Knowledge Commons
TR
HomeTopicsSeriesGuidesPopularAboutContact

Bubble or Debt? Meta's 145 Billion Dollars and the Class Character of the AI Bubble

Who Won While It Inflated, Who Pays If It Bursts?

35 min readAuthor: Oğuz Demirkapı
Bubble or Debt? Meta's 145 Billion Dollars and the Class Character of the AI Bubble

Bubble or Debt? Meta's 145 Billion Dollars and the Class Character of the AI Bubble

Who Won While It Inflated, Who Pays If It Bursts?

Dear Young Comrades,

This week a question came back to the table on Wall Street: is artificial intelligence a bubble? The question is not new; it has resurfaced every three months since autumn 2025 and been forgotten each time until the next earnings season. This time the occasion was Meta. On the evening of 21 October the company reported record revenue; the next day its shares fell 7 per cent and the Nasdaq suffered its sharpest daily loss since June.

Revenue breaks records, the share price collapses. To understand this contradiction we need a better question than "is it a bubble or not". In this piece we propose changing the question: who won while the bubble inflated, and who pays if it bursts? Because the bubble debate, as conducted in the mainstream, is an investors' debate; it revolves around whether to buy or sell the shares. Yet the 145 billion dollars Meta will spend this year is not a matter of share prices. That money comes from somewhere concrete and goes somewhere concrete: into electricity bills, pension funds, the wages of the thousands of workers laid off in May, the water of the towns next to the data centres.

First the summary, then the detail.

This piece is not investment advice. Figures are compiled from publicly available sources as of 24 October 2026; sources are linked in the text and at the end.

The Summary First: What Does This Piece Say?

For those in a hurry, the whole piece in nine points:

  • Meta reported record revenue; its shares fell 7 per cent. Third-quarter revenue rose 26 per cent to 64.6 billion dollars; but in the same quarter the company spent 34.9 billion dollars on capital expenditure and its free cash flow turned negative for the first time (−2.1 billion dollars). Its 2026 capex target was raised for the third time, to 135–145 billion dollars. That is double 2025.
  • The debate has moved from shares to bonds. The real signal came not from the stock market but from the debt market. CoreWeave's bond yield neared 16 per cent and its credit default insurance at one point priced in a probability of default above 60 per cent within five years; Oracle's credit rating fell to one notch above junk; even Meta's default insurance hit a record. This year alone, 466 billion dollars of debt has been issued for AI.
  • The money goes round in a closed circuit. Nvidia invests in OpenAI, OpenAI uses that money to buy Nvidia chips; Nvidia takes a stake in CoreWeave, CoreWeave rents Nvidia chips; 45 per cent of Microsoft's cloud order book comes from OpenAI. In one economist's words: how much of the revenue comes "from outside the ecosystem", and how much is "recycled capex"? Until that question is answered, the growth figures cannot be read.
  • There is a bubble not because the technology is useless, but because capital has over-accumulated. The railway was real too, and it still burst in the 1840s. In a Marxist reading the problem is not "does it work" but from whom so much capital, flowing so fast into a single field, will extract its profit. Today five companies carry 1.35 trillion dollars of debt on their balance sheets and 1.65 trillion dollars of commitments off them.
  • The depreciation debate is a class debate. Is a GPU's life three years or six? The answer determines the profit on paper. Marx called this "moral depreciation": the machine loses value before it wears out. Whether the shareholder, the lender or the worker will pay for this loss of value is an open question.
  • The price is already being paid. Meta laid off 12 per cent of its workforce in May and is investing 145 billion dollars in the same year: the investment is being funded out of the wage fund. In the US, 77 per cent of voters worry that data centres will push up their electricity bills; New York State has declared a statewide moratorium. AI debt now makes up 10 per cent of investment-grade bond funds; in other words, pension savings are in this bet too.
  • If it bursts, the bill goes down the ladder. That is what happened in 2000 and in 2008. The founders have long since sold their shares (CoreWeave's founders 2.3 billion dollars' worth since the IPO); the remaining risk lands on the pension fund, the small investor, the laid-off worker and the public money that will be mobilised for infrastructure that "must be rescued".
  • Seen from Türkiye, it is the same film. Between the scheme that gathered the savings of 455,000 people in the fund crisis with the promise of "20 per cent a month" and the data centre bond "yielding 16 per cent a year" there is a difference of scale, not of logic: while it inflates the insiders get out; when it bursts the outsiders are left holding it.
  • The stance of the class: neither wish for the bubble to burst nor fear it. Whether it bursts or not, computing power, data and models remain in the hands of a few monopolies; the problem is that ownership. The demand is not that data centres should not be built, but for whom and under whose control they are built.

