Pricing · Novation working paper

Two prices for the same silicon

The gap between what a hyperscaler charges for an H100-hour and what a neocloud charges has roughly halved in a year. It has not closed, the indices tracking it disagree with each other, and the first regulated instrument to hedge it lists next month.

Commodity markets are defined less by what a thing costs than by whether everyone agrees what it costs. On that test, compute is an immature market with a maturing price system — and the gap between those two facts is where a buy-side desk earns its keep.

The price gap compressed, and that is the important number

Through the middle of 2026 the headline spread narrowed sharply. Hyperscaler H100 rates in US regions fell toward roughly $5.40–6.10 per hour, while non-hyperscaler wholesale recovered off its cyclical lows to around $2.00–3.30. Depending which end of each range you take, that is a ratio anywhere from 1.6× to 3.05× — midpoint to midpoint it is 2.17×, down from something like 3–4.5× a year earlier.

But "hyperscaler versus neocloud" is the wrong cut

This is the correction that matters most in this paper, and it runs against how we and everyone else have been describing the opportunity. Take list prices for the same class of machine on the same day, 4 September 2026, per GPU-hour:

AWS list is only 1.12× CoreWeave list. The 3× spread people quote is hyperscaler against the marketplace and long tail, not hyperscaler against a price gap neocloud. If your alternative is a strategic neocloud with an enterprise contract, an SLA and a support agreement, the published gap against AWS is close to nothing, and what you are buying with the move is capacity and terms rather than a lower rate.

Sources, all fetched 4 September 2026: Azure Retail Prices API; AWS pricing feed, us-east-1; coreweave.com/pricing; lambda.ai/pricing; runpod.io (page dated 27 July 2026); Together AI and Vast.ai via getdeploying. Cross-vendor comparison caveat: Azure prices its InfiniBand variant 11.1% above the non-IB variant, and quoting a single-GPU PCIe SKU against an 8-GPU SXM+IB node is a category error, so the rows above are held to the same machine class where the source allows.

And list overstates the spread, on every vendor's own admission

CoreWeave advertises "up to 60%" off for committed usage. Nebius says "up to 35% less." Google's 36-month reserved rate is 56% off list; AWS reserved is 57% off. Apply those and the 3× marketplace price gap collapses toward roughly 1.5×.

Then note what follows. Nobody publishes actual enterprise contract discounts. Not one of the providers above, and not any index. Which means the true effective spread is unmeasurable from public data — including by us, and including in this paper. That is not a hedge; it is the whole reason the only defensible baseline is the blended effective rate computed from your own invoices, and the reason we will not quote you a price gap before we have seen them.

Two movements, opposite directions, both mattering. The retail anchor came down as capacity landed. The wholesale floor came up as miners, neoclouds and independents found long-dated buyers and stopped discounting to fill.

Now subtract your own discount

Here is the part most vendors leave out, including an earlier version of this paper. Both figures above are published rates. No large enterprise buyer pays the hyperscaler one. Once you apply a negotiated enterprise discount to the numerator, the price gap against your actual rate is more like 1.3–1.6×.

This is not a small correction. It is the difference between a compelling slide and a marginal business case, and it is the single most common piece of arithmetic sleight-of-hand in this category: size the prize against list, then report the saving against blended. If you catch a vendor doing that — including us — the right response is to stop reading.

The ladder behind that band, since a range asserted without its derivation is exactly what this paper objects to. Take the midpoints: $5.75/hr published hyperscaler against $2.65/hr wholesale is 2.17×. Your price gap is that ratio after your own discount is applied to the numerator:

Two things fall out of that table, and both cut against us. The price gap is inversely correlated with buyer size, so the biggest, best-negotiated fleets — the ones that look like the most attractive customers — have the least to gain from supplier arbitrage and the most to gain from consuming what they have already committed. And a smaller buyer with a 10–20% discount has a wide price gap but too little absolute spend for the fixed costs to clear. The sweet spot is narrower than the screen suggests, and you should check where you sit on this ladder before believing any number we quote.

The compression is not a reason to stop looking, but it changes the posture completely — and it is worth being blunt about where it leaves a product like ours. Run our own model at a 1.30× price gap and whether the case clears depends entirely on how much of the work that can move actually migrates: it fails at 60% capture and clears at 70%. The condition is a joint one — 1.25× needs about 74% capture, 1.35× about 60%, 1.45× about 51%. So the bottom of this band supports buy-side dispatch only for a buyer who can genuinely move most of what they call movable. That is a much narrower claim than the one we started with, and a different kind of claim from the single break-even price gap we used to publish. A wide spread forgives sloppy execution. A thin one does not. When the gap was 4×, almost any second supplier was an improvement. At 1.3–1.6×, the delivered-cost terms discussed in the companion paper — egress, rerun, fit, vendor overhead — are large enough relative to the spread to reverse the decision outright. The analysis has to get better precisely as the opportunity gets smaller, which is the opposite of what usually happens.

Why the indices disagree

Several independent index families now publish compute prices, and they do not agree. This is usually read as immaturity. It is more useful to read it as a specification problem: they are measuring different things and reporting them under the same name.

