There is a habit in AI coverage of speaking about "the technology" as if it were weather — a force that arrives, reshapes industries, and moves on, with no hand on the tiller. That framing is comforting and wrong. AI is not autonomous. It runs on chips that a handful of companies design, that one company in Taiwan manufactures, on machines that one company in the Netherlands builds. It is trained in data centers financed by a small club of hyperscalers and sovereign funds, using capital arranged by a few dealmakers who keep appearing in each other's press releases. The models themselves come from perhaps six labs that matter. And the rules that decide who gets the chips are set by a government that has quietly become a market actor.
Power in AI is not diffuse. It is unusually, historically concentrated — and it concentrates differently at each layer of the stack. The person who controls the silicon is not the person who controls the capital, who is not the person who controls the model weights, who is not the person who controls the export license. Those are different levers, held by different hands, and the interesting story of this industry is how they press against each other. Some people hold two levers. One man is reaching for all of them.
This is a map of those hands. For each person we name the lever they actually hold — not their reputation, not their vision-statement, but the specific thing they can withhold that would make the rest of the machine stall. Every hard number, stake, and claim is published as reported and flagged ⚑ unverified for a later fact-check, because in a field this hyped the figures move weekly. The point is not to mythologize the machine. The point is to name the people running it.
LAYER 1 — COMPUTE / CHIPS
The narrowest chokepoint in the entire economy. If you cannot get silicon, nothing above this layer happens.
Jensen Huang · Co-founder & CEO · Nvidia
Huang runs the company whose GPUs are the default substrate of the AI boom; Nvidia crossed a $5 trillion ⚑ unverified valuation and booked roughly $215.9 billion ⚑ unverified in fiscal-2025 revenue, about $62 billion ⚑ unverified of it data-center in a single recent quarter. He beneficially owns roughly 3.58% ⚑ unverified of Nvidia — about 870.6 million shares ⚑ unverified — a stake that has carried his net worth past $200 billion ⚑ unverified. His lever is allocation: in a market where advanced GPUs are supply-constrained, Nvidia decides who is at the front of the queue, and that decision shapes which labs and clouds can even compete.
Lisa Su · Chair & CEO · AMD
Su is the only credible Western challenger to Nvidia's accelerator dominance, and AMD's data-center segment posted around $5.8 billion ⚑ unverified in a recent quarter with MI300-class parts driving most of it. The MI400 series and "Helios" rack platform — plus multi-gigawatt commitments tied to OpenAI and Meta ⚑ unverified — are the wedge for AMD's stated goal of double-digit accelerator share within a few years ⚑ unverified. Her lever is optionality: AMD is the pressure valve that keeps buyers from being fully hostage to one vendor, which is why hyperscalers keep funding a second source.
C.C. Wei · Chairman & CEO · TSMC
Wei runs the true bottleneck. TSMC fabricates the leading-edge logic for Nvidia, AMD, Apple, Broadcom's custom chips and most of the rest, and its CoWoS advanced-packaging capacity — the specific step that assembles GPU and memory into a finished accelerator — is "sold out through 2026" ⚑ unverified by his own account, with plans to scale toward roughly 130,000–140,000 units ⚑ unverified of monthly capacity. His lever is manufacturing itself: there is no second foundry at this node, so TSMC's packaging schedule is, in effect, the industry's clock.
Hock Tan · President & CEO · Broadcom
Tan has made Broadcom the arms dealer of custom silicon — designing the ASICs that let Google (TPU), Meta (MTIA) and now OpenAI ("Jalapeño") build accelerators that route around Nvidia. Broadcom guided to roughly $56 billion ⚑ unverified in AI-semiconductor revenue for fiscal 2026 and has floated a line of sight past $100 billion ⚑ unverified in 2027. His lever is the exit ramp: every hyperscaler that wants to escape Nvidia's margins has to come through Broadcom's design teams to do it.
ASML (Christophe Fouquet, CEO) · Netherlands
ASML holds the most absolute monopoly in the entire stack: it is the sole maker of the extreme-ultraviolet (EUV) lithography machines required to print leading-edge chips, each costing hundreds of millions of dollars ⚑ unverified. Without ASML there is no TSMC leading edge, no Nvidia, no boom. Its lever is a physical chokepoint upstream of everyone — and it is also the pressure point export-control policy squeezes hardest, because denying EUV to a country freezes its advanced-chip ambitions in place.
