Compute’s 1983 Moment
Second Level Thinking · Issue 46
Who owns the meter for the new oil?
Last week, the world’s leading derivatives marketplace, CME Group announced it’s launching a futures market for compute. What stands to be cash-settled contracts referencing GPU rental prices, in partnership with an outfit called Silicon Data, a company backed by Don Wilson’s DRW. Pending regulatory review, CME Group plan to go live later this year, with their CEO Terry Duffy handing us the line on a silver platter: “compute is the new oil of the 21st century.”
It would be easy to file this as a financial story of another commodity going through its routine derivatives-market moment and move on. But I think there’s something underneath this worth thinking about: Compute is going through the same standards war that every commoditising input has gone through, just compressed into 24 months instead of 25 years. The interesting question isn’t whether compute futures will succeed. They will. The interesting question is who controls the meter underneath them. Currently, there are four very different camps competing for that position right now worth looking at.
On 30 March 1983, NYMEX launched the first crude oil futures contract, a light sweet crude, deliverable in Cushing, Oklahoma. The contract existed because OPEC’s actions in 1973 and 1979 had made oil prices wildly volatile, and downstream consumers, such as, airlines, utilities, refiners and manufacturers had no way to plan. There was no neutral hedging vehicle. Building a refinery was a 30-year capital decision. Buying jet fuel was a daily one, and the gap between the two wreaked havoc.
The contract solved that by creating a price the whole industry could agree on. And it created an enormous power centre in whoever set the reference price. West Texas Intermediate, Brent and the agencies that published those benchmarks — Platts, Argus, S&P Global Commodity Insights — became some of the most consequential institutions in global commerce.
The futures contract is what most people see. The benchmark behind it is where the real power lives. Don Wilson, in last week’s press release, said: “The exponential growth in spending on data centers has been hampered by the lack of a hedging vehicle.”
That sentence is a direct echo of the 1983 NYMEX pitch. Same structural moment. Same problem. Different commodity.
Why now: the data is genuinely staggering
The Big Five hyperscalers — Amazon, Alphabet, Microsoft, Meta, Oracle — have committed roughly $725 billion in capex for 2026. That’s up 64% from 2025 and nearly triple 2024. About 75% of it is AI-specific infrastructure: GPUs, servers, data centres, power, cooling. McKinsey’s published estimate for cumulative AI capex through 2030 is $6.7 trillion.
Microsoft alone has guided to $190 billion for calendar 2026, $25 billion of which the CFO openly attributes to component price inflation. Sundar Pichai’s framing on the Q1 call was that Google remains compute-constrained in the near term. Every hyperscaler CFO is saying the same thing: we are spending unprecedented amounts and we still can’t keep up.
Meanwhile, the underlying price has been violently fragmented. A recent dataset from AIMC Technologies, tracking 141,000+ pricing observations across 24 GPU marketplaces, found NVIDIA H100 hourly rentals listed anywhere from $0.72 to $15.14 within a single 24-hour window. A 21x spread. The history is just as wild: H100s went from $7–$10/hour at launch in 2023 to $2–$4/hour by late 2025, with a 10% spike in four weeks over the Christmas–New Year period. The newer B200 jumped 24% in March 2026 alone, while H100 prices stayed flat.
This is exactly what an unhedgeable, fragmented input looks like just before someone builds a market on top of it.
Matt Levine’s column last week framed the deeper point cleanly. The economics of AI roughly look like this: utilities, real estate developers, chip manufacturers and big tech spend trillions building computing infrastructure, then rent it to small teams of graduate students who met six weeks ago and just raised a billion of venture capital, who then build AI tools that — if everything goes right — generate trillions of dollars of value and pay back the infrastructure.
In aggregate, that’s a fine credit bet. But at the individual data centre level, it’s an awful underwriting problem. The bank wants to know who your anchor tenants will be and you don’t know. The AI companies that will be your customers in 18 months probably don’t exist today. The ones that exist today might be irrelevant by then. You’re building a 30-year asset to serve a tenant base with a half-life measured in months.
