Who really builds AI — and the insiders' headstart

01Who builds AI

Let me answer the plain question first, because it is the one people search for and the one most coverage dances around. The senior operators of the major AI companies — the ones whose decisions move the whole field — are, as of mid-2026:

• Sam Altman, CEO and co-founder of OpenAI.
• Dario Amodei, CEO and co-founder of Anthropic.
• Elon Musk, who controls xAI, Tesla, SpaceX and Neuralink.
• Demis Hassabis, CEO of Google DeepMind, under Sundar Pichai at Alphabet.
• Mark Zuckerberg, CEO of Meta, with Alexandr Wang as chief AI officer of Meta Superintelligence Labs.
• Jensen Huang, founder and CEO of Nvidia — the company that sells the shovels to everyone else.
• Satya Nadella, chairman and CEO of Microsoft.
• Alex Karp, co-founder and CEO of Palantir.
• Jeff Bezos, founder of Amazon, funding Blue Origin and a new AI venture, Prometheus.
• Ilya Sutskever, co-founder and CEO of Safe Superintelligence.
• Mira Murati, founder and CEO of Thinking Machines Lab.
• Liang Wenfeng, founder of DeepSeek, China's most disruptive lab.

That is the power set. Twelve people, a handful of companies, and a supply chain that runs all the way down to a single Dutch company — ASML — that makes the lithography machines every advanced AI chip is printed on. Understanding these individuals is not celebrity-watching. It is the fastest way to understand where the technology, and a good chunk of the world's capital, is going next.

02Why a map, and not a list

A list tells you who exists. It does not tell you that Google funds Anthropic and rents compute from a rocket company, or that Nvidia invests in the very customers who buy its chips, or that roughly a trillion dollars of announced AI deals circle around inside one small cohort. Those are relationships, and relationships are where the meaning lives. A database stores rows; a blockchain stores events; neither understands relationships. An ontology does — it treats every fact as a typed, sourced claim connecting one thing to another, each carrying a source, a date and a confidence.

So we did not argue this in the abstract. We built it. The KXCO Ontology Engine renders the AI sector as objects and the verifiable claims between them — roughly a hundred entities, from ASML and TSMC up through the chipmakers, clouds, labs, investors and the twelve people above. This week we added the leadership layer: each person is a node, wired by typed control, capital and supply edges to the empires they run. Everything I assert below about who controls what, who funds whom, and who depends on whom is a claim you can click on in that map and check for yourself. That is the standard I want analysis of AI held to, and it is the standard I am holding myself to here.

A value is a typed claim, not a bare number. "Nvidia → OpenAI, $100bn" means nothing until it carries its source, its date, and how sure we are.

For the deeper argument on why this modelling discipline matters — why meaning, not compute, is the real bottleneck for the coming agent economy — see the companion essay Ontology is the missing layer in agentic AI. Here, the ontology is simply the instrument: the lens that turns a dozen personalities into a structure you can reason about.

03The power set, one by one

Each profile below follows the same four lines — core themes, what they are building, what worries them, and, where it is sourced, health and longevity, because the same people building AI are conspicuously investing in living long enough to see it finished.

Elon Musk — the physical-world bet

Core themes: the only figure here whose AI thesis is about atoms, not just bits — intelligence becomes cheap, so the constraint moves to robots, energy and getting off the planet. Building: xAI merged into SpaceX in early 2026, folding the Colossus supercomputers into the same entity that builds Starship — a company that has since completed its own IPO and now trades as SPCX; Tesla is pushing Optimus toward mass production; and Neuralink had implanted twelve people by September 2025 and is moving to automated robotic surgery and a visual-cortex implant, Blindsight. Worries: he has put a rough one-in-five figure on civilisational catastrophe from AI, while arguing that building it himself is safer than ceding it. Longevity: publicly optimistic that ageing is solvable, though he channels it into species-level survival rather than personal biohacking.

