We’ve been investing time over the past few weeks learning about compute markets, the players in the space, and where innovation is heading.
Compute is big.
Data center capex crosses $1 trillion in 2026, roughly double what the big four hyperscalers spent in 2025. Jensen calls for $3 - 4T a year by the end of the decade, and maybe $10T by 2031.
As a share of the economy, AI investment is ~0.9% of world GDP today heading to ~1.4% by 2028, already past the ~1% telecom peak of the dot-com era, with 1880s railroads (~6% of US GDP) as the only bigger buildout in history.
Here’s the part almost nobody prices in: Fable 5 and GPT-5.6 were each trained on well under 2GW ($120B). Dario’s framing says it best: the 2023 model cost $100M and returned $200M; the 2024 model cost $1B and returned $2B; the 2025 model cost ~$10B and returned ~$40B, and the 2026 models cost ~$25–30B a year of training and, on this cadence, return $100B+.
Each model pays for itself. The labs only lose money because they’re always training the 10x bigger next one.
You can imagine what happens when training runs go from hundreds of megawatts to multiple gigawatts, and when every country and company decides it wants its own.
Pair that with the current trend: Sovereign AI. Why pay Anthropic $50 per million output tokens when you can rent a slice of a datacenter, and produce your own tokens through models you post-train and own yourself?
This article will try to answer this question.
Start with the token math. A well-utilized rented H100 slice serving an open-weight model produces output tokens at roughly $0.20–$1 per million, a 25–100x discount to frontier sticker prices ($25/MTok for Opus 5, $50 for Fable 5). Two honest caveats. First, utilization is everything: a single H100’s effective cost swinging from $0.21 to $15.25 per million tokens as utilization falls, a 36x idle penalty. Below ~60% utilization the API always wins.
Second, caching crushes sticker prices: SemiAnalysis measured a true blended ~$1/MTok on agentic Opus workloads despite the $5/$25 list.
So the arbitrage isn’t “self-host and save.” It is: own the post-trained weights.
A tuned open model substitutes for frontier-priced tokens at commodity-priced serving.
So...Why is this not everywhere yet?
Compute is unbuyable in the way that matters.
Like
mentioned: “Purchasing compute in 2026 looks more like buying drugs than enterprise procurement (or so we hear)”
Four observations:
There is no price of compute. The same H100 rents for ~$1.95/hr on marketplaces and $8-9/hr at hyperscalers. An Itô Markets’ quote harvester pulled 1,417 quotes in a single day across 120 SKU-region-term combinations and found material same-day dispersion for identical silicon. Different configurations.
High volatility, zero hedging tools. Hourly rates can swing 137% in a year. Any other commodity this volatile has had a futures curve for a century.
A $1.5 trillion financing hole. Global data center spend is estimated at ~$3T through 2028, with $1.5T needed beyond hyperscaler cash flows. Lenders want contracted revenue, buyers won’t commit to years 4-5. The unfinanceable tail of a datacenter loan, months 30-60, is the single clearest product gap in the market.
The buyer base is exploding. It used to be five labs. Now it’s all companies that want to own models (Ramp, Harvey, Cursor, Airbnb running on Qwen), and there’s no one servicing the smaller guys.
When a commodity gets this big, this volatile, this capital-hungry, with this many buyers, it financializes. Every commodity before it did: oil, power, freight, DRAM. Oil took ~15 years from the 1973 shock to Brent futures. Compute is doing it in four: ChatGPT launched November 2022; CME lists compute futures October 2026.
The labs and their hyperscaler deals, the largest bilateral contracts in the history of capitalism, signed in about 14 months:
OpenAI: ~$750B of planned compute spend through 2030, the number moved three times in nine months. Oracle ~$300B/5yr (~4.5GW), Microsoft $250B Azure, AWS ~$138B, CoreWeave $22.4B, Broadcom 10GW of custom silicon, AMD 6GW plus a warrant for ~10% of AMD at $0.01/share, and NVIDIA’s famous “$100B” letter of intent that landed as $30B of equity. Stargate: 9+GW planned in the US by 2029.
Anthropic: up to 1M Google TPUs (~1GW in 2026, >3GW in 2027), up to 5GW on AWS, $30B of Azure, and a $50B own-build with Fluidstack.
Meta:
of capex in 2026; Hyperion (2GW, scalable to 5GW) financed through with $27B, 80% off balance sheet.
xAI: Colossus 2 at ~2GW / ~555k GPUs, absorbed into SpaceX at a combined $1.25T, the largest private merger ever.
Oracle: $638B of RPO, +363% YoY, and BofA thinks more than half of it is one customer.
Neoclouds: ~200 exist. SemiAnalysis’s ClusterMAX census tracks 209 GPU clouds, rates 84, gives only 37 any medallion. How the business actually works: buy 8-GPU H100 servers at $40k deployed, rack them next to cheap power, rent at $2-4/hr, payback ~14-16 months at full utilization. Don’t sell spot: sign a 5-year take-or-pay with a lab, drop the GPUs and the contract into a SPV, and borrow 60-70% against the contracted cash flows.
