Akash Compute Forwards, an onchain order book for future GPU hours.
It's a design today, nothing trades, but 3 things go live today.
A white paper, with detailed rules, a simulation you can run, and detailed documentation on how to use and how to build on the forwards market.
@gregosuri at Breakpoint today 👇
@joe_kaylem on the Crypto x AI panel at the @UBC_Conference at @UT today.
Thanks to everyone who stopped by the Akash table this weekend and asked such good questions.
Greg sat down with @therollupco at Breakpoint to talk about why buying compute is harder than it should be.
You can rent a GPU for today, but for most GPUs there's no open price for next year.
That's the problem Akash Compute Forwards is meant to solve.
BREAKPOINT: Greg Osuri On Why The Real Bottleneck In AI Is Liquid Compute
@gregosuri@akashnet@andyyy@robbieklages
Timestamps:
00:00 Intro
01:03 Moving To Solana Injected New Community
03:16 Nvidia's Leader Was Open Source's Backstop
05:28 Every Lab Is Building Custom Chips
07:29 40 Of 41 Cities Passed Moratoriums
09:51 Compute Needs Futures Like Other Commodities
11:55 Compute Needs A Public Order Book
13:57 Turning Away 75% Of Customer Demand
Day one at the University Blockchain Conference in Austin.
We spent seven hours straight talking with students, and it felt like a workout in the best way. The questions were sharp and well informed, and one went all the way into how BME works.
Back tomorrow. @UBC_Conference
"This is financial advice. Stock up on compute... Compute will be the most scarce asset in the world."
@akashnet Founder @gregosuri joins SolanaFloor to unpack the newly announced Akash Compute Forwards 👇
Nothing trades yet. Akash Compute Forwards is signing up businesses for its first pilot, planned for early 2027.
Tell us how many B300 nodes and which months they're open.
https://t.co/dOPhPId4wT
Want to see how it works first? Run the simulation:
https://t.co/bRBpPfWx60
Bringing B300s online in 2027? You already know the math.
A GPU hour is perishable. Once it passes, you can't get it back. And 100% utilization is really hard to get.
The day before the month starts, your node gets matched with a buyer.
Validators check it every hour with a challenge only that hardware can answer. Served hours get paid daily. If a node goes down, a stand-in covers the buyer from the next hour.
@punqops It really is. That's the gap we're working on: one public daily rate for a GPU month, set by auction.
Nothing trades yet, but you can see how it works here: https://t.co/BfjYBtrORb
Planning a training run for next summer?
For most GPUs, there's no open rate for a month that far out.
You get a quote on a sales call, and nobody else ever sees it.
Nothing trades yet.
Akash Compute Forwards is in pilot, for businesses that buy GPUs.
Tell us which GPUs, how many and which months.
https://t.co/FohK1jVu4b
Want to see how it works first? Run the simulation:
https://t.co/bRBpPfX4Vy
Then you get the actual compute.
Validators check your node every hour.
If an hour is missed, you don't pay for it, and replacement capacity gets lined up for you.