A liquid futures market collapses that problem. You don’t have to know who your customer will be, you just need to know there’ll be one, at a price that pays back the loan. That’s the bull case for CME’s product, and it’s a strong one. It’s not just a hedging tool, it’s the missing financing primitive for the entire data centre build-out. Compute futures are what makes the next $5 trillion of AI capex bankable.
The second-level question
But here’s the bit nobody is asking: Whose benchmark wins?
Silicon Data, the index provider, is a private firm backed by DRW that sells pricing data via Bloomberg and API to paying clients. They claim to cover 80%+ of the global H100 rental market today. Their indices will be the underlying for CME’s futures contracts.
This is, structurally, a bid to become the Platts of compute, as there stands to be enormous economic rent in that position. Once an index becomes “the” reference price, the people who own it are doing the most quietly profitable job in the entire value chain. The exchange clears the contracts, the traders take the spread, but the index publisher sets the rules.
If you’ve thought about LIBOR and the years of work it took to migrate everything to SOFR after the manipulation scandals, you already know this story. The benchmark is the load-bearing pillar of the entire derivative complex. Get the benchmark right and the market works, get it wrong and you discover, painfully, that you’ve built a $300 trillion notional structure on top of a number being set by half a dozen traders over instant messenger.
So Silicon Data, in 2026, is making a bid to be the most important reference price in the AI economy. Whether they should be is a question, but the more interesting question is who else is going to fight for that position, because there are at least four distinct camps with very different answers.
Four roads forward
Camp 1: Financial infrastructure. CME, Silicon Data, DRW, BlackRock, ICE if they decide to enter. Their pitch is the one you already know: a centrally-cleared cash-settled futures contract referencing a private licensed benchmark. The benefit is liquidity, regulatory blessing, and the existing plumbing of the derivatives world. This is the camp most aligned with the existing financial system, which is both its strength and its limit.
Camp 2: Hyperscalers. AWS, Microsoft, Google, Meta, Oracle. They control the underlying physical supply, with most of the world’s GPUs sitting inside their data centres. Their natural incentive on benchmark transparency is the opposite of everyone else’s: they’d prefer pricing to stay opaque, because opacity preserves their pricing power against AI labs and enterprise customers. Watch carefully whether the hyperscalers actively support CME’s contract, passively tolerate it, or quietly lobby against it. My guess is mostly the second and occasionally the third.
Camp 3: Neocloud insurgents. CoreWeave, Nebius, IREN, Crusoe, Lambda. These are the specialist GPU clouds that emerged from the 2023-2024 shortage. They are the natural counterparties on a compute futures market — and the natural beneficiaries. A liquid forward curve is what makes their capex bankable, what lets them go public at credible multiples, what eventually lets them be repackaged as cash-flowing assets backed by sold-forward compute. Watch CoreWeave especially. They’re already public, already enormous, and have every incentive to support whatever benchmark emerges.
Camp 4: Decentralised alternatives. This is where it gets interesting. There are now multiple permission-less networks where compute pricing is structurally on-chain and public, not as a marketing choice but as a property of the architecture. Render Network for GPU rendering. Akash for general decentralised cloud. io.net aggregating GPUs on Solana. Gensyn for decentralised ML training. And most importantly, Bittensor, which is now genuinely worth taking seriously. Let’s explore that more deeply.
The Bittensor case
Bittensor is the most credible decentralised counter-move in the compute economy right now. It launched in 2021 and as of May 2026, its market cap sits around $2.6–2.9 billion. It runs 128 active subnets, each functioning as an independent market for a specific AI task: text generation, model training, inference, drug discovery, GPU rental. It’s structurally different from the centralised AI economy in that there is no company. Models compete for token rewards based on output quality, scored by validators.
The thing that makes this more than a thought experiment is what Subnet 3 called Templar did in March 2026. Seventy independent contributors, using commodity internet hardware, permissionlessly trained a 72-billion-parameter large language model called Covenant-72B. It scored 67.1 on MMLU, which is competitive with Meta’s Llama 2 70B, one of the most well-resourced AI labs on the planet. The result was confirmed in an arXiv paper, and Nvidia’s Jensen Huang and Chamath Palihapitiya publicly endorsed Bittensor at the time.