Now to the detail.

What Happened This Week?

Let us put the facts in order first. They are all connected, but none of them on its own says "bubble". Read together, a pattern emerges.

DateDevelopment
13–15 OctoberThe yield on the bonds that AI cloud company CoreWeave sold in June at a 9.625 per cent coupon jumps to 14.5 per cent; the share falls 13 per cent in a day. On 15 October the company's five-year credit default swap (CDS) hits a record 1,195 basis points; the market prices in a probability of default above 60 per cent within five years (Forbes, Naked Capitalism).
16 OctoberThe Nasdaq closes the week down 2.1 per cent, its worst week since June. CoreWeave loses 24 per cent over the week. Oracle's CDS approaches 300 basis points; S&P has already cut the company to one notch above junk. The 10-year US Treasury yield is at 5.1 per cent (CBS News, Motley Fool).
20 OctoberA note by Dario Perkins of TS Lombard circulates: the AI ecosystem's debt could reach 2 trillion dollars by the end of 2027; most revenue comes from spending within the ecosystem; the trigger for a crash may come not from inside the bubble but from outside it (interest rates, political backlash, "one big accident") (Business Insider).
21 OctoberMeta reports third-quarter results: revenue 64.6 billion dollars (+26%), capex 34.9 billion dollars, free cash flow −2.1 billion dollars, 2026 capex target raised to 135–145 billion dollars. A new class of non-voting shares is announced (Seoul Economic Daily, Investing.com call transcript).
22 OctoberMeta −7%, Amazon −4.6%, Nvidia −4.4%, Intel and AMD −8%, Broadcom −4.3%. Nasdaq −2.6% (its sharpest fall since 5 June), S&P 500 −1.2%. The Nasdaq is 5.8 per cent below its 6 October peak. The same day Meta's five-year CDS hits a record. Brent crude passes 114 dollars (Al Jazeera, Business Today).
23 OctoberBranko Marcetic in Jacobin: "Wall Street is getting very nervous about AI" (Jacobin). Asian markets open sharply lower in the morning and close flat.

The thing to notice in this table: it was not the stock market that started the fall, but the bond market. A share price is an expectation; a bond yield is a cost. When someone says "I want 16 per cent to lend to this company", they have said more about that company's business model than ten reports from an equity analyst.

Meta's Results: Record Revenue, Negative Cash

Let us look closely at Meta's numbers, because these results triggered the debate and the whole story is in them.

The revenue side really is strong. 64.6 billion dollars, up 26 per cent on the same quarter last year. Ad impressions rose 10 per cent, the average price per ad 14 per cent. Daily active users: 3.64 billion. Meta's advertising machine works; AI feeds it with better targeting. The company presents this as "our investments are paying off", and that part is true.

The cost side is growing faster. Total expenses 51.6 billion dollars, up 68 per cent. Of this, 9.9 billion is the one-off cost of the youth protection settlement with state attorneys general in August; the rest is infrastructure, depreciation, data centre operating costs and third-party AI "token" costs. Operating profit fell 37 per cent to 13 billion dollars; the margin is 20 per cent. Expense guidance for the full year is 175–179 billion dollars.

The investment side is the story itself. 34.9 billion dollars in three months; 135–145 billion for the year. Look at the journey of this figure: at the start of the year it was 115–135 billion; then 130–145; now 135–145. The third upgrade. In 2025 Meta spent 70–72 billion dollars; so a doubling in one year. More importantly, the company's non-cancellable contractual commitments were 700 billion dollars in July and are now 825 billion dollars. This is money that will be paid in the coming years whatever happens.

