Recently, not only the tech media but even mainstream media have been reporting on the Wall Street Journal’s recent report that Sam Altman thinks it will cost up to $7 trillion to build the capacity required to keep AI progress moving forward. Although the Journal report attributed the number (originally “as much as $5 trillion to $7 trillion”) to only one source, and didn’t state that Altman was actually trying to raise that money, other news outlets discarded all such nuance and described it as “Altman’s…$7 trillion AI chip venture” or noted that “Altman wants to raise up to $7 trillion for a new AI chip project.”
This probably isn’t what he actually meant. The Information subsequently reported that Altman was talking about the total capital investment required across the entire value chain, and wasn’t trying to fly home with a check for $7 trillion in his pocket. Altman himself poked fun at the $7 trillion number on X:
But one way or another, it’s pretty clear that Altman thinks we are going to need trillions and trillions of dollars in capital investment over the coming years to build the next generation of semiconductors required to keep AI moving forward.
So how much money is this, really? And do we really need that much?
The answer to the first question is that it’s a LOT. I mean, $7 trillion is a really, really, really large amount of money. I ran some tests, and it’s a very big amount of money.
News outlets have lovingly compared $7 trillion to massive expenditure in other fields: comparable to the US federal budget, more than the US spent on World War II (adjusted for inflation), more than the GDP of Japan, enough to provide universal health care, or education, or eradicate any number of diseases, and so on. This is actually helpful–investments on this scale change the allocation of resources across the entire economy, and so understanding the tradeoffs is important from a political and social point of view–but it doesn’t really tell us whether Altman’s figure makes sense, and what we would get if we spent it on semiconductors.
Let’s assume that Altman was indeed talking about the entire value chain, from land and power all the way up to an enormous tower of GPUs. How much would that cost?
First, most of the discussion here has been about the capital expenditure required to actually build the infrastructure we’re going to need–the buildings, the semiconductor fabs, and so on. That may indeed be what Altman meant, but it’s not the right way to think about it. While there will certainly be a lot–a lot–of capex, the running costs of all this equipment will be far greater. It’s one thing to build enormous fabs; it’s another to pay the thousands and thousands of highly skilled and expensive people who will be designing the chips that are built there. Similarly, it’s expensive to build data centers, but it’s a lot more expensive to run them. So let’s assume that we’re talking about annual running costs.
To do this, I’m going to start with the entire cloud-computing value chain as it stands today, which I’m going to call the “stack.” There’s a lot more going on in the cloud besides AI, of course, but it’s a handy measuring stick. I’m also going to invent a new unit called the “stack-year,” which is the cost of running the stack for–that’s right–one year. If we can work out the cost of one stack-year, then we can figure out what our $7 trillion is going to get us.
There are only three economically meaningful components of the stack. At the bottom, we have the data centers and their running costs (which will pick up land and power). In the middle, we have the chips. And at the top, we have all the AI people, the people building and training the LLMs (but not the people working in the middle two layers).
For the bottom level, we start with data center revenues (you can argue about whether revenues or costs are the better measure; revenues are more transparent and net out the impact of capex over time). There are many different estimates out there, but they tend to cluster around $200 billion annually. Most of these try to include the internal “revenue” generated by the vertically integrated tech giants, accounting for what Google would be paying itself to run the DCs to power its own search.
The middle layer is more difficult to estimate. At the upper end, the trade association World Semiconductor Trade Statistics has estimated 2023 semiconductor market revenues at $520 billion. This includes categories that I wouldn’t consider part of the stack, like chips for cars, refrigerators, and smartphones (this last category is arguable). Going bottom-up, Nvidia’s third-quarter revenue was $18 billion, or $72 billion annualized, and still growing quickly. Intel’s 2023 revenue is expected to come in at $54 billion. You could argue that these revenues reflect growth rates, not the installed base–the annual cost of the stack should be only the cost to replace chips as they age–but given the relatively short useful life of GPUs, I think it’s fine to use these revenues as a base. Let’s add in the revenues of TSMC, the world’s largest contract chipmaker, which are currently running at about $80 billion annualized–a lot of this will be double-counted with Nvidia in particular, but most of it won’t be, and we’ve ignored plenty of other companies providing products and services in this middle layer, as well as other forms of storage. These three companies get us to another $200 billion; I’m going to add another $100 billion because, you know.
