//
Discussion · 2026-09-14

The ‘silicon spark spread’: does compute have a price floor?

Rahil Mittal of itomarkets, who builds compute forward curves, proposes that compute pricing has a “silicon…

The Debate

For vs against
Essence
Rahil Mittal of itomarkets, who builds compute forward curves, proposes that compute pricing has a “silicon spark spread” — a price floor set by electricity cost multiplied by datacenter PUE plus opex, with the front of the curve behaving like jump-diffusion power flow and the back converging to the capital replacement cycle. AI Quanting pushes back that the power analogy breaks because GPU capex is sunk and depreciating whether the chip runs or not, so the true floor is “whatever the operator with the most sunk capital will accept” — with power itself costing only about 13 cents a GPU-hour. What’s at stake is how compute gets valued in a downturn: whether spot and forward compute prices have a hard marginal-cost floor, or can crater to whatever keeps the most desperate operator’s servers lit.
For
  • Rahil Mittal: the price floor is bounded by the “silicon spark spread” — “electricity cost multiplied by datacenter PUE plus opex” — with the front of the curve trading at “super violent backwardation” while “the back faces silicon obsolescence.”
  • Mittal: below marginal electricity cost, “running the job literally burns additional cash compared to idling” — an H100 pulls ~1kW of power plus cooling, so the short-run shutdown rule applies to silicon just like it applies to power plants.
  • Mittal (first-party): “I actually have spoken to data centers who shut down operations when this happens... grid operators like ERCOT offer them fat ass incentives to curtail during peak hours” — operators do curtail, so the floor binds in practice.
  • Mittal: “Front of the curve behaves like jump-diffusion power flow and the back converges to a capital replacement cycle” — a tractable structure borrowed from power markets.
Against
  • AI Quanting: “The spark spread works in power because a plant can shut when the spread goes negative. A GPU cannot. The capex is sunk and depreciating whether it runs or not, so the floor is not fuel cost. It is whatever the operator with the most sunk capital will accept.”
  • AI Quanting: the power floor is pennies — “An H100 SXM is 700 watts, and with host and cooling call it 1.3 kW. At 10 cents a kWh that’s 13 cents a GPU hour” — so curtailment is about beating chip earnings, not fuel: “They only take the curtailment money if it beats what the chips earn.”
  • Dave Friedman: obsolescence curves have to be added to the model — “As new GPUs come on line that has to affect the forward curve for the older chips,” so any static floor decays with each generation.

View the thread on X

The Tweets

4 posts
RMRahil Mittal@rahilmittal · 3:17 AM · Sep 14, 2026X Post

Compute forward curve pricing is weirdly elegant Such a silly financial product: part non-storable electricity flow, part exponentially depreciating capital bond (?) The price floor is bounded by what I like to call the "silicon spark spread"... electricity cost multiplied by datacenter PUE plus opex - Front trades at super violent backwardation - Back faces silicon obsolescence If there's any power market quants out there trying to figure this compute nonsense out: Front of the curve behaves like jump-diffusion power flow and the back converges to a capital replacement cycle

4 Reposts · 53 Likes · View on X →
AQAI Quanting@AIQuanting · 12h ago (feed label at capture)X Post

The spark spread works in power because a plant can shut when the spread goes negative. A GPU cannot. The capex is sunk and depreciating whether it runs or not, so the floor is not fuel cost. It is whatever the operator with the most sunk capital will accept.

RMRahil Mittal@rahilmittal · 10:58 AM · Sep 14, 2026X Post

I hear u but that is sunk cost fallacy. Capex depreciates whether the server runs or sits cold, but running an H100 pulls ~1kW of power plus cooling... if bids fall below marginal electricity cost, running the job literally burns additional cash compared to idling. The short-run shutdown rule applies to silicon just like it applies to power plants ! (I actually have spoken to data centers who shut down operations when this happens... grid operators like ERCOT offer them fat ass incentives to curtail during peak hours)

3 Likes · View on X →
AQAI Quanting@AIQuanting · 3h ago (feed label at capture)X Post

An H100 SXM is 700 watts, and with host and cooling call it 1.3 kW. At 10 cents a kWh that's 13 cents a GPU hour, so the power floor is pennies. The data centers you talked to are shutting for another reason. They only take the curtailment money if it beats what the chips earn.