Companies reach the top of the market capitalization table by selling to everybody. That is close to a law: oil, telephones, software, groceries, search, all of them built on billions of small transactions with people who will never meet anyone who works there.
The exception arrived on Tuesday. Nvidia passed Microsoft to become the most valuable public company in the world, at about three point three trillion dollars, and a very large share of its revenue comes from a handful of buyers.
This is a ledger of that: where the money comes from, who pays it, and what the concentration means for a valuation built on it.
What is being sold
Not consumer graphics cards, which is what the company was known for and is now a minority of the business.
The revenue is data center: accelerators bought in quantity by organizations building systems to train and run large models, together with the networking to connect them and the software layer that makes them programmable.
That software layer is the part most often left out and it is doing enormous work. A programming framework that has been the standard for well over a decade means an accumulated body of code, tooling and trained engineers that assumes this hardware. A competitor selling a faster chip still has to persuade somebody to rewrite all of it.
The hardware itself is also not a single product. An accelerator is sold as part of a system: the chips, the high-bandwidth memory that feeds them, the interconnect that links thousands of them into one machine, and the reference designs that let a buyer assemble it. Selling the system rather than the part is what keeps the margin, and it is why a competitor with a comparable chip is not a competitor with a comparable offering.
Who is paying
A small number of very large buyers, and this is the entry that makes the ledger unusual.
The dominant customers are the handful of companies operating hyperscale cloud infrastructure, together with a few model developers and a small number of governments. Company disclosures have repeatedly noted that a limited number of customers account for a substantial share of revenue.
Concentration of that kind has two properties and they pull in opposite directions.
It is efficient. Selling to five organizations means a sales operation that would be absurd at consumer scale, long order books, and visibility into demand a quarter or more ahead.
It is also fragile in a specific way. Those buyers are among the few organizations on earth with the capital and the engineering depth to design their own silicon, and several of them are doing exactly that. A customer base that can become a competitor is a different kind of customer base.
What the buyers are buying it for
Here is where the ledger stops being about this company.
The purchases are capital expenditure against expected future demand for services that, in most cases, are not yet generating revenue at anything like the scale of the spending. That is not a criticism; it is what building infrastructure ahead of a market looks like, and it is how railways, telephone networks and data centers have always been financed.
It does mean the revenue is a derivative of somebody else’s forecast. The chips are sold because buyers believe in demand that has not arrived. If those buyers revise, the revision reaches this income statement immediately and with no cushion, because there is no diversified base of small customers to absorb it.
Who collects
Shareholders, obviously, and the sums are large enough to have moved index arithmetic.
A company at this weight in the major American indices means that a very large number of people who have never chosen to own it do own it, through pension and retirement funds tracking those indices. The concentration is therefore not only in the customer base; it is in the portfolios of people with no view on any of this.
The manufacturing is collected elsewhere. The company designs and does not fabricate, and the fabrication is concentrated in a small number of facilities in Taiwan and a growing number elsewhere. That is a separate concentration with separate risks and it is not on this balance sheet at all.
The entry that is easy to miss: who is not paying
A ledger should record absences, and there is a large one.
The end users of what these systems produce are, in the main, not paying for them yet. Model developers are supplying capability at prices well below cost, cloud providers are subsidizing access to win position, and a great deal of consumer-facing usage is free.
The company sits at the end of a chain in which the money currently flows one way: from investors, through the buyers, into the hardware. Nothing about that is improper and it is the normal financing of an infrastructure build.
It does mean the revenue is not yet connected to customer payment at the far end. Until it is, the demand is an expression of confidence rather than of receipts, and confidence reprices faster than receipts do.
What the valuation is a claim about
Not this year’s earnings, which are real and large.
A valuation of this size is a claim that the current demand is the early part of a durable expansion rather than a build-out with an end. Those are different shapes: infrastructure cycles have historically involved a period of intense buying followed by a period of digesting what was bought, and the second period is usually painful for whoever sold the equipment.
Whether this is a cycle or a plateau is not knowable now and will be obvious in retrospect, which is the standing condition of every such argument.
One asymmetry belongs on the page before the forward look. The company’s position is strongest in training, where the largest systems are built and the software advantage is deepest, and weaker in inference, which is the part that scales with actual usage and where cheaper specialized parts compete. If the market’s center of gravity moves from building models to running them, it moves toward the half of the business where the moat is shallower.
What to watch
Not the share price, and not the quarterly revenue figure, which will be large.
Watch the customer concentration disclosure in the annual filing, which is a required statement and is where a shift from five buyers to twenty, or from five to three, becomes visible before it becomes a story.
Watch the large buyers’ own capital expenditure guidance, because that is the actual demand signal and it is published by other companies a quarter earlier.
Above everything else, look for a serious competitor to the software layer. Hardware advantage in this industry has never lasted more than a few years, and the only reason to expect this position to outlast that pattern is the code rather than the chips. The first credible sign that the code is portable is the signal worth acting on.