The cash side is negative for the first time. Free cash flow −2.1 billion dollars. Meta has 82.4 billion dollars in cash and 83.7 billion in debt. In other words, one of the most profitable advertising companies in the world can no longer cover its investment out of its operating profit. CFO Susan Li said on the call that "these investments will take longer to generate returns than those in our core business" and gave no figure for 2027: "We are still developing our budget and capacity plan." This is exactly where investors are fleeing: 145 billion in 2026, 2027 uncertain, returns "over a longer period".

And two moves on the financing side. The first is a new class of non-voting shares. One new share will be distributed for every existing share; technically a split, but the purpose is clear: to create a "currency" that can be used for capital raising, acquisitions and employee pay while preserving Zuckerberg's voting power. The second is off-balance-sheet financing. 27.3 billion dollars of bonds were issued for Meta's Hyperion data centre in Louisiana; 80 per cent of the project belongs to a private equity fund called Blue Owl, and 18 billion of the bonds are held by Pimco. These bonds run to 2049; but Meta's lease on the data centre renews every four years. The bonds now trade at 91 cents (Forbes). In other words, the market has started to price the gap between a promise to pay until 2049 and a four-year lease.

Put these side by side: revenue at a record, profit falling, investment doubling, cash negative, 2027 uncertain, commitments at 825 billion, debt moving off the balance sheet. This is not a picture of collapse; but it is a picture of the moment a company that grew on its own profits turns into a company that grows on debt. And this moment belongs not only to Meta but to the whole sector.

Last month, discussing what Zuckerberg actually wants, we wrote that Meta's promise of "personal superintelligence" is a project to carry its advertising monopoly over to the next interface. These results confirm that thesis: the return on AI spending arrives as higher ad prices, while "superintelligence" still stands on the call as a promise that "we will bring it to billions of people". Muse has 10 million users after six weeks, and its global launch cannot go ahead because there is not enough computing power.

A Bubble That Lives on Debt

The real development, overshadowed by the stock market debate, is how the financing of AI investment has changed over the past year.

In 2023–2024 the big technology companies paid for their data centres out of their own profits. That is why the "bubble" warnings did not stick: there was no debt, no leverage, nobody was going to go under. In 2026 the picture changed. According to Forbes's compilation, 466 billion dollars of debt flowed into the sector this year alone; total AI-related debt has passed 500 billion dollars. According to Nikkei's calculation from SEC filings, Alphabet, Amazon, Meta, Microsoft and Oracle carry 1.35 trillion dollars of debt on their balance sheets and 1.65 trillion dollars of burden off them (leases not yet started, purchase commitments).

The form of the debt has changed too. Let us count four layers:

The first layer: corporate bonds. Meta, Alphabet, Amazon and Oracle issue bonds directly. These are still investment grade; but Oracle is rated BBB−, one notch above junk, and its CDS has risen from 141 to 308 basis points since January. In the words of an official at S&P Global, Oracle's bonds are priced "like BB- and B-rated borrowers".

The second layer: project bonds. Debt issued through off-balance-sheet structures (SPVs) tied to a single data centre, such as Hyperion. Here the lenders are funds such as Pimco and BlackRock; the borrower is not Meta but a vehicle company to which Meta pays rent. Meta's balance sheet stays clean; the risk passes to the funds and to the pension savings behind them.

The third layer: private credit. "Non-bank banks" such as Apollo, Blackstone, Magnetar and Blue Owl. In the second quarter of 2026 Apollo and Blackstone set up a 35-billion-dollar vehicle to buy chips to be leased to Anthropic; UBS called it the largest private credit deal in history. Almost all of these loans are floating-rate. When interest rates rise, the debt automatically becomes more expensive; and the 10-year US Treasury is at 5.1 per cent, while the probability of a Fed rate rise in December went from 50 per cent to 74 per cent in a month.

The fourth layer: chip-backed loans. CoreWeave's model: put the GPUs in a subsidiary, pledge the rental income from those GPUs as collateral, borrow. 27 billion dollars across six deals. The collateral is a chip that loses value every year. The debt is five-year, the customer contracts three years on average.

