Compute Is A Trillion-Dollar Market Trading In Group Chats | Brett Harrison & Andrawes Bahou
Summary
Compute is already an approximately $1 trillion physical market, yet much of it still trades through calls, texts, Slack, and WhatsApp without a standard forward curve. Individual contracts can run from $100 million to $1 billion and sometimes $10 billion, while buyers may receive radically different quotes for nearly identical capacity. As Brett Harrison puts it, “most of the actual buying and selling of compute happens through text messages.”
The explosion from roughly three hyperscalers to some 500 GPU-service providers is commoditizing compute while creating a sprawling brokerage layer. Entry can require capital, GPUs, power, a site, and demand, but buyers must distinguish real supply from brokers reselling another broker’s promise. Andrawes Bahou calls the resulting pre-committed but nonexistent capacity “phantom power.”
Fast-growing AI companies face a brutal trade-off between paying up immediately and accepting years of capacity risk. In the host’s example of 5–8% daily growth—perhaps 10%—a company might pay 10–20% more, and in observed cases up to 2x, to avoid lost revenue, while a lower price can require a three- or five-year commitment. Futures and options could separate the compute contract from the risk, addressing the mismatch between “a specific need right now” and capacity sold years forward.
The reference-price opportunity depends on transactional data, not public cloud rate cards. Brett says compute prices have broadly risen since late 2025 as demand outpaced deployable supply, even while hyperscaler GPU margins compressed. The index operation therefore pays major neoclouds for access to actual billing-system prices and publishes H100, H200, B200, and B300 benchmarks on Bloomberg. Public on-demand prices are “overall, useless” for indexing the large, privately negotiated contracts that dominate the market.
Architect is betting that dated futures and options—not just crypto-style perpetuals—fit compute’s contract-driven cash flows. A data center earning $5 per GPU-hour today may need protection against renewal pricing falling to $2, perhaps through a $4 or $3.50 floor; an AI buyer uncertain whether it will need four or eight H200 nodes can instead pay for optionality. “Perpetual contracts are great” for indefinite exposure, Brett says, but they cannot hedge a cash flow tied to a specific future date.
Capital may be the most systemic bottleneck because lenders cannot confidently price or hedge the asset beneath their loans. Unless the borrower or offtaker is investment-grade, financing may cost 20–25% or remain unavailable; meanwhile, hyperscaler free cash flow has approached zero or turned negative as AI capex rises. The market is shifting from underwriting Google’s advertising cash flow to asking how to underwrite compute itself—Brett’s analogy is that a car with only “a gas pedal and a handbrake” cannot safely go fast.
Andrawes expects frontier labs’ pricing power to decline as open and more efficient models improve, though Brett says the several billion AI users generally do not understand that shift yet. Brett then compares approximately $1 trillion of physical compute trading with $2.8–3 trillion of physical crude oil, whose paper market is roughly 40 times larger. If compute itself grows another 10x, he sees potential for two or perhaps three orders of magnitude of notional growth: “People just don’t realize how big it is and how quickly we’ve achieved this.”
Deep dive
1. Compute has become a commodity without becoming fungible
Brett’s starting point: the AI economy’s principal expenditure is renting or buying GPUs, with large contracts ranging from $100 million to $1 billion and sometimes $10 billion. Buyers and sellers therefore carry enormous exposure to future compute prices, but today “that difference in potential risk” cannot be cleanly hedged.
Andrawes traces the market from Google, AWS, and Azure through the scaling-era realization that bigger computers could keep improving models. Capital requirements then pulled in lenders, while specialist GPU clouds such as FluidStack and Lambda offered machine-learning teams raw GPU access without a full hyperscaler cloud stack.
The structural break is proliferation: capacity once sourced from perhaps three hyperscalers can now come from roughly 500 tracked GPU-service providers. When many suppliers can enter with capital, equipment, power, and a site—and demand remains intense—Andrawes sees “a sign of something being commoditized.”
Brett’s Bitcoin-2017 analogy captures the opacity and geographic arbitrage. Andrawes stresses the crucial difference: one bitcoin is interchangeable and location-independent; compute is a service shaped by chip memory, network fabric, cooling, orchestration software, utilization, provider, and geography. It is a heterogeneous commodity, not a portable unit.
2. Procurement is an RFQ conducted through relationships and “phantom power”
A buyer may ask ten AI-founder friends for “your guy,” then hear that capacity is available in 16 weeks when it is needed today. Brokers span hyperscalers, direct providers, and intermediaries promising 40,000 GPUs in two weeks; the buyer must determine whether each offer represents real machines or “a broker for another broker.”
Andrawes’s “phantom power” describes clusters advertised before they exist. A developer first needs a customer to sign a three-year agreement; that commitment unlocks lender capital, which funds purchases from vendors such as Dell, Supermicro, and Nvidia, followed by installation in a data center. A medium cluster can take four to 16 weeks or longer to fulfill.
For a company growing 5–8% per day—perhaps 10%—speed may justify paying 10–20% more and, in observed high-growth inference deals, as much as 2x because forgone revenue costs more than impaired unit economics. The alternative is a cheaper three- or five-year commitment whose volume could be badly wrong once growth or AI demand changes.
3. Financialization separates capacity decisions from price risk
Brett’s commodity analogy is Hershey buying cocoa: the company needs beans, but it need not retain every risk associated with future cocoa prices. Banks and specialist trading firms can model, warehouse, diversify, or hedge that exposure, allowing producers and consumers to focus on their operating businesses.
The host pushes back that a startup should perhaps stay fully exposed—if AI assistants fail, the company fails anyway. Brett’s narrower answer is that hedging need not neutralize the venture thesis: if sourcing capacity will take two months and the buyer fears prices could double meanwhile, it can simply “fix the price” before securing the physical contract.