Sam Altman — the abundance salesman

Core themes: relentless optimism about abundance — intelligence and energy both trending toward near-free. Building: OpenAI sits at the centre of the largest infrastructure mobilisation in corporate history — the $500 billion Stargate joint venture with SoftBank and Oracle, up to $100 billion from Nvidia paid in GPUs, and a $300 billion Oracle compute deal. His most revealing bets are biological: he seeded Retro Biosciences with $180 million, now valued around $1.8 billion and running its first anti-ageing trial, and co-founded the brain-computer interface startup Merge Labs, putting him head-to-head with Musk's Neuralink. Worries: on record since 2015 that AI could be "the greatest threat to the continued existence of humanity." Longevity: diet, exercise, sleep and metformin — plus the biggest cheque of anyone here written directly at extending human lifespan.

Dario Amodei — the safety-first accelerationist

Core themes: AI will be extraordinarily powerful, and precisely because of that must be built carefully. His essay "Machines of Loving Grace" describes "a country of geniuses in a datacentre" compressing a century of biomedical progress into five to ten years. Building: Anthropic's Claude models, funded by up to $10 billion from Nvidia and a $30 billion Azure compute commitment; notably, a large and growing share of Anthropic's own code is now written by AI. Worries: catastrophic misuse — biological weapons, cyber, loss of control — and a society "dangerously underprepared." Longevity: private, but driven by a family illness that became treatable just too late.

Demis Hassabis — the scientist

Core themes: a 2024 Nobel laureate who frames AGI as an "infinity machine" still five to ten years out, because genuine scientific creativity remains unsolved. Building: Google DeepMind's Gemini models and Isomorphic Labs, the AlphaFold-derived drug-discovery company now moving toward clinical trials, inside Alphabet's ~$180 billion 2026 capital budget. Worries: that too much responsibility rests on too few people; he has called for a "CERN for AI."

Jeff Bezos — industry off-Earth

Core themes: century-scale — move heavy industry into space, treat AI as an engine of higher living standards. Building: Blue Origin, now raising outside capital for the first time; a new AI venture, Prometheus, aimed at an "artificial general engineer"; and, through Amazon, the single largest 2026 capital line at ~$200 billion. Longevity: a structured fitness regimen, and an investor in Altos Labs, the ~$5 billion cellular-reprogramming firm.

Mark Zuckerberg & Alexandr Wang — spend through the doubt

Core themes: "personal superintelligence" for everyone, owned end-to-end by Meta; out-spend and out-hire the field. Building: a stated $600 billion in US infrastructure through 2028 — including a five-gigawatt Louisiana data centre and a gas-powered Ohio site — plus a pivot to sell excess compute as a cloud. Zuckerberg installed Alexandr Wang, founder of Scale AI, as chief AI officer. Longevity: Zuckerberg funds disease research through CZI and predicts living past 100 will be normal by century's end.

Jensen Huang — the man who sells the shovels

Core themes: we are building "AI factories," and demand for them is structural, not a bubble. Building: Nvidia designs roughly 80% of data-centre AI compute by revenue and has taken strategic stakes in its own customers — up to $100 billion into OpenAI, up to $10 billion into Anthropic, and a stake in xAI. The irony the ontology annotates directly: Nvidia invests in the companies that then spend that money buying Nvidia chips.

Alex Karp — the sovereigntist

Core themes: the field's fiercest internal critic — Western governments and companies must own and control their models, data and compute; outsourcing national security to unaccountable labs is, in his words, indefensible. Building: Palantir's AI Platform and Ontology, deep defence and intelligence contracts, and a battlefield AI layer used within NATO, with Nvidia as a partner for sovereign AI. Longevity: the most disciplined of the group — high-intensity cardio, daily Tai Chi, eight hours of sleep, very low body fat.

