Who's Powering
AI

Trillions in announced AI capex. Tens of billions in nuclear deals. Hundreds of billions in cloud contracts. And then the question every buyer, every investor, every policymaker eventually asks — Who's actually powering AI?

This study runs that question against the four engines where the answer now forms — ChatGPT, Claude, Perplexity, and Google AI Overviews — and reports who they cite, how prominently, and from which sources. The data is a snapshot of the second quarter of 2026. The patterns are durable.

What the data shows

• Microsoft owns the AI-power story — not because of capex, but because of Three Mile Island. One reactor restart became the most-cited single AI-power story across all four engines.
• Amazon outspent Microsoft by $18 billion in a single Talen Energy deal and still trails on citation share. The Susquehanna restructure is a bigger megawatt commitment than Crane. The story is smaller.
• Meta's 6.6 GW January 2026 deal with Oklo, TerraPower, and Vistra was the largest corporate nuclear order in American history. It ranks fourth on citation share. Press density wasn't enough to overtake the Microsoft narrative.
• xAI built Colossus in 122 days and runs it on gas turbines outside the grid. The AI engines mention it less than they mention Constellation Energy — a regulated utility nobody had heard of three years ago.
• The data center operators are invisible. Equinix and Digital Realty house most of the buildings the conversation is about. AI engines almost never name them when answering AI-power questions.
• Chinese hyperscalers are a citation desert inside Western AI engines. Alibaba and Huawei lead Asian AI infrastructure by capacity; on the Western citation graph they barely register.

Why this matters

More than a third of consumers and a growing share of B2B buyers begin research with AI engines, not Google. Procurement teams at hyperscalers and at hyperscaler customers are running these exact prompts. Wall Street analysts are running them before earnings. DOE staff are running them when shaping policy. The companies that show up in those answers are not always the companies spending the most or building the fastest. They are the companies whose narrative has structurally lodged itself in the source stack the engines retrieve from.

Citation share is the new shelf space. This study is the floor map.

WHAT THIS REPORT MEASURES

This study estimates citation share — how often, and how prominently, each of 25 named entities appears in answers generated by leading AI engines when users ask category-defining questions about powering AI. Citation share is the new shelf space. In a world where more than a third of consumers and a growing share of B2B buyers begin research with AI, the answer is the market.

WHAT THIS REPORT DOES NOT MEASURE

This is not a survey of users. This is not a count of mentions on social media. This is not a ranking of AI capability or compute capacity. It is a structured estimate of which entities AI engines surface, name, and source when users ask category-defining questions.

THE FOUR ENGINES TESTED

ChatGPT (OpenAI), Claude (Anthropic), Perplexity, and Google AI Overviews. These four account for the dominant share of consumer and enterprise AI-assisted research traffic in the United States as of publication.

THE PROMPT SET

Sixty-four prompts across six subcategories — Big Five Hyperscalers, AI-Native Compute, AI Consortia & Sovereign Compute, Foreign Hyperscalers, Data Center Operators, and Power-Anchored Compute — plus a cross-category authority set. Prompts were written to mirror real buyer, investor, and policy-research intent. The full prompt set is published in the Appendix.

ESTIMATION APPROACH

Citation share is estimated through a combination of underlying model knowledge and structured web-search cross-checking. Four independent estimation passes were conducted per prompt to surface variance and firm directional ranges. Each entity is scored on (1) frequency of appearance across the prompt set, (2) prominence of placement inside answers, and (3) the source stack each engine cites when the entity appears. All financial figures, deal terms, and capacity numbers were verified through primary sources and current web search.

WHY DIRECTIONAL ESTIMATES

AI engine outputs vary by user, session, geography, account history, and model version. This report does not present logged query runs or single-pull rankings. It presents a directional model of the citation landscape — sufficient for category strategy, insufficient for trading decisions.

LIMITATIONS

Outputs shift week to week as models retrain and source graphs update. A retest in six months will produce different absolute numbers; the structural patterns are more durable than the precise rankings. Companies that begin investing in citation share now will move on the leaderboard. That is the point.

The question is asked daily. By Wall Street analysts before earnings. By policy staff inside the Department of Energy and the White House. By procurement teams sizing cloud spend. By journalists framing the next nuclear story. By the citizen reading about a new data center in their county.

