Don't Look Up

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large. My Hater's Guides To the SaaSpocalypse, Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater's Guide To Oracle, as well as the Hater’s Guide To Oracle (Part 2).Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week. On Friday, I'm going to pull together exactly how much money is needed to keep the AI bubble inflated in the next three years. It's gonna be a laugh-riot. Or very scary, one of the two.If you want to get in touch — and especially if you have any juicy information about Anthropic, OpenAI, or any other companies in the AI bubble — hit me up on Signal at ezitron.76. I’m also on IB on The Terminal. Last week I put out one of the most consequential newsletters I’ve written yet, pulling together multiple distinct financial analyst notes from Wells Fargo, Barclays, and UBS that directly estimated that 70% or more of the AI revenues of Microsoft, Google, and Amazon were from either OpenAI or Anthropic. To be clear, UBS estimated that next year, Anthropic and OpenAI’s compute spend would be 48% of all Google Cloud revenues — which means that they likely account for even more than 70% of its AI revenues, but I wanted to be fair. This was both a colossal pain in the arse and a story that I knew would piss off a lot of people, because of its huge ramifications. Some outright dismissed it as “doomerism,” while others insisted it was a good thing, because OpenAI and Anthropic are growing so fast.24 hours later, Bloomberg ran a story estimating, based on OpenAI’s $24.1 billion dollar contribution to Microsoft’s Fiscal Year 2026 revenues and previous statements, that OpenAI alone contributed to 70% or more of Microsoft’s AI revenues for the year. For some context, Microsoft has spent $261.3 billion dollars in capital expenditures since the beginning of 2022.Meanwhile, Apollo chief economist Torsten Slok said Friday that profit margins in AI are “...higher the further you get from the end user,” and then said something I think I’ve said maybe four times in the last three months:The bottom line is that the most profitable part of the AI value chain depends on the least profitable part continuing to grow revenue or raise capital. Capital can bridge the gap for a while, but not indefinitely. And therein lies the risk: will the ROI show up for AI's end customers fast enough to sustain the spending that is generating those upstream margins?Good bloody question Torsten! The answer is “probably not.”Let’s get real simple about this because everybody wants to make AI so complex.The Future Growth of Google, Microsoft, and Amazon Is Contingent On Anthropic and OpenAI Spending $200bn+ in 2027, Which Requires $250bn to $300bn in FundingSidenote: before we go any further, I need to be clear that AI is not the reason that these companies are growing, outside of the compute spend from Anthropic and OpenAI. AI is not “boosting other product categories” or “helping other categories grow,” because if it was, they’d tell you specifically. I get so many emails from people sending me the overall revenues of these companies, mostly from people that don’t appear able to read, but nevertheless, I want to add this note on the off chance they learn.If we assume, on the low end, that Jensen Huang is right and he’s going to sell $1 trillion or so of GPUs (roughly 30GW of billable IT capacity), that’s somewhere between $360 billion and $435 billion of annual compute revenue demand. Right now, there are (outside of hyperscalers buying compute for them, and whatever it is Meta is up to) two companies that spend more than $500 million a year on AI compute, namely Anthropic and OpenAI. Both are unprofitable, and both lose tens of billions of dollars a year.If we take OpenAI’s testimony from the Musk-Altman trial as gospel, it’ll spend around $50 billion on compute this year, and if we (kindly) assume Anthropic will spend $50 billion itself, that brings us to $100 billion. To get to that level of spend, Anthropic and OpenAI have raised a combined $217 billion in the first half of 2026. Every neocloud is effectively an outgrowth of this spend, either through direct contracts or by proxy via Microsoft or Google. Outside of hedge fund and investor Jane Street and NVIDIA, neoclouds do not have significant customers at the level that would warrant all this capex.So, the world is building AI compute capacity with the expectation of at least $360 billion in annual revenue, all while we struggle to find single-digit billions in AI compute spend. The only way all that compute gets used is if either A) Anthropic and OpenAI rent all of it or B) massive (and I’m talking multiple $10 billion-a-year customers) appear virtually overnight. In both those cases, the money to pay for