The AI Demand Bubble

Aug 04, 2026 10:51 PM - 3 hours ago 1

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, aliases $7 a month, and successful return you get a play newsletter that’s usually anyplace from 5,000 to 18,000 words, including vast, elaborate 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 basal to knowing our existent financial system, and my guideline to really OpenAI Kills Oracle pairs nicely pinch my Hater's Guide To Oracle, arsenic good arsenic the Hater’s Guide To Oracle (Part 2).

Subscribing to premium is some awesome worth and makes it imaginable to constitute these large, deeply-researched free pieces each week. On Friday, I’ll people the 2nd installment of the Hater’s Guide to Nvidia — wherever I’ll return a look astatine really the AI bubble transformed the institution from a axenic hardware subordinate to a purveyor of the financial acheronian arts.

If you want to get successful touch — and particularly if you person immoderate juicy accusation astir Anthropic, OpenAI, aliases immoderate different companies successful the AI bubble — deed maine up connected Signal astatine ezitron.76. I’m besides connected IB connected The Terminal.

Soundtrack: Tool - Forty Six & 2

The mobility I want to inquire anyone reference this who mightiness person invested successful aliases successful immoderate measurement backed the hyperscalers and the greater AI industry:

What is it you deliberation you’ve gotten yourself into? Because I deliberation you’re being sold a lie.

Last week’s tech net saw outlet aft outlet declare that Amazon, Google, and Microsoft’s AI bets were “paying off” arsenic their respective unreality segments reported grounds gross growth, casually ignoring that none of them person surgery retired their AI revenues. To adhd reproach to injury, Microsoft decided, aft sharing that it had a $37 cardinal AI tally complaint (about $3.08 cardinal a month) in Q3 FY2026, that it simply didn’t person to share thing astir its existent AI payoff successful Q4, realizing that its wide numbers would beguile reporters and analysts — particularly those pinch small liking successful what was actually going on arsenic agelong arsenic the topline worldly looked good.

To beryllium clear, each 3 of these companies’ unreality platforms person galore different customers paying for galore different things different than generative AI services aliases AI GPUs, and they’ve each engaged successful a operation of multiple outright value increases and changing their halfway subscriptions to unit AI features connected them arsenic a intends of boosting revenues and conning the thoroughfare into believing that “AI is paying off” each clip they non-consensually thrust it connected their customers, framing higher prices arsenic “better value” successful a measurement that fucks the personification to appease Wall Street.

Yet the biggest con of each is that a immense mostly of this gross maturation comes from the compute walk of Anthropic and OpenAI, some of whom relationship for the immense mostly of AI revenues and wide unreality maturation we’ve seen successful the past fewer years.

Every publication you publication correct now will show you that AWS and Azure and Google Cloud are increasing for illustration wildfire arsenic a consequence of the hundreds of billions of dollars they’ve invested successful AI GPUs and information centers, erstwhile the truth is acold simpler: their revenues are being buoyed by 2 unprofitable, unsustainable AI labs that cannot beryllium without being funneled tens of billions of dollars each year.

And a decent chunk of that money is coming from the hyperscalers themselves. In the past 7 months alone, Google has sunk $10 cardinal (and up to $30 cardinal more) into Anthropic, pinch Amazon funnelling $5 cardinal to Anthropic wrong a week of that finance and a full of $50 cardinal into OpenAI. For each the interest astir information financing successful the AI world, it’s astonishing that truthful overmuch attraction has (rightly, to beryllium clear) centered connected NVIDIA’s backstopping and backing of neoclouds, and little connected the truth that hyperscalers are propping up their now biggest customers, giving them rate that will yet migrate backmost to the hyperscaler.

I’d besides reason that the immense mostly of their capex exists to support these 2 load-bearing failsons. A fewer months ago, a Microsoft executive told the judge during the Musk-Altman proceedings that its OpenAI narration had costs it “over $100 billion,” including some the $13 cardinal it sunk into the institution and the associated infrastructure.

Microsoft has dedicated its Fairwater information centers (however overmuch really exists) wholly to OpenAI, overmuch for illustration Amazon has for Anthropic pinch nevertheless overmuch of its monolithic Indiana-based Project Rainier has really been turned on, and overmuch for illustration Google is successful talks to backstop a $15 cardinal information halfway task for Anthropic, on pinch information centers with Cipher Mining and TeraWulf and a $35 cardinal backstage credit-funded Broadcom-backstopped deal wherever Google will waste Anthropic its TPU AI chips, put them successful a Google-built information center, and rent them backmost to Anthropic.

I want to spell this out: erstwhile you region Anthropic and OpenAI’s compute spend, I americium not assured that Google, Microsoft and Amazon person overmuch of an AI business.

While galore group judge — mostly because the large 3 garbage to break retired their existent AI revenues aliases disclose their customer attraction — that they person AI revenues coming from a divers group of different customers, the reality is that their largest cloud customers, fto unsocial AI customers, are 2 companies that tin virtually not spend to salary them without a near-infinite travel of task superior aliases debt.

Per Ross Sandler of Barclays, Anthropic and OpenAI are estimated to dress up 73% of all of Amazon’s AI revenues successful some 2026 and 2027 and 75% of AI revenues successful 2028, pinch Anthropic spending $14.1 cardinal successful 2026, $25.3 cardinal successful 2027, and $35.8 cardinal successful 2028, and OpenAI spending $9 cardinal successful 2026, $15 cardinal successful 2027, and $20 cardinal successful 2028.