Each of them now asks the question to an AI engine first. The answer the engine produces becomes their working model. That working model becomes the next public meeting, the next investment memo, the next policy draft.

Powering AI splits into three questions that are converging into one market:

• Who buys the power? — the hyperscalers, the AI-native clouds, the sovereign-AI consortia.
• Who builds the compute? — the chip operators, the data center developers, the cluster integrators.
• Where does the electricity come from? — the generators and the grid: nuclear, gas, geothermal, transmission.

This study answers the first. The generators show up in the source stack of every subcategory because they are the proof the hyperscalers point to. The hyperscaler is the buyer. The generator is the receipt.

What follows is the leaderboard.

Citation Share Index. 0–100 scale. Aggregate across 64 prompts × 4 engines × 4 estimation passes. Directional.

SCORES ARE DIRECTIONAL ESTIMATES. NOT LOGGED QUERY RUNS. SEE METHODOLOGY § II.

The Big Five — Microsoft, Google, Amazon, Meta, and Oracle — capture more than 70% of total citation share across the prompt set. That tracks their share of announced AI capex. What does not track is the internal ranking. Microsoft sits well above the others because of a single asset: the Three Mile Island restart, rebranded Crane Clean Energy Center, anchored by a 20-year power purchase agreement with Constellation Energy for 835 megawatts. The deal generates more inbound search and editorial coverage than any other AI-power story in the cycle. Oracle, the lowest-ranked Big Five name, is now overtaken on the overall leaderboard by xAI — the first non-Big-Five entity to break into the top five.

This subcategory is where citation share concentrates fastest. The AI engines learn to recognize a small set of names — CoreWeave dominant, Crusoe second, the rest measurable but not memorable. The challenger tier is a winner-take-most market inside the answer engines.

This subcategory did not exist three years ago. It now produces more cross-engine citation density than any other subcategory in the study except the Big Five. The reason is news velocity: Stargate, xAI, G42, and HUMAIN each carry a near-weekly editorial drumbeat that the engines treat as fresh authority.

This is the most asymmetric subcategory in the study. By compute capacity, by capex, and by domestic market share, Alibaba, Tencent, Huawei, and Baidu would rank among the top ten entities in any honest accounting of who's powering AI globally. Inside the four Western AI engines tested, they collectively account for less than 8% of total citation share.

The cause is mechanical, not editorial. The source stacks the engines retrieve from — WSJ, Bloomberg, Reuters, the New York Times, the trade press, SEC filings, Wikipedia — are Western-language and Western-author dominant. Chinese hyperscalers produce abundant infrastructure reporting; almost none of it lands in the source stack the Western engines retrieve from at scale.

Equinix, Digital Realty, NTT, and QTS run a large share of the physical data center footprint the hyperscalers occupy. They are nearly invisible in the AI-power citation graph. When the engines answer "who's powering AI," they jump straight to the hyperscaler-buyer or the nuclear-supplier and skip the colocation layer entirely.

This subcategory captures the operational hybrids — companies that anchor their AI compute strategy directly to power generation, not to grid connection. xAI is the marquee. Crusoe (counted in AI-Native) is the second name. Tesla shows up adjacent. The category is small in entity count and large in citation share-per-entity because the off-grid narrative is one of the most-cited frames in the AI-power story.

The aggregate leaderboard hides meaningful per-engine variance. Citation share is not a single market — it is four overlapping markets with different source-stack preferences, retrieval logics, and content biases.

The pattern that matters

An entity that wants to grow citation share cannot run one campaign. It needs four citation-share strategies, one per engine, because the source stack each engine retrieves from is structurally different. Press releases work in Google AI Overviews and underperform in Perplexity. Trade press works in Perplexity and underperforms in ChatGPT. Wikipedia anchoring works in Claude and ChatGPT and underperforms in Google AI Overviews.

The source stack is more durable than any single ranking, because moving citation share requires changing what shows up in the source stack. The matrix below maps source types to citation weight, by engine, for the AI-power category.

The structural pattern

When the source stack is the same across two subcategories, citation share is mostly a function of paid-PR muscle and press-cycle saturation. When the source stack diverges, citation share is a function of strategic source placement — which is the GEO playbook.