Satya Nadella — the platform operator

Core themes: AI as the platform shift of the era; Microsoft must own the layer businesses run on. Building: the OpenAI partnership and a vast Azure buildout, hedged with in-house Microsoft AI models and custom Maia silicon so Microsoft is never wholly dependent on any single lab.

Ilya Sutskever & Mira Murati — the purists

Two of OpenAI's most important alumni have gone their own way. Ilya Sutskever, the former chief scientist, founded Safe Superintelligence with a single stated goal and no intention of shipping interim products, raising billions at no revenue. Mira Murati, OpenAI's former CTO, founded Thinking Machines Lab, gathering ex-OpenAI talent and raising one of the largest seed rounds on record. Both are bets that the frontier can be pushed by small, focused teams rather than only by hyperscalers.

Liang Wenfeng — the disruptor

Liang Wenfeng and DeepSeek are the reason no one in this group can relax. Funded off his quantitative trading fund, DeepSeek has repeatedly delivered frontier-level results at a fraction of the compute and cost, releasing open weights as a strategy — the single strongest piece of evidence that the moat around the frontier is thinner than the incumbents would like.

04What they are actually building

Step back from the personalities and a single shape emerges. The money is not chasing a clever algorithm. It is chasing infrastructure — and everything physical that infrastructure needs.

• Compute and chips. Combined hyperscaler capital expenditure is on track for roughly $700 billion in 2026 — Amazon near $200 billion, Alphabet $175–185 billion, Meta $115–135 billion — on top of the $500 billion Stargate project. Nvidia sells the chips; TSMC fabricates them; ASML makes the machines that make them possible. Whoever controls this stack controls the pace of everything above it.
• Foundation models. OpenAI, Anthropic, Google DeepMind, xAI, Meta and — from China — DeepSeek and the rest, in a race where the leading edge keeps moving but the gap to the cheap, open alternatives keeps narrowing.
• Energy. The quiet chokepoint. AI's power demand is restarting nuclear plants and prompting five-gigawatt campuses and gas plants built specifically to feed models. Energy, not talent, may be the binding constraint of the next five years.
• Robotics. Tesla's Optimus and companies like Figure AI push humanoid robots toward production — Musk's wager that the payoff from cheap intelligence is cheap physical labour.
• Defence. Palantir under Karp has made "AI for national security" its whole identity, with a battlefield layer inside NATO.
• Space. Musk's SpaceX and Bezos's Blue Origin are both, increasingly, AI-infrastructure plays — connectivity, and eventually compute, beyond Earth.
• Biotech and longevity. The pattern the mainstream coverage keeps missing: Altman's Retro Biosciences, Bezos's Altos Labs, Hassabis's Isomorphic Labs, Zuckerberg's CZI. They are using AI to attack ageing itself.
• Brain-computer interfaces. At the far edge, two of the most powerful people in AI — Musk with Neuralink and Altman with Merge Labs — race to connect the brain directly to machines.

Read across those eight sectors and the through-line is unmistakable: compute leads to models, and models lead into the physical world — robots, drugs, rockets, brains — with energy as the constraint everyone is scrambling to secure.

05The trillion-dollar loop

There is one structural feature you cannot see from a list of names, and it is the one that worries me most as an investor. The capital in this sector does not flow in a straight line from backers to builders. It flows in a circle.

Follow it. Nvidia invests up to $100 billion in OpenAI. OpenAI commits hundreds of billions to compute — much of it spent, directly or indirectly, on Nvidia chips. SoftBank and Oracle fund Stargate; Stargate buys Nvidia. Microsoft funds both OpenAI and Anthropic; both spend that money on Azure, which runs on Nvidia. Google funds Anthropic while renting compute from SpaceX, whose data centres are full of Nvidia GPUs. Roughly a trillion dollars of announced deals circulate inside one small cohort, and a striking share ends up back where it started.

