back to problem page
Whitepaper · v0.1·hard tech·open to refutation

An optimism.fun request for startups

AI datacenter power

The AI compute buildout is no longer chip-constrained. It is power-constrained, and the gap is $1.3 trillion.

Published

2026-04-24

Authors

optimism.fun

Status

Draft · v0.1

License

CC BY 4.0

§0weekly drop · week 1

Editor’s drop

Published 2026-08-20 · optimism.fun

blackpaper · the problem

Global data-center electricity demand grew 17% in 2025. AI-specific demand inside that grew 50% (Spheron Network / arXiv 2605.14109). Hyperscalers will spend $775-800 billion on AI infrastructure in 2026 alone (CFA Institute / Alcapital Advisory). None of that spend is bottlenecked by chips anymore. It is bottlenecked by power. Nearly 2,300 gigawatts of generation and storage capacity sits stuck in US interconnection queues right now (Quartz / Sightline Climate). The US builds roughly 888 miles of new transmission line a year against a need closer to 5,000 miles. High-power transformer lead times, once 24-30 months, now run five years at the major manufacturers (Hitachi Energy, Siemens Energy, GE Vernova), whose combined order backlog exceeds $180 billion with six-plus years of revenue already booked. Siemens Energy's own grid-technologies backlog hit a record 59 billion dollars in August 2026, up 28% year over year. The chip layer is fully captured. Nvidia alone is worth $5.27 trillion as of this writing (companiesmarketcap.com, live), holding 81-87% of AI datacenter chip revenue. A new company chasing "AI compute infrastructure" broadly is chasing a market three incumbents already own outright. That is not the opening. The power layer is different. McKinsey's own "$7 trillion race to scale data centers" report splits the AI-load datacenter buildout through 2030 into roughly $3.1 trillion for chips and tech vendors (captured, see above) and roughly $1.3 trillion for power: generation, transmission, cooling, and electrical equipment. That $1.3 trillion is not one company's market. It is not even one technology's market. It splits into at least four real sub-markets, moving at different speeds, gated by different constraints, some already spoken for and some still genuinely open. The honest framing, not the hopeful one: this is closer in shape to climate change than to a single winnable prize, a systemic, multi-actor, multi-decade buildout that no one company captures whole. What makes it worth building in anyway is that unlike climate change, several of its real sub-niches are small enough, young enough, and fast-moving enough that a single focused company can plausibly dominate one of them within a decade, the way Oklo, X-Energy, and Kairos are already trying to dominate the SMR-for-datacenters niche specifically, not the power buildout in general.