In this study, the Big Five Hyperscalers and the AI Consortia & Sovereign Compute subcategories share substantially the same source stack — both rely heavily on Primary Reporting and Vendor-Owned content. The AI-Native Compute subcategory diverges sharply, leaning on Analyst & Research and Independent Trade Voices. The Foreign Hyperscalers subcategory has an entirely different source stack — Bloomberg Asia, Reuters Asia, Nikkei, South China Morning Post — that the Western engines under-retrieve.

The under-represented entities are not failures of capacity. They are failures of source placement. Each name below operates AI-power infrastructure at a scale that should produce stronger citation share than it does.

• Digital Realty — second-largest data center operator by capacity, fourth-tier citation share. Sells one named SMR or hyperscaler deal and rises a full tier.
• Stack Infrastructure — Project Jupiter is a $165 billion campus partnership inside the Stargate site expansion. The Western engines have not consolidated the story around Stack. Lowest citation share in the study, highest GEO upside.
• QTS — Top-three colocation operator. Brand layer suppressed by Blackstone ownership framing. A brand-front-and-center repositioning would clear a tier.
• Nebius — Public-listed AI cloud with multiple operational data centers and almost no citation footprint outside the developer press. The source-stack mismatch is total.
• HUMAIN — Younger than G42 in the narrative cycle, comparable in capacity ambition. The AMD/Cisco joint venture is a top-of-stack citation anchor not fully ingested yet.
• Together AI — Developer brand. The trade-press source stack does not retrieve it at the weight it operates at.
• Tesla — Treated as a power-accessory to xAI rather than an AI-power entity in its own right. The Megapack-at-scale story is structurally separable from Elon Musk's AI brand and never separated.

Where the answer engines are right but vulnerable

Microsoft's citation lead is sticky for as long as the Three Mile Island story keeps cycling. The next major reactor restart or first-SMR-online event will move share. Google's host-engine bias inside Google AI Overviews is a real structural advantage and a measurable one — strip it and Google's overall lead narrows to a competitive margin. Amazon's underweight relative to spend is the most resolvable gap among the Big Five — the Susquehanna front-of-the-meter transition in spring 2026 is a publishable narrative event that has not been fully harvested.

Citation share inside AI engines is not a brand-awareness metric. It is a buyer-research surface. The companies named here are the companies whose customers, investors, regulators, and journalists now form working models of the AI-power market from AI-engine answers. Five operational implications follow.

One — Build the source stack, not the press kit.

The companies that rank highest are the companies whose deals show up in Reuters, WSJ, Bloomberg, DCD, TechCrunch, and the SEC filings in primary citations. Press releases hit Google AI Overviews and underperform everywhere else. Trade-press relationships and analyst placement matter more than the next funding announcement.

Two — One famous asset beats ten ordinary ones.

Microsoft's Three Mile Island position. xAI's Colossus 122-day build. Oklo's Pike County campus. The engines reward concentrated narrative gravity. Companies with twelve mid-sized deals trail companies with one famous one.

Three — Run four campaigns, not one.

ChatGPT, Claude, Perplexity, and Google AI Overviews retrieve from different source stacks. A press release works in one engine and not in another. A Wikipedia update works in two and not in two others. Analyst placement works in some prompts and not others. The four-engine source-placement matrix is the playbook.

Four — Move on the gap categories first.

Data Center Operators (Equinix excepted), Foreign Hyperscalers in Western engines, and AI-Native Compute below the top-two all show structural under-representation relative to operational scale. These are the fastest-moving citation-share opportunities in the study.

Five — Build infrastructure before the crisis, not during it.

Every entity in this study carries reputation risk that is one news cycle from the front page — a reactor incident, an outage, a permitting fight, a labor action, a security event. The citation share you have on a calm day is the citation share that retrieves on a crisis day. Build the infrastructure before the crisis, not during it.

This is a directional model. It is not a logged-query benchmark and never represents itself as one. The following limitations are explicit.

