I’ve heard from people in the past that my articles are too long, and I wanted to start by saying that, for the most part, they’re going to stay long, because I feel like the only way for me to make my arguments is to be as specific and detailed as possible about the things I’m talking about.
Then again, sometimes it’s just because I imagine arguments against my work in my head and want to pre-empt them.
Something about the AI bubble has made the boosters genuinely insane. They see these otherworldly declarations — hundreds of billions or trillions of dollars — and assume that nobody would say them in bad faith, and that the tech industry would never fail to live up to them, even though we’re barely a few years divorced from when Mark Zuckerberg burned $80 billion on the metaverse, what will one day be known as “the second-worst misallocation of capital in corporate history.”
When the boosters hear that OpenAI plans to spend $750 billion on compute costs through the end of 2030, they shrug their shoulders and say “it’ll work it out.” When they hear that hyperscalers have $1.65 trillion in off-balance-sheet obligations and debt, they nod approvingly, saying that “these are some of the richest and most-profitable companies in the world,” and that they will “simply keep raising debt.” It’s somewhere between number-blindness and make-believe — these are such unfathomably-large sums that it’s hard for the average person to assume anything other than that nobody would sign contracts agreeing to pay them without the confidence they’d be able to do so, even though it’s all very silly.
In any case, readers, I hear you, and today’s premium newsletter is going to be a shorter one, because it’s been an incredibly long week for me, including a day that started at 5AM with four different interviews — including my appearance on CNBC, which I encourage you to watch — that ended roughly 14 hours later, which means I’m a little depleted but nevertheless dedicated to you, the reader, and giving you value for your subscription.
So today I’m going to be pithier, and focus on hard numbers and harder truths about the AI industry, and specifically seek to answer a question: how much does the AI industry actually need by 2030?
Sidenote: I wrote this intro before I wrote the rest of the copy, and now this thing is over 7000 words. God damnit. I tried, I swear.
To be specific, I’m going to be focusing on the next three fiscal years for the companies that matter — the lead hyperscalers (Meta, Google, Microsoft, Amazon, Oracle), the two leading semiconductor firms making AI chips (NVIDIA, Broadcom), the main neoclouds (CoreWeave, Nebius, IREN, which I’ll cover in short), the main AI labs (OpenAI, Anthropic, and SpaceX) and the overall AI compute industry.
I’ve spent a great deal of time in the last few years explaining in detail why I think this will all collapse, but today’s goal is to show you, in hard numbers, exactly how much money the main players in AI need, based on consensus analyst estimates and my own research.
And god damn, do they need a lot.