whitepaper · the proposal

The $1.3 trillion power layer under the AI buildout decomposes into four real sub-markets. Each moves at a different speed, is gated by a different constraint, and is at a different stage of already being spoken for. Ranked from most open to least, with real numbers. 1. Behind-the-meter battery storage for datacenter power smoothing, most open. Size: $4.82 billion in 2026, growing to $10.23 billion by 2032 at a 13.4% CAGR (Precedence Research). GPU cluster load spikes whipsaw the grid in a way ordinary datacenter loads never did. That specific smoothing problem, distinct from grid-scale arbitrage storage, has no dominant player yet. FERC's co-location rules for behind-the-meter storage at datacenters were still being written as of 2026 (Davis Graham analysis), which is itself a signal the category is early, not mature. The gate is not capital, battery cells are a commodity, it is integration: fast-response power electronics tuned to the specific transient signature of a GPU cluster spike. A new entrant could plausibly land pilot deployments in 12-24 months. 2. On-site gas-to-power, the aeroderivative/used-turbine lane, open but narrower than it looks. Individual deals are already enormous: Chevron and Microsoft's West Texas "Project Kilby" is 2.67GW, roughly $7 billion, FID expected end of 2026, live by 2028. New heavy-frame gas turbine manufacturing (GE Vernova, Siemens Energy, Mitsubishi Power) is a captured oligopoly with backlogs stretching to 2029-2031, that reads as closed. ProEnergy's real model is the opening: buying or leasing repurposed jet-engine turbines and deploying them in 12-18 months, sidestepping the heavy-turbine backlog entirely because incumbents are turbine-supply-constrained, not EPC-constrained. The gate is securing turbine allocation, full stop. 3. SMR fleets for datacenters, looked open a year ago, reads as closed now. Over 9.8GW of committed nuclear capacity for datacenters across all signed deals (DCD / smrintel tracking). X-Energy (Amazon), Kairos Power (Google), Oklo (Equinix, Prometheus Hyperscale, Switch), plus Westinghouse's AP300 and GE Hitachi's BWRX-300, four to five named developers hold essentially every hyperscaler off-take deal that exists. First power for any current deal is 2030 at the earliest. Don't build here, this niche is effectively captured at the vendor level already. 4. Grid hardware and transformer manufacturing, the most quoted bottleneck, the least accessible. Siemens Energy's Grid Technologies backlog alone hit a record 51 billion euros (~$59B) in August 2026, up 28% year over year. The real constraint underneath the constraint: grain-oriented electrical steel, a five-company global oligopoly with structurally capped capacity. A new mill costs over $1 billion and takes four to six years to reach commercial production. Don't build here either, this is a capital-and-raw-material moat, not a market-openness one. The recommendation: behind-the-meter battery storage for datacenter power smoothing is the real opening. It is real systems engineering, not primarily capital deployment, the capital intensity is meaningfully lower than gas, nuclear, or transformers, which means it is actually startable rather than requiring a nine-figure check before day one. And it is still genuinely open, not spoken for the way SMRs already are. Second choice, if faster revenue and more capital risk is the tradeoff wanted: the gas-turbine aeroderivative EPC lane, ProEnergy has already proven the model works, the question is whether a new entrant can secure turbine allocation before the window narrows further. Confidence: the $1.3 trillion power-layer figure is McKinsey's own published split of a larger, real $7 trillion buildout estimate, medium confidence, it is a modeled projection through 2030, not a measured historical fact. The four sub-market sizes and deal figures are each independently sourced and higher confidence individually. The ranking of "how open" each sub-niche is reflects a snapshot as of August 2026, this is exactly the kind of fast-moving space where the SMR niche went from open to closed within roughly a year, so treat the ranking as time-sensitive, not permanent.
§1abstract

The four-axis ranking

We rank humanity’s most important problems on four quantifiable dimensions — quantity of humans affected, severity per capita, current solution quality, and addressable market size — and package each as a proposal in the spirit of Musk’s Hyperloop Alpha. This document is the proposal for ai datacenter power. Every number below is sourced and tagged with confidence. Every ranking is a conjecture, open to refutation.

Quantity · humans affected

8.2B

low

Severity · WTP / wealth

50%

low

Current solutions

2.5 / 10

low

Market size · TAM

$1.3T

med
§2problem statement

What we are trying to solve

Nvidia and the hyperscalers already captured the AI compute layer — Nvidia alone is worth $5.27T (live, companiesmarketcap.com), and hyperscaler AI capex will hit $775-800B in 2026 (CFA Institute / Alcapital Advisory). None of that is bottlenecked by chips anymore. It is bottlenecked by power: nearly 2,300GW of generation and storage sits stuck in US interconnection queues, transformer lead times have stretched from 24-30 months to 5 years, and the three grid-hardware majors (Hitachi Energy, Siemens Energy, GE Vernova) already carry a combined $180B+ backlog with 6+ years of revenue booked. McKinsey's own "$7 trillion race to scale data centers" report splits the AI-datacenter buildout through 2030 into ~$3.1T for chips/tech (captured) and ~$1.3T for power: generation, transmission, cooling, and electrical equipment. That $1.3T is not one company's market — it fragments into real sub-niches moving at different speeds, some already spoken for (SMR fleets: four vendors hold nearly every hyperscaler off-take deal), some genuinely still open (behind-the-meter battery storage for GPU-cluster load smoothing, still uncaptured as of 2026).