Grayscale already runs a trust (GTAO) listed on the NYSE, with a spot ETF filing pending at the SEC. Digital Currency Group’s Yuma subsidiary is actively contributing to 14 different subnets and this is no longer a fringe experiment.
Why does this matter for the meter question? Because Bittensor’s compute pricing is on-chain by default. Every subnet’s emission curve, every validator’s stake, every miner’s reward is all publicly readable. You don’t need a Bloomberg terminal to see the price of “decentralised LLM training compute” in May 2026. You need a block explorer. That’s not the same thing as Silicon Data’s H100 spot index. The pricing surfaces aren’t directly comparable. But it is, structurally, what an open-protocol meter looks like.
The interesting structural fact is that Bittensor isn’t alone. Render’s GPU rendering market, Akash’s compute prices, io.net’s GPU rates, Gensyn’s training compute — all of these are observable in real time by anyone with a node. Aggregate them and you have an alternative compute-pricing surface that no single entity owns.
The Munger inversion
The Charlie Munger move here is to invert. Don’t ask whether compute futures will succeed — they will, the data demands them. Ask instead: who would be most threatened by widely-observed, transparent compute pricing, and what will they do about it?
The answer comes in two layers.
The first layer is the obvious one: hyperscalers, because transparent pricing erodes their pricing power. Expect them to resist or co-opt rather than lead.
The second layer is more subtle: index publishers themselves. Silicon Data has every commercial reason to prefer a market where its private index is the only credible benchmark. If a decentralised alternative ever becomes a referenced benchmark — even in a small slice of the market — Silicon Data’s economic moat shrinks. So expect a quiet lobbying war over how regulators define “the” reference price for compute futures.
The third-order question is which camp gets the neocloud alliance? CoreWeave and Nebius are the natural counterparties of a futures market. They want price transparency but which kind of transparency — Silicon Data’s, or the on-chain version, or some hybrid?
I genuinely don’t know which way this goes. The gravitational pull of established financial infrastructure is enormous, so you could see Silicon Data becoming the de-facto benchmark for the next decade. Given the rate of change and the size of the constraint, it’s plausible the decentralised market could still offer honourable benefit at hefty valuations. Think Bitcoin today.
What I’m watching
The window for structural decisions in this market is going to close pretty quickly. Eighteen to twenty-four months is my guess. So here’s the things I’m watching:
Whether CME’s contract gets CFTC approval, and how Silicon Data’s index methodology gets disclosed. The more transparent the methodology, the harder it is for an open alternative to claim moral high ground. The less transparent, the larger the opening.
Whether any major neocloud — especially CoreWeave or Nebius — publicly references an open or decentralised compute benchmark alongside the private one. This would be the leading indicator that the camp 3 & 4 alliance is forming.
Whether the financialisation of GPU clusters as cash-flowing assets actually happens. Compute-backed bonds or securitised data centres. If this kicks off, the importance of the underlying benchmark becomes existential.
Whether Bittensor (or one of its peers) develops a genuine bridge to the institutional financial system. A spot ETF approval for TAO, a custodial integration with a tier-1 prime broker, a publicly-traded company using on-chain compute pricing in its accounts. Any of those would mark the moment the decentralised alternative becomes structurally competitive rather than philosophically interesting.
Whether the agentic economy actually shows up at scale. If a meaningful slice of B2B transactions starts running through autonomous agents in the next two years, every settlement-layer assumption gets re-priced. Compute pricing is just one part of a broader question about what kind of financial infrastructure is fit for purpose when the participants are not human.
Compute is going to be one of the defining economic stories of the next decade. How we measure it, price it, hedge it and trade it is currently being answered by a handful of people most of us have never heard of. That’s worth paying attention to, even if you never plan to buy a single futures contract.
The interesting investments, as always, are upstream of where everyone is looking in the present.
Parting Shot
Jens Grede on anxiety:
“The more you tailor your life to avoid situations that make you anxious, the more you get anxious in lesser situations. When you lower the bar, that becomes the new bar and you need to lower it again. And life gets smaller that way.”
