The Country Called NVIDIA: Its Internal and External Skeleton, and Where It Is Taking Us Together With Its Nodes

On 14 August 2026, Nvidia filed a routine document with the Securities and Exchange Commission. Form 13F is an inventory: every listed security an institution holds, with the share count and the value. Nvidia's ran to eight lines and $63,439,974,569.

Read as a portfolio it is unremarkable. Read as a diagram it is a map, and what it maps is not a company. It is a world under construction, with dates on it and most of the money already committed.

Civilisations are not usually legible while they are being built. The decisions that set their shape are taken privately, years before anyone can see the result, and they become visible only as concrete. This is one of the rare occasions when the plan is a public document, because the law requires the people executing it to file what they have bought, what they have promised and what they have guaranteed. The filings are dry. What they describe is not.

The structure has two skeletons. The internal one is the balance sheet: what has been promised, to whom, and in which direction the money runs. The external one is the set of companies it has bought a piece of, which is to say every part it needs in order to exist. This is a reading of both, of how the parts fit together, and of what the finished body does to the ground it stands on.

The internal skeleton

The scale first, because it fixes the order of magnitude. In the quarter to 26 July 2026 Nvidia reported revenue of $96.2 billion, of which $89.0 billion was data centre, net income of $59.7 billion and a gross margin of 75.0 per cent, and guided the following quarter to $108 billion. Over six months: $177.8 billion of revenue and $118.0 billion of profit, against $215.9 billion and $120.1 billion for the whole of the preceding year. Half a year of profit has very nearly matched twelve months.

The money runs in two directions at once.

Upstream, towards the factories, sit $279 billion of supply obligations: $92 billion falling due in the remainder of fiscal 2027, $87 billion in fiscal 2028, $88 billion in fiscal 2029 and $12 billion across fiscal 2030 and beyond. That is capacity bought before it exists, from a manufacturing base that cannot be enlarged quickly. That base reported $40.2 billion of revenue in its own June quarter at a 67.7 per cent gross margin, and has raised 2026 capital spending towards $60 billion to $64 billion, much of it for the advanced packaging that is the real constraint of the industry.

Downstream, towards the buyers, the filing discloses a different class of instrument. In August 2026 Nvidia entered into guarantees, capped at $105 billion, providing credit support on a land, power and shell buildout with affiliates of SB Energy; $3.5 billion more guarantees land, power and shell for AI clouds. Maximum gross exposure, as disclosed: $108.5 billion. Beside it sit $51.2 billion of holdings in privately held companies and $25 billion more committed to, in the filing's own words, "AI model makers, infrastructure financiers, and other private companies".

Note what those guarantees are for. Not chips, which the company makes, and not demand, which it has. Land, power and shell: the three things the structure cannot manufacture for itself.

The external skeleton, bone by bone

Eight listed holdings, as filed. What each one is, what it is actually trying to achieve, and where that points.

Intel: $29,989,261,126. The position is 214,776,632 shares, which at $23.28, the price at which the investment was announced, is $5 billion to the dollar; the private placement completed in December 2025 and the stake has not been touched. At $139.63 a share on 30 June it was worth six times what was paid. What Intel is trying to do is narrower and harder than being a chip company: it is trying to become a foundry that other people use. Its 18A process entered high-volume manufacturing at its Arizona and Oregon plants in late 2025 and now carries its own lead products. The decisive one is the next, 14A, whose defect density and transistor performance are running ahead of 18A at the same stage, with two prospective external customers evaluating test chips and commitments expected from the second half of 2026 into the first half of 2027. Capital spending is above $20 billion this year and planned significantly higher next, the vast majority inside the United States. The two companies are also building x86 processors that speak NVLink directly. Where it points: a second geography for advanced logic. Today the frontier is manufactured in one place; this line in the filing is an attempt to buy a second.