Amazon plans to walk $220 cardinal successful superior expenditures successful 2026 and moreover much successful 2027, and appears to beryllium doing truthful almost-exclusively to supply compute for a institution that had to raise $95 billion successful backing successful the abstraction of six months, pinch $5 cardinal of that coming from Amazon itself.

Editor’s Note: Just earlier I headed to property connected this piece, I recovered different statement from Stephen Ju (who you’re conscionable astir to study astir for the first time) astir AWS revenues, pinch the numbers a small different. He has estimated full AI revenues astatine astir $30.9 cardinal for 2026, pinch OpenAI and Anthropic’s compute walk sitting astatine 59% of those revenues ($18.3 billion) and the remaining $12.6 cardinal coming from Bedrock, the level from which Amazon sells entree to some TPUs and AI models from Anthropic and (more recently) OpenAI. This gross attraction improves to 55% successful the 2027 estimates.

Anyway, the remainder of this portion focuses connected Sandler’s numbers, arsenic I did not get a ton of clip to excavation complete these. These numbers, while different, do not meaningfully alteration my perspective.

As I’ll reason astir Vertex, making money by proxy of having monopoly support to waste OpenAI and Anthropic’s models perfectly counts arsenic gross related to Anthropic and OpenAI to me. A chunk of some labs’ gross comes from the resale of these models, easy money that besides becomes different measurement successful which hyperscalers provender their revenues backmost into the AI labs truthful that the AI labs tin walk the money connected compute. I will adhd that

Microsoft nary longer pays a gross stock to OpenAI.

In immoderate case, the viability, efficacy, and attractiveness of these models are still a merchandise of Anthropic and OpenAI’s ongoing work.

Google is successful a similar-position. Per Stephen Ju of UBS, “...Anthropic, OpenAI and Meta will relationship for 21%, 7% and 1% of 2026 Google Cloud revenues, respectively, and 44%, 5% and 1% of 2027 revenues,” or, put different way, 28% of each 2026 and much than 48% of each 2027 Google Cloud revenues are from Anthropic and OpenAI.

Ju besides estimates Meta will dress up a whopping 1% of Google Cloud revenues successful each year, and does not mention a azygous different customer, which heavily-suggests that location aren’t really immoderate ample ones.

Based connected Bloomberg Intelligence’s statement estimates for Google Cloud’s revenues successful 2026 ($105.9) and 2027 ($173.8), OpenAI and Anthropic correspond $29.4 cardinal ($7.4bn/$22bn) successful 2026 and $84.69 cardinal ($8.69bn/$76bn) successful 2027. To beryllium definitive here, this is all Google Cloud revenues. It is reasonable to judge that this represents astatine slightest 75% of Google’s AI revenue, if not more.

What’s crazy is that these numbers are really little than UBS’ estimates. As the floor plan beneath demonstrates, OpenAI and Anthropic’s walk is estimated to beryllium astatine complete $35 cardinal successful 2026, larger than some its full Google Cloud halfway non-AI business and Vertex AI exemplary rental business that is mostly boosted by Google’s expertise to waste Anthropic’s models.

Sidenote: Ju and Sandler look to disagree connected really overmuch of Anthropic’s compute walk that Amazon and Google get, which is fair, because some Google and Amazon separately declare to beryllium Anthropic’s superior provider.

Eagle-eyed readers will besides spot that Google’s non-AI unreality business is estimated to beryllium efficaciously level successful 2026, 2027, and 2028.

I besides don’t deliberation it’s communal knowledge that OpenAI is specified a ample customer of either Google Cloud aliases Amazon Web Services, spending astatine slightest an estimated $52.5 cardinal successful 2026 and astatine slightest an estimated $125 cardinal successful 2027.

In the Musk-Altman trial, OpenAI estimated it would walk $50 cardinal connected compute successful 2026, and based connected those estimates, that gives america astir $16.4 cardinal crossed Amazon and Google, leaving a apt $33.6 cardinal successful walk near for Microsoft Azure, though I’ll adhd that OpenAI continually underestimates its ain compute walk and losses.

And based connected a statement from Michael Turrin of Wells Fargo from May 31 2026, things are conscionable arsenic bad for Microsoft, pinch 70% aliases much of its AI revenues coming from Anthropic and OpenAI. While Turrin “expects investments astatine package & models layers [to] salary disconnected successful meaningful take complete time,” it’s difficult to reason that Microsoft has immoderate meaningful AI strategy extracurricular of OpenAI and Anthropic’s compute spend.

To make matters worse, based connected Wells Fargo’s estimates, it appears that Microsoft 365’s AI revenues are hardly — and I mean barely — beating the gross stock Microsoft gets from OpenAI’s sales.

Wells Fargo besides includes a adjuvant cheat expanse of its estimates for AI contributions, estimating that moreover astatine the very extremity of FY2027 (which began connected July 1 2026), OpenAI and Anthropic’s walk will correspond a melodramatic 74% of each AI revenues. Wells Fargo besides estimates that the 2 AI labs represented 23% of Azure gross successful FY2026, increasing to 35% successful FY27.

Considering Azure grew 41% year-over-year, this intends that 40% aliases much of Microsoft Azure’s maturation came from them — and remember, Azure sells far much than conscionable AI services.

This is an absolute fucking scandal.

The immense mostly of Microsoft, Google and Amazon’s AI revenues and gross maturation successful their typical unreality platforms are from Anthropic and OpenAI, and they are blatantly, unashamedly misleading investors by not disclosing that this is the case. We’re talking 73% of AWS’ AI revenues, 74% of Microsoft’s, and apt 70%+ of Google Cloud’s considering that just Anthropic and OpenAI’s AI spend is expected to beryllium much than 48% of all unreality revenues.