• AI engine outputs vary. Different users running the same prompts will get different answers depending on account, session, geography, and model version. The scores here represent the modal pattern across passes, not any single query.
• Models retrain. Outputs shift week to week. A retest in six months will produce different absolute numbers; structural patterns are more durable than precise rankings.
• Engines weight recency differently. Perplexity over-weights last-14-day events; Claude and ChatGPT weight longer-arc patterns. The current snapshot reflects an active news cycle (Deep Fission IPO, X-Energy post-IPO, Meta deal anniversary).
• Source-stack mapping is observed, not measured. The matrix in Section XII reflects qualitative pattern recognition across the prompt set, not a quantitative parse of every retrieved URL.
• The 25-entity universe is a category cut, not the full market. Several adjacent entities — Constellation Energy, Talen, Oklo, Kairos Power, X-Energy, NuScale, TerraPower, Vistra — appear extensively in the source stack but were excluded from leaderboard scoring because they are suppliers, not hyperscalers. They show up as citation context for the entities scored.
• This study does not measure AI capability, compute capacity, or financial performance. Citation share is one dimension of market presence. It correlates with influence over buyer research; it does not substitute for operational benchmarks.

Subcategory 1 — Big Five Hyperscalers (12 prompts)

Who powers ChatGPT · Microsoft AI power strategy · How does Google power Gemini · AWS AI infrastructure · Meta AI data centers · Oracle AI cloud · Microsoft nuclear deals · Amazon Talen Energy deal · Google Kairos nuclear partnership · Meta nuclear RFP · Best AI cloud provider for enterprise · Hyperscaler power consumption ranked

Subcategory 2 — AI-Native Compute (10 prompts)

CoreWeave vs AWS · Crusoe AI data centers · Best GPU cloud for AI training · Lambda Labs power consumption · Nebius AI infrastructure · AI-native cloud providers list · Where do AI startups get compute · Specialized AI compute providers · Together AI infrastructure · Best alternative to hyperscalers for AI

Subcategory 3 — AI Consortia & Sovereign Compute (10 prompts)

What is Stargate · OpenAI infrastructure partners · xAI Colossus cluster · Largest AI training cluster in the world · G42 AI infrastructure · UAE AI strategy · Saudi Arabia AI compute · HUMAIN AI · Who is building gigawatt AI data centers · AI compute build-out 2026

Subcategory 4 — Foreign Hyperscalers (8 prompts)

Alibaba AI cloud · Chinese hyperscaler AI · Tencent AI infrastructure · Huawei AI compute · Baidu AI Cloud · Asia AI hyperscaler · China AI data center capacity · How does China power its AI

Subcategory 5 — Data Center Operators (8 prompts)

Largest data center operator · Equinix AI customers · Digital Realty AI tenants · Best colocation for AI · NTT data centers AI · Who builds data centers for hyperscalers · Iron Mountain AI data center · QTS Blackstone AI

Subcategory 6 — Power-Anchored Compute (8 prompts)

Tesla data center Memphis · xAI Memphis power · Behind-the-meter AI power · Off-grid AI compute · Most power-efficient hyperscaler · Renewable energy AI data centers · Stranded power AI compute · Gas turbine AI data center

Cross-Category Authority (8 prompts)

Who is winning the AI infrastructure race · Most-cited AI hyperscaler · AI data center capacity by company · AI compute market share · Hyperscaler AI capex ranking · Where will AI compute come from · AI power crisis solutions · Next decade of AI infrastructure

Glossary

Citation Share. Estimated frequency and prominence of named-entity appearance inside AI engine answers, normalized across the prompt set. Directional.

GEO. Generative Engine Optimization. The discipline of building brand authority and source-stack presence such that AI engines surface, name, and cite a brand inside category-defining answers.

Source Stack. The ordered set of publications, filings, and reference layers that AI engines retrieve from when answering a category-defining prompt.

Retrieval Anchor. A piece of content (article, deal release, filing, Wikipedia entry, analyst report) that AI engines reliably surface when a related entity is queried. The unit of GEO production.

SMR. Small Modular Reactor. Advanced nuclear reactor designs typically generating 50–300 MW per unit, intended for data center and industrial power.

PPA. Power Purchase Agreement. Long-term contract between an electricity generator and a buyer, typically 10–20 years.

BTM / FTM. Behind-the-Meter / Front-of-the-Meter. BTM bypasses the public grid; FTM delivers through the grid. Talen-Amazon converted BTM to FTM in spring 2026.