Ed Zitron：AI 泡沫要「成功」有多难

If you liked this piece, you should subscribe to my premium newsletter, and you can subscribe on the following links: $70 a year, $18 a quarter, or $7 a month.

In return, you get a weekly premium newsletter that dives deep into the finances of the AI bubble. Over the past few months, I’ve published a two part guide to the world of AI debt, an analysis of the (dubious) impact of generative AI on the global economy, deep analyses of the biggest players in the AI infrastructure space like NVIDIA (part one and two) and Oracle (part one and two), how OpenAI will eventually kill Oracle, and how venture capital is killing itself. 

Friday’s piece will be the Hater's Guide to Blackstone — one of the most loathsome companies in existence, and one that’s staked its entire existence on the AI bubble. 

If you want to get in touch — and especially if you have any juicy information about Anthropic, OpenAI, Blackstone or any other companies in the AI bubble — hit me up on Signal at ezitron.76. I’m also on IB on your Bloomberg Terminal. 

Soundtrack: Royal Blood — Pull Me Through

At the end of July, Apollo Chief Economist Torsten Slok gave a blunt warning to the Washington Post:

“This AI thing better work out because if it doesn’t ... we’re going to have a problem,” said Torsten Slok, chief economist at investment firm Apollo Global Management.

Today, I’m going to quantify exactly how difficult it will be for the “AI thing” to “work out.”

As I’ve said before:

…everybody wants to make the AI bubble really complex, it’s actually super simple: hundreds of billions of dollars are being invested to make single-digit billions of dollars maybe, some day, if AI data centers actually get built at scale and OpenAI and Anthropic can afford to pay for their compute. 

The theoreticals and hypotheticals of this industry are so powerful that they’ve created one of the (if not the) largest equities bubbles in modern history, all heavily-centralized around the valuations of a handful of names related to the so-called AI trade, which translates to “buying or selling stuff to do with AI GPUs.” 

The AI Trade Is A Global Concentration Risk, With Growth Dependent On Endless Capex Spend and VC-Backed GPU Rentals

This has led to the single-highest inflow of foreign investments into US equities in history — larger than the Dot-Com Bubble, the leadup to the Great Financial Crisis, and even the dying gasps of the Zero Interest Free Era in 2021, per Apollo:

However you might hand-wave away the AI bubble’s economic issues by saying it’s “just like the Dot-Com Bubble” — when the NASDAQ 100 fell 77% from its peak — the consequences are now global, at a scale unseen in history, flowing into both direct stocks and indices.

NVIDIA, Google, Amazon, Meta, Microsoft, Broadcom, Micron, AMD, and CoreWeave Represent 46.5% of the NASDAQ 100 (and 39.3% of the NASDAQ Composite), With Values Inflated By The “AI Trade”

As of typing this sentence, NVIDIA alone makes up 11.31% of the NASDAQ 100, with the combined weight of NVIDIA, Google, Amazon, Meta, Microsoft, Broadcom, Micron, AMD and CoreWeave representing about 46.5% of the index’s value, or 39.3% of the entire 3000-company NASDAQ composite. 

These same names — replacing CoreWeave with Oracle — represent around 28.9% of the value of the S&P 500. 

To sustain the valuation of these stocks — let alone continue growing — will require every company involved to either spend or make hundreds of billions of dollars in perpetuity, increasing every single quarter, without fail. This will come from either selling ever-increasing amounts of AI chips and related hardware, or buying and using them to generate revenue either through rentals or building their own software products. 

While the centralization was more severe at the top of the Dot-Com era NASDAQ — with Cisco, Intel and Microsoft accounting for roughly 74% — the total value of the index was a mere $1.82 trillion ($3.58 trillion adjusted for inflation). 

Today’s top three — NVIDIA, Google and Microsoft — account for around 29% of the NASDAQ 100, for a total market capitalization of $13.5 trillion.

I’m afraid things only get worse.

Retail Investors Own 10% Of The Market, With 38%+ US Household Wealth Held In Stocks, and Hold at least 22.4% of NVIDIA and 21.7% of Microsoft’s Traded Stock 

Per Goldman Sachs (via Reuters), "...equity gains have been the dominant driver of household wealth accumulation ​and the main contributor to a positive wealth ​effect on consumer spending," and per data from the Fed, corporate equities accounted for 37.9% of all household wealth in Q2 2026, the highest number on record since Q4 2000. 

Economist Paul Kedrosky told me it might actually be higher:

Americans are near records in terms of equity exposure, over 50% [as in number of Americans holding stocks], ahead of real estate for the first time since WWII — and Americans are at record levels in terms of equities as a percentage of their liquid net worth — over 60%.

While the Fed’s data only goes back to Q4 2000, corporate equities as a share of household wealth only sat at 25.9% ($11.5 trillion, or $22.7 trillion adjusted for inflation), back at a time when total household wealth sat at $44.4 trillion, or about $85.2 trillion in today’s money. In Q2 2026, total household wealth sits at about $195.8 trillion, with corporate equities sitting at roughly $74 trillion, or roughly 86.8% of the entire US household wealth 26 years ago.