CoreWeave is where all four layers intersect, and that is why it is seen as the "canary". The company is a start-up that turned to renting out GPUs when crypto mining collapsed in 2019; in 2026 it has 51 data centres, roughly 52 billion dollars of debt and lease obligations, and a history of never having made a profit. Last quarter it paid 640 million dollars in net interest; that is a quarter of its revenue. According to UBS it must find another 102 billion dollars of debt by 2030. S&P writes that weaknesses in the company's internal controls cap its credit rating. And in January Nvidia invested another 2 billion dollars, nearly doubling its stake. Jensen Huang said this week that "CoreWeave's financial strength is growing every day"; the bond market said the opposite the same day.

Let us also note whose debt AI debt is. By Forbes's calculation, it now makes up about 10 per cent of the large investment-grade bond funds such as VCIT and VCLT; a few years ago it was 7 per cent. These funds are the "safe" part of ordinary retirement accounts in the US. So when CoreWeave's bond falls to 80, part of that fall comes out of a teacher's retirement account. Private credit has also started to be sold to small investors through so-called "interval funds", which limit withdrawals to 5 per cent every three months. The door is narrow; if the crowd wants out, it jams.

A sentence from an analyst called Hedgie sums this section up: "The AI bubble lives not in demand but in debt."

The Closed Circuit: Whose Money Goes Out, To Whom Does It Return?

The most repeated but least understood concept in the bubble debate is "circular financing". Let us show what it is through the facts.

Nvidia has committed to invest up to 100 billion dollars in OpenAI. OpenAI's CFO Sarah Friar openly acknowledged that this money will return to Nvidia as GPU purchases. 45 per cent of Microsoft's 625-billion-dollar cloud order book comes from OpenAI. OpenAI's own internal projection is a loss of 14 billion dollars in 2026 and cumulative losses of 115 billion dollars to 2029; profitability has been pushed back to the 2030s. Under the 500-billion-dollar headline of the Stargate project sit 52 billion dollars of equity; the remaining 448 billion is to be found through debt (R&D World). Nvidia is a shareholder in CoreWeave; CoreWeave rents Nvidia chips; CoreWeave's customers are OpenAI and Microsoft; Microsoft's customer is OpenAI.

Think of this not as a conspiracy but as a circuit. In our piece on physical AI we showed the same pattern at venture capital scale: start-ups funded by Nvidia, AMD, Amazon and Google pay that money back to Nvidia, AMD, Amazon and Google as GPU and cloud rent. Now the same circuit is running at trillion-dollar scale.

Dario Perkins's question reduces this circuit to a single sentence: is the revenue coming from outside the ecosystem, or is it "simple recycling of capex"? Gartner forecasts 2.7 trillion dollars of global AI investment for 2026; UBS calculates that the big cloud companies alone will spend 1 trillion dollars in 2026 and 1.45 trillion in 2027. According to the same UBS report, the capex of Amazon, Alphabet and Microsoft equals 102 per cent of their cloud revenue (24/7 Wall St.). In other words, the cloud business sinks more than it earns back into cloud infrastructure; the gap is covered by advertising and software profits, and now by debt.

The workings of this circuit are what Hilferding described in Finance Capital: monopoly does not abolish competition; it absorbs competition into its own circuit. Nvidia investing in its customers guarantees its chip sales; if the customer goes under, Nvidia's loss is the investment, while its gain is the chip price it has long since collected. Every link in the circuit is the creditor and the debtor of the next. That is why Perkins's warning is apt: a missed earnings figure or a cancelled order in a single link echoes through the whole circuit. If the big cloud companies choose not to renew their leases, the weak links are left exposed. The gap between Hyperion's four-year lease and its 2049 bond is the price of exactly this risk.

A Bubble? Three Answers and Our Question

Three answers circulate in the mainstream. Let us give all three their due, then say why all three fall short.

The first answer: yes, it is a bubble, and that is a problem. Last year Michael Burry wrote that Meta and Oracle were inflating profits by extending depreciation periods. Mozilla's CEO Laura Chambers said "there's a lot of junk being built very fast". In Jacobin, Marcetic writes that the sector is built on "artificially inflated hopes and bad accounting" and that investors are finally starting to act accordingly. In the Financial Times, Martin Wolf compared US valuations with September 1929 and noted that today's are higher; the top 10 companies in the S&P 500 make up 41 per cent of the index.