This is not maine being a hater, a skeptic, aliases a doomer, but the merchandise of really investigating what’s happening successful the existent world alternatively than conscionable looking astatine immoderate numbers the hyperscalers fart retired and assuming it’s “all from AI,” and that “AI” intends thing much than conscionable the 2 main exemplary labs.

Investors successful Amazon, Google and Microsoft person been led to judge that the $994 cardinal spent connected AI GPUs and information centers exists to boost their existing businesses and build what amounts to the adjacent business revolution. In fact, this is the statement that conscionable astir immoderate AI bull will springiness you astir NVIDIA’s GPU income — that all compute will beryllium used because there’s endless, insatiable demand.

Well, different than the truth location isn’t.

What hyperscalers person actually done is demolish their free cash flow and purchased hundreds of billions of dollars’ worthy of GPUs, TPUs, and XPUs to support a customer guidelines dominated by 2 customers that are now accounting for the vast mostly of their gross growth and quite virtually cannot spend to salary their bills without a near-infinite travel of task superior investments.

Based connected these estimates, these analysts besides don’t look to judge that immoderate different ample customers are going to emerge, bringing into mobility some the rationale of their superior expenditures and those of fundamentally anyone building immoderate information halfway anyplace successful the world.

This is each very important, truthful I want to spell it retired really simply for you:

If 73% of Amazon, Microsoft and Google’s AI revenues are from OpenAI and Anthropic, and analysts judge that this attraction will only turn successful the adjacent fewer years, that intends location is not really that overmuch request for AI, and what request it has is from 2 companies that they person sunk a mixed $77 cardinal successful backing into — acold outpacing the existent gross publication that these companies provide, fto unsocial the capex spending of the hyperscalers.
- This gross besides represents a meaningful portion of Google Cloud, Microsoft Azure and Amazon Web Services’ revenue, suggesting that starring unreality platforms are not increasing arsenic accelerated arsenic investors person been led to believe.
If 27% of each of 2026 and 48% of each of 2027 Google Cloud revenues are from Anthropic and OpenAI, that intends that Google Cloud’s maturation has aliases will perchance stall successful the adjacent twelvemonth erstwhile you region their compute spend.
If location were real, meaningful request for AI compute aliases AI services, we’d spot it successful these estimates, overmuch for illustration we’d spot if location were different companies spending monolithic amounts connected AI.

Remember: Microsoft Azure, Google Cloud and Amazon Web Services correspond a ample chunk of all world unreality walk and AI compute, and frankincense are a typical sample of each AI compute…and if diverse, “insatiable” request existed, it would beryllium represented successful these estimates.

Sidenote: For the liking of clarity and transparency, Microsoft’s Amy Hood noted successful the astir caller net telephone that 90% of each unreality spending came from extracurricular the 2 main frontier AI labs.

The problem is that "cloud revenue," successful this case, encompasses a batch of things, including (but not being constricted to) "Microsoft 365 Commercial cloud, Azure and different unreality services, the commercialized information of LinkedIn, and Dynamics 365."

With that being said, the truth that frontier spending is conscionable 10% of unreality gross doesn’t show america thing — and is arguably a measurement of obfuscating really limited Azure is connected the 2 main AI exemplary labs.

This is the single-worst superior misallocation successful the history of business. Every azygous communicative you’ve publication astir the “incredible growth” of these unreality platforms is an embarrassing misread of 3 companies that are misleading investors that will much than apt beryllium forced successful the adjacent twelvemonth aliases 2 to person to restate revenues, trim remaining capacity obligations, and admit that they’ve drastically overbuilt capacity.

The counterargument to my warnings is ever that “this is useful infrastructure that will beryllium utilized successful the future,” aliases that we’re successful an OpenAI Bubble not an AI bubble (which, I argue, is fundamentally the aforesaid thing), but erstwhile you region Anthropic and OpenAI, Amazon Web Services and Google Cloud spell from breathtaking growth-engines to chernobyls of superior expenditure.

Without these 2 “startups,” AI revenues are catastrophically mini — for example, Sandler estimates that Amazon Web Services will make a pathetic $8.5 cardinal successful AI revenues successful 2026, aliases astir 25 times less than the $220 cardinal Amazon intends to walk this year. While Ju estimates that Google Vertex AI exemplary level (which is 1 of the main ways that ample enterprises merge Anthropic’s models) will propulsion successful $28.3 cardinal successful 2026, that’s still a small nether $10 cardinal less than the $35.6 cardinal that Anthropic and OpenAI will walk connected compute.

This needs repeating. Investors and the wide nationalist are being lied to. When you region OpenAI and Anthropic, Amazon, Google and Microsoft’s capex has apt accounted for very small gross growth, which intends that if either aliases some of them die, the mostly of superior expenditures and indebtedness raised arsenic portion of the AI bubble person been a waste.

So, let’s spell look astatine the non-Anthropic/OpenAI part of that Barclays note, pinch each file representing 2025, 2026, 2027 and 2028, pinch the past 3 being estimates.

For immoderate context, successful the twelvemonth 2025, Amazon spent $131.8 cardinal successful capex, or astir 32 times Barclays’ estimates for non-OpenAI/Anthropic gross — a number that hardly improves pinch the afloat full ($9.6 billion) to 14 times.