If you’re a bitcoin miner, you convert: Miners have flipped ~7GW across 19 deals worth >$135B in two years.
Oil didn’t start as a market. Standard Oil set prices by fiat, then the Seven Sisters did, Saudi crude was $2.18/bbl in 1947 and $1.80 in 1970. Less than 5% of oil traded spot.
Then in 1973 price quadrupled, 1979 spot cargoes at $50 against $13 official. The Rotterdam spot market exploded from ~5% to a third of world trade, netback pricing killed the posted price in 1986, and paper arrived: NYMEX WTI futures in 1983, Brent in 1988. But compute is electricity, not oil. You can’t store a GPU-hour, idle compute is gone forever. No storage means no inventory arbitrage, no floor, the right analogy is power markets: non-storable, heterogeneous by location (an H100-hour in Virginia is not an H100-hour in Jakarta, just as power at node A isn’t power at node B), prone to negative prices for stranded capacity and 100x scarcity spikes.
Power solved this with hubs, nodal benchmarks, and above all the PPA, a long-term creditworthy offtake contract a bank will lend against. That’s how every independent power plant since 1978 got financed, and it’s precisely how neoclouds are financed today: the take-or-pay compute contract is a PPA.
What exists today:
Indices: Silicon Data’s SDH100RT (H100 rental, live on Bloomberg since May 2025, ~3.5M data points/day, DRW-backed) plus a 1–36 month forward curve; Ornn’s OCPI (H100 printed $2.93 on Aug 28); SemiAnalysis’s contract-price index; Compute Desk’s indexes on 350k Bloomberg terminals. There is already a live term structure, and it’s in backwardation (H100 spot ~$2.72 vs ~$2.38 36-month), pricing in supply growth and chip succession exactly like an oil curve prices in balances.
Futures: CME lists cash-settled H100 and B200 monthly rental futures on NYMEX October 5, 2026, pending regulatory review (settling on Silicon Data). ICE announced its own GPU futures on Ornn’s index. Architect, is going for perps on GPU and DRAM prices plus an exchange-for-physical bridge so futures can convert into actual delivered capacity. Three exchange groups announced listings for the same new commodity within three weeks of each other this May.
Spot venues: SF Compute runs an order-book market where compute contracts are resellable. Compute Exchange runs auctions across 100+ providers and opened a used-GPU secondary market in July.
Speculators and desks: Kalshi’s GPU markets did ~15x Polymarket’s volume and published a market-implied forward curve; Hyperliquid already trades H100 perps; Goldman and JPMorgan are exploring GPU futures desks. Itô Markets is building the agentic OTC desk layer.
Credit: >$20B of GPU-collateralized loans outstanding, from CoreWeave’s facilities down to onchain:
lends at up to 70% LTV against tokenized warehouse receipts on the actual hardware, GAIB tokenizes GPU financing yield. Apollo and Blackstone built a $35B chip-backed SPV for Anthropic with Google payment guarantees.
Insurance: Lloyd’s syndicates are absorbing GPU residual-value risk; Forward Compute places outage cover, RV guarantees and fixed-for-floating compute swaps under ISDA docs; American Compute writes residual-value floors up to 3 years so equipment lenders can do balloon structures.
The hedge
Every functioning commodity market is a machine for moving risk between two natural sides.
The natural shorts are datacenters and neoclouds: they’re long physical capacity financed with debt, and their lenders demand contracted revenue, so they need to sell compute forward, especially the tail years no buyer will commit to. The natural longs are AI labs, sovereigns, and every company that owns models or sells fixed-price AI products with floating inference costs, they need to buy forward to make compute budgetable. Speculators warehouse the mismatch. What’s still missing is options, liquidity beyond ~12 months, standardized physical settlement, inference-token futures (hedge $/MTok, not $/GPU-hour, power-compute spread products (the spark spread of AI), and a credit-ratings layer for counterparties. Each of those gaps is a company.
You decided to own your models. You post-trained an open-weight model on your workflow, it beats the frontier on your benchmark at a fraction of the cost, and now compute is your largest COGS line.
What do you need? Six slices, each with an early-stage winner to be picked:
Price discovery: what does compute actually cost? (indices, benchmarks, forward curves)
Procurement: where do I buy it without a 12-month lock-in? (spot markets, auctions, brokers)
Predictability: how do I make my biggest cost line budgetable? (OTC desks, fixed-rate forwards)
Hedging: how do I protect the P&L when spot moves? (futures, perps, options)
Financing: who funds the hardware on the other side of my contract? (GPU-backed credit)
Protection: what happens when the cluster fails or the chips depreciate? (insurance, residual-value guarantees)
Our internal database of ~11,000 researched deals surfaced 51 companies building compute market structure. 24 financialization plays (futures/forwards/perps, insurance, indices, OTC desks, structured financing) and 27 marketplaces/brokers. A striking share are crypto-native: on-chain settlement, tokenized collateral, perps. The crypto capital-markets toolkit finding its first real-world commodity.
