🌿 Friday's Climate Infra Brief: Predictability, speed, access, flexibility
News: At 5 p.m. on July 2, PJM’s grid was pulling roughly 163 GW, a hair under the all-time record of 165.6 GW it set two decades ago. Spot power blew past $2,500 a megawatt-hour. And the DOE did something it almost never does: it authorized transmission owners to force any load >= 50MW, data centers included, onto backup generation within 15 minutes of an emergency signal. Wow, imagine getting a text from your internet provider saying they’re cutting you off in 15 minutes because the utility can’t serve its load.
My read of this: the value of a MW is decomposing into three products, energy, capacity, and flexibility, but only the first two are being priced. Capacity payments buy availability, a power source shows up in a scarcity hour. Flexibility payments buy fast response/willingness to be interrupted. On capacity, we have the data. The PJM capacity auction for the 2025/26 delivery year cleared at $269/MW-day, a 10x jump from the previous year. Then 2026/27 hit the regulatory cap of $329, setting a new record high. This is still below the true price as the governor sets a cap out of protection for ratepayers. On flexibility, we don’t have direct pricing data yet. But intuitively, some large loads are far more willing and more prepared to be paid to go off-grid than others - we’ve all sat on an overbooked flight where they offer $1,000 credits for someone to voluntarily take the next one, and there is always that someone who says yes.
Firm power for AI training data centers is a crowded bet. What other bets we can make here, assuming this market repricing is structural and durable?
Structuring backup power and who pays what. If you’re 50 MW+ in PJM, you’re a dispatchable resource whether you wanted to be or not. The obvious response is to build your own firmness. I’d be honest about what’s filling that gap right now: gas plus storage. The auction cleared 45% gas, and turbine order books at GE Vernova and Siemens are sold out to the end of the decade. The clean-firm answers (nuclear, advanced geothermal) are real but late, mostly 2028 and beyond. So near-term backup is a gas genset paired with a battery that handles the fast-response piece the turbine can’t. If the hyperscaler self-provides to protect uptime, it’s capex on their balance sheet. If the grid procures it on behalf of people, ratepayers pay for it . The interesting structure combines both: A data center that installs backup (BESS or gensets) can monetize it as a grid asset, via capacity, demand response, power price arbitrage, and ancillary. For someone who structures deals, there is value to convert volatile capacity revenue into contracted cashflow that is bankable.
Grid flexibility assets. The DOE order is essentially unpaid, forced demand response. The paid version is a real, fast-scaling asset class, and it may be the cheapest firmness on the board.
VPP - On June 24, Sunrun, Renew Home, and Tesla announced a framework to deliver more than 16.8 GW of flexible capacity straight to hyperscalers and utilities. It is so cool to orchestrate hundreds of thousands of home batteries alongside more than 8 million smart thermostats and EVs, with no new hardware, interconnection, water, or land. When a data center gets told to throttle during the worst hour of the year, these companies can sell it the flexibility to stay on, and spare ratepayers the bill for another gas peaker that runs only 50 hours a year anyway. That’s the retail version. The wholesale version lives on the C&I side. Voltus managed 8.1 GW of flexible load in 2025, steel mills, cement plants, cold storage, big-box retail, hotels, and paid its customers a record $240M to power down on cue.
C&I storage as-a-service market is fast growing. A developer drops a battery behind the customer’s meter at no upfront cost, the customer signs an energy-services agreement shaped like a solar PPA, and the developer keeps the value stack: demand-charge reduction (often 30–70% of a C&I bill), arbitrage, and grid-service revenue.
Storage as a transmission. Site a battery at a constrained node and the utility contracts it to defer a wires upgrade. The battery is a fraction of the cost and online in 18 months instead of 7–10 years.
And the biggest flexible load of all is the one everyone is racing to build - can data center itself be flexible? First, you can manage the workload - training pauses, batch inference reschedules, etc. Second, setting up a market place. SF Compute is running a spot market for GPU time, betting on compute can trade like a commodity and can therefore chase cheap power across time and geography.
The list can go on. But my biggest takeaway from this gold rush to power is more philosophical: nobody in this market is really selling turbines, batteries, or GPUs. They’re selling predictability, access, speed, and flexibility, and we are financing exactly those. The hardware and contract structures are just how those get delivered. May we all find predictability, access, speed, and flexibility in our lives. :)