Retails are also making up an historically-large amount of the volume of trades on the market, per Bloomberg Intelligence, making up 20.3% of trading in 2026 to date, approaching the dangerous heights of the worst excesses of the end of the Zero Interest Rate Policy era.

Retail Investors Have Over $4.65 Trillion In Exposure To AI-Related Stocks

Retail investors also directly own about 29% of Amazon’s, 26.2% of NVIDIA’s, 22.6% of AMD’s, 21.8% of Microsoft’s, and 18.6% of Meta’s stocks, though these numbers are not inclusive of index funds like SPY or QQQ (a fund tied to the NASDAQ). In totality, retail investors own equity assets (at least as of May) equivalent to about 10% of the market.

In fact, Goldman Sachs had retail volumes significantly higher:

In a note to clients, analyst Daniel Chavez explained that Goldman Sachs trading desk data shows retail volumes have risen 28% since mid-April, with a basket of retail-favorite stocks tracked by the firm gaining 29% over the same period.

By comparison, Dot-Com era retail trading was much, much lower — less than 2% of the NYSE’s trading volume according to the University of Michigan’s Alicia J. Davis in a paper that begins with the phrase “the American retail investor is dying.” 

When all of this is over, I think retail investors might wish they were dead, as they directly own somewhere in the region of $4.65 trillion worth of NVIDIA, Amazon, Microsoft, Google, Meta, Broadcom, AMD, Micron, Oracle, and CoreWeave stocks. The number is likely much smaller than the true exposure retail investors have, as it isn’t inclusive of investments in index funds.

Sidenote: For some comparison, the total market capitalization of Microsoft, Cisco and Intel on March 10, 2000 (the peak of the market) was around $1.36 trillion. 

Another point is that, unlike 2000, retail investors have easy access to exotic securities contracts that weren’t previously available (or, if they were available, were expensive or otherwise obscured from the casual view of most retail investors), and that have the potential to amplify any potential losses.  

RobinHood, for example, allows retail investors to buy event contracts (where “investors” speculate on an event with a yes/no outcome), to trade futures, to short stock, and to invest on margin, or with borrowed money. All of this can be done (assuming the investor meets the criteria to access these products) with a few taps on a phone, effectively eliminating the friction associated with making a potentially life-altering financial decision.  

A 20% Drawdown Across NVIDIA, Amazon, Microsoft, Google, Meta, Broadcom, AMD, Micron, Oracle, and CoreWeave Would Match or Exceed The Worst Day Of The Dot-Com Bubble

Sidenote: CoreWeave isn’t on the S&P 500 and Oracle isn’t on the NASDAQ, just for comprehensiveness.

On April 14, 2000, the NASDAQ composite dropped 9.67% and the S&P 500 5.78%, with a little under 97% of the S&P 500 constituents falling on the day, per TIME and The Washington Post, led by hot inflation numbers and overall anxiety over the tech sector’s valuations:

There were virtually no exceptions to the decline. The correction, analysts said, was taking no prisoners. The crash came at the end of an extraordinary week, where volatility and huge drops were the rule rather than the exception. And market analysts said they expect the declines to continue next week--and possibly in the coming weeks and months--as overvalued stocks come down to earth.

Yet the Post had one telling quote, emphasis mine:

Market momentum--which has trended up for the longest bull market in history for much of the past decade--has now turned down. The recognition that many technology stocks can't possibly grow quickly enough to merit the heights they achieved through momentum investing may have pierced the bubble of overvaluation in recent weeks, analysts said, as momentum changed direction. That turnaround has sent the Nasdaq composite index down 35 percent since its high on March 10.

While it seems unlikely, a drop of around 20% across NVIDIA, Amazon, Google, Meta, Broadcom, AMD, Oracle, Micron, and CoreWeave would have the same effect as the entire market downturn on the worst day of the Dot-Com Bubble crash. 

In truth, it wouldn’t even have to be that bad. 

Per CNN Moneyline News Hour transcript from April 14, 2000, Qualcomm fell 21%, Yahoo fell 20%, Intel a little over 10%, and Microsoft a mere 5% — and that’s all it took to make the market bottom out. 

Brief NVIDIA selloffs have already caused big market declines:

• On Jan 27, 2025, NVIDIA dropped 17%, its worst daily percentage loss (per Forbes) since March 2020, with the S&P 500 dropping 1.5% and the NASDAQ 3.1%, though its share of the S&P 500 at the time was about 1.9% smaller.
• On June 5, 2026, NVIDIA dropped about 6.2%, with the NASDAQ ending the day down 4.18% and the S&P 500 down 2.64%.

Anyway, remember that quote about momentum investing, overvaluations and growth? Here’s a quote from Bloomberg from last week (October 14), emphasis mine:

Chipmakers fell for an eighth straight day, the longest slide since 2008, on worries over whether the AI infrastructure buildout has outrun the ability of firms to get a return on their investments. CoreWeave Inc., a cloud-computing services provider whose bonds recently hit levels that typically signal distress, sank 12%. The S&P 500 lost 1.1%. Oil climbed, weighing on Treasuries.

