Risk Doesn’t Disappear, It Gets Transferred: Ishanee Nagpurkar on Building a Futures Market for Compute
Our guest on Where Capital Meets Compute is Ishanee Nagpurkar, Head of Compute Markets at Architect, a CFTC regulated futures exchange focused on the markets that move the AI industry. Architect first tested the idea with an H100 perpetual on its offshore exchange earlier this year. Interest was strong, but the feedback was consistent: you cannot really hedge with a perp. Below is a breakdown of the conversation with Kony, in the order it unfolded.
Why futures, not perps
A futures contract lets you lock in the price of, say, an H100 six months from now at $2.50. Architect starts with cash settlement: if the price has moved to $3.00 by expiry, the long collects the 50 cent difference. Physical delivery may come later.
Perps suit front month speculation, but hedgers pay funding and rebalance margin constantly. And compute contracts are signed three to five years out, so the real hedging need sits much further along the curve.
The real hedging comes from much further ahead expiries. The price discovery needs to happen through futures.
Why the GPU hour is the unit
Kony pushed on the unit, since many operators are becoming token factories. Ishanee’s answer was elimination. Hardware prices are effectively set by NVIDIA. Closed model token prices are set by a handful of labs and can shift the moment subsidies stop. Neither supports real price discovery.
GPU rental hours are priced by a fragmented set of neoclouds, and they touch the widest range of participants: data centers, neoclouds, AI labs, and even inference buyers, for whom bare metal is still a line item somewhere in the stack.
It’s the most blanket, catch all unit you can have at the moment in the industry.
Heterogeneity is basis, not a bug
The common objection is that no two H100s are alike. Ishanee argues this comes from an industry raised on fungible equities. Commodities are heterogeneous by definition. Grain contracts specify moisture content. A GPU contract specifies a deployment standard, and anything better trades at a premium while anything worse trades at a discount.
Just normal commodity market rules apply here. The confusion is that the standards are not set yet.
Choosing the index
There are roughly three index providers today: Silicon Data, Ornn and Compute Desk. Architect uses Compute Desk. Of roughly $600 billion in GPU CapEx, about half sits with hyperscalers. In the remaining free market, Ishanee estimates around 85 percent is locked in reservation contracts, with on-demand pricing carrying a 15 to 40 percent premium and far more volatility.
Compute Desk ingests both types of pricing but outputs a normalized one year reservation price, which neoclouds told Architect they would be comfortable hedging against. She was candid about the limits of every index, given how little deal data private neoclouds share.
The indexes we have currently are fine, but they’re not something you would trust trillions of dollars of hedging on.
New chips face the same problem. Architect would not list Vera Rubin today: only a few suppliers have access, with a handful of deals each.
What the contracts are actually for
If the futures do not become a leading indicator for physical pricing, Ishanee says exchanges are not doing their job. Beyond that, the core use is hedging, and B300 prices spiking this quarter, a year after launch, shows the volatility at stake.
The most interesting use case was cost of capital. Lenders financing a non investment grade neocloud are really underwriting a startup off taker for three years. A liquid futures market gives them a price floor at which to exit, which should support better loan terms for the operator.
Kony asked how this fits alongside NVIDIA’s backstops and residual value insurance. Ishanee’s point: those cover hardware value, while the bigger number is three years of rental revenue. For a lender, pulling and reselling a deployment is the worst outcome. Swapping in a new off taker is far better.
It’s kind of like real estate. You’re not going to take out all the furniture. If you can find someone who wants an apartment with furniture for rent, I’m just going to give that instead.
Who is long and who is short
Neoclouds are structurally short, hedging their books. AI labs are structurally long. Lenders and insurers lean short, since they may need to liquidate. The long side broadens into speculators, hedge funds, ETF issuers and enterprises running sovereign AI on their own hardware.
How a startup wins an exchange market
CME and ICE hold sticky customers, but there have been almost no new commodity markets in decades. Compute is Silicon Valley’s first meeting with commodity markets, and those customers have no exchange loyalties. Architect’s edge is focus: this is its main thing.
Running a regulated exchange is capital intensive, from compliance to market makers, KYC and margin. The heaviest lift is education.
Turn that concept of downside protection into a math equation. That’s the actual job here.
Next on the roadmap: power, other compute denominations, supply chain metals and possibly regional contracts for Asia or Europe if demand shows up.
Accelerating the bubble, or defusing it?
Kony closed by asking whether a compute exchange accelerates an AI bubble or reduces leverage. Ishanee framed it as risk transfer: bubbles hurt most when risk is concentrated, and an exchange exists to disperse it. She sees a real chance of trouble, with operators borrowing at 20 to 25 percent against startup off takers. Whether a liquid market is live when it happens may decide how the aftershock spreads.
Risk doesn’t disappear. It is transferred.
The goal is not to remove risk from AI infrastructure, but to price it and put it where it can be carried.
Listen to the full episode of Where Capital Meets Compute for the complete conversation.