Space Exploration Technologies: $20,975,594,582. SpaceX absorbed xAI on 2 February in an all-share transaction valuing the combined company at $1.25 trillion, then listed on 12 June, selling 555,555,555 shares at $135.00 to raise $75.0 billion, with a thirty-day underwriters’ option on a further 83,333,333 shares that would take the total to about $86.2 billion if taken up in full, the largest flotation ever completed; Nvidia's 122,764,805 shares were worth $170.86 each eighteen days later. One company now holds the launch vehicle, the satellite constellation and a frontier laboratory. Its stated ambition is the one that breaks the frame: it has filed with the Federal Communications Commission for a constellation of up to one million orbital data centre satellites, and projects that placing a million tonnes a year into orbit would yield on the order of 100 gigawatts of computing. The first article, shown in June, is a computing platform flanked by solar arrays and radiators, averaging around 120 kilowatts, to be built at a dedicated factory in Bastrop, Texas; the launch rate has to climb from roughly 2,500 tonnes a year towards a million in about three. Where it points: computation that needs no grid connection, no water, no permit and no sunset.

CoreWeave: $4,699,617,158. On the surface a landlord: it buys Nvidia hardware and rents it out, expecting to spend $30 billion to $35 billion this year, roughly doubling active power to more than 1.7 gigawatts against about 4.2 gigawatts contracted and a backlog reported near $88 billion, with new contracts struck at approximately $40 million per megawatt. Microsoft, OpenAI and Meta are expected to account for more than 80 per cent of this year's revenue, on contracts reported at $22.4 billion, $21 billion through 2032 and similar magnitudes. But its real innovation is financial. It has raised an $8.5 billion facility described as the first investment-grade debt secured on GPU clusters and the customer contracts attached to them, rated A3 and A (low), with a fixed tranche near 5.9 per cent maturing in March 2032, and has since sold a $3.1 billion leveraged loan on the same principle. Where it points: the accelerator becomes a mortgageable asset. Computation joins real estate, aircraft and shipping as something banks will lend against, which is the step that lets the buildout grow faster than anyone's profits.

Coherent: $3,072,195,870. The least visible and among the most constraining. It makes the optics that move data inside and between data centres: transceivers at 800 gigabits and 1.6 terabits, optical circuit switches, co-packaged optics. Quarterly revenue of $1.7 billion, with the datacentre and communications side at $1.2 billion and growing by a third year on year. Its current project is unglamorous and decisive: moving indium phosphide production from three-inch wafers to six, which yields more than four times the devices at less than half the cost, with the first transceivers containing six-inch parts already shipped. Where it points: the limit on how large a single machine can be is no longer the processor but the connection between processors. Push that limit into light and a hall of racks stops being a building full of computers and starts behaving as one computer.

Nokia: $2,209,650,581. The position is 166,389,351 American depositary shares, which at $6.01 is exactly $1 billion; it is now worth more than twice that. What Nokia is attempting is to change what a mobile network is. Future base stations will carry Nvidia's ARC platform, pairing Grace processors with Blackwell accelerators, with pilots this year and commercial availability targeted for 2027, and a path to 6G delivered as a software upgrade to installed hardware rather than a replacement of it. Where it points: inference stops happening in a distant hall and starts happening at the tower, a hundred metres from the person using it.

Synopsys: $2,150,823,302. The software in which the world designs its chips, and since the $34.9 billion acquisition of Ansys closed in July 2025-$17.6 billion in cash and $17.1 billion in stock, also the software in which the physical object is simulated. Third-quarter revenue of $2.477 billion against $1.740 billion a year earlier, with design automation alone up 53 per cent. The stated aim is to join the two: semiconductor design on one side, multiphysics simulation and system-level engineering on the other, in one environment. Where it points: the chip is designed inside accelerated software, and so is the engine, the airframe and the turbine that the chip will later control. Authorship of the designed world migrates into the same stack that computes it.

