The Gonzo Size of the HyperScaler DataCenter Investment Boom: Chart of the Day
The current $1.5T AI arms race: are hyperscalers building utopia, building dystopia, building digital god, or simply lighting trillions of dollars on fire in a dollar auction? When “free cash flow...
The current $1.5T AI arms race: are hyperscalers building utopia, building dystopia, building digital god, or simply lighting trillions of dollars on fire in a dollar auction? When “free cash flow” meets a once‑in‑a‑lifetime tech shock, balance sheets stop looking like software and start looking like railroad lines…
People talk about this. But even the people who talk about this—most of those that I run into do not have an understanding of the sheer scale at which things are going on:
That is, at the current pace, $1.5T for 2026: $0.5T, $0.45T, $0.35T, and $0.3T for Amazon, Google, Microsoft, and FaceBook, respectively. That is 1/4 of all US capital investment, 1/20 of global investment, 1/20 of US GDP, and 1/100 of world GDP.
This is just the biggest four. But the biggest four are the overwhelming bulk of it. My guesses, going further down the list:
$35B :: Oracle
$18B :: Apple
$8B :: Tencent
$8B :: Anthropic
$7B :: AliBaba
$6B :: SpaceXAI (computers only)
$6B :: Equinix
$5B :: Digital Realty
$4B :: SalesForce
$4B :: IBM
$3B :: SAP
$2B :: Adobe
Call that an extra $100B in 2026. And then we have $30B from the biggest chip designers:
$25B :: NVIDIA
$4B :: Qualcomm
And on the hardware side, $200B in 2026, in response to the current dire shortage of foundry/IDM capacity and thus of processor and RAM chips:
$54B :: TSMC
$40B :: Samsung
$24B :: SK Hynix
$18B :: Micron
$18B :: Intel
$15B :: Broadcom
$7B :: AMD
$6B :: Applied Materials
$5B :: ASML
$5B :: Tokyo Electron
$4B :: LAM Research
$3B :: Cisco
$3B :: Dell
$3B :: HP
Two of these numbers are sore thumbs standing out. NVIDIA and Apple:
What in the world is NVIDIA supposed to be spending $25B on this year? It is, overwhelmingly, not datacenters or foundries with “NVIDIA” on a sign on the front. It is, overwhelmingly, prepaying for, committing to, and co-funding foundry capacity at TSMC and Samsung.
No: I have no idea whether and how much there is of double-counting here. Thus I have no idea whether actual foundry capacity expansion at TSMC and Samsung will total closer to $94B or $119B for 2026. Somebody knows this. Not me.
Apple is only at $18B because they have no inference-engine services that are ready to stand up to try to sell to anybody: they are, effectively, on-device voice recognition, grammar-spelling-correction, and minor picture-tweaking only.
No: I do not know whether being such a comprehensive loser at the race to write working software in this area is a huge misstep for Apple or a Xanatos Gambit <https://tvtropes.org/pmwiki/pmwiki.php/Main/XanatosGambit>—the exact opposite of a Pyrrhic Victory. And I do not think anyone does right now.
With respect to Apple, the very sharp M.G. Siegler sees it as an incredibly risky “binary bet”:
M.G. Siegler: Apple’s Binary Bet <https://spyglass.org/apple-crazy-capex/>: ‘It’s just such a wild break from their peer group. And it keeps getting more wild. It seems like the most binary bet imaginable. Either Apple is right and the rest of Big Tech will have lit hundreds of billions – perhaps trillions when all is said and done – of dollars on fire, or Apple is going to be in big trouble. Obviously, there will be some nuance…
He is very sharp indeed. And I am not sure I disagree. But let me propose an alternative interpretation, amping up the nuance to the maxxxx:
First, it is not a conscious and chosen bet—they could not do otherwise than they are doing, given their failure at software writing. Other companies might be lighting hundreds of billions of dollars on fire. Apple would be lighting hundreds of billions of dollar on fire were it to try to deploy its LLM software models on datacenters at scale. Thus if it is a successful gambit, it is a pure Xanatos Gambit.
Second, Apple’s lane here with “Apple intelligence” AI was never in trying to build Digital God or the most accurate possible simulation of the typical internet s***poster. It was always on-device, taking advantage of Apple Silicon’s extraordinary power performance features.
