{
"version": 1,
"stories": [
{
"id": "circular",
"title": "The circular financing machine",
"tagline": "NVIDIA funds the companies that buy its chips - and the hyperscalers underwrite the middlemen.",
"steps": [
{
"view": "atlas",
"groups": {
"supply": false,
"capital": true,
"owns": false,
"rivals": false
},
"fit": true,
"say": "Money is the hidden wiring of the AI boom. Turn on just the capital edges and the map changes character: every amber line is money flowing - short dashes are equity investments, long dashes are capacity commitments. Notice how many of them radiate from a single company in the silicon band."
},
{
"view": "atlas",
"focus": "nvidia.1",
"say": "NVIDIA is the sun of this solar system. Its investments reach labs (SSI, Thinking Machines, Reflection, Cohere), inference clouds, robotics - and crucially the neoclouds. Many of these companies take NVIDIA's money and hand it straight back as GPU orders. That's the circle: vendor financing at industrial scale."
},
{
"view": "atlas",
"focus": "coreweave.1",
"say": "CoreWeave shows the other half of the loop. Microsoft and Meta don't just buy from it - they commit capacity spend (the long-dash lines), effectively underwriting the datacenters that CoreWeave builds to rent NVIDIA GPUs. It also bought Weights & Biases, and its storage runs on a $1.17B VAST Data deal."
},
{
"view": "atlas",
"focus": "fluidstack.1",
"say": "Fluidstack is the newest - and starkest - example. Anthropic made it the primary partner in a $50B US datacenter build-out, while Google backstops $1.8B of its leases (and negotiates equity) to spread TPUs. A company most people have never heard of, load-bearing for two giants at once."
},
{
"view": "atlas",
"groups": {
"supply": false,
"capital": true,
"owns": false,
"rivals": false
},
"fit": true,
"say": "Why it matters: when the same dollars cycle from chipmaker to cloud to model lab and back, revenue signals get circular too. This map keeps equity and capacity commitments as separate edge types precisely so you can trace which is which. Click any amber line for its source."
}
]
},
{
"id": "prompt",
"title": "How a prompt becomes an answer",
"tagline": "Follow one request from the app you type into, down to the lithography machines that made it possible.",
"steps": [
{
"view": "inference",
"fit": true,
"say": "This view is the inference stack - the same dataset, sliced as: apps at the top, then agents and gateways, model labs, serving providers, hyperscale clouds, GPU neoclouds, and silicon at the bottom. A prompt you type falls all the way through it in under a second."
},
{
"view": "inference",
"focus": "cursor.1",
"say": "Start where you type. Cursor, the AI code editor, routes your keystrokes to frontier models - Anthropic and OpenAI among them - and serves parts of its stack through Baseten. Apps like this are heavy inference buyers; they own the user, not the GPUs."
},
{
"view": "inference",
"focus": "openrouter.1",
"say": "Some traffic goes through a router. OpenRouter owns no GPUs at all - it's a switchboard across hundreds of models, moving 25 trillion tokens a week. Gateways like this (and Vercel's AI Gateway) turn 'which model?' into a runtime decision."
},
{
"view": "inference",
"focus": "together.1",
"say": "Serving providers do the actual computing. Together, Fireworks, Baseten, Modal and friends load open-weight models onto rented clusters and sell tokens. Notice their rents from edges down to the neoclouds - they're tenants, not landlords."
},
{
"view": "inference",
"focus": "coreweave.1",
"say": "The landlords: neoclouds. CoreWeave, Nebius, Crusoe, Lambda build and operate the GPU datacenters everyone above rents. Their edges point down to one company: nearly every accelerator here is NVIDIA's."
},
{
"view": "inference",
"focus": "tsmc.1",
"say": "And the floor under everything: silicon. NVIDIA designs, but TSMC fabricates - for NVIDIA, AMD, Google's TPU, Amazon's Trainium, Qualcomm, Etched. One foundry, fed by ASML's EUV monopoly and HBM from SK Hynix, Micron and Samsung. Your prompt ends here, in a cleanroom in Taiwan."
