🌿 Friday's Climate Infra Brief: It is not you, it is us

A concept got me intrigued this week: compute-power coordination, “ 算电协同”. The idea is that data centers coordinate with the grid through bidirectional sensing, scheduling, and pricing, instead of just sitting on it as rigid load. In China, it’s written in the national policy. In the U.S., the market is arriving at the same answer on its own, through research, product launches, entrepreneurship, and venture money. The way I see it, this concept has three layers.

Start with siting. Data centers have always been sited along fiber backbone on powered land. What changed is scale. A 10 MW cloud DC plugs into existing grid infra without anyone blinking. A 500 MW AI training campus overwhelms what local transmission was built to serve, even at well-connected sites. That’s why PJM and ERCOT interconnection queues stretch 5–7 years and hyperscalers are signing 20-year nuclear PPAs or building gas turbines behind the meter. Memphis xAI is the canonical version: hundreds of megawatts of on-site gas because waiting wasn’t an option. The catch is that on-site generation solves the queue problem and creates a different one — a 24/7 firm load gives you nothing to flex, which undercuts everything in the next layer.

Next is operations, and this is where bidirectional sensing matters. Today, the data center tells the grid what it needs and the grid serves it. One direction. Bidirectional means the grid pushes real-time signals — price, congestion, frequency, carbon intensity — and the data center pushes back its state and how much it can flex right now. Closer to a smart thermostat than a static breaker.

Two physical realities make this hard. AI training has a weird load shape: GPU clusters alternate every few seconds between compute phases (full draw) and communication phases (reduced draw). At 100 MW scale, that’s tens of MWs swinging on a sub-minute timescale. NERC has already flagged voltage excursions traced back to training DCs. The hardware fix is short-duration batteries sized not for grid arbitrage but as power conditioners — smoothing spikes before they hit the transmission line. These cycle thousands of times a year at durations measured in seconds (!!!), not the 4-hour lithium discharge profile most U.S. BESS projects gets built around. Different product, different economics.

The compute side has to bend too: schedulers that pause, migrate, or delay jobs based on grid signals, not at the rack level but at the AI cluster level, which is much harder. Google has been doing a version of this through their carbon-intelligent computing program since 2020, shifting non-urgent tasks toward cleaner, lower-demand windows. Tyler Norris — now Google’s Head of Market Innovation — quantified the upside from his earlier Duke research: if data centers curtail for a quarter of 1% of their annual uptime, the existing U.S. grid can absorb roughly 100 GW of new load, covering most of the AI demand forecast through 2030.

Emerald AI is a Series A company directly building on this layer. Their Conductor platform orchestrates AI workloads against grid signals in real time. Phoenix demo with Oracle: 25% power reduction over three hours without breaking workload SLAs. Their chief scientist’s peer-reviewed paper puts the flexibility envelope across AI workloads at 18–55% depending on job type. They closed $25M in late March led by Energy Impact Partners with Nvidia, and have a 96 MW commercial flexible AI factory in development at Digital Realty’s Aurora site in Manassas with EPRI, Dominion, and PJM. If the grid-DC orchestration play becomes the mainstream, they want to be the operating system where everyone plugs in.

The third layer is the market, and I love this one the most. The pitch is that AI APIs should be priced like electricity, cheaper when the grid is clean and abundant, more expensive when it’s stressed. Time-of-use pricing for tokens. Sounds speculative until you notice it just shipped. On April 2, Google Cloud introduced two inference tiers for Gemini: Flex and Priority. Flex requests can queue for up to 15 minutes and cost up to 50% less. Priority gets guaranteed throughput at full price. I have seen this before, it is called Uber Surge Pricing! Google frames this as cost-and-reliability tradeoffs for AI developers, not grid pricing pass-through. But the structure creates the power-to-token surface through which grid signals could eventually flow. Norris did join Google after all.

Regulatory framing is moving the same direction. SPP has filed a High-Density Interruptible Load tariff proposal. Pennsylvania’s PUC is advancing flexible tariff options for large industrial loads. None of this reshapes the market alone, but the direction is regulators are beginning to frame large loads as participants, not simply demand to be served.

Constellation’s CEO said this at CERAWeek: we don’t have a supply problem, we have a peak problem. If Norris is right, the next 50 GW of U.S. AI load gets accommodated not by more generation but by making compute behave for a few hundred hours a year. I feel that is as much a project finance question as a technical question. As a financier, we always love negotiating risk and reward allocation!

The real bottleneck might be the shortage of people who understand both grid and compute. Then again — in 2026, what can you not learn?