The second answer: yes, it is a bubble, but a good one. The formula of Lyft CEO David Risher: "We're in a financial bubble, but the technology is transformational." The dot-com bubble burst too, and Amazon and Google were left. A bubble weeds out the weak; capital is wasted but the infrastructure remains. This view is the story Silicon Valley tells itself, and it is partly true: the British railway mania of the 1840s burst, and the rails remained.

The third answer: there is no bubble; demand exceeds supply. Microsoft President Brad Smith: "We have more demand than supply." Nvidia, ARM and Google say the same. Meta's Muse cannot launch globally because there is not enough computing power; this too is counted as proof of demand.

What the three answers have in common is this: all of them answer the investor's question. "Will the share price fall, will the debt be repaid, who will be left with the infrastructure?" None of them asks: for whom is this infrastructure being built, and from whom is its cost already being collected?

Pay particular attention to the second answer, because it looks the most reasonable. "The railway remained" is true; but who was left with the railway? In the crash of the 1840s small shareholders, the middle class and thrifty artisans of the day, lost their savings; the rails were concentrated in a few large companies; tens of thousands of workers employed in construction were left unemployed when the sites stopped. The infrastructure "remaining" is not the same as that infrastructure belonging to the public or to the workers. After the dot-com crash the fibre cables remained; but Google and Amazon bought them up for pennies, the employees of the companies that laid them were left unemployed, and the pension funds holding their shares wrote them off. The "good bubble" thesis is the name for capital re-concentrating out of its own destruction.

That is why our question is different: who won while the bubble inflated, who pays if it bursts, and who pays if it doesn't? Pay attention to the third question; because even if the bubble does not burst, a price is paid, and that price is already being paid.

A Marxist Reading: Over-Accumulation and Fictitious Capital

This section needs a little theory; read it patiently, as the sections that follow lean on it.

Over-accumulation. The core of what Marx wrote about crisis in the third volume of Capital is this: capitalism must reinvest its profit; but not all capital can find a profitable field of investment. Accumulated capital becomes "excess" the moment it can no longer reproduce itself at a profit, and it looks for a new field. In the early 2020s the largest corporate cash pile in history had built up in the world: Apple, Microsoft, Alphabet and Meta did not know where to use hundreds of billions of dollars and were buying back their own shares. AI offered this capital an outlet. The data centre is the physical form of that outlet. So the question is not "does AI work" but "if this much capital flows this fast into the same field, from whom will it extract its profit?"

Fictitious capital. Marx calls fictitious capital all forms of value that rest on a future gain and are bought and sold today: shares, bonds, promissory notes. Fictitious capital is not useless; it finances production in advance. But it can inflate independently of production, and the more it inflates, the wider the gap between it and real value grows. In the first piece of the Stock Market File we showed this at Turkish scale: while the factory was closing, the share price multiplied 25 times. In the AI ecosystem the same gap looks like this: OpenAI will lose 115 billion dollars by 2029, yet today it is valued at more than 1 trillion dollars; Anthropic aimed for 2 trillion for its IPO and has now come down to 1.5 trillion. These valuations rest on the profits of the 2030s. If those profits are not there in 2030, today's price was fictitious.

Moral depreciation. Here there is a concept that is little discussed but decisive. Marx writes that fixed capital undergoes two kinds of wear: physical wear (the machine gets old) and moral depreciation (the machine loses value before it wears out, because a better one has come out). A GPU's physical life is perhaps six years; but when Nvidia releases a chip twice as fast every year, a three-year-old chip becomes unrentable. The depreciation debate Burry pointed to is exactly this: if Meta and Oracle depreciate their chips over six years, profit looks high; if they ought to depreciate them over three, part of today's profit is in fact a hidden loss. Gartner calculates that if AI revenue falls short of expectations, 3.5 trillion dollars of infrastructure could be written down by 2030. "Write-down" is an accounting term; its class equivalent is this: someone will pay the difference. The shareholder (the share falls), the lender (the bond goes bad), the worker (cost cuts) or the public (a bailout)? The depreciation debate is a class debate, because it decides where the loss lands.