And here’s another quote from CNN from a day later, emphasis mine:

The global retreat in chip stocks followed a sharp sell-off on Wall Street on Wednesday, after a strong US inflation report bolstered worries that rising interest rates will make it more expensive for companies to fund AI investments. Many AI companies are already heavily indebted.

It’s Not “Like The Dot Com Bubble,” It’s Worse, With Valuations Inflated By Hype-Based Capex Spend Rather Than Underlying Revenues

History may not repeat, but it rhymes. Every problem that burst the Dot-Com Bubble is here, but it’s worse, bigger and uglier, with global contagion risk that I basically never see discussed.

For example, the ongoing equity price of TSMC on the Taiwanese Stock Exchange makes up about 40% of the entire market, with the ODMs that have seen booming earnings making AI servers for hyperscalers representing another 3% or so. On The Tokyo Stock Exchange, the combined weight of SoftBank, RAM maker Kioxia, and the two major banks (SMBC and MUFG) backing AI data centers make up approximately 11.24% of the TOPIX Index. Samsung and SK Hynix, both of whom have seen remarkable RAM sales (see my Hater’s Guide To The Memory Crisis for more), make up more than 50% of the entire value of Korea’s KOSPI.

Their elevated valuations come from a belief that they’ll grow in perpetuity, much like the American markets.

And let’s be clear about something: all it takes is the vibe souring to the point that people lose faith in the AI trade for everything to start to collapse.

So, what will it take to keep the party going?

The answer is simple: perpetual, parabolic growth in both AI revenues and AI capex.

Gartner Says We Have More Compute Capacity Than We Have Demand, and 80% Of Neoclouds Will Be Dead By 2030

Per CNBC, well, exactly what I’ve been saying, emphasis mine:

“At the moment, we have more compute capacity than we have demand, and we have a lot of the tokens and processing power being given away in freemium solutions,” [Gartner’s John Lovelock] told CNBC. “Everyone in the AI race is running for a last-man standing position. Pretty much every market that Gartner is tracking in our AI race is likely to collapse down to a duopoly.”

Neoclouds would not be going out of business en masse, Lovelock reiterated, but could be subject to consolidation, such as takeovers and mergers. 

As a result of their profitability analysis, Gartner predicted last week that 80% of neoclouds, which are AI-dedicated computing power providers, could cease functioning as independent companies.

Oh! 

Are…are you sure about that? Isn’t AI the biggest, most hugest thing ever? It’s transformative! It’ll change everything! Aren’t we all excited-

Lovelock also said Monday that AI is a technology as opposed to a market, comparing it to table salt that will be sprinkled into existing commercial operations rather than creating an inherently new field in which companies will sell differentiated and competing products.

Okay, well, check out Doctor Doom over here, I bet he’s gonna be so surprised when those profitable returns come in-

In their Wednesday report, Gartner researchers found that overall “AI revenue will be insufficient to support profitable returns on all planned AI capital expenditures” between 2027 and 2030.

Well, at least there’s a diverse revenue base for GPU sales-OH MY GOD!

NVIDIA and Broadcom’s Continued Growth Is Contingent On Hyperscaler and Neocloud GPU/TPU Sales, Which Is Contingent On Debt and AI Revenues

NVIDIA has guided to a 70% increase in its next fiscal year — I assume FY2028, starting February 2027 — or over $670 billion in annual revenue, over three times its FY 2026 sales, with said sales dependent on whether Microsoft, Google, Amazon, Oracle, Meta, SpaceX (who collectively make up 50% to 60% of NVIDIA’s data center revenues), and the various neoclouds like CoreWeave, Nebius, and Nscale keep buying more of them every quarter, or at the very least that a willing vessel can step in to pick up the slack if they stop. If demand wavers, NVIDIA, Broadcom, Micron, AMD, and any associated silicon or hardware counterparties will see their revenue growth stumble, which will have a brutal knock-on effect on their stocks. 

NVIDIA already has its own revenue concentration problem — 44% of its first half FY2027 revenues came from three customers, and in Q2 FY2027, 70% of its accounts receivables were from three customers, with 16% of its revenues coming from one. One. How is this not more worrying to people?

Anyway, to buy those GPUs (or TPUs, in Broadcom’s case), everyone outside of Microsoft will have to raise tens of billions of dollars of debt at a time when underlying interest rates are growing and the overall debt markets have chilled around AI data center debt. 

Said chill comes from the world finally waking up to what I said in March 2025 — that CoreWeave is a time bomb of expensive debt that’s now selling off as people realize that making any of its supposed $104 billion in backlog (at least 50% of which is OpenAI and Anthropic) will require, per UBS, about $102 billion in debt in the next three years. CoreWeave’s long-dated bonds (by which I mean maturing in 2031) had an effective yield of as much as 16% in the last week. 