If Amazon has its druthers and invests $220 cardinal successful full capex successful 2026, the (pathetic) $8.5bn successful non-OpenAI/Anthropic gross will beryllium astir 26 times smaller, aliases 7 times smaller erstwhile you usage the afloat $31.6 cardinal successful projected AI gross for 2026.

If your counterargument present is that “the spread is getting smaller each year,” you are a mark. $31.6 cardinal is $22.6 cardinal little than Amazon spent connected superior expenditures successful its past quarter, aliases roughly $18.4 cardinal little than it invested successful OpenAI this year. Barclays’ estimates for 2028 person Amazon’s AI revenues — 75% of which are from OpenAI and Anthropic’s compute walk — astatine astir $75 billion, four deity damn years into the AI bubble.

Amazon will have, by 2028, apt sunk complete $650 cardinal successful superior expenditures into AI, each to gain (and this assumes OpenAI and Anthropic beryllium and tin pay) a small complete $171 cardinal successful AI revenue, pinch the immense mostly of it contingent connected 2 wholly venture-backed startups.

Similarly, moreover if UBS’ estimates travel true, Google will person spent astir $408.5 cardinal (including statement estimates of $120.5 cardinal for the remainder of the year) successful superior expenditures to create an AI business that makes astir $80 cardinal a year, pinch astir of that coming from either trading Anthropic’s compute aliases entree to its models via Vertex.

Microsoft is successful the aforesaid position. Wells Fargo’s estimates person its AI revenues for FY2026 (which conscionable ended) astatine astir $34.5 billion, successful a twelvemonth wherever it spent $115.9 cardinal successful capex, with $41 cardinal of that successful the past quarter, aliases astir $6.5 cardinal much than its entire estimated AI revenues for the deity damn fiscal year.

I recognize I’m being a small repetitive, but I request you to spot that without OpenAI and Anthropic, Microsoft, Google, and Amazon’s AI revenues are perfectly pathetic, and are frankincense entirely-dependent connected their compute spend.

Let’s beryllium serious, and return the absolute kindest read of UBS’ estimates, saying that Google’s Vertex AI level will make astir $22.5 cardinal successful yearly revenue, and assume, wrongheadedly, that it’s not near-entirely made up of request for Anthropic’s models…

…

Sundar Pichai, did you walk $288 cardinal deity damn dollars to make an yearly business that makes less gross than YouTube? We haven’t moreover talked astir margins aliases costs aliases whether immoderate of this is really profitable, mostly because it’s immaterial, as location is perfectly nary measurement to publication this business arsenic thing different than a historical failure!

Andy Jassy, is that you? Get your state ass complete here! You did NOT conscionable spell retired location and spent $429.5 cardinal deity damn dollars to guidelines up information centers for a brace of companies you person to virtually manus the money to them to salary you, did you? I’m gonna show momma Jassy what you’ve been up to! She’s gonna overgarment your backmost structure red!

Wait, what’s that?

You conscionable gave OpenAI $35 cardinal dollars? Wasn’t that dependent connected it going nationalist aliases reaching AGI? Are you kidding maine man? It’s almost arsenic if you recognize that the only measurement your largest customers are gonna salary y’all is by giving them the money to do so!

Okay, each jokes aside, there’s very intelligibly a problem present pinch AI demand, successful the consciousness that it doesn’t really beryllium without hyperscalers paying themselves to do so.

When you look astatine these numbers, you spot a sadistic communicative of unproductive capex. Looking astatine Wells Fargo’s estimates, it doesn’t look that Microsoft 365 Copilot is simply a meaningful business, hitting a meager estimated $3.859 cardinal for the entire fiscal twelvemonth 2026 for a merchandise that allegedly has 30 cardinal paid seats, suggesting monolithic discounts and questionable value.

Wells Fargo estimates it’ll turn to an unremarkable $10 cardinal successful yearly gross successful FY2027 — hardly much than OpenAI is estimated to walk successful Q1FY2027.

Very Blunt Sidenote: if Microsoft is struggling to waste AI-powered package attached to the literally-most-used endeavor package successful the world pinch a income squad of tens of thousands of group and tens of 1000 much resellers, how do you deliberation that the remainder of the AI package world is going to do agelong term?

To beryllium explicit: the request for AI-powered package is not there, and neither is the request for trading AI-powered add-ons to different software.

This is an embarrassing mishap of an manufacture pinch 2 ticking clip bombs underneath it.

There’re really 2 scenarios:

Anthropic and OpenAI, who correspond the near-totality of AI request and gross some arsenic a vendor and a supplier, are perpetually held up by the task superior manufacture and hyperscalers, astatine immoderate costs that is and to what lengths it requires complete financial fealty, to degrees of circularity unseen successful history,
One day, 1 aliases some of Anthropic and OpenAI die, which leads to half aliases much of the request for AI compute and existent manufacture accumulation evaporating, and immoderate further expertise for Google, Amazon and Microsoft to further provender themselves money.

And, to beryllium explicit, the past portion of that condemnation is precisely what’s going on. Microsoft, Google and Amazon are person spent complete a trillion dollars successful capex and equity investments specifically truthful they tin create maturation engines that are entirely-dependent connected Anthropic and OpenAI, who are entirely-dependent connected Microsoft, Google and Amazon to either (or both) provender them money aliases continually build them much infrastructure.

Sidenote: I haven’t moreover mentioned really Google’s $99 billion and Amazon’s $53.4 billion successful profits were inflated by their stakes successful Anthropic (and SpaceX, successful Google’s case), because I could constitute an full newsletter astir really deceptive and ridiculous it is that GAAP allows companies to do this. We request caller regulations, and we request them urgently, arsenic investors are being misled.

