Energy sector - AI as demand and as tool#

← III.B - Physical · Part III · Index

Part I treats power as something AI consumes. This page is the other half: power is also the sector AI most changes - operations, markets, planning, exploration, and the R&D path to the next generation of generation.

Those two roles fight each other. The same technology that adds inflexible GW of load is the best available tool for running a grid that has to absorb them. Analysis that only prices the demand side misses where productivity and surplus land.

Four layers, ordered by ground-truth cost#

This is the master asymmetry again. The dispatch desk moves first. The substation and the new reactor do not.

Layer 1 - Markets and operations (the quiet win)#

Wholesale power is close to an ideal AI domain: abundant time-series, clear loss functions (imbalance, uplift, curtailment), and decisions that reprice every five minutes.

Expect:

• Better short-term load and renewable forecasts - already deployed; residual error still costs real money
• Automated trading and hedging at the edge of what market monitors allow
• Maintenance that is predictive rather than calendar-based - transformers and turbines are the long-lead items in Part I; extending their life is GW that does not need permitting

by 2030, the majority of US ISO/RTO short-term load and renewable forecasts used operationally are model-assisted or model-primary. The visible effect is lower reserve margins for the same reliability target - or higher reliability at the same cost - not a consumer-facing "AI grid" brand.

Surplus accrues to whoever owns the scarce complements: interconnection rights, generation, and balancing capability - not to the software vendor in competitive markets. → Game 3

The one place this layer could produce something other than a quiet efficiency gain is flexible demand, and it is underrated because it is boring. A grid's cost is set by its peak, and large AI training loads are among the few multi-hundred-megawatt consumers that are genuinely interruptible on the timescales that matter, since a checkpointed training run can pause in a way a smelter or a hospital cannot. Load that can be curtailed on price is worth vastly more to a system operator than load that cannot, and it converts the datacenter from purely a problem into partly a balancing resource. The obstacle is commercial rather than technical: interruptibility has to be contracted and priced, and the tenant's economics reward utilization above almost everything else. Inference load, which is user-facing and latency-bound, is not flexible in this way at all, so the flexible share falls as the mix shifts from training to serving. → Inference economics

Failure mode for the quiet-win claim: operations gains are measured against a counterfactual nobody observes. A grid that avoided an outage produces no evidence, so this layer's value is systematically under-recorded, and the same opacity makes vendor claims here unusually hard to audit. Treat operational AI as high-confidence in direction and low-confidence in magnitude.

Layer 2 - Planning hits the institutional wall#

Interconnection queues and transmission siting are not compute problems. They are study processes, stakeholder processes, and legal processes. AI can draft studies, screen contingencies, and propose topologies faster. It cannot hold a county hearing or rewrite NEPA.

So planning software compresses the analysis inside a timeline set by state capacity and Uncertainty 2. The tightest loop in the corpus still holds: using AI to clear permitting backlogs relieves the energy constraint on AI itself. That is administrative, not algorithmic.

Hyperscalers becoming power companies (Part I) is the private workaround: leave the queue, own the generation. Utilities and ISOs that adopt the same tools without ownership still face the political allocation of who pays for upgrades.

That workaround has a political cost that is accruing quietly. Behind-the-meter generation lets the largest new loads exit the queue and the cost-allocation fight at the same time, which leaves the residual grid's fixed costs spread over a customer base that no longer includes the fastest-growing consumer of power. Ratepayers see bills rise while the visible cause of the growth is not on their side of the meter, and that is a durable political grievance rather than a transient one. The corpus expects the response to arrive as cost-allocation rulemaking and special tariff classes for large loads rather than as a ban on datacenters, and the leading signal is regulatory proceedings on who pays for interconnection upgrades, not headlines about moratoria. Where the politics land determines whether the private workaround stays available, which makes it a constraint on Part I's capacity story rather than a solution to it.

Layer 3 - Exploration and physical plant#

Upstream oil & gas and mining already use inversion and interpretation models; AI extends that. New resource finds and better recovery matter for gas peakers and for the materials in turbines, transformers, and batteries - the industrial inputs with no learning curve that Capital flags as rising $/MW.

Construction and field labor remain robotics-gated. An optimized drill plan still needs a rig and a crew. Planning compresses; handling does not - same split as logistics and medicine.

Layer 4 - Next-generation generation#

Fusion, advanced fission, geothermal, long-duration storage: AI helps design and simulation where ground truth is computational; it does not remove licensing, materials, and first-of-a-kind construction risk.

• SMR and advanced reactor design loops compress; NRC and supply-chain clocks do not
• Fusion control and design benefit early; net-energy deployment remains a physics-and-plant problem
• Materials discovery for magnets, cladding, electrolytes is a science automated-lab story

AI is a multiplier on the research throughput of next-gen generation, not a substitute for a decade of demonstration plants. Claims that "AI solves energy" by inventing a reactor are category errors - they confuse the design layer with the atoms layer.

Who captures the value#

The sectoral prediction matches the rest of the document: adoption is mandatory where margins are competed; surplus sticks to inelastic complements - wires, water rights, generation sites, interconnection positions, and licensed operators.

Interaction with the demand story#

Failure modes#

• If OT cybersecurity becomes the binding brake on grid AI, the operations layer stalls while the demand layer does not - net worse congestion. → Cybersecurity
• If load flexibility (interruptible training) is widely priced, the effective interconnection queue shortens without this page's planning story mattering as much.
• If a major grid incident is attributed to automated control, the 18-month Game 2 window writes restrictive OT rules that freeze Layer 1–2 gains.

Two pages, two objects#

Part I energy is AI as load. This page is AI as tool inside the industry that serves that load. Confusing them produces nonsense: "AI will solve the power shortage" mixes a demand story with a supply-side R&D story. Operations AI can raise effective capacity of existing assets; it does not pour concrete for a substation. Score Layer 1–2 (ops/planning) on congestion and outage metrics; score Layer 3 (generation R&D) on FOAK timelines measured in years. Never credit a chatbot demo against either.

OT cybersecurity can freeze Layer 1–2. A grid incident blamed on automation writes restrictive rules fast. Pair ops-AI adoption claims with OT incident and exclusion trends - cybersecurity.

Related: Energy - the constraint · Uncertainty 2 · State capacity · Science · Geopolitics · Prices

View markdown source