AI Debt Is Becoming Too Expensive To Handle, With Many Bonds Secured By GPUs and Infrastructure With Questionable Post-Bubble Value

With endless delays and hundreds of billions of dollars of chips sitting in storage, AI data centers have dramatically turned from the panacea for all infrastructure investment to fixed-income nightmare fuel. For whatever reason, nobody thought that lending hundreds of billions of dollars to a selection of different dodgy SPVs and neoclouds to build the world’s most ambitious infrastructure projects was a problem until, well, everybody, per the WSJ:

There has been a surge in issuance of bonds to finance the construction of data centers for AI computing. This year through Oct. 16, data-center issuers have raised more than $100 billion in the U.S.-dollar bond market, or about 25% of AI-related gross issuance in U.S. currency, according to data from Dealogic and Goldman Sachs credit research analysts. That is up from around 16% the prior year.

Per the Journal, this debt heavily relies on both the continued value of the data center and critical IT gear and whether the bloody thing actually gets built:

The recent AI data center bonds typically have structures that are different from a company’s bonds. They can be secured by the data center being built, while corporate debt is unsecured. How much security that provides, though, also rests on the completion of the data center, and on the value of the data center to another tenant. Effectively, investors are betting that overall demand for data infrastructure means the data center will be of use to someone, even if the current tenant runs into trouble.

In other words, the value of that data center debt is heavily contingent on whether the underlying assets retain their value, which in turn is contingent on whether there are customers that exist. For whatever reason, nobody wants to consider the most obvious problem: that if one customer can’t pay, other customers can’t pay, and if other customers can’t pay, there’s going to be less demand for AI data centers in general, which means there will be a glut of data center infrastructure, which will mean those assets you intend to resell if everything goes poorly won’t be worth as much.

I sure hope that financial institutions have a plan…

“With a data-center bond, you’re underwriting assets, and our view has been that when it gets completed, that can be a meaningful catalyst for spreads to compress,” says Andrew Keches, who is tech, media and telecom credit analyst and head of high-yield research at Barclays.

Oh okay, the problem with overly-expensive AI data center debt is simply that the markets will feel better when the underlying data centers (which are overwhelmingly delayed) get built? Awesome! It’s just that simple! I’m sure somebody out there will deal with the massive delays for the talent, materials and power connections necessary to complete them!  

As of writing this sentence, CoreWeave’s bonds are trading somewhere between 14.1% and 15.1%, and the only way it can calm investor anxiety will be building AI data centers at a rate that has proven next to impossible.

Part of the problem is that recent pushes by smaller and somehow-dodgier neoclouds like Firmus (which just pulled its IPO) and Nscale (whose investors are pushing for it to nearly halve its IPO float to $16 billion and just lost Microsoft as a client due to power delays) have made investors intimately aware of the overpromises of AI data center companies. 

Both Nscale and Firmus have the same customer concentration problems as CoreWeave, with 85% of Nscale’s backlog coming from Anthropic and Microsoft, and 90% of Firmus’ coming from Meta and OpenAI. Again, these are problems that CoreWeave, IREN and Nebius have had for years, but for whatever reason everybody has decided to wake up. Everybody has the same backlog customers — Microsoft (for OpenAI), Meta, OpenAI, Google (for Anthropic and OpenAI), Amazon (for Anthropic), and, of course, Anthropic. 

All this ire towards AI data centers means that the price of borrowing for them is skyrocketing on top of ten-year-dated Treasuries stubbornly sitting above 5%, which makes the overall economics of AI data centers increasingly problematic, as nobody can seem to measure A) how much they make and B) how much they cost.

Or, indeed, how any of this will pay off.

We Already Have An Oversupply of GPUs, With Amazon Warehousing Billions of Dollars Of GPUs “With No Clear Forecasted Path” For Them To Be Installed 

Congratulations to Eugene Kim of Business Insider for getting one of the biggest scoops of the AI bubble, confirming my estimate that billions of dollars of NVIDIA GPUs were being warehoused in this incredible story, emphasis mine:

Some AWS data center sites had room for more racks but not enough available power to turn them on, people familiar with the effort said, leaving the equipment to be staged at intermediary warehouses until the facilities were ready. These people asked not to be identified discussing sensitive matters.

Usually, Amazon doesn't pay for racks until the gear is put into a data center, leaving suppliers to carry the cost of the equipment in transit, according to the internal documents obtained by Business Insider. This helps the company conserve cash.

However, the intensity of the AI data center building boom has made this type of coordination harder. Amazon has to compete more for scarce resources as it jockeys with rivals for access to Nvidia GPUs and other AI components.

Under the new "Early Rack Receipt" process, Amazon now takes ownership of these racks earlier in the supply chain, rather than waiting until the equipment reaches a data center.

That means spending cash sooner. And any delays getting the racks installed and plugged into a reliable power source potentially pushes back the moment when revenue starts flowing from these assets.

At least one AWS supply chain team did not have "a clear forecasted path" for getting racks into data centers, according to one of the emails reviewed by Business Insider.

