Hyperscalers Just Signed 38 GW of PPAs. Your Scheduler Needs to Catch Up.

What happened

Ember's April 2026 dataset shows wind and solar generated more electricity worldwide than gas for the first time in a calendar month. The headline number is the milestone, but the structural story underneath it is more useful: hyperscalers signed roughly 38 GW of new corporate PPAs in the trailing 12 months, with Amazon, Microsoft, Google, and Meta accounting for over half of that volume. That's not a sustainability press release — it's procurement at the scale of national grids.

The crossover happened because the cheapest marginal megawatt in most ISOs is now solar-plus-storage, not combined-cycle gas, and the entities buying that megawatt at industrial scale are the same companies running your workloads. In CAISO, ERCOT-West, and large chunks of EU ENTSO-E, midday clearing prices regularly hit negative territory while 7-9pm peaks clear at 8-12x the daily average. The duck curve isn't a forecast anymore. It's the bill.

The Electrek piece frames this as a clean-energy milestone. For anyone running infrastructure, it's a scheduling problem that just escalated from "nice ESG slide" to "line item your CFO can see."

Why it matters

The gap between average grid carbon intensity and *marginal* grid carbon intensity has widened sharply over the last 18 months. Average tells you what the grid looked like. Marginal tells you what your incremental kWh actually caused to be burned. In a grid that's 70% solar+storage at noon and 55% gas at 8pm, running a batch job at the wrong hour can be 4-6x more carbon-intensive than running it three hours earlier — even though the "average grid mix" number on your dashboard barely moves.

Google has been public about shifting non-urgent compute (video transcoding, ML training checkpoints, some BigQuery batch) toward low-carbon hours using their internal carbon-intelligent platform. Microsoft published similar work on Azure's carbon-aware Windows updates. Both companies have quietly pushed the same math into their commercial billing: Spot instance availability in solar-rich AZs is now correlated with irradiance forecasts, not just demand. If your autoscaler treats `us-west-2` as a uniform resource pool, you're leaving 15-25% on the table during shoulder hours.

The community reaction on Hacker News split predictably. One camp pointed out that gas is still the swing producer and that solar's nominal capacity overstates real delivered energy — both true. The more interesting comments came from grid operators and SRE leads who noted that intraday price volatility has roughly doubled in solar-heavy ISOs since 2024. ERCOT saw 47 hours of negative pricing in April alone. CAISO's evening ramp now requires 13 GW of dispatchable capacity to swing online inside 90 minutes. That volatility is the actual operating environment your cloud provider is buying against.

The dev-facing insight: "carbon-aware" and "cost-aware" have converged. Optimizing for one now optimizes for the other within a few percentage points. That wasn't true even 18 months ago, when carbon-aware scheduling required explicit trade-offs against latency or spot-pricing. The PPA buildout collapsed the trade-off curve.

What this means for your stack

Three concrete actions worth considering this quarter.

First, audit your batch boundaries. Any job that doesn't have a user staring at a progress bar is a candidate for time-shifting. Nightly ETL, model training, log compaction, backup verification, image transcoding, search index rebuilds — these are the obvious targets. If your orchestrator is Airflow, Argo, Temporal, or even cron-on-a-VM, the change is adding a "defer if grid intensity > threshold" gate. The [WattTime](https://www.watttime.org/) and [Electricity Maps](https://www.electricitymap.org/) APIs both expose marginal intensity per AZ-equivalent region with 5-minute granularity. Free tiers cover most use cases.

Second, stop treating regions as fungible. `us-east-1` and `us-west-2` have radically different carbon profiles at any given hour, and the gap is growing, not shrinking. If your multi-region failover policy is "closest to user," you're optimizing for a constraint that matters less than it used to for latency-tolerant work. Google's `us-central1` is roughly 50% lower marginal carbon than `us-east1` on average, with the gap widening during midday hours. For workloads with no hard latency floor, region selection is now a cost lever, not just a compliance one.

Third, price the curve, not the average, in your capacity planning. If your reserved-instance strategy assumes flat power costs, you're modeling a grid that stopped existing about two years ago. Spot pricing in solar-rich regions is increasingly bimodal: deep discounts during oversupply windows, sharp spikes during evening ramps. Workloads that can interrupt cleanly (Spark with checkpointing, Ray clusters with elastic workers, anything containerized with graceful shutdown) will see materially better economics than they did in the flat-grid era. Workloads that can't — stateful databases, latency-bound APIs, anything with sticky sessions — should be on reserved capacity in the regions with the most stable evening ramp profiles.

Looking ahead

The April 2026 number will be revised. The structural shift it represents won't. Hyperscaler PPA pipelines through 2028 imply another ~80 GW of contracted renewables coming online, most of it solar-plus-storage in the same handful of ISOs where compute is densest. That means the intraday carbon and cost curves get steeper, not flatter, before they normalize. The teams that wire grid-awareness into their schedulers this year will spend the next two years compounding the savings. The teams that wait will eventually be told to retrofit it by a finance org that just figured out the line item exists.