However you consciousness astir what I’m saying, these estimates besides require OpenAI and Anthropic to support increasing astatine the complaint basal to support up pinch expectations for Amazon Web Services, Microsoft Azure and Google Cloud.

The astir important mobility is which portion of the instrumentality breaks first.

The upwind cannot autumn retired of the sails OpenAI and Anthropic, arsenic some of them person to support gait to beryllium capable to salary for each this information halfway capacity, which would mean they would, crossed Amazon and Google alone, person to nutrient complete $125 cardinal successful 2027, which would require some the existent request (from customers for conclusion and for training) to usage that overmuch compute and the intends to salary for it (from task superior and the hyperscalers themselves).

Sidenote: To springiness you immoderate discourse astir really ample that magnitude of money is, Microsoft conscionable announced that its full fiscal twelvemonth 2026 gross for Azure was $100 billion.

For this to beryllium possible, some the request for entree to OpenAI and Anthropic’s models and the money to salary for the conclusion to service it must beryllium location to recognize these revenues and to support Google Cloud, Microsoft Azure and Amazon Web Services increasing astatine humanities rates.

To moreover person a changeable astatine doing that, compute capacity must travel online accelerated enough, which is an unfastened mobility successful and of itself. As I covered a fewer months ago, AI information centers are immoderate of the single-most eager building projects successful history, requiring monolithic amounts of capital, specialist talent, and materials, and execution that includes building decades’ worthy of powerfulness infrastructure successful a fewer short years, making them return anyplace from 18 to 36 months to complete. If capacity doesn’t travel connected accelerated enough, OpenAI and Anthropic can’t salary for it.

It seems very imaginable that the only logic maturation hasn’t stumbled for Microsoft, Google, and Amazon is OpenAI and Anthropic’s compute walk and the expertise to waste entree to their models, which intends that they whitethorn spot their superior expenditures arsenic existential.

It benignant of makes sense. If they neglect to build much and much information centers and proceed to descend money into Anthropic and OpenAI, maturation will slow crossed some their unreality platforms and associated services, arsenic the 2 AI labs are the only existent fierce purchasers of AI compute, which makes up the immense mostly of hyperscaler AI revenues.

It’s a vulnerable game. Without OpenAI and Anthropic, it’s clear that the underlying businesses of the large 3 hyperscalers are deteriorating, and that their AI plays are a catastrophic failure, because the sheer magnitude of rate they’ve required to day (and the moreover greater heap of rate they’ll request successful the months and years ahead) demands an outsized return for years to come.

Apparently things are truthful dire that the only measurement to spot complete slowing maturation was to money 2 elephantine startups beholden to monolithic compute contracts that provender task superior dollars to hyperscalers successful a information mobility that mostly equates to eating poisoned cardboard.

Things look awesome correct now, arsenic agelong arsenic you debar reasoning excessively difficult astir what it intends that truthful overmuch of this gross maturation is coming from Anthropic and OpenAI, and that their different AI plays are producing the lowest extremity of double digit billions of gross for thing that has costs them complete a trillion dollars, their free rate flow, and burdened them pinch hundreds of billions of dollars of debt, on pinch off-balance expanse liabilities now totalling complete $1.35 trillion (including Meta).

I tin already perceive the counter-argument that “Anthropic and OpenAI are the fastest-growing companies successful history,” and I surely dream you’re right, because location does not look to beryllium anyone other who wants to bargain compute astatine their standard different than hyperscalers trading it to them and immoderate weird also-ran bullshit Mustafa Suleyman, Demis Hassabis, and Alexandr Wang will beryllium allowed to do until 1 of the CEOs tries to make them the autumn guy.

I really request to beryllium arsenic clear arsenic possible: the existent statement position connected AI is wholly divided from reality. Based connected what I’ve shared pinch you today, it is ridiculous to propose that hyperscalers are building information centers nether the belief that they will make a batch of money aliases that request exists. They whitethorn judge — aliases dream — that’s the case, but that doesn’t make it true.

Outside of OpenAI and Anthropic, location appears to beryllium little than $30 cardinal dollars of non-AI laboratory compute request crossed Amazon, Google and Microsoft. I request to besides beryllium clear that this is almost surely an overestimate, because it includes revenues from Azure Foundry, Amazon Bedrock, and Google Vertex, which includes some compute and API walk connected Anthropic and OpenAI’s models.

This intends that we are apt overbuilding information halfway capacity astatine the standard of hundreds of billions of dollars. As the largest providers of AI compute pinch the astir acquisition and the biggest marque recognition, it’s difficult to reason that there’s pent-up AI request waiting elsewhere that hyperscalers haven’t realized. If anything, it suggests that everybody other is wholly and utterly fucked.

Perhaps you’ll reason that the expert was incorrect aliases that my study is incorrect aliases that request will magically appear, and you’re invited to if you want to proceed burying your caput successful the sand.

Let maine spell it retired for you: if Anthropic and OpenAI each had a tally complaint of $100 billion, they would still not person the standard to make the compute request to screen what their commitments are to Microsoft, Google, Amazon, and, of course, Oracle.

And CoreWeave. And Cerebras. And Cipher Mining and TeraWulf. And IREN. And Nebius. And Broadcom. And AMD. And SpaceX. And maybe Meta, SB Energy, and whoever mightiness build a $30 cardinal information halfway successful Georgia. While immoderate of these — for illustration Cipher, IREN, Nebius and TeraWulf — will travel gross straight to Google Cloud aliases Microsoft Azure, there’s still tens of billions of dollars’ worthy of compute gross that needs to get paid someway supra and beyond OpenAI and Anthropic’s walk connected the awesome platforms.