Monopoly capitalism. The finance capital Lenin described in Imperialism was the fusion of banking and industry. Its form today: private equity funds such as Apollo, Blackstone and Blue Owl buying chips and leasing them to the monopolies; Pimco holding data centre bonds; Nvidia becoming a shareholder in its customers. Production, finance and supply are in the same few hands. That is why the expectation that "the bubble bursts, the market is cleansed and competition returns" is empty. It did not happen in 2000 either; the product of the crash was fewer and bigger monopolies. This time too, computing power, data and models are concentrated in three to five companies; even if the bubble bursts, this ownership structure remains, and the infrastructure of the weak links may even pass cheaply to these companies. In the Distillation File we spoke of "openness that serves those who own the chips"; a crash that serves those who own the chips is also possible.

One last theoretical note. Those who speak of "techno-feudalism" will look at this picture and say "this is no longer capitalism but a rentier order". Yet the picture shows exactly the opposite: pressure on the rate of profit, over-accumulation, credit expansion, speculative inflation, moral depreciation. These are the most classic laws of motion of capitalism. A rentier lord does not invest 145 billion dollars; a capitalist who must reinvest his profit does. What we are living through is a fixed capital cycle of monopoly capitalism, and it will end the way such cycles end: write-downs, concentration, the bill passed down.

Who Is Already Paying the Price?

The greatest ideological function of the bubble debate is to defer the cost into the future: "it would be bad if it burst." Yet the cost is already being paid. Let us count four addresses.

The wage fund. Meta closed the third quarter with 66,000 employees; a 12 per cent fall on the previous quarter. The cut was made in May 2026; in the same year capex doubled. Let us put it plainly: the data centre is paid for out of the payroll. What Marx said about machinery applies here too: investment in fixed capital is made to save on living labour. A company investing 145 billion dollars and laying off 12 per cent of its staff in the same year is not a contradiction; it is the plan itself. Moreover, this cut was part of the operation in which Meta liquidated workers with their own data in May. This was also the WSWS's warning in July: the expected return on the investment rests on AI "cutting costs" across industry, that is, on mass job liquidation. If the return does not come, the shareholder loses; if it comes, the worker loses. Either way, the worker loses.

The electricity bill. In the US, data centres have become one of the main issues of the 2026 midterm elections. According to a Reuters/Ipsos poll, 77 per cent of voters worry that data centres will push up their electricity bills; 64 per cent think rapid construction is not a good thing; only 14 per cent would be willing to live near a data centre. New York has imposed a statewide moratorium on new large data centres; the governor of Pennsylvania has introduced new rules for the permit process; the governor of Texas has called for oversight. In Utah a 9-gigawatt, 40,000-acre project was protested in front of the state legislature (Brookings). The mechanism is simple: the data centre buys capacity from the grid, the capacity price rises, and the price is passed on to all subscribers. Louisiana's electricity customers pay for the power line to Meta's Hyperion; Pimco's bondholders take Hyperion's profit. This is one of the most direct forms of transfer of surplus value, and it appears on no balance sheet.

Pension savings. We showed this above: AI debt is now 10 per cent of investment-grade bond funds. Since CoreWeave's IPO its founders have sold 2.3 billion dollars' worth of shares; the company has never made a profit. The fund that bought Hyperion bonds at 100 now holds them at 91. The same happened in the dot-com crash: between 2000 and 2002 most of the loss was written off against retirement accounts and small investors; the founders and the intermediary banks had long since got out. The same happened in the fund crisis in Türkiye: while the list of those who exited the funds on 13–16 September was protected by a "right not to be smeared", the losses of 455,000 victims were locked in at the liquidation price.

Public money and public consent. Data centres get tax exemptions, cheap land, priority grid connections and fast-tracked permits. In the event of a crash, declaring them "critical infrastructure" and rescuing them will also come onto the agenda; in our piece on Anthropic's activist monitoring we discussed what this "critical infrastructure" move means. The logic that rescued the banks in 2008 as "too big to fail" is waiting ready for data centres in 2027. Perkins's warning that "political backlash could be the trigger" in fact records that public consent is starting to be withdrawn. Capital reads this as a risk; we should read it as an opportunity.