Nebius: $328,773,757. The smallest of the infrastructure lines and the fastest moving. Cloud revenue of $575 million in the second quarter, up more than fivefold year on year, annualised revenue of $3.0 billion against a target of $7 billion to $9 billion by the end of this year, on a capital plan of $16 billion to $20 billion. It is building nine new sites across seven locations in the United States and in Europe and the Middle East (Missouri, Alabama, Oklahoma, Minnesota, France, Israel, the United Kingdom), for sixteen in total, with more than two gigawatts contracted and 800 megawatts to a gigawatt expected to be connected by the end of 2026. It has delivered contracted capacity to Microsoft and to Meta on schedule. Where it points: capacity that sits inside European and American borders and is marketed as such, while being assembled from the same components as everything else in this list.

Generate Biomedicines: $14,058,193. The smallest line by a factor of more than twenty, and the only one that ends inside a human body. Its model, Chroma, generates proteins directly: entirely new ones with no close natural analogue, multi-chain complexes, redesigned sections of existing proteins. This is no longer a demonstration. The first Phase 3 patient was dosed on 26 January 2026, one candidate received Fast Track designation from the Food and Drug Administration three days earlier, five molecules are in clinical stages, and the company listed on Nasdaq this year raising roughly $400 million. Where it points: medicines that were designed rather than discovered, now being tested in people.

Beyond the filed eight lies a private layer that no 13F records. CNBC reported in September that Nvidia's equity investments had reached roughly $99 billion in value against about $7 billion a year earlier, some $30 billion of it in OpenAI and up to $10 billion in Anthropic, with robotics, physics simulation and quantum computing alongside. Those are press figures rather than filed ones and should be read as such. The eight above are filed.

That private layer matters structurally rather than financially, because it is the demand. The laboratories are the only participants whose purpose is to consume capacity rather than supply it, and their commitments are what the rest of the structure is built against. Anthropic announced on 6 April an expanded arrangement with Google and Broadcom for multiple gigawatts of next-generation tensor processors, with capacity arriving from 2027, on top of its existing use of Amazon and Nvidia hardware. Note what that tells you. Even the attempt to buy something other than Nvidia silicon is denominated in gigawatts and dated to next year, which is the clearest available measure of how little slack exists anywhere in the system.

How they cohabit

Set them in a line and the structure stops looking like a portfolio.

Intel makes the logic and is trying to make it somewhere other than one island. Coherent supplies the light that joins the parts into a single machine. CoreWeave and Nebius buy those machines, wrap them in power and sell the result by the megawatt. Nokia pushes a slice of the same hardware to the edge of the network. SpaceX proposes to lift the whole category off the grid. Synopsys is the room in which the next generation of all of it is drawn. Generate turns the output into a molecule. And in the private layer, the laboratories consume the capacity and generate the demand that pays for the rest.

Each one's output is another's input, and only one node appears in every link.

The circuit is financial as well as industrial, and it runs in both directions. CoreWeave's $30 billion to $35 billion of spending is largely Nvidia hardware; Nvidia holds $4.7 billion of CoreWeave and separately guarantees land, power and shell for buildouts of that kind. Nvidia holds some $30 billion of a laboratory that commits to buy capacity from landlords that Nvidia also holds. Revenue leaves as investment and returns as revenue, which is not improper and is entirely disclosed, but it does mean that demand inside the circuit is not independent evidence of demand outside it.

The arrangement has an unusual property: internal competition does not threaten the centre. If Intel takes foundry share, the centre gains a second supplier. If Nebius takes customers from CoreWeave, the hardware is the same hardware. If the laboratories displace each other, they do it on the same racks. The structure is indifferent to who wins any particular square, because it holds a position on most of them.

One absence in the list is more revealing than any of the entries. None of the eight is the company that actually manufactures the chips. The structure's most critical dependency is handled not by ownership but by contract: $279 billion of purchase obligations stretching to 2029, placed with a manufacturer whose advanced packaging capacity is the industry's binding constraint. Beside that sits $29.99 billion in the one company attempting to become a second source of the same thing. The arrangement is legible as a sentence: buy the capacity you cannot replace, and separately fund the only plausible alternative to it.