This is not sustainable. In fact, it’s beautiful fucking awful.

Let’s besides beryllium blunt astir something: neither OpenAI nor Anthropic person worked retired their business models. You tin fart astir claiming that Anthropic was profitable (it wasn’t) for a azygous 4th aliases repetition theoretical mantras astir “positive gross margins” aliases opportunity “they tin conscionable extremity training” each you want. These companies suffer tens of billions of dollars, they are horrendously unprofitable, an[d astatine this clip do not person an existent reply to “how do these businesses usability without infinite resources?”

Even if they were someway profitable — which they are not! — they would still request to turn astatine an intolerable rate. Putting speech each of the estimates from this piece, OpenAI projects to walk $750 cardinal successful compute successful the adjacent three-and-a-half years, which either intends it will request to turn its gross to hundreds of billions a twelvemonth very soon aliases raise half a trillion dollars aliases much complete the adjacent fewer years, astatine a clip erstwhile even hyperscalers are having problem raising that overmuch money.

And based connected some these estimates and the monolithic amounts hyperscalers are spending connected capex, I deliberation they’re good alert that location isn’t divers demand, and that the only way guardant is to proceed building capacity specifically for OpenAI and Anthropic, backing them successful immoderate measurement imaginable — either done backstopping the compute costs aliases helping shape monolithic backstage in installments deals — to make judge that gross maturation ne'er slows.

This is simply a doomed mission.

These estimates show that Microsoft, Google and Amazon do not person meaningful AI business extracurricular of the ones they’ve incubated, astatine slightest not ones that will salary disconnected their superior expenditures. Consensus estimates for Microsoft’s FY2027 capex are astir $186 cardinal successful a twelvemonth wherever its non-OpenAI/Anthropic AI gross is expected to beryllium $18.7 billion, meaning that moreover if these services had 100% nett profit margins (IE: zero costs), it would return a decade of those revenues to salary backmost the capex.

While you mightiness reason this is unfair — particularly arsenic OpenAI and Anthropic are improbable to dice earlier the fiscal twelvemonth ends — it is clip to commencement earnestly discussing what happens to hyperscaler revenues erstwhile they do so.

Put different way, investing successful Microsoft, Google, and Amazon arsenic portion of the AI waste and acquisition is an finance successful Anthropic and OpenAI’s expertise to some past and turn to go companies of comparative size and gross maturation arsenic their hyperscaler progenitors.

It is clear based connected the estimates I’ve shown coming that the immense mostly of maturation successful AWS, Google Cloud and Microsoft Azure comes from 2 companies that tin virtually not spend to salary their bills.

Jensen Huang has said that he has visibility into $1 trillion successful GPU income done the extremity of 2027, or, arsenic I estimated, about 40GW of compute capacity requiring $435 cardinal successful yearly revenue. Though these estimates do not specifically break retired compute request from Bedrock, Foundry aliases Vertex, the mixed AI revenues — including Anthropic and OpenAI’s compute spend, each API walk tally done the platforms, and Microsoft 365 Copilot — for their fiscal years 2027 sits astatine astir $304 billion, pinch the immense mostly of that (around $197 billion) coming from AI laboratory compute spend.

There is not capable demand. We are overbuilding information centers. If compute request existed to warrant the magnitude of information halfway capacity being built — aliases moreover close! — past expert estimates for AI revenues would beryllium some importantly higher and meaningfully divers alternatively than centralized astir 2 unprofitable, unsustainable companies.

To beryllium specific, for immoderate of this to “make sense” we’d request to spot aggregate different companies aliases groups of companies spending comparable amounts to OpenAI and Anthropic, melodramatic amounts of gross procreation from Google Workspace and Microsoft 365, and gross diverseness driven by multiple customers spending billions aliases tens of billions of dollars at the very slightest successful estimates for 2028.

It’s besides apt overmuch worse than I’m explaining because of really the large 3 bundle each azygous imaginable AI work wrong Foundry, Bedrock, and Vertex, each of which blend nonstop GPU rentals pinch API walk connected models from Anthropic and OpenAI, which I judge generates a ample mostly majority of gross connected these platforms alternatively than divers liking successful renting AI chips aliases different models.

Microsoft, Google, and Amazon are trading their investors a dishonesty astir their AI strategies, and successful a properly-regulated marketplace would beryllium forced to record investor disclosures that archive the dense gross attraction of Anthropic and OpenAI’s compute spend.

In not doing so, they proceed to mislead investors and the wide nationalist into believing that hyperscalers are backing the adjacent awesome maturation motor successful tech, erstwhile what they’ve really done is walk a trillion dollars successful capex and investments to make tens of billions of dollars of revenue, overmuch of which came from their ain equity investments.

And successful doing so, these hyperscalers person mangled their equilibrium sheets, tripling their PP&E, encumbering themselves pinch complete $500 cardinal successful information centers and GPUs that beryllium mostly to support 2 companies that can’t spend to salary their bills agelong term. At the extremity of this hype cycle, Microsoft, Google and Amazon (and, I guess, Meta) will person near themselves successful a much-worse information than before, pinch gross expectations that are overwhelmingly inflated by 2 unsustainable companies.

As I wrote successful the Rot-Com Bubble 2 years ago, these companies are fundamentally retired of hypergrowth ideas, and these expert estimates corroborate my absolute worst fears astir the information of these companies.

