1/ People in AI still seem to think GPUs can’t support a liquid futures market because most large, long-term compute contracts are bespoke.
This fundamentally misunderstands how commodities markets work. Plenty of mature commodities trade through customized physical contracts alongside standardized futures.
The key is having a benchmark that captures enough of the underlying price exposure to make hedging useful.
A quick thread on this…
3/ This is where standardized benchmarks become useful.
Our compute indices aggregate executed on-demand rental prices by GPU model, across providers and configurations. ICE’s H100 and B200 futures settle in cash against monthly averages of those indices.
A cloud provider with 70% of its fleet under bespoke five-year reservations and 30% sold on demand could:
- Short futures to hedge on-demand revenue
- Go long to hedge future capacity purchases
- Trade calendar spreads as reservations expire
- Trade H100/B200 spreads to manage relative price exposure
None requires standardizing its physical contracts
4/ This is all pretty familiar territory for power traders. In my conversations, they’ve been by far the quickest to grasp the dynamics of compute markets (by far)
This makes sense given power routinely involves managing bespoke exposures against standardized futures, while pricing and trading the remaining locational, shape, and tenor risks
I suspect there are some sharp power folks out there who will a whole lot of money in compute over the next few years, while the rest of the industry catches up
People tend to assume that the more GPUs you rent, the cheaper each GPU gets, but that's not really how the market works
Prices often bottom out around 32–64 GPUs, then rise as larger, interconnected clusters become harder to find and more expensive to operate.
At a certain size (~2,000 GPUs+), network designs even require an additional layer of switches, which can create massive incremental cost
Cluster size deserves a lot more attention in compute pricing negotiations
Charlie Sheiner retweeted
Given the current state of AI financing, I wanted to make a simple live tracker for the state of the AI money machine.
Check it out at compute.arinjaff.com. If I missed anything there's a submission form per-company, feel free to contribute, will be updating this live!
Charlie Sheiner retweeted
Yesterday, the @CFTC held its inaugural Frontier Forum on super intelligence where attendees discussed the growing role SI has to play in our derivatives markets and the future of agentic trading.
Thank you to all the panelists, moderators, and guests who contributed to this insightful and productive event.
Charlie Sheiner retweeted
Blackstone’s Jon Gray says neoclouds that can get to the other side will find compute pricing up 50% and legacy chips worth more than people thought, and the market is confusing deliverability with demand
“I think what we’re seeing in the market is a little bit of a conflation between deliverability of compute and demand for compute. And at our companies, as an example, we’ve seen a 20-fold increase in spend on models over the last 12 months. It’s pretty extraordinary."
"We have a joint venture with Anthropic to deploy these models. The company’s called Ode, and it can’t keep up with the business that’s coming in the door. So we’re seeing demand really growing for compute."
"And at the same time, the constraint, of course, is delivering that supply. And that’s hard because of amassing capital, data centers, power, chips, all of that."
"And so for the neoclouds, if they can get to the other side, the pricing of compute’s up 50%. Their legacy chips are worth much more than people thought. Their data centers are leased below market."
"Their challenge is, do they have the capital, the debt and equity and the quantum that’s needed to deliver?"
"I think some of these companies have very talented CEOs. CoreWeave’s a good example of this. They’ll find a way through it, but that’s the challenge they face. The underlying business is actually quite strong.”
Gavin Baker says a trillion dollars of AI capex on a six-to-nine month cycle is a lot of wasted capital, and Jensen Huang responds that DSX pods could cut deployment from months to weeks, which is "game-changing" for AI factories
BAKER: “I do think we should be, you guys touched on it a little bit earlier, but the DSX pod, I actually think it’s pretty important here. If we think, let’s use a trillion dollars as a nice round number of AI CapEx, seems like we’re going to go much higher very soon, but a trillion dollars, if it’s a six-to-nine month cycle time, that’s a lot of wasted capital.”
HUANG: “That’s right.”
BAKER: “It’s unproductive. So maybe talk about the DSX pod and how that changes.”
HUANG: “The moment it gets shipped, somebody has to pay for it. But the moment then you have to install these AI factories and they’re gigantic. And putting these AI factories in the service takes months. If we could reduce that months, you know, from the moment that the systems are shipped to the moment that it’s deployed and being operational, if we can reduce it from months to weeks or even a month, that will be completely game-changing for all of these AI factories. Having a trillion dollars, having a few hundred billion dollars sitting idle while you’re waiting for revenues is obviously a challenge.”
BOSA: “And can give your customers a little bit more control over that sort of very pressing, very important financing question.”
GPUs can be extremely valuable and cloud providers can still go bankrupt, not a contradiction.
If delay penalties pile up and the provider runs out of cash, lenders could end up with $11 per GPU-hr B300s much sooner than expected...
ornncompute.substack.com/p/a…
Charlie Sheiner retweeted
We are happy to announce that the contract specifications for our @ICE_Markets compute futures are now available.
Find more information here: ice.com/publicdocs/futures_u…
Charlie Sheiner retweeted
Really interesting.
Ornn is the only compute exchange delivering capacity.
I asked people: "where should neolabs go for GPUs?"
15 votes: SF Compute (@evanjconrad @jeztek @Ethan_is_online @EricMenees @apagajewski)
11 votes: CoreWeave (@BrianVenturo @branninmcbee)
10 votes: Modal (@bernhardsson @akshat_b)
10 votes: Nebius (@romanchernin)
9 votes: Lambda (@stephenbalaban)
5 votes: Fluidstack (@garywu @CesarMaklary @jimcox25)
4 votes: RunPod (@flash__singh)
3 votes: Crusoe (@ChaseLochmiller @Electron_Cowboy) (also received one vote against)
3 votes: Together AI (@vipulved @ce_zhang @tri_dao @percyliang)
2 votes: Baseten (@tuhinone @amiruci @saltyph @defpan)
2 votes: Google Cloud (@sundarpichai)
2 votes: Oracle (@ClayMagouyrk)
2 votes: Ornn (@bavaria_kush @wayne_nelmz)
People couldn't pick a company they work at or founded. Disclosure: I'm a small investor in SF Compute. I didn't vote.
Charlie Sheiner retweeted
An NVIDIA GPU is a high-yield asset.
An H100, bought on the secondary market in Sep 2025 for ~$19k:
Resale today: ~$20.5k
Net rent earned: ~$13.5k
Total: ~$34,000, +76% in 1 year.
A B300, bought in Jul 2026 for ~$54k:
Resale today: ~$51k
Net rent earned: ~$14k
Total: ~$65,500, +21% in under 3 months
What allows these assets continue to earn? Resale value has held as rental rates have increased and demand has grown across all hardware generations.
Resale rates have held within −5% to +7%, while Ornn's H100 index is $2.91/hr, +49% from a year ago, and our new B300 index tracks at $11.32/hr, up 66% from June. H100 marketplace utilization is 87% today. B300 is at 90%.
Ornn's forward curves predicts that in 2 years, by Sep 2028:
H100 ~$63,500 (+227%)
B300 ~$186,000 (+244%)
NVIDIA Hopper and Blackwell hardware both hold incredible residual value.
Charlie Sheiner retweeted
NVIDIA compute is fungible, durable and highly rentable. It is a productive, revenue-generating asset.