What If It Bursts, What If It Doesn't?

Let us read both scenarios from a class point of view; because in both there is something that does not change for us.

If it bursts. The first losses fall on the weak links: debt-fuelled cloud companies such as CoreWeave and Nebius, enterprise AI start-ups, data centre construction companies. The funds holding their bonds write them down; retirement accounts shrink. The second wave hits the monopolies that place leases and orders with these companies: if 45 per cent of Microsoft's order book depends on a single loss-making company, the book shrinks when that company's financing dries up. The third wave is in supply: chipmakers, memory, energy. The South Korean stock market has halted trading 37 times this year; Samsung fell 10 per cent on the day it reported record profits. The fourth wave is in labour: in 2001 the US technology sector lost hundreds of thousands of jobs in two years; this time the workers in data centre construction, chip factories and energy projects, and the white-collar staff hired for "AI transformation", are next in line. And the fifth wave is in ownership: the infrastructure of the weak passes cheaply to the strong. The product of the crash is not a more dispersed market but a more concentrated monopoly.

If it doesn't burst. That is, if demand really comes, if OpenAI turns a profit in 2030, if Gartner's 3.5 trillion becomes "value" rather than "write-down". In that case we have to ask where the revenue will come from. The answer is written in these companies' own investor presentations: from businesses' "efficiency" spending, that is, from the wage fund. If AI spending is to be profitable, the source of that profit is the living labour it replaces. A bubble that does not burst may be heavier for the worker than one that does: ownership still stays in a few hands, job liquidation happens on top of that, electricity bills stay high, and "success" legitimises this model.

What stays constant in both scenarios is this: computing power, data and models are owned by a few companies. The bubble debate does not discuss this ownership; it discusses its price. We should discuss not the price but the ownership.

Seen from Türkiye

Looking from here, two things are visible.

First, the same mechanism at a different scale. Recall the fund crisis we have written about over the past month: the promise of "20 per cent a month", the closed-circuit fund cycle, the price inflating before the regulator's eyes, the class line between those who got out on 13–16 September and those who stayed. Now look at the 16 per cent yield on CoreWeave's bonds, the Nvidia–OpenAI–Microsoft circuit, valuations inflating before the eyes of the SEC and the rating agencies. The scale is a thousand times larger, the logic the same: while fictitious capital inflates the insiders get out; when it bursts the outsiders are left holding it. To read the fund crisis in Türkiye as "a corruption peculiar to us" is to ignore the same mechanism at world scale. The difference is this: in the US a bubble that bursts is rescued by the Fed and the Treasury; in Türkiye a fund that bursts is covered by "the liquidation's own resources".

Second, the data centre debate will come to Türkiye too. In our pieces on Türkiye's Artificial Intelligence Action Plan and the Directorate General of Public Artificial Intelligence we discussed the state's enthusiasm for AI infrastructure. In a country dependent on energy imports, managing electricity prices through subsidies and short of water, the question of whom data centre construction will serve and who will pay the bill is heavier than in the US. In the US, 77 per cent of voters worry about their electricity bills; in Türkiye the question has not even been asked yet. We are the ones who will ask it.

The Market's Question, the Class's Question

What the market asksWhat the class should ask
Will the share price fall?If the share price falls, whose pension savings melt away?
Will the debt be repaid?Who lent the money, and whose electricity bill and whose rent is the collateral?
Is the demand real?If the demand is real, which wages is the revenue replacing?
What should the depreciation period be?Will the shareholder, the worker or the public pay for the loss of value?
How much will Meta spend in 2027?How many people will Meta lay off in 2027?
Is data centre investment sustainable?Are the water and electricity of the town next to the data centre sustainable?
What happens if the bubble bursts?What happens to the worker if the bubble doesn't burst?
Which company will win?Who will own computing power, data and models?

So What Is to Be Done?

