--- title: "More Server Rooms, but the Keys Went to Two Tenants First" date: 2026-11-05 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-compute-math-11-1105 language: en --- # More Server Rooms, but the Keys Went to Two Tenants First > On November 3, at a conference called Runtime 2026, Dylan Patel, founder of the semiconductor research firm SemiAnalysis, took the stage and spoke about three things: computing power, capital, and On November 3, at a conference called Runtime 2026, Dylan Patel, founder of the semiconductor research firm SemiAnalysis, took the stage and spoke about three things: computing power, capital, and who can actually put machines into customers' hands. He began with a table of totals. At the end of 2024, global computing capacity was about 60 gigawatts; by the end of 2026, it exceeds 100 gigawatts. Two years, nearly double, as if the stock built up over decades had been built once more. A gigawatt is a unit of power, equal to one million kilowatts. It is used to measure computing because today's data centers are limited first by electricity: only as much power as you can draw determines how many machines you can install. Power is the ceiling. The conversion is not precise. The same kilowatt paired with different chips yields different amounts of computing. But in the industry it is the common ruler for comparison across companies, and everyone uses it when talking about scale. ## Doubling in Two Years Doubling sounds abstract until you turn it around. The total power of the world's data centers grew by single digits for a long time; in the past two years it suddenly lifted, and the curve is almost vertical. The immediate cause is large models. Training one model means running thousands of chips at full load for months, and afterwards keeping a large batch of machines answering users' questions. The meter runs from the start of the year to the end without a break. Patel did not draw the curve into a long-range forecast. He gave one near point: by 2028, AI labs alone will need to add about 30 gigawatts a year. Two years ago, few would have dared put that on a slide. ## Two Buyers Took Thirty Percent The most concentrated buyers in the increment are two companies. The computing power held by OpenAI and Anthropic was about 4 gigawatts combined at the end of 2025, about 10 gigawatts in 2026 and more than 20 gigawatts in 2027, more than quadrupling in three years. That curve is steeper than the global total. Look at the increment and it is clearer. In 2027 alone, the two will add about 12 gigawatts, while the world adds about 40 gigawatts that year, roughly 30 percent. Patel gave a range of 25 to 40 percent, with 30 percent in the middle. What does 30 percent mean? Of every ten new data centers in the world, about three are being prepared for these two companies. The other seven are left for cloud providers, startups and other industries, and they must pay first. That also explains why so many companies complain they cannot buy computing. It is not that machines cannot be built; it is that the batch being built already has an owner before it leaves the factory, and latecomers can only wait at the door for the next batch. Patel mentioned another ratio: of the computing already contracted for next year, about 65 to 70 percent goes to hyperscale cloud providers and AI labs. What is left for the outside world is the remaining small slice. ## Two Dollars Eighty an Hour Price is another thread. Patel said the one-year contract price for Nvidia's H100 chip rose from 1.70 dollars per hour at the October 2025 low to about 2.80 dollars. This is not a retail price but a wholesale one, negotiated for long-term use of tens of thousands of chips. Mind the unit. This is not the price of buying a card but the price of renting one for an hour, and on a one-year contract, where the usage for the next twelve months is locked in at signing and cannot be cancelled midway. He gave a transaction range of 2.40 to 3.20 dollars, with 2.80 the midpoint. At the trough, one could still find 1.45 dollars, usually when a cloud provider's service was worst and its machines hardest to use. The reason for the rise is not complicated. More people are willing to pay high rents for chips to sell inference services, and rents were pushed up. In this market, price is not set by cost alone; it is also set by the people in the queue. ## Three of Four Gigawatts One detail deserves separate mention. Anthropic currently has about 4 gigawatts of computing, and the entire Claude Code product uses only about 1 gigawatt of it for inference, with the other 3 gigawatts elsewhere. Put the two numbers together and the ratio is close to one to three. In other words, three quarters of the machines are not answering users' questions. They are training the next generation of models, or preparing to. The small amount of intelligence a user sees rests on a great deal of electricity not yet spent on them. Patel's formulation was that the compute spent on inference is shockingly small compared with training. That explains why model companies keep buying chips: it is not that the current product is short of capacity, but that the next version is not finished. For an ordinary person, this means the computing consumed by each question is only the tip of an iceberg. The part below the water is the bulk of the bill, and the thing these companies are really fighting over. ## The People Borrowing to Build Chips cost money; building data centers costs more. Patel noted that US Treasury yields have risen sharply in the past two months, with the 10-year yield climbing above 5 percent in early October. Borrowing is getting more expensive. When rates rise, young cloud providers that build on borrowed money feel it. They lack the cash flow of big companies, interest starts accruing before machines are racked, and the financing bar rises day by day. He also did the arithmetic for building oneself: a one-megawatt B300 cluster, spread over five years, costs about 9 million dollars a year to operate; renting the same computing from a cloud provider, 14 million dollars a year would count as lucky. Nvidia's approach is to backstop some projects. Patel said it has provided some form of guarantee for about 6.5 gigawatts of computing, but that this is a small share of the capacity added before 2030. How small, he did not put a percentage on, using only a phrase: a drop in the bucket. That judgment is more direct than any number, and it shows the gap is not something one company can fill. The guarantees come in several forms: some cover only land, power and the building; others commit to renting the cluster at a set price. ## Computing Is Not Electricity Put these numbers together and they all say one thing: computing is turning from a commodity you can buy at any time into capacity that must be booked years ahead. That sounds like common sense, but in this industry it has been true only for the past two years. Electricity can be generated and used on the spot; a data center cannot. From site selection and grid connection to construction, installation and testing, every step takes time. Money can arrive overnight; concrete and workers cannot. Approving a transmission line can take longer than putting up a building. So those gigawatt figures are not forecasts in essence; they are orders. Whoever locks in power and land first gets next year's machines; whoever is late pays a higher price in the secondhand market. The word "book" is turning from an adjective into a verb. Having laid this out, Patel's advice was direct: if you can find a B300 at six dollars an hour, sign it and go sell inference. The price is absurd, he said, but it is the right decision, because the revenue from selling inference is higher. The key is whether you can lock down the machines and the power before anyone else. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-compute-math-11-1105 Free to quote with attribution and a link to the original.