The memory makers — SK Hynix, Samsung, Micron
High-bandwidth memory (HBM) is the co-bottleneck alongside packaging: GPUs are useless without it, and the three suppliers are sold out through 2026 at elevated prices ⚑ unverified. SK Hynix leads with roughly 56% ⚑ unverified share in early 2026, Samsung holds a bit over a quarter, Micron the rest, with SK Hynix reportedly committing much of its supply to Nvidia ⚑ unverified. Their collective lever is scarcity of a component no accelerator ships without — which is why memory pricing now swings the economics of the whole industry.
LAYER 2 — HYPERSCALERS / CAPEX
The buyers. They convert cash into compute at a scale no one has attempted before — and their spending is the demand the entire chip layer is betting on.
Satya Nadella · Chairman & CEO · Microsoft
Nadella turned an early, aggressive OpenAI partnership into Azure's AI franchise and is steering Microsoft toward roughly $190 billion ⚑ unverified in 2026 capex. Microsoft is both a top Nvidia customer and a designer of its own Maia silicon, and it sits inside the AI Infrastructure Partnership financing vehicle. His lever is capital-plus-distribution: Microsoft can fund a lab, host it, and push its models to enterprise desktops worldwide — a full-stack grip few can match.
Andy Jassy · President & CEO · Amazon
Jassy is guiding Amazon toward the largest capex line of the group — around $200 billion ⚑ unverified in 2026 — anchored on AWS, custom Trainium/Inferentia chips, and a multibillion-dollar stake in Anthropic ⚑ unverified. His lever is the default-cloud position: a huge share of the world's AI workloads still runs on AWS, and Amazon's chip program is a direct bid to own the economics rather than rent them from Nvidia.
Sundar Pichai · CEO · Alphabet / Google
Pichai commands the only company that is fully vertically integrated in AI — its own frontier lab (DeepMind), its own accelerator (TPU, via Broadcom), its own cloud, and its own distribution through Search and Android — with 2026 capex guided around $175–185 billion ⚑ unverified. His lever is independence: Alphabet is the one player that needs nothing from Nvidia, OpenAI, or Microsoft to field a frontier model, which makes it the structural counterweight to the whole OpenAI-Nvidia axis.
Mark Zuckerberg · Founder, Chairman & CEO · Meta
Zuckerberg controls Meta through supervoting stock ⚑ unverified, which means his roughly $115–135 billion ⚑ unverified 2026 capex plan answers to essentially no one. He has spent it on open-weight Llama models, enormous GPU clusters, and aggressive talent raids. His lever is unaccountable capital married to open distribution: by releasing weights, Meta can undercut the paid-API business models of rival labs at will, a strategic weapon no other hyperscaler wields.
Larry Ellison · Co-founder, Chairman & CTO · Oracle
Ellison's Oracle reinvented itself as the compute landlord of the AI boom, renting infrastructure to OpenAI and others; cloud-infrastructure revenue jumped around 93% ⚑ unverified in a recent quarter, and Oracle is a lead partner in the $500 billion ⚑ unverified Stargate buildout. Ellison's own fortune swung by more than $100 billion ⚑ unverified in 2026 on Oracle's volatility, a vivid measure of how leveraged he is to this one bet. His lever is contracted capacity: Oracle's backlog is a claim on future compute that helps make the whole demand story look real.
The neoclouds — Michael Intrator · Co-founder & CEO · CoreWeave
CoreWeave is the archetype of the "neocloud" — GPU-only data-center operators that exist to rent Nvidia capacity. It reported roughly $2.08 billion ⚑ unverified in Q1-2026 revenue with a backlog near $100 billion ⚑ unverified, took a $2 billion ⚑ unverified Nvidia investment, and signed multibillion-dollar deals with OpenAI (~$22.4 billion) ⚑ unverified and Meta (~$14–21 billion) ⚑ unverified. Its lever is speed-to-capacity — but it is also the single clearest node of circular financing, since Nvidia funds the buyer that buys Nvidia's chips, which is precisely why the neoclouds are where skeptics look for the first crack.
LAYER 3 — AI MODEL LABS
The names the public knows. They hold the weights — the actual trained models — but almost none of the compute or capital underneath them. That dependence is the whole drama.
Sam Altman · Co-founder & CEO · OpenAI
Altman runs the lab that defined the era with ChatGPT and is reportedly heading toward an IPO at a valuation that could top $1 trillion ⚑ unverified, after raising tens of billions from SoftBank, Nvidia, Amazon and MGX ⚑ unverified. Famously, he holds little or no direct equity ⚑ unverified; his lever is not ownership but orchestration — he is the central dealmaker of the boom, simultaneously OpenAI's product visionary, its chief fundraiser, and the counterparty on the Stargate, Oracle, Broadcom and chip commitments that anchor the demand side of the entire market. If any one person is the keystone of the circular-financing arch, it is Altman.