Indices & data
Silicon Data: the Platts of GPUs. Its H100/B200 rental indexes are what CME’s futures will settle against, distributed on Bloomberg. ($4.7M seed + $30.5M Series A, DRW, Jump Trading)
Ornn: OCPI index on Bloomberg + the ICE futures partnership + GPU residual-value swaps. The most horizontally ambitious player, index, derivatives and insurance surface area at once. ($5.7M seed, Crucible, Vine; reported $33M from a16z)
Internet Backyard: standardizes datacenter capacity into tradeable units with metering/billing rails underneath. The unsexy plumbing every other slice quietly depends on. ($4.5M pre-seed, Basis Set Ventures, Crucible, Operator Collective)
NATIVX: its COIL Index normalizes GPU prices to a stable energy unit, stripping regional power-cost noise out of the benchmark — the purest power-compute-spread primitive yet, and ICE picked it as its second compute-futures partner alongside Ornn. Founded by Vapor IO founder Cole Crawford. (no disclosed round, founder-led)
Marketplaces & spot
SF Compute: build data centers, and have an order-book spot market where compute contracts are resellable. ($40M Series A, DCVC, Wing)
Compute Exchange: Marketplace for compute from neoclouds and others. Reserved-capacity auctions (1-36 months). (funding undisclosed, co-founded by DRW founder Don Wilson)
Stoa: A gated venue discovering clearing levels for enterprise GPU hardware itself, financializing the box, not just the hour. ($500K, Y Combinator S26)
RunPod and Vast: Airbnb-of-GPUs spot marketplaces. Spot prices, you can be evicted mid-run, but get it as sport. (RunPod: $20M seed, Intel Capital, Dell Technologies Capital; Vast: no disclosed VC)
Computable: buy, sell and redeem dedicated H100 nodes by the calendar week, sealed-bid auctions, posted sell-back quotes, forward weeks priced into 2027, plus an open-source GPU price index. Built by ex-Jump Trading and Citadel Securities market-microstructure people. (YC 2026 batch)
Desks & risk transfer
Itô Markets: an agentic OTC desk, quotes a single fixed GPU-hour rate with an SLA, stands as principal, sources across 10+ providers and lays the risk off. (Alliance-backed)
Anera Labs: a clearinghouse for AI risk, a forward capacity market for deliverable inference that executed the first institutional block trade on a US-regulated GPU instrument, as an embedded hedge for one of its own providers. The earliest visible compute-native trading desk. (~$500K pre-seed, a16z CSX, Anagram, Orange DAO)
Exchanges & derivatives
Liquid Compute (fka Pluto): the only team pursuing both a CFTC-regulated exchange and clearinghouse for compute, with cash and physical settlement. Slow road, but if it lands, it’s the picks-and-shovels of the whole category. (~$3M seed, Y Combinator W25, early Polymarket backers)
MNX: Manifold’s founders building non-custodial H100 perps plus AI-lab valuation futures on MegaETH, deliberately no token. The degen on-ramp to compute exposure, and the first venue where you can short a lab. ($6.4M pre-seed, Village Global, Cambrian, North Island)
Architect: Brett Harrison’s (FTX) regulated venue for GPU/DRAM perps with an exchange-for-physical bridge into real delivered capacity. ($52M total; $35M Series A, Miami International Holdings, Tioga Capital)
Castle: hedge the unhedgeable, markets for the embedded exposures companies cannot trade today, datacenters named first: a data center is a wager on permits clearing before a moratorium lands. Ex-SIG quant founder; launched August 2026 with a CNBC exclusive. (Ribbit invested, no public raise disclosed)
Credit & financing
USD.AI: GPU-collateralized stablecoin credit, non-recourse loans at up to 70% LTV against legally repossessable, warehouse-receipt-tokenized hardware. $225M+ locked and a fresh $100M Bullish facility ($13.4M Series A equity, Framework Ventures, Coinbase Ventures, Dragonfly). “compute as collateral”.
GAIB: tokenizes GPU and robotics financing into a yield-bearing synthetic dollar. (~$15M total; $5M pre-seed, Hack VC, Faction, Hashed)
Aravolta: telemetry-based collateral monitoring for GPU lenders, real-time verification and depreciation-aware valuation of the actual hardware behind the $20B+ GPU-debt stack. ($5.1M seed, Y Combinator, Topology, Wischoff, Crucible)
Insurance & protection
Forward Compute: Lloyd’s-placed outage insurance, GPU residual-value guarantees, and fixed-for-floating compute swaps under ISDA docs. (seed, undisclosed; KDX, Octopus Ventures, Yonder, Forum Ventures)
PRINCEPS: Underwriting infrastructure that turns SLA terms, market prices and failure modes into coverage carriers can actually bind. Insurance is how the tail years of datacenter financings eventually get de-risked, someone has to price the risk first. (YC S26, amount undisclosed)
American Compute: GPU residual-value insurance written through surplus-lines paper, with declining guaranteed floors up to 3 years, the product that makes GPU balloon financing possible. (no VC funding; insurance-capital backed)
Feel free to DM us if you’re building in compute markets, or to get access to the full 51-company list we made.
Thanks for reading! Subscribe for free to receive new posts and support my work.