Third, and if Apple does turn out to have cloud-stuff to do at scale? The iPhones of the world are all connected, almost always dark all the time, and yet have about 10% of the world’s total computational capabilities that Apple could, with very minor revisions to terms-of-service, draw on at will.
Fourth, nobody save Anthropic has a product people are eager to pay for at scale.
Nilay Patel: Interview: Joanna Stern & Nilay Patel <https://www.theverge.com/podcast/926752/joanna-stern-i-am-not-a-robot-new-things-media-youtube-ai-automation>: ‘You open Google, you get some cheap-to-run AI model in your face doing AI overviews…. Google had to do that because they felt very threatened by ChatGPT. You open free ChatGPT. You get some cheap-to-run AI model that has a bunch of engagement prompts at the end of every query…. Yes, people are using [AI], and the experiences that are being foisted upon people look like slop….
Buying an iPhone was a thing people chose to do because they were excited about that product. You and me both lived through that entire moment…. These products, the free products that are in front of people—these aren’t actually great. They have not become great in the 3-4 years since ChatGPT was released….
Is there a point where [you get to] “this is definitely good enough: this is great!” the way that the products that we came up with as tech reviewers were just obviously great? The iPhone was an absolutely great product…. Is there, has there been, or will there be a killer consumer AI product?… The internet, especially when it came to smartphones, was just so obviously how everyone wanted to do everything…
Joanna does offer some pushback here:
Joanna Stern: Interview: Joanna Stern & Nilay Patel <https://www.theverge.com/podcast/926752/joanna-stern-i-am-not-a-robot-new-things-media-youtube-ai-automation>: ‘People have figured out [some] other uses cases where AI is now helping them…. I actually coined this term at the end of the book: AEI, Artificial-Enough Intelligence. We don’t need AGI….Tools that we already have are good enough…. They just have to be applied better…. Someone smart… needs to… [find] the best way for a consumer to actually want to interact with this stuff. Some companies, I think, have gotten there, though I think a lot of them just end up being acquired and then sitting in the basement of Meta…
Bare natural-language fluency for interfaces is of great value everywhere—but that can already be done on-device. Central enterprises need and find value in datacenter-scale very big-data, very high-dimension, very flexible-function classification and prediction. Coding for programers. Possibly administrative-secretarial (via something like OpenClaw). Possibly research (via something like Andrej Karpathy’s LLM Wiki or Steven Johnson’s NotebookLM). But that may—not will: may—be where the mass-market use cases stop.
And the cost is the automated production of infinite amounts of AI-slop polluting the infosphere.
Nobody save Anthropic may ever have a product people are eager to pay for at scale.
We also have, giving their view on all this:
Ryan McMorrow, Rafe Rosner-Uddin, Stephen Morris & Hannah Murphy: Big Tech’s $725bn AI Spending Spree Sends Free Cash Flow to a Decade Low <https://www.ft.com/content/b3dfaba9-17a2-4fac-90fe-4ab3ca7c9494?syn-25a6b1a6=1>: ‘It is a striking turn for companies that have rapidly transformed from relatively asset-light cash generators into some of the world’s biggest investors in physical infrastructure. “This is the deepest industry-wide capex cycle they have had,” said Justin Post, an internet analyst at Bank of America. “They see it as a once in a lifetime opportunity.”… Andy Jassy, Amazon’s chief… told investors that the AI build-out was reminiscent of the group’s early bet on… AWS…. “The free cash flow and [return on invested capital] for these investments are cumulatively quite attractive a couple of years after being in service,” he said. Jassy added that in periods of “very high growth, like now” capital expenditure meaningfully outpaces the growth in revenue from those investments, meaning “early-year free cash flow is challenged”….