If you’re wondering exactly how much it’s warehousing…

One Amazon email projected the company would take delivery of about $5 billion of Nvidia GB300 racks in August, followed by another roughly $3 billion each month "for the foreseeable future."

That’s right folks — Amazon, which spent $54 billion in capex last quarter — is taking $3 billion worth of GPU deliveries every month “for the foreseeable future,” despite having such a supply glut that it cannot install them, and in some cases lacks an idea of when it may do so.

NVIDIA Appears To Be Channel Stuffing, Forcing Its Customers To Take Early Delivery Of GPUs  

As I discussed a few weeks ago, neoclouds and hyperscalers have over $374 billion in “construction in progress” assets, and I estimate that $200 billion or more of them are GPUs, primarily from NVIDIA. One quote from the piece stood out to me, emphasis mine:

Usually, Amazon doesn't pay for racks until the gear is put into a data center, leaving suppliers to carry the cost of the equipment in transit, according to the internal documents obtained by Business Insider. This helps the company conserve cash.

However, the intensity of the AI data center building boom has made this type of coordination harder. Amazon has to compete more for scarce resources as it jockeys with rivals for access to Nvidia GPUs and other AI components.

It appears that NVIDIA is channel stuffing. The phrase “compete more for scarce resources” doesn’t make sense in the context of GPUs that you don’t have enough power or space to install…unless, of course, NVIDIA is saying that if you don’t order and take delivery right now, the GPUs won’t be available by the time you actually build that data center.

The booster counter-argument will be that “Amazon is just planning ahead,” but riddle me this — if Amazon, Google, Microsoft (CEO Satya Nadella said he had chips “sitting in inventory he couldn’t plug in) and many other companies do not have the power to plug in the chips…why are they scarce? What’s making them scarce? If companies cannot install the chips (or, put another way, do not need them yet) what is creating the scarcity, other than something that NVIDIA itself is demanding its partners do?

Then there’s that curious matter of the whole “Amazon doesn't pay for racks until the gear is put into a data center.” How long has that been the case? 

Let’s get specific about the verbiage — “...suppliers…carry the cost of the equipment in transit…” heavily suggests that these racks (which had GPUs in them, by the way!) were not actually shipped to customers until the customers needed them…until some indeterminate point at which the “Early Rack Receipt” program took place.

This, if anything, proves that the “demand for AI chips” is far more a result of NVIDIA’s deliberate channel stuffing and antagonistic approach to doing business than anything about how many chips the world needs right now.

Sidenote: Fuck Jensen Huang, man. This means that the increased price of RAM and consumer electronics was a result of an artificial "supply chain crisis" created by Jensen Huang pushing NVIDIA's customers to take early delivery of GPUs. Jensen Huang is responsible for making life worse for everyone. What a greedy, joyless, soulless oaf of a man. Doing this requires you to be a certain kind of low that no money or status can ever fix.

Jensen Huang is a piece of shit. Every hyperscaler CEO is too. These scumbags drove the price of everything up so that they could all help Jensen Huang make more money. We did not need him to build all these GPUs. He just wanted to sell them to boost earnings. Everyone who participated in the early shipment of GPUs is a scumbag.

I would like to take this opportunity to apologize for taking a break from calling these people what they were. I decided to temper some of my more acerbic tendencies because I figured it should take a back-seat to my financial analysis. I won’t make that mistake again.

What these people have done is morally reprehensible, on top of the already-vast damage done by the very existence of generative AI. This destructive, pointless void of an industry has made everything it touches worse, all to prop up the vague dreams and excesses of the wealthiest, most powerful and utterly fucking dull pieces of shit I’ve ever had the misfortune of witnessing. Everything is more annoying, expensive and worse on some level because of generative AI, down to the price of fucking US treasuries. Everyone who aided in the ascent of generative AI should know they have profited off of making the world worse. 

A divide has emerged that is much, much larger than data centers. How into AI were you? How defensive of the industry were you? Were you normal about LLMs, or did you act strangely defensive of them? Why did you do that? How gullible were you about the financial or technical promises of AI? These questions will likely be asked of you in the next few years, and I believe your answers will be illuminating. Even if LLMs somehow survive in an open source form, the industry as it exists today is held together through cultism and private equity excess and is rigged to blow. The world will be different afterwards, eventually for the better. 

Anyway, even if hyperscalers installed all the fucking things, it’s unclear who would actually pay for them at a volume that would match analyst expectations.

Analysts Expect Big Tech To Add $1.2 Trillion In Cashflows In The Next Two Years — And It’s Unclear Where That Money Will Actually Come From, As Their Customers Are Expected To Have Far, Far Less Cash

Per Slok at Apollo, financial analysts covering the tech sector seem to be pricing in cash flows that do not match with reality, emphasis mine: 

The analysts covering tech expect the sector's operating cash flow to more than double to roughly $2.4 trillion by 2028, an increase of over $1.2 trillion, see chart below. Meanwhile, the analysts covering the other sectors in the S&P 500, which are tech's customers, expect those companies to add much less operating cash flow.