Dario Amodei · Co-founder & CEO · Anthropic
Amodei runs OpenAI's closest capability rival and its philosophical opposite — a safety-forward lab that nonetheless raised a reported $65 billion ⚑ unverified round near a ~$965 billion ⚑ unverified valuation and filed to go public, backed heavily by Amazon and Google ⚑ unverified. His lever is frontier capability with an enterprise-and-safety posture that wins the customers and regulators wary of OpenAI, plus a seat at the policy table (the G7, Senate testimony) that shapes the rules everyone else must follow.
Demis Hassabis · Co-founder & CEO · Google DeepMind
Hassabis directs the research arm of the only fully self-sufficient AI company, giving him frontier models (Gemini) without the existential compute-and-capital dependence that shadows OpenAI and Anthropic. A Nobel laureate for AlphaFold ⚑ unverified, he commands unmatched scientific credibility. His lever is capability backed by Alphabet's balance sheet and TPUs — he can pursue the frontier without asking Jensen Huang or Masayoshi Son for permission, a freedom no independent lab enjoys.
Meta AI (Alexandr Wang, Chief AI Officer) & the open-weight camp
Meta's superintelligence effort — with Scale AI founder Alexandr Wang installed atop it after a large acqui-hire ⚑ unverified — is the open-weight wildcard. Its lever is Zuckerberg's capital plus a release strategy that gives models away, dragging down the price of everyone else's proprietary API and setting the floor the closed labs must price above.
Mistral (Arthur Mensch, Co-founder & CEO) · France
Mistral is Europe's flagship lab and the standard-bearer for sovereign, open-weight AI outside American control. Smaller in capital than the US giants, its lever is geopolitical: it is the model European governments and firms can adopt without routing their data and dependence through Washington's tech stack — which gives it influence disproportionate to its balance sheet.
LAYER 4 — GOVERNMENT / POLICY / STATE
The lever nobody voted to give them. States now decide which chips cross which borders — which makes them market actors, not just referees.
Michael Kratsios · Director · White House Office of Science and Technology Policy (OSTP)
Kratsios, Senate-confirmed to lead OSTP ⚑ unverified, co-authored the administration's AI Action Plan — the deregulatory blueprint that fast-tracks data-center permitting and pushes federal AI adoption ⚑ unverified. His lever is the federal policy framework itself: the plan sets the default posture (build fast, regulate little) that shapes where and how quickly the physical buildout happens.
David Sacks · former White House AI & Crypto Czar (resigned March 2026) ⚑ unverified · now PCAST co-chair
Even after stepping down from the czar role, Sacks defined the era's signature policy reversal: reopening advanced-chip sales to China and brokering the deal letting the UAE buy on the order of 500,000 advanced chips ⚑ unverified. His tenure drew conflict-of-interest scrutiny given extensive AI and crypto holdings ⚑ unverified. His lever was — and through PCAST partly remains — the export valve: the single most consequential control in the stack is who gets to buy the chips, and that is a government decision.
The export-control apparatus (Commerce / BIS)
Beneath the named principals sits the Bureau of Industry and Security, which writes and enforces the chip-export rules. Its lever is the licensing regime on Nvidia/AMD accelerators and ASML EUV tools — the mechanism that can, with a rule change, wipe billions off a chipmaker's China revenue or hand an ally a strategic AI campus. This is the clearest proof that the state is now a price-and-allocation actor inside the market, not outside it.
Sovereign players — UAE (MGX / G42) and Saudi Arabia (PIF / HUMAIN)
Gulf states have become first-order capital and demand actors. Abu Dhabi's MGX closed a ~$49 billion ⚑ unverified AI fund, aiming past $100 billion ⚑ unverified AUM, and has taken positions in OpenAI, Anthropic and xAI while co-owning the AI Infrastructure Partnership and Stargate ⚑ unverified. Saudi Arabia's HUMAIN signed roughly $23 billion ⚑ unverified of deals with Nvidia, AMD, AWS and Qualcomm. Their lever is patient sovereign capital plus privileged, policy-granted chip access — they can finance the buildout at a scale and time-horizon private markets cannot, which is why every lab and cloud now flies to the Gulf.