Christian Leuz… said that because “free cash flow” is not defined in standard accounting rules companies have… discretion in how they… treat [things like] share-based compensation or leased data centres. “The real free cash flows of many hyperscalers are probably worse than what they call their free cash flows,” he said. The AI spending surge is flowing into a strained hardware supply chain…. Leuz said Big Tech’s AI build-out resembled the capital cycles seen in cyclical industries… such as telecoms or chemicals, where over-investment eventually leads to overcapacity, depressed margins and weak returns. But tech bosses feel compelled… [by] a technology they believe will be transformational. “They have to invest when their competition invests,” he said. “It is essentially a prisoner’s dilemma [and] this in turn reinforces the capital cycle”…
As I have said, do say, and will say over and over again: the normal over-investment cycles that come during the build-out of new technologies of positive but unknown long-run value (as in say, DeLong (1990) <https://www.nber.org/system/files/working_papers/w3546/w3546.pdf>) are greatly reinforced by the fact that the hyperscalers believe that their current monopoly platform profits are at risk should somebody else develop a better natural language interface to what they think of as their rightful service flow property. Optimal investment is very lumpy in rational response to shocks to expected growth:
The imperfect-competition platform-monopoly and contestation-disruption additions to the situation make investment lumpier, and on average larger. This is great for user surplus—as long as it is genuine user surplus that customers are buying, rather than malevolent brain-hacking attention capture. But for the hyperscalers it looks more like a dollar auction to me <https://en.wikipedia.org/wiki/Dollar_auction>.
Still, the economy as a whole should be OK, as long as all of this is equity or quasi-equity financed.
Hah!
References:
DeLong, J. Bradford. 1990. “‘"Liquidation’ Cycles: Old-Fashioned Real Business Cycle Theory & the Great Depression”. Cambridge, MA: Harvard University xerox. <https://www.nber.org/system/files/working_papers/w3546/w3546.pdf>.
“Dollar Auction”. 2026. Wikipedia: 25 Years of the Free Encyclopedia. Accessed May 12. <https://en.wikipedia.org/wiki/Dollar_auction>.
McMorrow, Ryan, Rafe Rosner-Uddin, Stephen Morris & Hannah Murphy. 2026. “Big Tech’s $725bn AI Spending Spree Sends Free Cash Flow to a Decade Low”. Financial Times. May 7. <https://www.ft.com/content/b3dfaba9-17a2-4fac-90fe-4ab3ca7c9494?syn-25a6b1a6=1>.
Patel, Nilay, & Joanna Stern. 2026. “Interview: Nilay Patel & Joanna Stern”. Decoder. May 11. <https://www.theverge.com/podcast/926752/joanna-stern-i-am-not-a-robot-new-things-media-youtube-ai-automation>.
Siegler, M.G.. 2026. “Apple’s Binary Bet”. Spyglass. May 1. <https://spyglass.org/apple-crazy-capex/>.
“Xanatos Gambit”. 2026. TV Tropes. Accessed May 12. <https://tvtropes.org/pmwiki/pmwiki.php/Main/XanatosGambit>.
If we look at aggregated earnings, we are booking the revenue for chipmakers now, but the hyperscalers' capital expenses won't appear for years. The depreciation won't even start until the data center is finished, and those are lagging. The OBBB allows immediate tax expensing of investment. Earnings are front-loaded like never before. Our policies have made equity bubbles and malinvestment more likely.
Someday a lot of this investment will stop. All at once. Equity values for hyperscalers, chip makers, electricity producers, and H/VAC firms will plummet, not just in the US, but also Korea and Taiwan. Then foreign investment in US equities plummets. Because imports of semiconductors fall, so does the trade deficit, reducing further the demand for US equities and US dollars. Lower equity prices means lower consumption by wealthier households. Combined with lower investment in equipment, the US enters a recession and needs to issue even more Treasury debt. I'm not saying a Global Financial Crisis, but it's going to leave a mark.
You write:
"if Apple does turn out to have cloud-stuff to do at scale? The iPhones of the world are all connected, almost always dark all the time, and yet have about 10% of the world’s total computational capabilities that Apple could, with very minor revisions to terms-of-service, draw on at will."
You have mentioned this before, but I suspect that it would be *extremely* difficult to draw on this resource at any reasonable scale.
The problem I see is that iPhones (and all smartphones, really) are *designed* to be "dark" the vast majority of the time. I am not an iPhone user, but I understand from others that making the phone really work hard runs down the battery very quickly. And I have frequently heard users *already* complaining about running the battery down too quickly.
Thus, using the installed iPhone base as a gigantic distributed computing cluster would require either a) some significant change in the power usage for "computation"; or b) not really doing very much computation on each device - which would drive up the cost and complexity of building the "cluster".