In other words, the tech silo is betting on a future in which demand for AI and tech services explodes, while the silos covering the companies that would pay for those services see a much more modest outlook. Both cannot be right at the same time.

The bottom line is that either tech's customers will generate a lot more cash than their analysts expect, or tech's cash flow forecasts are too optimistic, which raises the question of who exactly will be writing all those checks to buy AI services.

So, in case you didn’t know, every time a company reports earnings, financial analysts at big financial institutions set expectations — sometimes years in advance — of their expected revenues and cashflows. Right now, the analysts covering tech companies are guiding towards them adding $1.2 trillion in operating cash flows, all as the rest of the S&P 500 (ostensibly its customer base) doesn’t add anywhere near that much money.

This is one of those very simple points that nobody seems to be able to answer about AI: where will all the money come from to pay for those revenues? 

As Torsten said, tech analysts are expecting a parabolic surge of revenues across the entire tech stack, but the analysts covering them are saying the cash won’t be there.

What’s the plan, exactly?

The answer, it appears, is “Anthropic and OpenAI,” and maybe “open source” sales of indeterminate value.

Sadly, there’s another wrinkle in the plan.

Nomura Says That $60 Billion A Gigawatt Is The Most You Should Spend On A Data Center, And Anthropic Makes About $20 Billion Per Gigawatt, Which Suggests A Massive Mismatch In Data Center Economics

Japanese financial services firm Nomura put out a worrying take a few days ago, emphases mine and from the poster I found it from:

Anthropic hosted a presentation for investors on 15 October. The Financial Times reported sales at Anthropic of $16.8bn in Jul–Sep, versus $11.5bn in Apr–Jun and $4.8bn in Jan–Mar. We have not verified this. We regard ARR at Anthropic as a useful barometer of data center capex, as we think it is currently able to monetize data center capacity better than any other frontier AI lab. For example, (1) assuming that Anthropic has data center capacity of 5GW and ARR of $120bn at end-2026, we estimate that 1GW in compute resources translates into annual sales of $20bn. In addition, (2) if 60% of this revenue is spent on purchasing computing resources, we estimate that it will be able to spend $12bn a year on 1GW of data center capacity, and (3) if data center equipment is depreciated over five years, $60bn per GW represents about the most that can be allocated to data center construction (while noting that data centers also incur operating costs, and that the depreciation scheduled for power infrastructure depreciation is longer than five years). Increased ARR per GW is likely to boost data center spending, and the opposite is likely to hold true too. Assuming that growth in ARR at Anthropic slowed q-q in Jul–Sep, as the Financial Times article suggested, we think this might encourage a slightly more cautious outlook towards data center investment. Moreover, even if rivals were to lose market share to Anthropic, we note that revenues per GW could come under downward pressure across the industry as a whole.

We expect uptake of multi-agent systems (MAS), in which agents collaborate among themselves to execute prompts, and think demand for computing resources will rise once cost-effectiveness has been demonstrated. We think it makes sense to gauge whether the current cost of data center construction is commensurate with the economic benefits that AI models generate by looking at ARR generated by Anthropic and other frontier AI labs, as well as at new services and inventions.

To simmer it down, Nomura is saying based on how much a data center costs both to build and operate, you shouldn’t spend more than $60 billion given the amount of revenue you could hope to generate. 

It’s hard to get a precise measurement of how much one costs. The CEO of IBM said $80 billion a gigawatt at the end of 2025. Foxconn said $47 billion per gigawatt in September 2026. Back in June, Jensen Huang said future data centers could cost $80 billion to $100 billion a gigawatt. It is completely insane we do not have a consensus here.

Nomura also throws cold water on the many people excitedly suggesting that Anthropic could make $50 billion to $75 billion a gigawatt, suggesting instead that the number is around $20 billion. In other words, if Anthropic is paying anything more than $20 million a megawatt, it’s losing money hand over fist just on the compute alone.

Anyway, I sure hope that the people building these data centers aren’t dependent on Anthropic’s revenues…

Hyperscaler and Neocloud Revenues Are Contingent On Anthropic and OpenAI’s Continued GPU/TPU Rentals

More-importantly, outside of the companies selling the hardware, we know very little about the actual revenue generated from renting AI GPUs. Amazon, Google, Microsoft, and Meta refuse to disclose their actual AI revenues — but analysts estimate that more than 70% of Amazon, Google and Microsoft’s AI revenues come from Anthropic and OpenAI’s compute spend, which is contingent on their continued ability to raise tens of billions of dollars. 

As I’ve discussed in the past, I see around $22 billion in non-OpenAI/Anthropic compute spend globally, with an indeterminate amount of that coming from AI startups that, well…

OpenAI and Anthropic’s Revenue Growth and Fundraising Is Contingent On Unprofitable AI Startups’ (80% Of Revenue) Continually-Increased Token Spend, Which Is Contingent On Venture Capital Funding

Said billions are contingent on the AI labs’ continued revenue growth, and considering that 80% of Anthropic and OpenAI’s enterprise revenues come from loss-making AI startups allowing their customers to burn many times their subscription price in AI tokens (as much as $1,300 on Cognition’s $200-a-month plan, for example), meaning that Anthropic and OpenAI’s revenue growth is contingent on venture capital’s continued ability to fund their customers.