§3why it persists

The gap between the world and the world that is physically possible

Today: AI datacenter buildout is power-constrained, not chip-constrained: 2,300GW stuck in US interconnection queues, 5-year transformer lead times, and grid-hardware majors already booked 6+ years out. Real capital is chasing SMRs and gas, but nobody has meaningfully attacked GPU-cluster-specific load smoothing yet.

Current solution quality is rated 2.5 / 10 (low confidence) — meaning there is substantial unclaimed ground between what exists and what is possible. estimated — real capital is flowing (SMR/gas/battery deals signing), but grid hardware and interconnection reform lag years behind compute demand growth (17% overall, 50% AI-specific in 2025, Spheron Network / arXiv 2605.14109).

§4existing alternatives

Who is already working on this

No companies have yet been tagged to this problem in the dataset. If you know one, open a PR.

§5proposed direction

If we solve this, here is the world we get

After · 10-15 years

Power stops being the binding constraint on AI compute growth. Interconnection queues clear in months, not years. A real market exists for datacenter-specific power smoothing, grid-scale storage, and fast-deploy generation, not just chip supply.

Requests for startups · 1 concrete companies to build

Behind-the-meter battery storage for datacenter power smoothing

GPU cluster load spikes whipsaw the grid in a way ordinary datacenter loads never did. Build the fast-response battery system purpose-built for that exact transient signature, not adapted grid-scale arbitrage storage.

why now
FERC's co-location rules for behind-the-meter storage at datacenters were still being written as of 2026 — the regulatory window is open now, not closed, and no player has captured this specific niche yet.
shape
A battery-plus-power-electronics company selling fast-response, GPU-cluster-tuned smoothing systems directly to hyperscalers and colocation operators, distinct from general grid-scale storage.
success
Datacenter operators treat GPU-cluster power smoothing as a standard purchased line item, the way they already buy UPS systems, and this company is the default vendor.

full rubric + framing on the Requests for Startups page.

§6cost & scale

What the market can pay

The world is already paying $1.3T per year against this problem (McKinsey, "The cost of compute: A $7 trillion race to scale data centers" — power/generation/transmission/cooling slice of the AI-datacenter buildout, distinct from the ~$3.1T chip/tech slice; med confidence).

A successful solution does not need to capture more — it needs to redirect a meaningful slice of existing spend, plus the latent willingness-to-pay implied by the severity score above. The cost ceiling for a real solution is bounded by this number; everything cheaper is dominated, everything more expensive is a non-starter.

§7safety & considerations

What could go wrong, and how we know we are not wrong

Section in progress

Failure modes, ethical considerations, and the conditions under which this whitepaper would be falsified are being authored as the weekly cadence ships. The Deutschian commitment: every claim above is a conjecture; we publish the conditions under which we would update. New whitepaper sections ship with each Monday newsletter drop. Subscribe to get the upgrade, or contribute on GitHub.

§8suggested investors

Who would back this

Section in progress

No capital allocators have yet been tagged to this problem in the dataset. New whitepaper sections ship with each Monday newsletter drop. Subscribe to get the upgrade, or contribute on GitHub.

§9sources & criticism invite

Where this is wrong, tell us

Every number on this page carries a source and a confidence tag. Every section open to refutation. If a citation is wrong, a number is stale, or a conjecture is unfounded — file a correction.

corrections → use the feedback widget in the nav · open issue at github.com/adamtpang/optimism.fun

Weekly whitepaper drop

One problem.One whitepaper. Every week.

Each week we ship a deep-dive whitepaper on a top-ranked humanity-scale problem — built in the spirit of Musk’s Hyperloop Alpha and the transformer paper. Problem + market size + before/after vision + a proposed solution + the investors who would back it. Humanity’s Requests for Startups, sourced and sorted quantitatively. No spam. No marketing.

optimism.fun uses your email address only to send low-volume project updates through Resend. You can unsubscribe with one click at any time; read the privacy policy.