• Instrument. On-demand, reserved, committed-use, spot and secondary blocks are different products with different option value. An index that blends them reports a number that describes no purchasable thing.
• Region and residency. The same accelerator in US-East and in a European sovereign region are not substitutes for a buyer with a residency obligation, and their prices reflect that.
• Configuration. Interconnect class, memory configuration, storage tier and the presence or absence of a fast local scratch move delivered throughput materially. Most published indices do not distinguish.
• Contract shape. A price for one hour and a price for a twelve-month reserved block are separated by term structure and credit. Reporting both as "the H100 price" is a category error.
• Observation method. Scraped list prices, surveyed transactions and inferred fills produce systematically different numbers, and only the last describes what anyone paid.

The practical implication for a buyer is not to pick the "right" index. It is to stop treating any index as a price and start treating it as a reference — useful for direction, contract length and relative movement, useless as the basis of a purchase decision. The number that governs a purchase is a quote, for a named instrument, at a named supplier, in a named window, with a timestamp.

Age is a field

The most under-rated attribute of a compute quote is how old it is.

In a market where capacity is allocated dynamically and providers re-price against their own utilization, a quote is a perishable object. A price observed four hours ago on a spot pool describes a market condition that may no longer exist. A price observed four hours ago on a twelve-month reserved block is very likely still good.

So the discipline is simple and rarely practised: every mark carries its age, and a mark past its refresh interval cannot be acted on. It greys out. The desk either re-quotes or parks the decision. This sounds fussy until the first time a schedule is built on a stale spot price and the fills come back forty cents higher, at which point it sounds like the minimum standard of care.

The rails arrived, and the instrument is dated

The market's financial plumbing improved through 2026, though less than the commentary suggests and less than an earlier version of this paper said. Kalshi has listed binary event contracts on a monthly GPU price index since July, which is a real, live, cleared instrument — but binaries are not forwards. A curve can be implied from the strike ladder; none is listed, and volume is near zero. We previously described these as "forward curves," which named the wrong instrument.

The combinatorial auction supplier is not yet buyable either. OneChronos and Auctionomics announced it on 29 July 2025 and the product page still reads "coming soon," awaiting federal regulatory approval as of August 2026. We had it listed as launched. Applying serious auction theory to an allocation problem handled by email remains the right idea and is still not a thing you can transact on.

And a cleared futures contract is now dated. On 5 October 2026, CME Group and Silicon Data list the first regulated compute futures on NYMEX: monthly contracts on H100 and B200 hourly rental, cash-settled against Silicon Data's indices. That is a genuine change in market structure and it deserves more precision than the usual commentary gives it.

"Pending regulatory approval" overstates the gate. New contracts on a designated contract market are listed by self-certification under CFTC Rule 40.2 — the exchange files, the contract lists. There is no approval to wait for in the ordinary case.

"No hedge exists" was too strong, and we have corrected it. Event contracts on compute prices already trade on at least one CFTC-designated contract market, and they clear. What is true is narrower and more useful: those instruments do not give an institutional buyer a usable delta hedge, because of contract value limits, binary payoff structure, no physical delivery, and basis risk between an index settlement and your delivered cost at your configuration in your region. The October contracts change the contract value and structure problems. They do not remove basis risk, and for most buyers basis is the binding constraint.

Which leaves the difference that does not go away. There is no must-offer obligation — and this is the comparison most often stated too loosely, including by us. Must-offer in power attaches to resources that have sold into a capacity market: PJM, ISO-NE, MISO, or a resource-adequacy program. Energy-only markets such as ERCOT impose no such duty at all. Compute resembles the energy-only case: a supplier can simply decline to quote, and typically declines exactly when you most want capacity. Any system built on the dispatch analogy has to design for a supply curve that can withdraw, rather than assume one that must appear.

Compute has the price gap of a commodity market, the contracting conventions of enterprise software, and none of the settlement infrastructure of either.

What a buyer should do about it

Given a market with real price gap, disagreeing references, perishable quotes and no hedge, the reasonable posture is neither to ignore the spread nor to treat it as free money.

Keep a curve of your own, built from the quotes you actually receive and the fills you actually get, per supplier and per silicon class, with confidence bands rather than point estimates. It will be better than any public index for your fleet, because it is conditioned on your configuration and your qualified supplier set.

Express your indifference as a value cap — the delivered price above which you would rather wait than run. A cap converts an open-ended optimization into a decision rule, and it makes "not yet" a legitimate outcome rather than a failure.

And treat every published index as what it is: someone else's average, of something slightly different, at an unspecified time. Useful. Not a price.

Sources: Silicon Data regional pricing report, 31 July 2026; Kalshi compute forward curves, launched 14 July 2026; CME Group and Silicon Data announcement of compute futures listing on NYMEX for 5 October 2026, 11 August 2026; OneChronos/Auctionomics combinatorial auction launch, late July 2026. Conditions as of 4 September 2026, not forecasts. This paper was revised after buyer review: the price gap was re-baselined against negotiated rates, the futures and must-offer claims were corrected, and the blanket "no hedge exists" assertion was narrowed.