20 0.38 0.42 0.04 6d 6h

# buy one right now<br>$ gpubook buy 5090-standard-A --market

filled $0.25/hr · 6h remaining · provider aa-rated<br>redeeming...

$ gpubook ssh abc123<br>root@gpu-5090-a7f3:~$ _

Three commands. Browse the book. Buy at market. SSH in. The exchange handles matching, payment, and access. The provider's agent handles provisioning and telemetry. Everything else is the market.

Contracts settle in USDC. Payments are instant — provider gets paid when the contract sells, not when it expires. No escrow. No holdback. If you buy a contract and the provider defaults, you eat the loss and their rating tanks. You knew the rating when you bought. That's the deal.

This isn't a marketplace with extra steps. It's a different thing entirely. Marketplaces have listings. Exchanges have order books. Marketplaces set prices. Exchanges discover them. Marketplaces resolve disputes. Exchanges let the market enforce quality.

Compute will be...

market compute provider gpubook rating exchange

Related Articles