As I've established again and again, we are in an AI bubble, and no, I cannot tell you when the bubble will pop, because we're in the stupidest financial era since the great financial crisis — though, I hope, not quite as severe in its eventually apocalyptic circumstances.
By the end of the year, Microsoft, Amazon, Google and Meta will have spent over $400bn in capital expenditures, much of it focused on building AI infrastructure, on top of $228.4bn in capital expenditures in 2024 and around $148bn in capital expenditures in 2023, for a total of around $776bn in the space of three years.
At some point, all of these bills will have to come due.
You see, big tech has been given incredible grace by the markets, never having to actually show that their revenue growth is coming from selling AI or AI-related services. Only Microsoft ever bothered, piping up in October 2024 to say it was making $833 million a month ($10bn ARR) from AI and then $1.08 billion a month in January 2025 ($13bn ARR), and then choosing to never report it again.
As reported by The Information, $10bn of Microsoft’s Azure revenue this year will come from OpenAI’s spend on compute, which, also reported by The Information, is paid at “...a heavily discounted rental rate that essentially only covers Microsoft’s costs for operating the servers.”
It’s absolutely astonishing that such egregious expenditures have never brought with them any scrutiny of the actual return on investment, or any kind of demands for disclosure of the resulting revenue. As a result, big tech has used their already-successful products and existing growth to pretend that something is actually happening other than Satya Nadella standing with his hands on his hips and talking about his favourite ways to use Copilot, a product that so unpopular that only eight million active Microsoft 365 customers are paying for it out of over 440 million users.
Sidenote: Speaking of unpopularity, Microsoft is currently being sued in Australia for raising the cost of the personal versions of Office 365 to reflect the integration of Copilot and other features, and hiding the fact from users that they could continue using the software they were already using, and at the same price.
This stuff is so unpopular, the world’s biggest and most powerful software company — and one with a virtual monopoly on the office productivity market — had to use dark patterns to get people to pay for this stuff.
Earlier in the week, OpenAI announced that it had “successfully converted to a more traditional corporate structure,” giving Microsoft a 27% position in the new entity worth $130bn, with the Wall Street Journal vaguely saying that Microsoft will also have “the ability to get more ownership as the for-profit becomes more valuable.”
Said deal also brought with it a commitment to spend $250bn on Microsoft Azure, which Microsoft has booked as “remaining performance obligations” in the same way that Oracle stuffed its RPOs with $300bn dollars from OpenAI, a company that cannot afford to pay either company even a tenth of those obligations and is on the hook for over a trillion dollars in the next four years.
But OpenAI isn’t the only one with a bill coming due.
As we speak, the markets are still in the thrall of an egregious, hype-stuffed bubble, with the hogs of Wall Streets braying and oinking their loudest as Jensen Huang claims — without any real breakdown as to who is buying them — that NVIDIA has over $500bn in bookings for its AI chips, with little worry about whether there’s enough money to actually pay for all of those GPUs or, more operatively, whether anybody plugging them in is making any profits off of them.
To be clear, everybody is losing money on AI. Every single startup, every single hyperscaler, everybody who isn’t selling GPUs or servers with GPUs inside them is losing money on AI. No matter how many headlines or analyst emissions you consume, the reality is that big tech has sunk over half a trillion dollars into this bullshit for two or three years, and they are only losing money.
So, at what point does all of this become worth it?
Actually, let me reframe the question: how does any of this become worthwhile?Today, I’m going to try and answer the question, and have ultimately come to a brutal conclusion: due to the onerous costs of building data centers, buying GPUs and running AI services, big tech has to add $2 Trillion in AI revenue in the next four years. Honestly, I think they might need more.
No, really. Big tech has already spent $605 billion in capital expenditures since 2023, with a chunk of that dedicated to 5-year-old (A100) and 4-year-old (H100) GPUs, and the rest dedicated to buying Blackwell chips that The Information reports have gross margins of negative 100%:
Big tech’s lack of tangible revenue (let alone profits) from selling AI services only compounds the problem, meaning every dollar of capex burned on AI is currently putting these companies further in the hole.
Yet there’s also another problem - that GPUs are uniquely expensive to purchase, run and maintain, requiring billions of dollars of data center construction and labor before you can even make a dollar.
Worse still, their value decays every single year, in part thanks to the physics of heat and electricity, and NVIDIA releasing a new chip every single year.