This is one of the core issues with measuring AI demand — that the vast majority of it comes from selling services that do not make any economic sense.

And their spend is contingent on their backers funding them in perpetuity.

That won’t be easy.

Continued Venture Capital Fundraising And Exits Are Hitting Historic Lows, With Investment Heavily Centralized In EBITDA-Negative AI Startups With No M&A or IPO Hopes

This, in turn, is contingent on venture capital’s continued ability to fund raise, and as I discussed in last week’s premium, VC has been on an historic losing streak since 2017, with the median net TVPI (total value put in, as how much you get in return on a per-dollar basis) sitting between 92 cents and a dollar, with 2025 fundraising dropping by 35%, and 2026 fundraising for first-time funds at its lowest rate in over a decade, per Murph Capital, which adds that despite representing only 6% of funds closed, funds of $500 million or more took 78.1% of all capital raised in 2026 to date.

AI startups, which are predominantly unprofitable, unsustainable, and require near-constant flow of venture capital, represented 64% of all venture capital investment in Q3 2026, with 50% of all global venture flowing into AI in 2025. 

As discussed both last year in the Enshittifinancial Crisis and last week’s premium, AI startups represent a devil’s deal for venture — by virtue of being buzzy “AI” companies, their valuations are naturally high, and by virtue of their capital-intensive businesses, they require more rounds of venture capital, with each round increasing valuations to unsustainable heights that make their IP-light, EBITDA-negative businesses impossible for either an IPO or potential acquisition.

In fact, both venture capital and private equity are having real problems selling their investments, with IPOs becoming effectively non-existent as a share of exits for PE, per S&P Global.

Meanwhile, per DealRoom, venture-backed from 2026 seem high, but are inflated by SpaceX’s IPO ($75 billion), the PE-backed take-private of gaming company Electronic Arts ($55 billion), SpaceX’s acquisition of Cursor (all stock, $60 billion), NVIDIA’s acquisition of Hugging Face ($13 billion), and Google’s acquisition of Wiz ($29 billion). 

And since 2015, the $2.7 trillion in venture capital invested has only managed to increase its value by about 1.6x — to around $4.4 trillion in value (not returns), or around a third of what you’d get if you’d have invested in the S&P 500.

In other words, venture capital is not particularly good at creating value, and has, in a period of poor value creation, centralized the majority of its investments in a single industry, with the majority of those investments going to companies related to or powered by Large Language Models that have yet to prove their economic worth.

And in this historically-low period of fundraising and exits, venture capital has invested a historically large amount of money, with the vast majority of it going into one bet: AI.

What’s that phrase? “You should always keep your eggs in one basket, so that you know where they all are”? I think that’s it, right?

Too Many Things Have To Go Perfectly Out For AI To “Work Out”

At this point, the sheer scale of the AI bubble is too large and unwieldy for things, to quote Mr. Slok, “work out,” and yes, I believe we “have a problem.” I say this with no sarcasm, as I think Torsten has done fantastic work posing reasonable, logical questions that nobody seems to want to answer, like “how does this work out?” and “where does the money come from?” even as most media outlets and analysts nod and smile at basically any statement said by an AI lab, hyperscaler, or neocloud.

For everything to “work out,” the current equities surge will have to continue in perpetuity, which requires both hyperscaler AI revenues — which are already depressingly small — and hyperscaler capex to expand infinitely, along with neoclouds, sovereign clouds and AI labs’ vague infrastructure plans to also expand infinitely. 

All the while, there can be no perceptual shift against generative AI, because any further souring of the debt markets will have neoclouds and would-be data center operators paying 12% to 15% in interest to build data centers sometime three years in the future for a customer base that’s pretty much entirely made up of Anthropic and OpenAI. 

As we’ve seen in the last few weeks, AI debt markets can quickly sour even without any bad news, and conditions will only improve in connection to an incredible earnings season for CoreWeave and other AI infrastructure players.

Said “incredible” earnings would require A) data centers to be built and B) said customers to actually pay for them, which roughly translates to “OpenAI and Anthropic’s continued ability to spend more and more money.” Anthropic’s 47% quarter-over-quarter growth rattled the markets because this company has to grow fast enough to cover $413 billion of non-cancellable compute contracts over the next few years, and everybody kind of knows deep down that the bubble can only continue with parabolic growth, which is why everybody freaked out when OpenAI reported a lower annualized run rate (a made up number) than expected, because growth is the only thing keeping this going.

And said growth rate is predominantly based on whether AI startups continue to let their customers spend anywhere from $2 to $10 in tokens per dollar of subscription, which in turn is dependent on venture capitalists’ continued status as the world’s most braggadocious paypigs, and at some point that venture money will run out, much like it did during the Dot Com Bubble.

