The Other Side of the Spark Spread

The headline is the OpenAI posting, and it deserves the attention: it is the first role at a frontier lab whose responsibilities read like the mandate of a merchant commercial desk. The hire will “own commodity hedging strategy and execution across OpenAI’s data center power portfolio,” quantify exposure “by market, site, load shape, tenor, tariff, and supply structure,” and execute “fixed-price supply, forwards, swaps, options, retail supply products, congestion and basis risk mitigation.” It sits in a Power & Land team, carries no direct reports at the outset, and lists ISDA and risk-limit experience as a plus rather than a requirement. Natural gas is named explicitly, which matters for a reason we come back to in section 07.

Read the rest of the market, though, and the picture changes. Anthropic’s Data Center Energy Lead pays a higher cash band than OpenAI’s trading seat — the one pay comparison worth making, because it says the labs still rank securing megawatts above managing what they cost — and its mandate is to “secure power capacity and accelerate energy delivery”: multi-hundred-megawatt procurement, interconnection acceleration, ISO and utility relationships, commercial frameworks. It does not mention a hedge, a forward, or a swap. Neither do Google’s Energy Market Development negotiators, Meta’s Energy Manager for Commercial Supply, or CoreWeave’s Energy Procurement Manager. Oracle’s Energy Risk & Procurement role names hedging and mark-to-market but is a procurement seat with a risk vocabulary. The one lab role besides OpenAI’s that is squarely about derivatives is Anthropic’s Director of Infrastructure & Energy Accounting, which wants ASC 815 and embedded-derivative experience on PPAs: the back office for a book that has not yet been announced.

In merchant terms: the labs have been hiring originators and interconnection people for two years, and have just begun to hire the desk. Figure 1 sorts the postings by the function they would occupy on a merchant commercial floor.

A merchant generator’s commercial operation is organized around one idea: the fleet is a portfolio of options on the spread between the price of power and the price of fuel, and the desk’s job is to decide how much of that optionality to sell forward and how much to keep. Vistra’s 10-K describes the company as a ~44,000 MW fleet combined with “commodity risk management capabilities” and a retail platform, and says the integration “mitigates the impact of commodity price fluctuations.” Constellation’s spin-era policy, stated in its FY2021 10-K, was to hedge the prompt three years on an approximate rolling 90% / 60% / 30% basis — though it applied only to merchant revenues not already hedged through state programs, and has not appeared in a filing since 2023. Its 2026 Outlook shows the CCGT fleet 80% / 70% / 60% hedged for 2026–2028, with a further 20% a year under contracted offtake, so the genuinely open position is nearer 0 / 10 / 20%. Vistra reports 100% / 94% / 72% hedged for 2026–2028 as of August 3, 2026, up from 100 / 84 / 58 as of February 18. NRG pairs ~25 GW of generation with ~8 million customers, generation against load.

Inside that envelope the floor splits into five functions, and every one of them has an analog, a mirror image, or a gap on the lab side.

The spread the whole floor trades

Everything above reduces to one number. A combined-cycle plant burning gas at a 7,000 Btu/kWh heat rate turns $3.50/MMBtu gas into power at a fuel cost of $24.50/MWh. If ERCOT North is at $45/MWh the spark spread is $20.50/MWh, which is almost exactly the $21 ERCOT spark spread Constellation assumed for its 2026 gas fleet. The market heat rate, power over gas, is 12,857: any unit more efficient than that runs. The plant is a call option on that spread with a strike at its own heat rate, and the desk’s only real decision is how much of the option to sell and at what tenor.

Hold that shape in mind. The lab has the same spread, with one more leg in it and the sign reversed on the first one.

A merchant converts fuel into power through a heat rate fixed by the turbine. An inference lab converts power into GPU-hours through a heat rate fixed by the silicon, and then converts GPU-hours into tokens through a second heat rate that is not fixed by anything: it is set by how hard the fleet is run. That is the whole mapping, and it is worth building out with real prints.

Leg one, power to GPU-hour. A DGX H100 draws about 10.2 kW at full load across eight GPUs, so roughly 1.28 kW per GPU; at a 1.2 PUE the site draws about 1.5 kWh per GPU-hour. At $45/MWh that is $0.069 per GPU-hour. The Silicon Data H100 rental index printed $2.53/GPU-hour this month, so power is about 2.7% of what a GPU-hour sells for on the neo-cloud market. This is the first structural difference from the merchant: fuel is roughly half the price of the power a CCGT sells, while power is a few percent of the price of a GPU-hour. The rest is capital recovery on the chip. The “compute spark spread”, rental price less power less everything else, is overwhelmingly a capex spread, which is why the forward curve that matters for this leg is the one CME is about to list on Silicon Data’s index (GPU1/GPU2, October 5, 730 GPU-hours per contract), not the ERCOT curve.