LAYER 5 — FINANCE / CAPITAL
Someone has to pay for the concrete, the transformers, and the GPUs before the revenue shows up. These are the people arranging the money — and the ones deciding when to stop.
Masayoshi Son · Chairman & CEO · SoftBank
Son is the boldest financier of the boom: chairman of the $500 billion ⚑ unverified Stargate project, a reported ~11% ⚑ unverified owner of OpenAI after roughly $41 billion ⚑ unverified of investment, plus tens of billions more in Ampere, robotics, French data centers and energy ⚑ unverified. His lever is willingness to write the largest, riskiest checks — Stargate's demand story leans on SoftBank's balance sheet, which means Son's risk appetite is load-bearing for the whole edifice. If he flinches, the biggest single buildout commitment wobbles.
Larry Fink · Chairman & CEO · BlackRock
Fink turned the world's largest asset manager into an AI-infrastructure financier: the AI Infrastructure Partnership (with Microsoft, Nvidia, MGX) targets up to ~$30 billion ⚑ unverified of equity, and a BlackRock-led consortium bought Aligned Data Centers at a ~$40 billion ⚑ unverified enterprise value. His lever is the pipe between ordinary savings and AI concrete — Fink is explicit that retirement and index money will fund data centers ⚑ unverified, which quietly makes the general public a leveraged stakeholder in the outcome.
The mega-VCs and private-credit / infra funds
Beneath the named giants sits a layer of venture and private-credit capital — the funds writing growth rounds into labs and, increasingly, the debt financing the physical buildout as hyperscalers move data-center spend off balance sheet. Their lever is the marginal dollar: they set the terms and the appetite that determine whether the next gigawatt gets built, and private credit's exposure is now a channel through which an AI stumble could transmit into the broader financial system.
The circular-financing dealmakers
The most important financial fact of the era is not a person but a structure: an estimated $800 billion-plus ⚑ unverified web of reciprocal deals in which Nvidia invests in OpenAI, OpenAI commits to Oracle and CoreWeave, and those firms buy Nvidia chips — money that leaves one balance sheet and returns as another's "revenue." The people arranging these loops (Huang, Altman, Ellison, Son and their deal teams) hold a subtle, dangerous lever: they can make demand look real by financing it, and the honest question no dashboard answers cleanly is how much of the reported growth is genuine end-demand versus capital going in a circle.
LAYER 6 — THE SKEPTICS / SHORT SIDE
The only people in this map whose lever is being right. They hold no chips and no capital — just a thesis, a position, and a willingness to say the number is wrong.
Michael Burry · Founder · Scion Asset Management
Burry — of The Big Short — has made the AI bubble his signature bet, reportedly concentrating as much as 80% ⚑ unverified of Scion's disclosed book into put options on Nvidia and Palantir with underlying notional near $1.1 billion ⚑ unverified. Disclosures point to puts on roughly 1,000,000 Nvidia shares (≈$110 strike, 2027) ⚑ unverified and a layered Palantir short ⚑ unverified; he frames Palantir's fair value in "single digits to low double digits" ⚑ unverified and says he is shorting the business model's unsustainability, not just its multiple. His lever is credibility-as-catalyst: when the analyst who called 2008 puts real money against the trade and publishes the math, he moves the narrative — the one countervailing force to the boom's self-reinforcing story.
Jim Chanos · Founder · Kynikos / Chanos & Co.
The veteran short-seller (famous for Enron) calls today's AI-capex cycle "the same setup" as the 2000 dot-com and telecom bust, arguing the data-center buildout rests on the same demand myth and predicting record equity issuance to fund it ⚑ unverified. His nuance is sharp and worth noting — he favors "the chip magic" and is bearish on the data-center trade specifically ⚑ unverified. His lever is a rigorous, historically-grounded short thesis that separates the durable winners from the financing-driven froth.
Gary Marcus · AI scientist and critic
Marcus attacks the boom at the technical root rather than the financial one, arguing that large language models lack the world-models needed for reliability and that the economics therefore cannot hold ⚑ unverified. His lever is intellectual: he supplies the capability-skeptic case that undergirds the financial bears, and if he is right about the ceiling, the demand assumptions above collapse from the top down.
The institutional warnings — BIS, and the MIT "GenAI Divide"
Not all skepticism is a trade. The Bank for International Settlements flagged an AI-capex bust, opaque circular financing, and record sovereign debt as interlocking systemic risks ⚑ unverified, and an MIT study cast doubt on how much enterprise value generative AI is actually delivering in daily workflows ⚑ unverified. Their lever is legitimacy: when central-bank economists and MIT researchers, not just short-sellers, put the bubble question in writing, it changes what regulators and allocators are willing to say out loud.
