ArticlesEnergy and infrastructure

The gigawatt and the dollar

The speed gap between AI and the grid also exists in dollars, and it is bridged the same way: with other people’s money.

Published on 20 October 2026 · updated on 28 October 2026

~$1tn
invested in AI infrastructure in 2026, according to Gartner
~$117bn
of annualised revenue for OpenAI and Anthropic combined
This article followsThe megawatt and the degreeWhy the real danger is not the speed of AI, but the gap between its speed and the grid’s.

From the physical gap to the financial gap

In “The megawattMegawatt (MW)A unit of power: what must be supplied at every instant. 1 MW = 1,000 kW. Not to be confused with the megawatt-hour, which measures an amount of energy used over time.See the glossary and the degree”, we argued one thesis: the danger is not the speed of AI, but the gap between how fast it calls for power and how fast a grid can respond. We had held one counter-argument in reserve. If the data centre pipeline deflates, the danger will be financial before it is physical.

Three weeks later, that counter-argument has become the central scenario. On 8 October, OpenAI told its investors that its annualised revenue was approaching $50 billion, not $70 billion. On the 14th, Gartner put possible write-downsWrite-downReducing the book value of an asset that is worth less than expected, for example servers that have become obsolete.See the glossary at more than $4 trillion by 2032. On the 15th, Anthropic announced $16.8 billion of quarterly revenue and its first adjusted profit, a few weeks before an IPO targeted at around $2 trillion.

So the market is doing very well and very badly at the same time. To get out of this contradiction, we propose tracking four indicators, the way one tracks a grid connectionGrid connectionHooking a site up to the public power grid, with the guaranteed capacity that comes with it. It takes years to obtain and is often the scarcest asset.See the glossary queueInterconnection queueThe list of projects (power plants, storage, large consumers) waiting to be connected to the transmission grid.See the glossary : the capex-to-revenue ratio, the gap between announced and operating gigawatts, the share of supplier financing, and the gap between private valuation and IPO price.

1. The capex-to-revenue ratio

In 2026, the sector invests about $8.50 in infrastructure for every dollar of revenue its two main labs generate. This is the first indicator, and the bluntest.

ScopeInvestmentRevenueRatioSource
Sector, 2026~$1tn of AI infrastructure~$117bn annualised (OpenAI ~50 + Anthropic ~67)~8.5Gartner via The Register, TechCrunch
OpenAI, 2026-2030$856bn of compute and infrastructure$840bn cumulative (assumption: ×10 in five years)~1.0FT via Reuters
Anthropic, no stated horizon$518bn of cloud and infrastructure obligations~$67bn annualised~7.7 years of current revenueDraft prospectus via Reuters; “more than $500bn” in the final prospectus
Sector, 2027-2030revenue needed for a 12% ROIC: $8.5tnexpected revenue: $3.6tn~2.4Gartner via The Register

Three readings stand out.

First, OpenAI’s ratio is the most telling because it comes from its own presentation. Even in its base case, its cumulative revenue does not cover its compute bill alone, before salaries and R&D. Hence the $278 billion of negative free cash flow it projects.

Second, the sector ratio overstates the gap. Part of AI’s revenue sits elsewhere (hyperscalerHyperscalerThe cloud giants that build and run their own data centers at very large scale: AWS, Microsoft, Google, Meta, Oracle.See the glossary cloud, software licences, advertising), and a data centre is depreciated over several years. But even halved, it stays above four.

Third, the two labs’ figures are not directly comparable. Anthropic counts sales made by its cloud partners in its annualised revenue; OpenAI does not, and accuses it of inflating its revenue this way. Anthropic replies that it follows established accounting practice (Reuters, via Hindu BusinessLine). Adding the two together, our ratio therefore gives an order of magnitude, not an exact measure.

Since then, Anthropic has published audited accounts, in the prospectus filed on 26 October. In the first half of 2026: $16.3 billion of revenue, against $1.2 billion a year earlier, and an operating loss of $693 million, against $2.1 billion. The business itself is therefore close to break-even. The net loss reaches $80.7 billion, but about 90% of that is a non-cash accounting entry: the rise in Anthropic’s valuation increases the value of its convertible notes and SAFEs. The more the company is worth, the bigger that loss gets. Compute remains the largest cost item, about $9 billion over the half-year. It has to be said plainly: this is a more nuanced picture than that of a lab burning through its cash.

2. The gap between announced and operating gigawatts

An announced gigawatt is not an asset: it is an option. In mid-2026, OpenAI had only 0.3 GW in operation out of more than 9 GW announced for Stargate, about 3%.

At OpenAI and Anthropic, fewer than one announced gigawatt in ten is in operation

Share of announced capacity actually in operation, mid-2026

  • Off-grid generation (US)2% · 2 / 90 GW
  • OpenAI (Stargate)3% · 0.3 / 9 GW
  • Anthropic9% · 1.4 / 15 GW
  • North American data centres19% · 39 / 203 GW
  • xAI (Colossus)50% · 1 / 2 GW
Sources: Epoch AI (Stargate, April 2026), MeasuredAI (Anthropic, August 2026), Wikipedia and Cleanview (xAI, off-grid generation), Sightline Climate (May 2026), as cited in “The megawatt and the degree”. Anthropic leases most of its capacity.

