quant research @TradeADMIS, @ucberkeleymet assoc president 🌱 lets talk ai, energy, compute, and markets (all opinions are my own; not investment advice)
San Francisco, CA
Joined February 2016
- Tweets515
- Following1.6K
- Followers1.6K
- Likes2.8K
Pinned Tweet
For those wondering why an AI company needs a power trading desk, I wrote about this shift back in December: as compute scales, managing electricity exposure becomes increasingly tied to protecting the economics and utilization of the underlying GPU fleet. (link in comments)
3 years of Power Bytes and we can say @TradeADMIS the FCM for Compute!
Commodity brokers are getting ready to trade AI compute futures ahead of the CFTC decision
ADM Investor Services, BGC Group and Nanhua USA are setting up to execute orders for hedge funds, traders and corporates, per Bloomberg.
CME plans two cash-settled contracts priced off Silicon Data's GPU rental benchmarks, which track hourly rates for Nvidia H100 and B200 chips. Each covers 730 GPU-hours a month.
The CFTC took public comment on compute futures from August, with a November 9 review deadline tied to CME's launch.
ICE is developing contracts on the Ornn and NativX indices. Nodal Exchange has partnered with Compute Desk.
Compute needs a hedge market for the same reason power did: buyers are signing multi-year capacity deals
Read more: bloomberg.com/news/articles/…
Aishwarya Mahesh retweeted
This is a brilliant thread of comments per usual but one of the main questions I have is this would structurally positive for compute demand but negative for compute prices near-term, no?
I’d assume Labs generate the most Revenue per GW of anyone today. If I use Ant as the leading example at $77B ARR by call it 4GW (they’re getting to 5 by YE26 is SA estimate) = $20B blended or so. Assuming 40% inference means they’re monetizing at $50bn avg across the 2GW of inference.
Elon and others have talked about prices that imply $30-$50B of compute. This means whoever is purchasing that compute needs to make at least $50B but in reality $62.5B assuming 80% API margins on that inference.
So if labs continue to remain the best monetizers of compute and they structurally start to be able to pay less, frontier/scarcity compute prices would need to move down? I’m sure we could find a natural equilibrium where compute, labs, etc can make acceptable returns as Gavin is arguing but it does mean one or multiple need to be true: i) compute prices have to go down from the high end we’re seeing in some illustrative quotes ii) labs allocate more compute to inference while pricing stays flat, or iii) monetization per GW needs to move up.
I’m sure there’s plenty of demand as prices compress over time but seems like that means scarcity pricing seen in market could move lower which might cool near-term irrational behavior we’ve seen in compute markets.
All of this could mean longer and healthier cycle vs traditional boom/bust but it makes me think that also all the rapid expansions + price increases slow down a bit too as prices rationalize throughout the value chain?
Separately, I hadn’t considered NVDA or another party stepping in to fund training but it’s an obvious point in hindsight! Good insights
1) Trying again.
TLDR: Everyone can win. Open-source, frontier labs, infrastructure, applications and most importantly the American people.
I understand all of the jitters about Anthropic losing share, real rates, spreads and second derivatives. Still think the most likely future outcome is a hybrid, multi-model future where almost everyone wins, open-weight tokens are likely the overwhelming majority of tokens, the dominant frontier model companies still do well albeit perhaps not the outcome that seemed possible in May and AI infrastructure/applications benefit immensely from the increased fragmentation, competition and lower margins at the model layer.
Lower margins at the model layer mechanically drive higher compute demand as long as someone is willing to pay for the training. And Jensen is obviously willing to be the “trainer of last resort.” As below-market compute contracts signed in 24/25 roll-off and reprice higher, cloud and neocloud revenue should continue to accelerate which will help address financing fears; I think we may be near peak external financing for this buildout. Noteworthy that using data from 3Fourteen Research, B200 clusters will generate a 37% IRR *if* their rental price declines 40% per year. Prices are still going up, but that IRR math is too ridiculous to print. 37% IRRs available at scale should get funded.
Re: this weekends controversy about MOUs, Anthropic’s $413 billion of real contractual commitments sounds scary. Reality is that it works out to roughly $31 billion of annual cloud spend, which is easily internally financed given their reported ARR, and $161 billion of spend at Broadcom which will be financed by Broadcom and private credit.