Elon Musk — the cross-layer player
Everyone else holds one or two levers. Musk is reaching across the whole board.
Musk is the singular figure who operates at every layer at once, which is why he does not fit inside any single one.
- Models (Layer 3): He founded and controls xAI, maker of the Grok models (Grok 4.3, with a reported ~117 million monthly active users) ⚑ unverified — a genuine frontier lab competing directly with OpenAI, the company he co-founded and now litigates against.
- Compute (Layers 1–2): He owns Colossus, described as the world's largest AI supercomputer at 200,000+ Nvidia GPUs, with a "Colossus 2" targeting up to 1 million GPUs and gigawatt-scale power ⚑ unverified. That makes Musk not just a model-maker but a compute baron in his own right — he builds the data centers rather than renting them.
- Capital / corporate structure (Layer 5): xAI was absorbed by SpaceX in a February 2026 all-stock deal valuing xAI at ~$250 billion ⚑ unverified and the combined entity near ~$1.25 trillion ⚑ unverified, with a subsequent SpaceX listing reportedly valuing the parent above $2 trillion ⚑ unverified — even as xAI reportedly burned on the order of $1 billion ⚑ unverified a month against a fraction of that in revenue.
- Distribution & narrative: Through his ownership of X, he controls a global distribution platform for both Grok and his own messaging.
- Policy (Layer 4): His proximity to political power — and his public feuds with it — give him informal influence over the regulatory environment that no other lab founder has held.
No one else combines model, compute, capital, distribution and political reach in one person. That concentration is either the boom's most efficient vertical integration or its most dangerous single point of failure — and it is held by its most volatile operator.
The Concentration
What the whole roster reveals when you stand back from it.
Money, compute, and models are held by different hands — and that is the only real check in the system. The capital (Son, Fink, the Gulf funds) is not the compute (Huang, Wei, the hyperscalers) is not the weights (Altman, Amodei, Hassabis). Each layer depends on the one below it and can be starved by it. The labs everyone fears hold the least durable position of all: they own brand and weights but rent their chips and their capital, which is why they spend their days fundraising rather than dictating terms.
The true chokepoints are astonishingly narrow. One company packages the chips (TSMC/CoWoS). One company makes the machines that make the chips (ASML/EUV). Three companies make the memory, one of them dominant. A rule change in one government office reroutes the whole flow. The public conversation fixates on the model labs, but the unbreakable monopolies sit two layers below them, in Taiwan and the Netherlands — and they are physical, not digital.
A single vendor sits at the center of the money-go-round. Nvidia sells the chips, invests in the buyers, and books the revenue when they spend it back — the pattern repeated with CoreWeave, OpenAI, and the neoclouds. An estimated $800 billion-plus ⚑ unverified of reciprocal financing means a meaningful share of reported "demand" may be capital circulating rather than end-customers paying. That is the boom's load-bearing ambiguity, and no one on the inside is incentivized to resolve it.
The state stopped being a referee and became a player. Export licenses now determine which companies and which countries can buy compute at all. Sovereign funds are among the largest single buyers of AI equity and infrastructure on Earth. The result is that "market" outcomes at the top of the stack are increasingly policy outcomes — set in Washington, Abu Dhabi, and Riyadh as much as in Santa Clara.
Who benefits most is not who you'd guess. Not the model labs, which burn cash and depend on everyone. The steadiest winners are the pick-and-shovel monopolists insulated from which model "wins" — TSMC, ASML, the memory makers, Broadcom, and Nvidia itself — because they get paid whether or not any given lab's economics ever close. The most exposed are the labs, the neoclouds, and the private-credit financing the concrete.
The unwind risk is global, not sectoral. Because the same names recur across chips, clouds, labs, capital, and sovereign funds, and because ordinary savings are now piped into the buildout through index and private-credit channels, a stumble at any layer transmits outward. The skeptics — Burry, Chanos, Marcus — and institutions like the BIS are all describing versions of the same fear: that a structure this concentrated, this reciprocally financed, and this dependent on demand that hasn't fully arrived is not a contained bet but a systemic one. The levers are held by very few hands. That is exactly why it matters when even one of them lets go.
Every figure, stake, valuation, and hard claim in this piece is marked ⚑ unverified and pending fact-check. Titles and roles reflect 2025–2026 reporting and are subject to change; attributions describe positions as disclosed or reported, not asserted fact.