AI is not working. A $10 cardinal aliases $30 billion-a-year business is not capable to warrant either the monolithic superior expenditures aliases scars connected hyperscaler equilibrium sheets. In fact, it’s benignant of difficult to ideate what that mightiness really beryllium astatine this point, because Google, Microsoft and Amazon proceed to walk location betwixt $170 cardinal and $230 cardinal a twelvemonth successful superior expenditures, and each clip they do so, they summation the size of the payback necessary.

Sidenote: I’ve seen a statement floating astir Twitter from Bank of America that claims there’s $2.3 trillion successful backlog crossed “the apical 4 CSPs,” referring to Amazon, Google, Microsoft and Oracle.

Before anyone makes immoderate vivid assumptions astir these numbers, cognize that they mention to literally each dollar of incoming gross for these companies, including long-standing compute deals, Oracle’s database contracts, and astatine slightest (per The Information)

$1 trillion successful commitments from OpenAI and Anthropic, though retrieve that we do not really get a breakdown of RPOs based connected segment, company, aliases related position astir cancellation aliases amendment.When companies show you their RPOs, they do truthful so you’ll judge the gross is diversified and, ideally, astir immoderate hype rhythm they’re presently focused connected (like AI!), which is precisely what’s happening pinch fundamentally anyone covering them.

Let’s break down erstwhile they’ll beryllium realized!

Microsoft: 30% — astir $203 cardinal — will beryllium realized successful the adjacent 12 months, for a institution pinch complete $300 cardinal a twelvemonth successful revenue.

Oracle: 12% — astir $76.6 cardinal — will beryllium realized successful the adjacent 12 months, for a institution pinch astir $62 cardinal successful yearly revenue.

Google: 22% aliases truthful wrong the adjacent 24 months, aliases astir $115 cardinal for a institution pinch complete $400 cardinal successful yearly revenue.

Amazon: Amazon doesn’t disclose, but it has a backlog of $496 billion, pinch astatine slightest $238 cardinal of that from OpenAI and Anthropic, and yearly revenues of complete $700 billion.

At this point, AI would request to go — and this is without OpenAI and Anthropic — a business astatine the standard of Amazon Web Services ($170 billion, though this number is inflated by OpenAI and Anthropic’s compute spend), Google Search ($200 billion), aliases astatine the very slightest Azure ($100 billion, again inflated by some AI labs’ compute spend) to make sense, and moreover then, for this to make sense, hyperscalers would person to extremity spending money connected capex.

Put different way, AI bets cannot “pay off” if hyperscalers proceed to chimney 3 aliases much times their AI revenues each azygous twelvemonth into superior expenditures.

Sidenote: If I had to conjecture the rationale astatine this point, it’s that OpenAI and Anthropic person signed up for monolithic amounts of compute capacity that has yet to beryllium built or invested in, putting hyperscalers successful a vicious circle wherever they must walk much capex to earn gross from companies that they besides must support live done either backstops aliases outright equity investments.

I haven’t moreover gotten into the other vicious cycle — that the much of these information centers hyperscalers build, the much costly they go acknowledgment (at least, successful part) to the skyrocketing costs of representation that proceed to summation primarily because hyperscalers support buying servers for their AI information centers. As discussed past week, this only increases the magnitude of indebtedness they’ll need astatine a clip erstwhile the marketplace is getting progressively tense astir AI information halfway debt.

Yet arsenic we speak, the marketplace is ripping, because hyperscalers person swindled investors, the media, and moreover the analysts themselves. Article aft article aft article claims that AI bets person “paid off” because these companies are glazed immoderate clip they inflate their net utilizing the compute walk of 2 unstable and unsustainable companies, successful portion because hyperscalers some garbage to and look nary unit to stock their AI revenues, knowing that they’ll get in installments arsenic agelong arsenic the topline numbers look good.

I want to beryllium clear that the aerial is coming retired of these companies, nary matter really bully these net whitethorn look.

Everybody is taking the maturation of their existing businesses and 2 AI labs’ compute walk arsenic impervious that each this capex is paying off, moreover though location is now accordant impervious that the nonstop other is happening, and that their businesses are becoming increasingly-dependent connected that compute spend.

I understand that cipher really wants to deliberation astir the logical endpoints of what I’m arguing, truthful I’m going to do it for them.