Continue to believe fragmentation at the model layer is the best outcome for ordinary Americans and humans everywhere. As I have said before, do not trust any 5-20 individuals to make the best possible decisions about the direction of this technology. I want as many AIs as possible to maximize the odds that they encompass the diversity of human values across the world and richness of human variation. And I want an AI that shares my own personal values. Safety, freedom and the maximization of human agency through democratization, distribution and diffusion vs. centralization. History has spoken on the dangers of the latter path.
Finally, from my perspective anything that slows this buildout down is net positive over the long-term as the biggest risk and worst outcome is an *actual* overbuild. We are structurally short watts, wafers and increasingly capital. That is all good. God forbid we were still in ZIRP, TSM ramped capex to 3x their current plan and we weren’t short turbine blades: there would almost certainly be an overbuild and a bubble. That would be terrible. Slower and smoother for longer is better both for investors and for the world. A terrible cyber event is the most plausible near term bear case IMHO, but AI cyberdefense ramping quickly and I was happy to hear Sholto say that cyber probably becomes defense dominant in the near future.
Longer, threaded version follows:
Aishwarya Mahesh retweeted
aishwarya points out that high spend in a compute market doesn't mean high spend in a derivative compute market.
i think this boils down to a couple things, aside from what she argued (that long term contracts *are* sometimes the best hedge):
1. the high cost of compute goes beyond the resulting GPU hours in a single multi-year GPU lease. it needs to take into account the cost of capital of the hardware, upfront and recurring costs of colocation, development and purchase of power and power generation assets, and the rising costs of each of those as scarcity prices and reprices each asset, which aren't necessarily (or immediately) reflected in the "market cost" of a GPU hour. this gets more and more complex as agreements and terms become more bespoke and geopolitical and regulatory risks make each deal more complex down to the dirt itself. there are few things that will solve this complexity but its highly unlikely a standard futures market solves it automatically. the "transaction cost" as aishwarya pointed out goes up significantly but so do the coordination problems - the cost of *not* incurring those transaction costs also goes up
2. aside from companies that have a long credit history (of which there are few), high spend in compute currently mostly comes from venture equity and venture debt. being able to buy compute often means raising more venture capital. if/when that slows down, even just because of a different capital season or falters in consensus, the capital for derivatives markets also dry up accordingly, unless the futures market quickly matures into enabling real financing products (startups are not great customers of hedging products!)
Aishwarya Mahesh retweeted
1/ Nice piece by @ash_locked on the future(s) of compute.
I agree with the central point: listing a futures contract doesn’t magically create a commodity market.
You need credible benchmarks, commercial hedgers and a connection to the physical market.
A few thoughts:
2/ Compute is not one homogeneous product.
An H100 hour depends on cluster size, networking, geography, uptime, contract length and more.
But commodities have never required perfect fungibility.
Oil has grades. Freight has routes.
3/ The solution is a trusted reference price—and a way to price the differences around it (basis).
It is a benchmark that represents a clearly defined exposure and is actually useful to buyers, sellers and financiers.
4/ On hedging demand, I’d make an important distinction.
A buyer with a fixed-price, multi-year capacity contract has largely locked in that contract’s rental rate.
But the ecosystem around that contract has not eliminated its economic risks.
5/ Neoclouds still face utilization, renewal, residual-value and refinancing risks.
Lenders face collateral and credit risk. Buyers care about performance and availability.
GPU rental futures won’t hedge all of these—but the ecosystem still needs risk infrastructure.
6/ On benchmarks: “How much executed volume backs the index?” is a fair question, but volume alone is not a quality score.
In a fragmented, thinly traded market, transactions can cluster in a narrow segment and fail to represent where the broader market is priced.
7/ A useful benchmark needs broad, diverse observations; normalization for contract terms and specs; transparent rules; independent governance; and tests of how well it tracks real commercial exposure.
More prints aren’t automatically better if you’re measuring the wrong thing.
8/ The piece cites low correlation between two H100 indices.
That is not, by itself, proof that either is wrong.
Different source universes and methodologies can measure different exposures.
The question is: which benchmark actually reduces basis risk for the hedge user?