There is one further thing the circuit cannot produce for itself, and the filings say it plainly by what they guarantee: land, power and shell. Everything else is inside the body. Electricity, ground and consent are outside it, and have to be obtained from the world.

How that leaves the Earth

This is where the abstraction lands, and it lands on specific places.

The unit of account is no longer the chip. It is the megawatt. Capacity is contracted in gigawatts and priced near $40 million per megawatt, which means the question a data centre asks a territory is not how many people live there but how much current can be delivered and how fast. One operator carries 4.2 gigawatts contracted; another expects a gigawatt connected by the end of this year. Those are national quantities of electricity being allocated by commercial negotiation.

The planetary total is now large enough to be measured against countries rather than industries. The International Energy Agency's projections put global data centre consumption in the region of 1,000 terawatt-hours this year, comparable to the annual electricity use of Japan; Goldman Sachs estimates United States data centre demand rising from about 31 gigawatts last year to about 41 this one. A single rack of training hardware draws 120 to 140 kilowatts, with the next generation expected to pass 200, which is why these buildings cannot simply be plugged into an existing network and why the guarantees in the filing are written over land, power and shell rather than equipment.

Such loads go where land, water and permits are cheapest, which is generally where people have least, and they increasingly arrive with their own generation attached because the public grid cannot carry them. Transmission gets rebuilt around industrial demand rather than households, and the price of electricity becomes the central political question in counties that previously had none. The most consequential negotiation a local government conducts is no longer with a factory. It is with a substation.

When ground and transmission run out, the work goes up. A constellation requires no county's consent, no water table and no connection queue, and above the atmosphere the sun does not set. A million satellites at roughly 120 kilowatts each is 120 gigawatts, and SpaceX projects on the order of 100 gigawatts from placing a million tonnes a year in orbit. That is the night sky as industrial estate, and a material share of the world's computing resting where no parliament has legislated and no court has obvious reach.

Underneath all of it sits a geography that no amount of capital can move quickly. Roughly nine in ten of the world's most advanced chips are manufactured on a single island, and by the estimate of the research firm Counterpoint, cited this year in reporting on the subject, close to 99 per cent of the chips used to train frontier models originate there. That concentration is now an instrument of statecraft in its own right: Bloomberg reported in June that Taiwan was weighing tighter export controls on artificial-intelligence chips bound for China, in part to address the diversion of servers. Advanced packaging capacity is being duplicated in Arizona, but on announced timelines it arrives towards the end of the decade. Until then, the most consequential industrial process on the planet has one address.

On the ground, intelligence becomes ambient rather than remote: with accelerators in the base stations from 2027, inference happens at the tower, and between towers and constellation there is no inhabited point without it.

And the work changes shape before the robots arrive. Chips are designed in accelerated software, physical objects are simulated before they are built, molecules are designed before they are synthesised and are already in Phase 3 trials. Expert labour moves from producing designs to approving them. Humanoid machines are real but still small in number, and the industry's own assessments put factory-scale adoption in the 2030s; the deeper change is the earlier one, because it moves the authorship of the physical world rather than its assembly.

Add it up and the Earth that comes out of this cohabitation is organised around the socket. Land is valued by its connection, the sky is annexed as the next site, the network thinks, design migrates into simulation, and the price of power becomes a line inside the price of the drug, the engine, the route and the diagnosis.

The reign of hardware over software

Everyone is leaning towards software, although in the future software will be able to be made by hardware without fail. So the companies that focus on being Nvidia's competition will themselves be the companies that touch the summit, which today is basically free access to success, although also extremely costly. Deep technology triumphs above everything. Media ecosystems and everything that has a community will go on breathing, but what does not, will simply die by natural selection.