Pricing evidence remains fragmented. GPU-compute margins at hyperscalers have compressed, supporting the commodity thesis, while average prices across data-center GPU types rose from late 2025 because demand expanded faster than supply. Yet nearly equivalent capacity can still produce widely separated quotes because no universally watched benchmark anchors negotiations.
4. Billing-system data is the foundation for a tradeable index
Public AWS-style on-demand rates and scraped API prices do not capture the market’s center of gravity: privately negotiated, long-duration contracts for thousands of GPUs. Brett says the decisive price lives “in the minds of the person who bought it, the person who sold it, and the billing system that recorded it.”
The index operation pays major neoclouds for access to actual billing records, aggregates completed transaction prices, and publishes reference prices for H100, H200, B200, and B300 compute on Bloomberg. The intended role resembles a crude-oil price-reporting agency turning dispersed bilateral trades into a benchmark such as WTI or Brent.
The exchange pays the index provider through a conventional minimum-fee and revenue-sharing arrangement, letting the data provider participate in exchange volume and open interest. Brett argues that index design is decisive: a benchmark based on on-demand rentals will not hedge a spot business conducted through three-year contracts for 5,000 or 10,000 GPUs.
Brett also points to Nvidia’s announcement of a 25% discount on these contracts as an attempt to supply a forward curve and establish a floor, arguing that this should be the market’s function rather than Nvidia’s.
5. Dated contracts match compute’s duration better than perpetuals
Architect acquired a designated contract market license and submitted proposals to the CFTC, which is considering index reliability, manipulation controls, and cash versus physical settlement. Brett accepts CME as a direct competitor but argues that Architect’s direct onboarding, cloud delivery, APIs, and equal access better fit commercial compute users unfamiliar with traditional futures infrastructure.
Perpetuals suit exposure without a natural end date—someone can remain long Apple until they change their mind—but compute buyers usually have contracts ending on known dates. Like loans and interest-rate risk, those obligations require futures or options that hedge a particular duration rather than perpetual price exposure.
One seller receiving $5 per GPU-hour may face renewal at $2 in 18 months, threatening its ability to service debt; a future or option could instead establish a $4 or $3.50 floor. The product transfers renewal risk without forcing the data center to alter the underlying customer contract.
Brett gives a structuring example: an AI company wanted four H200 nodes—each node comprising eight GPUs—for one year, but might need eight nodes after 60 days, all in one data center. A four-node annual contract combined with a put option on four nodes and a call option on eight nodes, giving it flexibility; the premium cost less than committing prematurely to unused capacity.
On manipulation, Brett argues that public data and tools can normalize and smooth risk, while major crises have emerged from opaque over-the-counter markets. The CFTC comment period is addressing how to keep compute indices reliable and resistant to manipulation.
6. Credit markets are being forced to underwrite compute itself
Brett says borrowers without investment-grade backing may face 20–25% rates or no financing because lenders lack operating histories, residual-value certainty, and confidence that today’s chips or algorithms will remain relevant. Subprime-style GPU lending can fund smaller operators, but he explicitly compares its danger to mortgage excesses two decades ago.
At the top end, firms such as Apollo, KKR, Blackstone, and Brookfield—alongside JPMorgan and Goldman Sachs—finance large operators with creditworthy offtakers such as Microsoft or Meta, often starting around $1–3 billion. Middle-market firms may underwrite hundreds of millions to $1 billion, while asset-leasing companies serve the long tail.
Andrawes identifies the catalyst for today’s sudden financialization: hyperscaler free cash flow has fallen toward zero or negative under AI capex. Capital markets can no longer assume Google’s advertising economics absorb every risk; they must ask how to value, insure, and hedge compute as an asset class because “the economy is now much larger than any single player can absorb.”
7. Constraints are local, but capital determines how quickly supply adapts
Bringing GPUs online requires capital, power, and equipment, with the critical path changing by project. High-memory systems and InfiniBand training fabrics can carry longer lead times; gigawatt campuses face multiyear grid interconnections, prompting behind-the-meter natural-gas generation, while smaller deployments hunt scattered megawatts through brokers earning “crazy money.”
Brett calls capital the largest systemic constraint: lenders demand long commitments because they cannot price or hedge compute. His analogy is a car with a gas pedal and handbrake—adding a normal brake lets it travel faster. Brett’s broader counterpoint is that demand will route around any bottleneck, as China’s constraints encouraged efficient algorithms competitive on older hardware.
No global ranking of power, chips, memory, or permitting survives geography: American companies are accepting capacity in the Philippines or Iceland, and hyperscalers once interested only above 100 megawatts may answer calls at 30. Andrawes, noting that only five weeks remained before the November elections, expected election-related caution to lift afterward; he joked that NIMBY opponents may prefer exposure “not in my yard, but maybe in my brokerage account.”
Andrawes expects frontier labs’ pricing power and dominance to decline as open-source and more efficient models improve. Brett pushes back only on how widely that shift is understood: niche AI circles discuss open models, but the several billion AI users generally do not. He then returns to the scale argument: compute is not merely “the new oil”; approximately $1 trillion already trades physically, versus $2.8–3 trillion for crude, whose derivatives market is about 40 times physical. Another 10x expansion in compute could imply two or three orders of magnitude of notional growth.
Verification Notes
- The transcript assigns ownership of the index/data operation inconsistently: the host identifies Andrawes as the index-business builder, while Brett says “my company” publishes the indices and describes paying neoclouds for billing data. The digest therefore refers to the index operation without assigning ownership.