I’m not adjusting the middle layer to reflect the fact that $1 billion spent in, say, 2030 will no doubt create far more compute and storage than it does today. This means that we don’t really know what we’re going to get for our $7 trillion when we eventually spend it, but if we try to adjust for the increase in processing power, it ends up being highly arbitrary, and you can get to very counterintuitive conclusions, like saying that spending will start to decline because computers will be so much more powerful by the time we get there.
Finally, what about all the people running around using this stuff? Amazingly, even under the most aggressive assumptions, they aren’t a material expense. When OpenAI very briefly fired Altman in November, it was widely reported that almost all of the 770 employees had signed a petition saying that they would leave if he weren’t reinstated. Perhaps there are 1,000 employees now? They couldn’t be paid an average of $1m per year (aside from the possibility of windfall equity gains), but if they were, that would bring OpenAI’s total people costs to $1 billion annually. And we have Meta, Google, Amazon, the big Chinese platforms, and so on, who might be spending $1 billion each on the people involved in AI. And plenty of startups. And the emerging-market sweatshops doing the RLHF. But I struggle to get to a credible figure of $20 billion, far less than my fudge factor of $100 billion on the middle layer. So I’m going to reduce my middle-layer fudge factor to $80 billion and add $20 billion in people costs at the very top of the stack.
So I’m going with $500 billion as the cost of a stack-year. This is not the current amount the world is spending on AI each year, of course–a stack-year includes storing all of your vacation photos in the cloud, and people buying GPUs to shoot zombies, and scanned files from the 1990s that a Midwestern insurance company is keeping just in case.
So Altman’s “as much as $5 trillion to $7 trillion” works out to 10 to 14 stack-years. Is that a lot?
Obviously the biggest question is the time period over which we’re going to spend this money. If we’re going to spend it over 10–14 years, then Altman is essentially saying we’re going to need to double our current cloud computing infrastructure to get to where we want to with artificial intelligence. This would mean significantly more than doubling the actual storage and compute, given that technology will improve between now and then, but in simple terms, it means twice the number of data centers full of equipment.
What’s the shortest this time period could be? Depending on your source, people quote between three and five years as the time period required to build a semiconductor fab. Data centers are much quicker and can be built in parallel, usually one to two years, and so won’t be the limiting factor. If we want to build new power plants for these data centers, that would be the slowest link in the chain of all–it depends on the type of plant, and most of all on the permitting process, but it could easily take more than a decade just to get new plants on line. But unless we’re very stubborn about where we want the DCs to go, we can probably just tuck them in where the power already exists. Data centers today consume an estimate of 1 to 1.3% of all global electricity production, so doubling or even tripling it doesn’t seem impossible (although it could have a significant environment impact–an important but separate topic).
So the middle layer of the stack will probably be the slowest to build out. The Journal reported that Altman wanted to work with TSMC to build several new fabs to produce chips–given that fab construction won’t start the moment the deal is signed, and that even when the fabs are up and running, it will take time to produce the chips, it seems that a 5–7 year period to spend the $5–7 trillion is the absolute best case. Spending $1 trillion a year is indeed a lot of money; it would be tripling the current $500 billion annual spend on the stack.
But it’s not an insane pipe dream. These back-of-the-envelope calculations suggest that what Altman is proposing works out to somewhere between a doubling and a tripling of our current spend on cloud infrastructure and the associated storage and compute required to train our models. We can argue about whether that’s the right number, and we can certainly argue about whether Altman’s proposed consortium approach as reported in the Journal is the right one, but it’s not an insane figure, the way it has been made out to be in the popular press.
To put it another way, global GDP is currently about $100 trillion. If Altman had said that we need to dedicate between 0.5% and 1% of global GDP to building out more computing resources over the next several years in order to accelerate the growth of AI, expecting that the investment would deliver productivity gains greater than that 0.5–1% investment, it wouldn’t have seemed crazy. It would have seemed ambitious–which it is, and rightly so.
So “as much as $5 trillion to $7 trillion” may or may not be the right amount of money to invest in accelerating AI. It may or may not be the right way to do it. But it’s a reasonable proposal that we should debate on its merits, and not dismiss because of the shock value of the headline number.