This is not necessarily folly — vendor financing is an old strategy — but it means the sector's most important numbers are, to a degree, marking their own homework. When a chipmaker invests in the customer who then reports record chip purchases, demand signals get harder to read from outside. The ontology annotates these loops explicitly — there is a filter on the live map that shows only the circular flows — because they are precisely the kind of relationship a headline reports as growth and a graph reveals as a loop. And the same concentration that gives a dozen people an informational lead in capability gives them a lead in understanding the real economics of this buildout: they are on both ends of the transaction, and the rest of us are reading press releases.

06The insiders' headstart

Here is what I believe, stated as plainly as I can: the people in this article are working with AI that is materially more capable than anything the public can access, and they have it first.

I do not mean a conspiracy. I mean something more mundane and more important. When a lab trains a new frontier model, it does not appear in your chatbot the next morning. It is used internally for months — to write code, to run research, to design the next model — before it is ever released, if it is released at all. The most advanced version of the technology is always the one being used by the people building it, pointed at the problem of building more of it.

The honest version of the theory

I want to be careful here, because this is exactly the kind of claim that gets inflated into nonsense, and I would rather be right than loud.

The strongest evidence comes from METR, an independent evaluation organisation that, in 2026, was given access to the actual internal frontier models of Anthropic, Google, Meta and OpenAI — not the public versions. Their finding cuts both ways. It confirms the headstart is real: the labs do run more capable models internally than they release, and deploy them heavily on their own R&D. And it puts a number on the gap that should keep the theory honest — internally, the frontier was on average only about two months ahead of what the public could document, and none of the internal models were dramatically, secretly superhuman.

So the extreme version — a lab sitting on a hidden AGI years beyond public knowledge — is not supported by the best evidence we have. But the reasonable version is not just supported, it is confirmed by the labs' own transparency framework, which explicitly warns that "developers may have extremely capable internal models that are not shared with the outside world," and treats "how many months behind is your public model?" as a metric worth disclosing.

Two months does not sound like much. But compound it — cycle after cycle, best model building the next — and it is not a gap that closes on its own. It is a gap that widens.

This is why I keep returning to the same conclusion. The advantage is fair — they built it, they paid for it, and in most cases they took real risks to do so. But fair is not the same as safe to ignore. An advantage this concentrated, this compounding, and this invisible to outsiders is precisely the kind of thing a healthy society keeps a close eye on. Not to punish it. To understand it, and to be ready.

07The China question

No honest map of AI power stops at the American coast. The most important structural fact of the last eighteen months is that China has built a near-complete parallel stack: SMIC fabricating chips, Huawei designing them, and a cluster of labs — DeepSeek, Alibaba, ByteDance, Baidu, Tencent, Zhipu, MiniMax, Moonshot — training models that increasingly rival the American frontier. Liang Wenfeng's DeepSeek is the emblem: open-weight models at a fraction of Western training cost, released to erode the incumbents' pricing power.

This changes the headstart argument. Inside the United States, the frontier is concentrated in a dozen people who share, however grudgingly, a common regulatory environment and a common set of independent evaluators. Across the Pacific, a second frontier is advancing that answers to none of those. US export controls on advanced chips and lithography — the ASML and Nvidia chokepoints — are the single biggest lever anyone has over the pace of that second stack, which is why the fight over them is so bitter. The "insiders" are not one group but two, in two systems, and the gap between them is now measured in months, not years. The ontology maps both stacks side by side for precisely this reason: you cannot reason about one without the other.

08Read the map yourself

Everything above is a set of relationships that changes every week. You cannot hold it in a static article, which is the whole reason it lives on a live ontology. The KXCO Ontology Engine now includes these twelve individuals alongside the ~100 companies, chokepoints and capital flows they connect to — and you can interrogate it directly.

Open it and ask the questions built into the engine:

• Who holds a capability headstart the public can't see?
• How far ahead are the labs' internal models?
• Who rents compute from a rival? Who funds a company they compete with?
• What does the whole sector rest on? (Follow the chain down to ASML.)
• How many AI capital loops can you find?