I didn’t expect this piece to be so sobering, but when I sat down to write it it just kinda flowed out of me how profoundly difficult the AI industry’s future has become. The fundamental disconnection between the perception of AI growth inflating the value of hyperscaler stocks and the worrying fact that we don’t really know exactly how much AI revenue they make means that the S&P 500 and NASDAQ are held up by a kind of childish excitement that will, at some point, turn to existential fear.

Those reflexively responding with “they’ll get bailed out!” neither understand the markets nor the situation at hand. The AI bubble was inflated by hyperscaler revenue growth (without any disclosure of AI revenue) and NVIDIA’s GPU sales, meaning that the moment one of those slows, everybody freaks out. The US government cannot bail out NVIDIA, Microsoft, Google, Amazon, CoreWeave or Meta’s future misses of analyst expectations, nor can it cover every single bankrupt data center, nor can it bail out investor sentiment when it realizes the AI “boom” was yet another tech-based bust. I went into significantly more detail about it in my premium piece about how AI Isn’t Too Big To Fail.

Remember: even with the bank bailout in 2008, the market still crumbled, and that was before we had over $40 trillion in debt, stubborn inflation, and a stock market that has fundamentally disconnected from reality. 

The underlying weakness of the AI trade is that it isn’t really related to anything other than NVIDIA and the hyperscalers’ earnings. 

We don’t know how many GPUs are in operation, we don’t know how much money the hyperscalers make from using them or renting them, we don’t know how many active gigawatts of AI capacity they have, we don’t know their actual customer concentration outside of analyst estimates, we don’t really know how much people pay to rent GPUs, and we don’t even know how much it costs to run a GPU on an hourly basis. 

OpenAI and Anthropic are horrifyingly unprofitable, deeply unstable, and as I’ve discussed, 80% of their revenues come from companies that literally can’t afford to pay them, and yet the markets never react, because this information doesn’t really matter as long as the right numbers keep going up.

This is why Anthropic’s IPO travails are so profoundly dangerous for the AI trade. Investors are now suggesting that it prices its IPO at $1.5 trillion, meaning a significantly lower payoff for investors and much less capital raised on listing, meaning that it will have to take on more and more debt — like its 10% interest chip financing deal with Broadcom and Blackstone — that will provide a brutal education in what a high-burn-rate, negative-EBITDA company really looks like when you remove the layers of annualized run rates and hype.

And, per Irrational Analysis, Anthropic’s investor day didn’t go particularly well:

Broadly speaking, coding ARR has plateaued, or at least slowed significantly.

OpenAI seems to have no plan, as Sam struggles to raise a new private round at $1.4T valuation. Middle East investors (of all people lol) are pushing back on the valuation.

Anthropic apparently thinks bio (drug discovery) is the next big market for them. Yes, the most regulated and slow-to-revenue industry is what will fund all those "compute commitments”.

I know several people who were at the private Antrhopic (sic) buyside investor day last week. It was apparently a shitshow. Nobody asked an intelligent question. Anthropic was vague and gave little information. Original valuation plan was $2T, dropped to $1.6-$1.8T, and has been dropped again to $1.5T allegedly… in less than 10 days.

Any further delays to Anthropic’s IPO will be somewhere between catastrophic and fatal, and at this point I think it’s fair to question whether it ever makes it to the public markets.

If it does, a $1.5 trillion valuation would spell disaster for OpenAI, which has (as mentioned above) been trying to raise $30 billion at a $1.4 trillion valuation. Closing that round will be difficult to impossible at that price even without Anthropic’s drooping numbers, and if it somehow IPOs at something approaching $1 trillion, I have no idea how Clammy Sammy and OpenAI make it out of this one alive. 

And really, does any of this look like the future? Does any of this feel like the next industrial revolution?

Or does it feel like the run-up to another burst bubble and resulting financial crisis?

The difference here is that AI has, for the most part, been sold as a panacea for the rich and powerful — a way to control, to steal, to acquire and to expand, to take and take and take as they always have, occasionally vaguely muttering something about “abundance” as both the media and financial institutions help force through yet another doomed cycle of “innovation.” 

The people to blame are the tech CEOs — Sam Altman, Dario Amodei, Sundar Pichai, Andy Jassy, Elon Musk, Hock Tan, Jensen Huang, Sanjay Mehrotra to name a few — and the financiers that have helped them pump this bubble, with the worst of them being that rat fuck Stephen A. Schwarzman, CEO of Blackstone, a private equity (or “alternative asset”) firm that has funded basically every name I just wrote down on some level, funded by everyone from the Saudis to the California state pension system. 

This week’s premium will be the Hater’s Guide To Blackstone, a company that vibrates with evil, and has staked its entire future on whether AI works out.

It’s a stupid fucking bet from an evil fucking company, and I can’t wait to tell you all about it on Friday.

If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, $18 a quarter, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 10,000 to 18,000 words and provides vast, detailed analyses of the biggest events and companies in the AI bubble.

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.