Flexible compute: the largest battery the grid already has

Grids are discarding record volumes of clean energy for want of flexibility while data-centre demand grows faster than wires can be built. Virtual power plants (VPPs) — software-dispatched aggregations of distributed resources that trade like power stations — are already operating at tens-of-gigawatts scale, and Bitcoin mining has quietly become the most proven flexible load on the planet, earning real grid revenue at sub-minute response times. The open question is not whether flexible compute works but how far AI workloads — the next order of magnitude of demand — can inherit that flexibility, and what architecture lets one site sell both compute and grid services.

• The bottleneck is flexibility, not wires. In Ireland — the sharpest case — 14% of all-island wind was dispatched down in 2024 (2.2 TWh, per EirGrid/SONI; worth an estimated €450M to consumers, per Aurora Energy Research), with 96% of the Republic's wind curtailment logged against minimum-conventional-generation limits, while data centres already take 22% of national electricity (CSO, June 2025) and head toward ~30% by the early 2030s on EirGrid's projections.
• VPPs are present tense, not future tense. North America runs ~37.5 GW of behind-the-meter VPP capacity (Wood Mackenzie, September 2025); the US DOE projects 80–160 GW by 2030. Kraken passed 2 GW of aggregated residential flexibility in 2025; AGL orchestrates ~1.5 GW; Sunrun's CalReady delivers ~250 MW average per event from 56,000+ California homes.
• Bitcoin mining is a de facto ~20 GW global VPP, dispatched by price. The receipts are real: ERCOT-related power credits paid Riot Platforms $71M in 2023 ($31.7M in August 2023 alone) and $33.7M in FY2024. ERCOT's mining fleet reached ~4.6 GW of price-responsive demand in 2025, though formal registration lags behaviour: ~0.24 GW enrolled as Controllable Load Resources plus ~1 GW in the Emergency Response Service (Potomac Economics).
• AI is not one load but three flexibility classes. Training can checkpoint, power-cap to 60–80% TDP and shift geographically — all in production; latency-bound inference cannot be curtailed without breaching SLAs; batch/async inference behaves like training. A grid-facing control system must treat these as separate workload classes.
• The integration, not the components, is the innovation. A 500 MW co-located campus — 150 MW firm inference / 200 MW semi-flex training / 50 MW batch / 100 MW Bitcoin buffer, first off and first on — operates as a two-quadrant grid resource. In Satscryption's model (v1.6) the power-side stack is roughly break-even on its own: the AI anchor lease drives the returns, and the Bitcoin buffer is what makes the over-built renewable PPA economic.

Infrastructure, energy and data-centre strategy leaders should:

• Treat load flexibility as a monetisable product line, not a compliance concession — price curtailment response, capacity value and ancillary services into every new data-centre business case.
• Bitcoin operators: productise what you already run — firmware-level curtailment, stateless restart, multi-tier hosting and energy-advantaged sites — as grid services and as hosted flexibility for AI tenants.
• AI and data-centre operators: segregate workloads into firm / semi-flexible / batch classes inside the control plane, and engineer training for checkpoint-and-shift from day one. Do not promise the grid flexibility from latency-bound inference.
• Grid operators and policymakers: procure flexibility explicitly — ERCOT's controllable-load programmes are the template; curtailment driven by system-stability limits is a flexibility problem that wires alone will not fix.
• Be willing not to proceed where the economics do not close: without an anchor tenant the campus power stack is near break-even, and without the buffer tier the over-built renewable PPA does not pay.

The gap: clean energy discarded beside load that cannot flex

Renewables dominate new generation but sharpen net-load ramps; transmission takes four to eight years to build (IEA), transformer lead times have doubled, and PJM's capacity auction has cleared at its price cap in consecutive auctions. The result is a grid that pays to discard clean energy at the same time as it struggles to host new load. Ireland states the problem in three numbers: 14% wind dispatch-down (2024), 22% of electricity already consumed by data centres, and an all-island DS3 system-services budget capped at €235M a year by the SEM Committee — a cap spending has in practice been permitted to overrun since 2023, which only sharpens the point that stability services are scarce and expensive. Marginal capacity is now cheaper to aggregate than to build.

What a VPP actually is

A VPP aggregates distributed resources, dispatches them in software within seconds, and trades the result like a power plant — a control system with a balance sheet, not a building. The standards exist (OpenADR 3.0, IEEE 2030.5, IEC 61850); the distinction that matters is a closed, market-aware control loop versus an annual interruptibility contract. At 37.5 GW deployed in North America alone, the model no longer needs proving.