9/ I’d elevate verification from a technical detail to a core market function.
“8 H100s” doesn’t tell you how a workload will perform.
Interconnect, design, uptime and throughput matter.
That’s why we’re building SiliconMark alongside Silicon Data’s price indices.
10/ Cash-settled futures aren’t physical procurement.
They may hedge a price move, but can’t guarantee GPUs, configuration or delivery dates.
That takes verified supply, market access and secondary liquidity—part of what we’re building at @computeexchange.
11/ I see compute as a stack:
Physical trade and execution → independent price discovery → performance verification → forward curves and residual values → derivatives and financing.
These layers reinforce each other. Neutrality and governance are critical.
12/ What should we watch? Not launch-day headlines.
Look for recurring commercial hedgers, open interest rolling across expiries, activity further out the curve, benchmark robustness and whether real buyers/lenders use these references in decisions.
13/ The opportunity isn’t just “GPU futures.”
It’s making compute measurable, priceable, financeable and hedgeable as an asset class.
Great to see serious market-structure thinking here.
aishwaryamahesh.substack.com…
Oil, freight power, and even cocoa have solved a lot of this before. What's unique to compute is that a GPU-hour can't be stored and the chips keep aging.
Based on that this is how I'd draw the market:
Before I knew about the AI grid announcement, I knew @amppublic reminded me of ISO-like entity. Will be interesting to watch how they frame their role in this market. @AnjneyMidha is cooking
Aishwarya Mahesh retweeted
There isn’t yet a single fungible unit of compute. An “H100 GPU-hour” tells you the hardware and duration, but cluster size, networking, location and contract terms can materially change what that capacity is worth.
@ash_locked does an excellent job explaining how this carries through to the indices themselves:
>Silicon Data adjusts pricing observations for rental term, cluster size, interconnect and geography. Ornn uses executed on-demand rentals and excludes reserved and long-term contracts. They measure different parts of the market, even when both reference H100s.
>She cites a study finding just 0.17 weekly correlation between their H100 indices over the overlapping sample. The benchmark you choose becomes part of the exposure you’re taking.
>A futures contract can trade actively and still provide a poor hedge if its index doesn’t move with your actual procurement costs or rental revenue. Differences between your capacity and the benchmark create basis risk.
>Total compute spending also overstates the potential hedge market. Buyers with fixed-price capacity agreements have already locked in that price, while cash-settled futures won’t solve a shortage of the configuration they need.
Her comparisons with oil, freight and power are particularly useful. Those markets developed workable benchmarks around heterogeneous products. Compute can follow that path, but still has to prove that its references track commercial exposures well enough to attract repeat use.
Really enjoyed this piece! She brings commodity-market experience to a discussion that needs more of it, and gives readers concrete ways to judge whether compute futures are becoming useful. Well worth reading.
circling back
aishwaryamahesh.substack.com…
When your X feed is all compute and commodities, it's time to put that Series 3 to use.
My latest Power Bytes, Future(s) of Compute is my take on whether compute futures can do what commodity markets were designed to do. 🧵
🤖 Made with AI
Shoutout @friedmandave @ShanuMathew93 @hrojantorse and other great industry friends for their feedback and support!
After launch, I'm watching whether these five functions actually show up. Until then, we are still in the early innings of compute markets …
Starting with all the benchmarks! Ornn, Silicon Data, SemiAnalysis and Compute Desk each make a different, defensible claim on what to measure.
Two H100 indices showed a 0.17 weekly correlation. Neither are wrong, but rather they capture different slices of the market
ALT Fernández-Fuertes & Grégoire, 2026
First, the commodities ABCDs. A corn farmer and an ethanol plant could haggle every term from scratch.
Instead, markets standardize what's cheap to measure (grade, size, delivery) into a benchmark, price the rest as basis, and use futures to hedge the price risk in between.
really great work from a great friend! power systems innovation at the most upstream layer! :)
there’s many ways to optimize intelligence per watt — even outside of GPU architecture or improving data center power efficiency (PUE). we ran a fun experiment to see how @crusoeAI’s inference stack itself can save energy a layer above the silicon, and the results were pretty cool!
crusoe.ai/resources/blog/how…