}
]
},
{
"id": "chokepoints",
"title": "The three chokepoints",
"tagline": "EUV lithography, leading-edge fabrication, and high-bandwidth memory - where the whole industry narrows to a point.",
"steps": [
{
"view": "atlas",
"groups": {
"supply": true,
"capital": false,
"owns": false,
"rivals": false
},
"fit": true,
"say": "The AI industry looks wide at the top and is terrifyingly narrow at the bottom. Hundreds of companies up in the application bands; a handful of irreplaceable ones in the semiconductor supply chain. Three of them are true chokepoints."
},
{
"view": "atlas",
"focus": "asml.1",
"say": "Chokepoint one: ASML. The only company on Earth that makes EUV lithography machines - the tools that print leading-edge chips. Every fab that matters is equipped by it. No substitute exists at any price; a single export rule about ASML reshapes the whole map."
},
{
"view": "atlas",
"focus": "tsmc.1",
"say": "Chokepoint two: TSMC. Nearly every leading AI chip - NVIDIA, AMD, Google TPU, Trainium, Etched - is fabricated here. Samsung is the only credible rival at the edge, and the gap is real. This is why a single island's geography is a systemic risk for the entire industry."
},
{
"view": "atlas",
"focus": "skhynix.1",
"say": "Chokepoint three: high-bandwidth memory. A GPU is only as fast as the memory feeding it, and HBM comes from exactly three companies - SK Hynix in the lead, Micron and Samsung behind. The 2024–25 HBM repricing multiplied these companies' values several-fold; the map's buys memory from edges show who depends on them."
},
{
"view": "atlas",
"groups": {
"supply": true,
"capital": false,
"owns": false,
"rivals": false
},
"fit": true,
"say": "The takeaway: every trillion-dollar valuation in the upper bands rests on three narrow gates near the bottom. When you read AI news - export controls, capex announcements, shortages - check it against these three nodes first."
}
]
},
{
"id": "ownership",
"title": "Who owns whom",
"tagline": "Divisions, acquisitions, and the reason this map never values a subsidiary separately.",
"steps": [
{
"view": "atlas",
"groups": {
"supply": false,
"capital": false,
"owns": true,
"rivals": false
},
"fit": true,
"say": "Turn on only the ownership edges and the map becomes an org chart. Dashed-outline nodes are divisions: real products, deliberately unsized - because a chip programme inside a trillion-dollar company has no separable valuation, and pretending otherwise is how datasets lie."
},
{
"view": "atlas",
"focus": "alphabet.1",
"say": "Alphabet is the classic conglomerate case. DeepMind builds the models, the TPU programme builds the silicon, Waymo drives the cars, Google Cloud serves it all. One market cap covers everything - so that's the only number this map shows."
},
{
"view": "atlas",
"focus": "coreweave.1",
"say": "Acquisitions become divisions here. CoreWeave bought Weights & Biases in 2025 ($1.4B per the SEC filing - rumors said $1.7B). The moment a company is absorbed, its bubble loses its size and joins its parent: the map shows structure, not stale price tags."
},
{
"view": "atlas",
"focus": "marvell.1",
"say": "The same rule caught two more in 2026. Marvell closed its Celestial AI acquisition in February (~$3.25B), and Cloudflare quietly picked up Replicate for $57M - a fraction of the $350–550M the rumor mill claimed. The SEC filing beat the press; the map records the filing."
},
{
"view": "atlas",
"focus": "xai.2",
"say": "And the strangest structure of all: xAI. Merged with SpaceX at a combined $1.25T, with Grok riding on the Colossus cluster - while SpaceX also acquired Cursor for $60B. Corporate boundaries in AI are moving faster than most org charts can be redrawn; the ownership edges are where this map keeps score."
}
]
}
]
}