that compute will have to come from somewhere.Remember: for Anthropic and OpenAI to be able to afford their current (and comparatively meager) spend, both have had to raise nearly a quarter of a trillion dollars in this year alone. The vast, vast majority of the world’s compute revenue — I’d wager anywhere from 70% to 90% — is contingent on venture capital propping the AI labs up, and to make matters worse, it is no longer sufficient for them to just “grow fast,” but to grow so fast that they can spend (per estimates from Wells Fargo, Barclays, and UBS) $197 billion on compute in 2027 just on Google Cloud, Amazon Web Services and Microsoft Azure. This does not include the billions that both will spend on CoreWeave, Cerebras, or Oracle.Sidenote: The numbers are all in this newsletter. I will also add that these estimates only include Microsoft Azure for the first two quarters of 2027 (as estimates are only up to the end of Microsoft’s FY27, which runs from July 1, 2026 to June 30, 2027). It’s likely that the number is more like $220 billion. It’s likely more.Also, before you ask: revenue concentration appears to be getting worse over time, because these 70% estimates mostly rely on the continued growth of overall compute and AI model rental platforms like Vertex, Foundry and Bedrock.Sorry, I got too complex again. Anthropic and OpenAI are only set to spend $100 billion on compute this year, and had to raise over $200 billion to do it, which makes it likely they’ll have to raise $150 billion each leading up to or in 2027. And let’s be clear about something: the future growth trajectories of Amazon, Google and Microsoft (not to mention Oracle, CoreWeave, and every other neocloud) are contingent on the continued ability for Anthropic and OpenAI to raise and have the demand necessary to spend that money. Let’s get specific! Stephen Ju of UBS estimates that Amazon’s AI revenue — 73%+ of which is OpenAI and Anthropic’s compute spend and revenue share (per Barclays) — accounts for 26% of AWS’ 2026 revenue and 30% of AWS’ 2027 revenue. Michael Turrin of Wells Fargo estimates that 25% of Microsoft Azure’s (calendar year) 2026 revenues come from AI (of which OpenAI is an estimated 70%). Brad Zelnick of Deutsche Bank estimates that AI will contribute 33% or more of Azure’s revenues in FY2027. As I mentioned last week, UBS estimates that 48% of Google Cloud’s 2027 revenues will come from OpenAI and Anthropic. Zelnick of Deutsche Bank also projects in its most-likely scenario that AI revenues will make up 37% of all Microsoft’s cloud revenues in FY2029.I feel like I need to spell this out more. These are analysts from major banks and financial institutions. Their estimates, which are based on detailed financial models, inform Wall Street and investors’ expectations, as well as informing Bloomberg Intelligence’s consensus estimates for revenues. These are serious numbers and Wall Street will be mad if they are not met! The other problem is that cloud is becoming an increasingly-larger part of the revenues of these companies. See the below chart that bakes in consensus analyst estimates up to 2029:I, again, will simplify: if more and more of the revenues of these three companies are coming from cloud segments that are increasingly-dominated by AI revenues mostly driven by two unprofitable, unsustainable companies, then the literal future of Microsoft, Google, and Amazon is whether Anthropic and OpenAI can pay them. This is not complex, it’s not contrived, it’s not doomerism or hating, these are the estimates from analysts and what they require to stop them from putting executives in The Wicker Man. You can cut this situation in any way you want, but there’s no getting away from the fact that we’re four years in and the vast majority of demand comes from two companies that can’t afford to sustain it, and won’t be able to even under the most mold-poisoned of booster projections. Hyperscale growth is contingent on the success of their AI plays, and at 70% of AI revenues, “AI plays” refers to “two unsustainable AI labs.” Perhaps another visualization would help! Below is a chart of the expected percentage of year-over-year growth that cloud revenues are estimated to contribute on a quarterly basis to revenues. Cloud revenues are the lynchpin of growth for Microsoft, Amazon and Google, though for whatever reason analysts estimate that YouTube and Google Search will re-accelerate.This is a huge issue when 33% of Azure revenue, 48% of Google Cloud, and (per Ken Gawrelski of Wells Fargo) 60% of AWS revenue growth is coming from companies that have been, assuming all the money crosses, sent a combined $115 billion from Amazon and Google in 2026 alone. I realize I’m repeating myself, and I’m sorry, but it’s all so insane! The future