Every answer is a typed claim you can inspect — a subject, a predicate, an object, each carrying a source URL, an as-of date and a confidence level. Where a fact is undisclosed, the engine says so rather than guessing. That is the difference between being told what to think about AI and being able to check it. If you are building agents that need to reason over this kind of structure, the model behind the map is described in Ontology is the missing layer in agentic AI and Chain, Quantum, Ontology, AI.

09What we claim, and what we don't

In the spirit of the ontology, let me be explicit about the boundaries of this piece.

• We claim that a dozen people exert outsized control over AI's direction; that they are building across eight sectors at once; that ~$700 billion of 2026 capex and ~$1 trillion of circular deals are real and largely concentrated; and that frontier labs run more capable models internally than they ship. Each of these is a sourced claim on the live map.
• We do not claim a hidden, generation-ahead superintelligence, a conspiracy, or that any of these people are acting in bad faith. The measured internal-vs-public gap is about two months, and we say so.
• Some figures move. Private valuations (Anthropic, OpenAI, xAI) and deal terms shift week to week; where sources conflict, the ontology carries a confidence level rather than a false precision, and this article defers to it.
• KXCO is software. The ontology is a public reference that models the sector as verifiable claims. It does not make investment recommendations, and nothing here is investment advice.

10What to watch

Three signals, over the next year, will tell you whether the headstart is widening or closing.

1. The IPO wave. SpaceX has already listed — it completed its IPO and now trades as SPCX, the first of the cohort to go public and the first real read on demand. OpenAI and Anthropic are still to come; when they file, we get the numbers that show how much of that $700 billion in spending is validated by demand.
2. The internal-versus-public gap. Watch whether independent evaluators like METR keep getting access, and whether that two-month gap grows. If the labs stop sharing, that itself is information.
3. The challengers. Watch DeepSeek and the open-weight models. Every time they match the frontier for a fraction of the cost, the insiders' headstart shrinks.

The people in this article are not villains, and they are not wizards. They are a small group of extraordinarily capable operators who got to the frontier first and are using every advantage of being there. Understanding them — precisely, not breathlessly — is the first step to making sure the rest of us are not left reacting to a world that a dozen people already finished building. That understanding starts with a map, and the map is live.

11Frequently asked questions

Who are the senior people building AI in 2026?

Sam Altman (OpenAI), Dario Amodei (Anthropic), Elon Musk (xAI/Tesla/SpaceX/Neuralink), Demis Hassabis (Google DeepMind, under Sundar Pichai), Mark Zuckerberg (Meta, with Alexandr Wang as chief AI officer), Jensen Huang (Nvidia), Satya Nadella (Microsoft), Alex Karp (Palantir), Jeff Bezos (Amazon/Blue Origin), Ilya Sutskever (Safe Superintelligence), Mira Murati (Thinking Machines Lab) and Liang Wenfeng (DeepSeek).

Do the people who build AI have access to more powerful AI than the public?

Yes, with a caveat about degree. METR's 2026 evaluation of the labs' actual internal models confirmed they run more capable models internally than they release — but the measured lead is about two months on average, not a hidden, years-ahead superintelligence. The advantage is real and compounding, and worth monitoring.

What is an ontology and why use one to map AI?

An ontology is a formal map of things and the relationships between them, where every fact is a typed claim carrying a source, a date and a confidence. A list of companies cannot show that Nvidia invests in its own customers; a graph can. The KXCO Ontology Engine models the sector this way so every claim can be checked.

How much are AI companies spending in 2026?

Roughly $700 billion of combined hyperscaler capex — Amazon ~$200B, Alphabet ~$175–185B, Meta ~$115–135B — on top of the $500B Stargate project and Nvidia's direct investments.

Which company is most important to AI?

By dependency, three: Nvidia (≈80% of data-centre AI chips by revenue), TSMC (fabrication) and ASML (the sole maker of EUV lithography). The whole sector rests on that spine.