Right now companies are taking advantage of the last breath of the collective ignorance of humanity in the face of artificial intelligence, but it will not last them long, and when it passes they will fall flat on their faces, together with all the investors and investment funds that bet on them. It has been clear for more than a hundred years: media wins, and deep technology wins.

I hope this piece corrects the course of the market towards the reign of hardware over software for absolutely everyone. Time is short and that is the reality for everyone.

Even so, it remains a fact that certain pure-software companies will go on living, depending on where their path is oriented, but as hardware advances it will be easier to reach them, because after all it is also a fact that not only the hard machinery matters, but the intellect that is put into it. So it could be an obvious surprise that small companies without much capital put out specific artificial-intelligence systems that completely destroy the big ones, above all if the big ones do not optimise down to the last cent and keep outdated perspectives of only code, and not real intelligence and genius in their models.

What the structure rests on

None of it is funded out of profit. It is funded out of order books, credit and guarantees, and a good deal of the hardware is pledged as collateral against the loans that bought it. The arrangement turns on one question: whether revenue from using the capacity justifies what the capacity cost to build.

Nvidia's own accounts are where that question first becomes visible. Cash from operations over six months was $74.4 billion against $118.0 billion of reported profit, or 63 cents of cash per dollar of earnings, with the $43.6 billion difference split between receivables, which rose from $38.5 billion to $63.1 billion, and $24.1 billion of other income, net, almost all of it non-cash gains on equity securities. Measured against quarterly revenue that is 59.6 days of sales outstanding, against 51.4 days at the close of the previous year. The filing supplies the mechanism without being asked: for investment-grade customers it provides terms "ranging from 90 days up to one year to assist customers with large data center builds". Five direct customers account for 22, 14, 13, 11 and 10 per cent of what is owed. On the evidence of the structure itself, those 59.6 days matter more than the $89 billion.

The arrangement is not new. General Electric financed the utilities that bought its turbines; equipment makers financed the telephone companies that bought their switches; railways were built by the people who sold the rails. Supplier-financed infrastructure built much of the modern world, and it has historically resolved one of two ways: the network is eventually regulated as a public utility, or it is separated into parts. Both took decades, and both arrived long after the concrete was poured.

Which is the durable point, and the reason to read the filings rather than the forecasts. The substations, the campuses and the orbits are not predictions. They are the collateral of the present. Whether the body being built works or fails, the ground and the sky have already been committed, and neither can be given back.

It is key to remember that companies are abandoning genius and creativity for the economic supercash. Things are simply absurdly dire in spite of having hundreds of billions behind them. Please work harder at it, and show who humanity is instead of creating dumb beings in software. If you have cash behind you, use it well, down to every cent, because at the end of the day, launching the first thing that occurs to you is, with nothing more to say, a fiasco.

Attentive and grateful for your time. Until next time.

Sources

Filed documents. Nvidia's Form 13F for the quarter ended 30 June 2026, filed 14 August 2026 (information table). Nvidia's Form 10-Q for the quarter ended 26 July 2026 (filing), for receivables, inventories, supply obligations, non-marketable securities, guarantees, payment terms and receivable concentration. Quarterly results releases on Form 8-K for Q2 FY2027, Q1 FY2027 and Q4 FY2026. TSMC's own second-quarter 2026 results. Anthropic's announcement of its expanded arrangement with Google and Broadcom, 6 April 2026.

Press reports, identified as such in the text. CNBC on the scale of Nvidia's equity investments (4 September 2026). CNBC on the SpaceX acquisition of xAI (2 February 2026). Data Center Dynamics on CoreWeave's contracted capacity and capital plan (report). Rest of World on the concentration of advanced manufacturing, citing Counterpoint Research (12 May 2026). Bloomberg on Taiwan's consideration of tighter export controls (9 June 2026). Projections of data centre electricity consumption are those of the International Energy Agency; the United States demand figures are estimates published by Goldman Sachs.