Bitcoin mining: the existence proof

Seventeen years of Bitcoin have produced a ~1 ZH/s network drawing roughly 20 GW continuously — about 138 TWh/yr (Cambridge, April 2025 revised methodology), with a measured energy mix of ~52% renewables plus nuclear. Because difficulty re-targets and every operator dispatches against price block by block, the network self-balances toward the cheapest joules anywhere — a globally distributed, autonomously dispatched, market-aware load with the properties of a VPP baked into the protocol. ERCOT is the live laboratory, and operating deployments show the model runs at scale: Riot in ERCOT; Crusoe's flared-gas sites, whose energy-first model now powers AI training at the 1.2 GW Stargate Abilene campus; IREN's BC-hydro campuses (built for mining, since substantially repurposed toward AI — the pivot itself proving the thesis); and Lancium's Abilene Clean Campus, pioneer of ERCOT's load-only controllable-load registration. Mining curtails in under a minute at zero SLA cost and zero restart cost — the gold standard against which other flexible loads are measured.

AI: the customer who pays to scale it

To the grid, AI training is not a flat load: compute and gradient synchronisation alternate on millisecond timescales, and at gigawatt scale the modelled swings reach hundreds of megawatts. But training can checkpoint on a ~10-minute cadence, power-cap to 60–80% TDP at modest throughput cost, and shift between regions — all in production today. Duke University's February 2025 analysis found ~98 GW of US grid headroom for new load willing to forgo just 0.5% of its maximum annual energy consumption — and most of that curtailment is partial throttling, not shutdown. Latency-bound inference is the hard case: sub-100 ms service objectives cannot nap for frequency response. The trapdoor is batch and asynchronous inference, which behaves like training and can be deferred for hours. One connection, three flexibility classes; the control plane must know which is which.

From Bitcoin site to Super-VPP

The growth path starts from sites that exist today — co-located renewables, mining and an aggregator at sub-100 MW scale — and adds the AI anchor tenant. The destination design: a 500 MW campus on one 220 kV substation (300 MWp solar, 200 MW wind PPA, 100 MW/400 MWh battery), with a four-tier load hierarchy — firm inference (150 MW, never curtailed) · semi-flex training (200 MW) · batch (50 MW) · Bitcoin buffer (100 MW, first off, first on). On a grid event the battery responds in milliseconds, mining sheds within two seconds, batch pauses at an epoch, training checkpoints inside two minutes — over 350 MW shed with the firm tier untouched. The streams earn from each other: without the buffer the PPA over-build is uneconomic; without the anchor tenant there is no engine; without the battery the training oscillations damage the grid.

Where it works, and where it does not

Risks and caveats

The campus economics are design-stage modelling, not operating results: in the v1.6 base case the power stack alone is near break-even and returns depend on the AI anchor lease — tenant risk is the dominant risk. Training's electrical signature ("AI flicker") is a real interconnection concern at scale. Latency-bound inference must be engineered around, not promised as flexibility. Network-level Bitcoin statistics carry estimation uncertainty, and mining revenue as a buffer tier varies with hashprice. Capacity-market de-rating for flexible compute remains an open regulatory factor in most markets.

1. EirGrid/SONI, Annual Renewable Energy Constraint and Curtailment Report 2024 (April 2025) — all-island wind dispatch-down 14.0% (2,181 GWh); 96.4% of Republic wind curtailment under the High-Frequency/Minimum-generation reason code. €450M consumer-cost estimate: Aurora Energy Research.
2. CSO Ireland, Data Centres Metered Electricity Consumption 2024 (June 2025) — 22% of national metered electricity (5% in 2015).
3. Wood Mackenzie (September 2025) — ~37.5 GW North American behind-the-meter VPP capacity; US DOE, Pathways to Commercial Liftoff: Virtual Power Plants — 80–160 GW by 2030.
4. Cambridge Centre for Alternative Finance (April 2025) — ~138 TWh/yr; ~52% renewables plus nuclear.
5. Riot Platforms disclosures (2023–2024); Potomac Economics, 2025 State of the Market Report for ERCOT — mining demand ~4.6 GW; ~240 MW registered CLR; ~1.05 GW via ERS.
6. IEA, Energy and AI (April 2025) — 415 TWh (2024) → ~945 TWh (2030). AI share of data-centre power 5–15% → 35–50%: Kamiya & Coroamă (EDNA / IEA 4E TCP, March 2025).
7. Norris et al., Rethinking Load Growth (Duke Nicholas Institute, February 2025) — ~98 GW at a 0.5% average annual curtailment rate.
8. Meta engineering disclosures; SemiAnalysis (June 2025); Microsoft/OpenAI/NVIDIA, Power Stabilization for AI Training Datacenters (arXiv:2508.14318).
9. Satscryption, VPP Financial Model v1.6 (June 2026) — 500 MW campus techno-economics.
10. Satscryption keynote film, The Largest Battery the Grid Already Has (8 June 2026) — youtu.be/ue8tEJeHCWk.

BESS — Battery Energy Storage System · CLR — Controllable Load Resource (ERCOT) · DERMS — Distributed Energy Resource Management System · DS3 — Delivering a Secure Sustainable Electricity System (EirGrid) · ECRS — ERCOT Contingency Reserve Service · IRR — Internal Rate of Return · PPA — Power Purchase Agreement · SLA — Service Level Agreement · TDP — Thermal Design Power · VPP — Virtual Power Plant.