Reading and watching
  • Watch the bonds, not the shares. A share price is sentiment; a bond yield and a CDS are cost. Follow the CDS of CoreWeave, Oracle and Meta, and the price of project bonds such as Hyperion's. That is where you will see where the bubble is leaking.
  • Read next week's results with a class eye. The big cloud companies' results and the 28–29 October Fed meeting are coming up. In every set of results look at three lines: capex, free cash flow, headcount. Put the three side by side; the story tells itself.
  • Keep a map of the "circular financing". Who invests in whom, who buys from whom? Draw the Nvidia–OpenAI–Microsoft–CoreWeave–Oracle circuit on a sheet of paper. Add arrows as the news comes in. You will see the weak link of the circuit for yourself.
Talking and organising
  • Turn the "bubble" debate into a "cost" debate. If people around you are talking about the AI bubble, put next to the question "will it burst" the question "whether it bursts or not, who is paying". The wage fund, the electricity bill, pension savings: three concrete addresses.
  • Talk to your friends who are tech workers. Meta's 12 per cent cut is not a Meta affair; the same logic runs in every technology company. Read the organising debate in The Computing Worker's Handbook alongside these results: in a company that cuts jobs while its investment rises, the demand for a collective agreement becomes the most legitimate demand of all.
  • Open the data centre debate in Türkiye early. When a data centre project is announced, have the questions ready: who pays for the electricity, who pays for the water, how big is the tax exemption, how many jobs will it create, whose data is it, whose model is it? These questions should be put on the agenda now in local government, professional chambers and unions. The city council tools in our Taking Part in Local Government guide are there for exactly this.
Collective alternatives
  • Demand public computing infrastructure. The problem is not that data centres are built but for whom they are built. Computing infrastructure under public control, running open models, that universities, public institutions and cooperatives can share, is not a utopia; it partly exists in some European countries. At the Odak 2026 conference in December, Bilgi Müşterekleri will talk about exactly this under the title "From the People's Data to the People's Model"; let us develop the debate together until then.
  • Ask where pension savings are. Which bonds and which funds hold Türkiye's private pension (BES) funds, its auto-enrolment scheme (OKS) and its public pension investments should be a matter for the unions. The claim in the fund crisis that the Unemployment Insurance Fund had gone into the "20 per cent a month" funds showed why this is necessary.
Dear Comrades,

This week Meta reported the highest quarterly revenue in its history and its shares fell 7 per cent. This contradiction is a contradiction of capitalism: profit can only exist by being reinvested; reinvestment carries the risk of destroying the profit. Marx called this capital's "own barrier". The AI bubble debate is the name of that barrier in 2026.

We do not know whether the bubble will burst; nobody does. What we know is this: those who won while it inflated are known (the founders, the chip monopoly, the private equity funds), those already paying the price are known (the laid off, those whose electricity bills are rising, those whose pension savings are held at 91), and the address the bill will go to if it bursts is known. That address has never changed: in 1847, in 1929, in 2001 and in 2008, it went down the ladder.

Dear young comrades, if only one thing stays with you from this piece, let it be this:

A bubble is presented as a question of price; in fact it is a question of ownership. The price falls or rises; computing power, data and models stay in the same few hands. Our debate is not whether the 145 billion dollars will come back, but who will own the infrastructure built with that money. The machine is new, the chain is old; and what breaks the chain is not the price but organisation.

Sources

Meta's results
Markets and debt
The bubble debate
The data centre backlash
Theory
  • Karl Marx, Capital, Volume III, Part Three (the tendency of the rate of profit to fall, over-accumulation) and Part Five (fictitious capital); Volume II (the moral depreciation of fixed capital)
  • Rudolf Hilferding, Finance Capital (1910)
  • V. I. Lenin, Imperialism, the Highest Stage of Capitalism (1916)
Related pieces

Bilgi Müşterekleri — Knowledge Belongs to Everyone. This piece is not investment advice. Figures are compiled from publicly available sources as of 24 October 2026; company valuations and projections are as reported by the relevant sources. CC BY-SA 4.0.

Topics:Artificial Intelligence and TechnologyLabour and OrganisingEconomy, Capitalism and EcologyWorld and ImperialismSocialism and the Left
Tags:#artificialintelligence#finance#monopolycapitalism#capitalism#class#labour#meta#openai#marxism#usa

Related Posts

Newsletter