Leg two, GPU-hour to token. This is where the analogy earns its keep. In our inference-spark-spread work we measured an inference “heat rate” from a published vLLM sweep of Llama-3 70B on H100s: 2.88 million total tokens per GPU-hour when lightly loaded at 10 requests per second, rising to a median of 11.16 million when saturated, at the trace’s own 0.73:1 input-to-output mix. Price those tokens at a Luna-tier list of $0.20 input / $1.20 output and the revenue per GPU-hour runs from $2.24 to $8.68. Against $2.53 of GPU rent the inference spread is −$0.29 at 10 req/s and +$6.15 at saturation. Same chip, same power draw, same model, a swing of $6.44 per GPU-hour determined entirely by utilization.

So what does the power trader hedge?

Not the margin. At 0.8% of token revenue, a doubling of power prices moves inference economics less than a 10% change in utilization does. What the trader hedges is the absolute dollars and the tail. A 1 GW portfolio at 80% load factor burns about 7 TWh a year, $315M at $45/MWh, and moves $70M for every $10/MWh. At the 10 GW scale OpenAI has discussed for its Ohio campus the same move is $700M a year, more than three times the illustrative sensitivity Constellation shows for its CCGT fleet in 2027 — up to about $200M of upside per $10/MWh against roughly $150M on the downside, on a fully hedged book with fuel held flat. The lab is not more exposed per megawatt-hour than Constellation; it is less hedged, and it is on the other side.

And the tail is where the load-side seat differs most from the generator. From our hourly-power dataset (374,550 hourly settles, July 2023 through July 2026): summer hour-ending-20 real-time prices at ERCOT North averaged $205/MWh against $21 overnight; the 99.7th percentile of the daily top-four-hour block was $3,133/MWh; real-time exceeded day-ahead in 41% of ERCOT North hours. A generator sells those hours. A flat inference load buys them, every one, unless it has a shape hedge or can move. The 168 hourly futures Nodal plans to list on August 31 (24 hours across seven hubs, including all four ERCOT hubs) are the first exchange-listed instrument that matches the load-side desk’s actual problem, which is not the calendar-year price but the 4 pm to 9 pm price in August.

Move the inputs and watch which spread moves. The left column is the merchant’s spark spread. The middle and right are the lab’s two legs. The portfolio row at the bottom converts the power price into dollars at the site scale you choose, which is the number the OpenAI hire will be asked for first.

Figure 5 takes each merchant desk function and asks three questions of the lab seat: what is the analog, is the position the same sign or the opposite, and what has been added that the merchant never dealt with. The “sign” column is the quickest way to see why a good generator trader is not automatically a good load trader: half the instincts transfer and half invert.

1. Utilization is the heat rate. The merchant’s heat rate is a nameplate; the lab’s second heat rate is a management variable that swings the spread from negative to +$6 per GPU-hour on the same hardware. The power trader does not control it, but every number the trader produces, power per token, cost per million tokens, carbon per query, is divided by it. In the calculator above, dropping throughput from 11.16M to 2.88M tokens per GPU-hour raises power’s share of revenue from 0.8% to 3.1% without the power price moving. A capacity-factor problem, in the energy audience’s own vocabulary.

2. The load moves. No generator can relocate a turbine for the evening peak. A lab can, for the deferrable part of its book, send the training job to the site where power is cheap tonight, and the ISOs are starting to price that: Google’s demand-response contracts with five utilities “limit or shift a portion of machine learning workloads.” The load-side RT desk therefore has an instrument the merchant never had, inter-regional and inter-hour load arbitrage, with the constraint that the inference half of the book is not movable at all. Knowing the split between the two is the first thing the trader needs from engineering.

3. One-sided until licensed. A merchant’s book is two-sided by construction. A lab without market-based-rate authority can only buy: surplus behind-the-meter generation, a curtailed hour, or an over-hedged site cannot be sold into the market by the lab itself, only through a marketer or a utility buy-back. Meta’s Atem Energy filing, and the Microsoft Energy and Amazon Energy EQR books, show the direction of travel. Until then, “hedging” for most labs means retail and financial products executed by others, and the ISDA the OpenAI posting mentions is the gateway to doing it directly.

4. Flexibility is becoming a legal obligation. SB6’s remote-curtailment requirement, PJM’s curtail-first proposal for new large loads after June 2027, and FERC’s June 2026 show-cause orders asking every RTO to revise or justify its large-load rules, including transmission services for flexible large loads that can limit withdrawals during grid stress, all point the same way: the lab will be told when it may consume. The merchant’s must-offer and capacity-performance obligations are the closest analog, and the penalties for non-performance transfer too. A risk book that does not carry a curtailment-probability term is missing a line.

5. Three legs, three venues, one of them missing. Power has ICE, Nodal and CME. GPU-hours will have CME on Silicon Data’s index from October 5 and ICE on Ornn’s pending approval, with Kalshi’s forward curves already printing. Tokens have no settlement-grade index: Silicon Data’s SDLLMTK is an expenditure-weighted barometer that moves on model mix, as we argued in the Silicon Data product-suite piece. The power trader will be the only person in the building who has run a hedged book against a liquid curve, and the first thing that experience will reveal is how much of the firm’s exposure sits on legs that cannot yet be hedged. That is the same “contest can’t hedge” conclusion we reached on the compute side in Hedging the Race, now arriving at the meter.