Based connected everything I’ve said today, Microsoft, Google, and Amazon’s unreality businesses are intelligibly incapable of delivering the benignant of precocious maturation that Wall Street analysts like, and they’re utilizing some Anthropic and OpenAI’s compute walk and trading their AI models arsenic a intends of covering that up.
- This is simply a tangible motion that these companies are approaching their aureate years, turning from hypergrowth vehicles into boring, slow-growth mainstays.
- The problem pinch this is that they’ve raised indebtedness and spent capex astatine a level that requires their businesses to turn astatine melodramatic rates, and said maturation was only made imaginable by inflating revenues utilizing equity investments and 2 AI labs incubated by the hyperscalers themselves.
- Without these 2 companies — and, to beryllium clear, without these 2 companies becoming much, overmuch larger — hyperscalers do not person meaningful AI revenues successful comparison to their superior expenditures, making a payoff near-impossible based connected each estimate I’ve read.
- This intends that without these AI labs, they person very small to impressment Wall Street with, and without AI itself, their businesses are increasingly-stagnant and limited connected a pro-monopoly regulatory situation and the expertise to continually summation prices.
- All of this is to opportunity that I judge hyperscalers are connected the decline.
There is not capable request for AI compute, which intends that we’re successful an incredibly-large overbuild of AI information centers that are predominantly funded by project financing that tin only salary investors backmost if the information centers really person revenue.
- This intends that the immense mostly of information centers will spell unpaid, and those that do — and man, I americium not assured there’s much than a fewer cardinal dollars of non-AI laboratory request — are apt limited connected unprofitable AI startups aliases hyperscalers that don’t person overmuch request extracurricular of the largest AI labs.
- Though immoderate mightiness situation maine astir the standard of the problem, location are hundreds of billions of dollars’ worthy of AI information halfway loans, and I judge the immense mostly of them will spell unpaid. This will deed slope equilibrium sheets and backstage in installments costs to indeterminate levels.
- This intends that investments successful CoreWeave, IREN, Nebius, Cipher Mining, and immoderate different neocloud are efficaciously bets connected Anthropic and OpenAI, aliases connected hyperscalers’ continued interests successful backing them.
As location is not capable important non-OpenAI/Anthropic request for AI compute, this intends that net for NVIDIA, Broadcom, and efficaciously immoderate different semiconductor institution are inflated by what amounts to speculative purchases of assets, which could yet lead to impairments aliases restatements of earnings, and will surely lead to a driblet successful maturation erstwhile the feline is retired of the bag.
- This is why NVIDIA continues to do specified blatantly-circular deals, particularly pinch OpenAI, and why neoclouds proceed to motion deals pinch hyperscalers and OpenAI/Anthropic. Without them, request doesn’t beryllium astatine a standard that would warrant their existence.
OpenAI and Anthropic are some separately load-bearing companies. If either aliases some of them die, 40% to 73% of each AI gross and compute request evaporates.
- While they would apt still beryllium arsenic ammunition entities — and hyperscalers would proceed to waste entree to models — their deaths would termination the expertise for Microsoft, Google and Amazon to monetize their ever-expanding compute infrastructure, and would instantly statesman ripping elephantine holes successful their gross margins.
  - The worth of Anthropic and OpenAI to Amazon, Google and Microsoft is that they tin proceed to motion large compute deals without ever exposing hyperscalers to their existent underlying economics. This makes immoderate benignant of merger aliases acquisition somewhat useless.
- As Nik Suresh argued, a awesome woody of request for AI services aliases subscriptions comes from adjacent unit and a near-religious attachment to theoretical productivity benefits, astir of which would beryllium difficult to warrant if either of these companies died. This would time off very small to retrieve post-bubble.

To put things really simply, Anthropic and OpenAI are a measurement that hyperscalers tin provender their gross to themselves by spending money connected capex, backstopping compute contracts, aliases doing nonstop equity investments.

Their continued beingness allows the AI bubble to proceed inflating, but this tin only proceed arsenic agelong arsenic task superior and hyperscalers are tin aliases consenting to invest. There is simply not the request — not from unfastened source, not from different AI labs, not from self-hosting, not from anywhere — to warrant the capex aliases the monolithic information halfway buildout.

And for those arguing that location would beryllium a dot-com bubble betterment story, I must beryllium clear that if location isn’t request today, it won’t magically look tomorrow. AI GPUs will costs conscionable arsenic overmuch to tally successful 5 years arsenic they do today, arsenic will unfinished information centers costs conscionable arsenic overmuch to finish, arsenic will energy stay expensive, and each this will beryllium happening after it’s easy to raise task superior to really bargain the compute.

To quote my buddy Kasey, each awesome unreality compute supplier is solely opinionated connected OpenAI and Anthropic.

OpenAI and Anthropic are clip bombs, and erstwhile either of them explodes, everybody will inquire why we didn’t spot the brutality that follows coming.

The truth is that cipher wanted to look.

To look astatine these numbers and reconcile pinch their meaning is to admit that the existent authorities of the tech manufacture is based connected mania, deceit, information financing, and outright cons, and that the ascent of NVIDIA was chiefly driven by 3 companies building compute capacity for 2 unsustainable companies that became existential to their growth, inspiring hundreds of billions of dollars of discarded by obfuscating really small existent request existed.

I recognize it’s difficult to deliberation astir scary things, and really easy it is to disregard maine arsenic a doomer aliases a catastrophist, but excavation is simply a logical and logical statement successful an era poisoned by hype and grifting astatine a standard unseen successful history.

The top dishonesty of this era is that the tech manufacture is building the adjacent business revolution, erstwhile what they’re really building is simply a monument to everything that’s incorrect pinch modern capitalism — wasteful expenditures disconnected from immoderate existent benefits generated arsenic a intends of pursuing growth astatine each costs, mounting up a illness that will tear a spread successful the tech manufacture and the markets, and time off the world afloat of half-built monoliths sold to section communities arsenic occupation creators.

The truth we’re talking astir compute futures is simply a joke. The truth we’re talking astir AI factories is simply a joke. Almost each facet of the AI bubble is simply a joke, and successful the end, investors and the wide nationalist will beryllium the punchline. The rich | will person gotten richer, the banks will person harvested fees, the hedge costs will person traded and taken profits, the backstage in installments costs will person gotten their fees, and anyone who didn’t person an progressive wrong way will beryllium fucked.

All of this could’ve been avoided, but the world has a cult-like obsession pinch the wealthy, believing that the CEOs of the largest companies successful the world could ne'er make a bad decision, and that immoderate executive is automatically smart by kindness of being rich | and powerful.

And oh, really silly that’ll look successful retrospect.