This indicator links the two articles directly. In the first, we quoted Sightline: a data centre announcement is a request for electricity, not a commitment to build. Here, it is also a request for financing. And announcements keep swelling faster than concrete. In August, OpenAI signed a 20-year lease for a 10 GW campus in Ohio, developed by SB Energy, SoftBank’s subsidiary (SiliconANGLE). Nvidia, for its part, puts OpenAI’s existing and planned commitments to its chips through 2030 at about 12 GW (AI News).

An announcement can also run into a permit. On 7 October, the Finnish environmental authority ordered Google to suspend preparatory work on two data centres, in Muhos and Kajaani: its subsidiary had started clearing land without the mandatory impact assessmentImpact assessmentThe file that assesses a project’s effects on the environment (water, air, noise, biodiversity), mandatory above certain thresholds.See the glossary , over more than 300 hectares of forest according to the Finnish Association for Nature Conservation. These sites are part of the €13 billion announced in September, Google’s largest investment in Europe. They remain halted until the assessments are complete, and the police have opened a preliminary investigation into Muhos (Al Jazeera, Yle).

Or on the grid. On 18 October, Microsoft pulled out of the data centre Nscale was planning in Loughton, Essex, presented by the British government as the country’s largest, citing the power supply (The Guardian, City AM). Nscale, which lost $1 billion in the first half on $141 million of revenue, is nonetheless negotiating a second $45 billion contract with Anthropic, according to Bloomberg (TechCrunch).

The gap matters for two reasons. On the physical side, every announced gigawatt enters a connection queue and weighs on capacity auctions, even if it is never built. On the financial side, valuations are calculated on promised capacity, while revenue comes from installed capacity.

The case of xAI shows the other side: 50% of its announced capacity is in operation, but at the cost of a private 1.2 GW gas plant. The gap only closes quickly when you bypass the grid.

3. The share of supplier financing

About a quarter of Nvidia’s business next year will come from labs it funds itself: its chief financial officer told analysts so on 26 August (AI News).

The mechanism is simple. The supplier invests in a lab, or guarantees its credit. The lab builds a data centre full of the supplier’s chips. The purchase becomes revenue for the supplier, which reinvests. Nvidia itself uses the phrase circular financing, in order to dispute it.

SupplierCommitmentFormStatusSource
Nvidia~$50bn already invested in labsequity stakesdoneAI News
Nvidiaover $500bn of third-party capital (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR)financing platformssubject to definitive agreementsAI News
Nvidiaguaranteed lease of part of neocloudsNeocloudA cloud provider specialised in renting GPUs for AI, such as CoreWeave, Nebius or Nscale. Often young and debt-financed.See the glossary ’ capacityrevenue floor for their lendersin placeAI News
Broadcom$42bn of senior debt for Anthropic, plus $18bn of junior debtdebt backed by Broadcom and the chips’ residual valuejunior debt conditional on published accountsMotley Fool
Broadcomup to $50bn for OpenAIfinancing for custom chip purchasesunder discussionMotley Fool

The textbook case: Google’s $200 billion machine

The most elaborate set-up is not Nvidia’s. It is the one Google has built around Anthropic, revealed by the Financial Times on 4 August: about $200 billion of contracts, 80% of them tied to the chips themselves, to supply more than $150 billion of TPUs to the lab (FT, translated by Expansión).

The starting problem is simple: nobody wanted to carry tens of billions of chips on their balance sheet. Google is already funding record capex. Broadcom agrees to buy the hardware, but does not want to become a financier. The solution borrows the model Boeing and GE invented to sell aircraft and engines.

Google, Broadcom and Wall Street carry Anthropic’s chips

About $200bn of contracts, 80% of them tied to the chips

sellsresellsleasesGoogledesigns the TPUswith Broadcomsells 4.5 GWBroadcombuys the TPUs$128bn of purchasesover 2027-2028Compute SPVset up by Morgan StanleyApollo, Blackstone debt$35bn for ~1 GWAnthropicleases the chips$125.2bn over 5 yearspays the rentData centresTeraWulf, Cipher,Hut 8: formerbitcoin minersguarantees the leasesup to $44bn,$815M on its booksguarantees ~$30bn of 35if Anthropic does not payhost the chips: 2.4 GW secured for 4.5 GW soldflow of chips and capacityguarantee: the risk shifted onto a third party
Sources: Financial Times, 4 August 2026 (Expansión and Folha translations); Anthropic’s prospectus, via CNBC, 1 October 2026.

Each player carries a piece of the risk: Google the data centres, Broadcom the chips, private credit the financing. Anthropic carries only the lease payments.

The prospectus adds one more piece to this arrangement: 48% of Anthropic’s revenue goes through the cloud platforms of Amazon, Google and Microsoft. Google and Amazon invest in Anthropic, Anthropic buys compute from them, and nearly half of its revenue reaches it through them. The loop closes on the revenue side too.

The first batch went through the machine in June. A special-purpose vehicle, Compute SPVSPVA company set up to hold a single project and its debt (special-purpose vehicle). It keeps that debt off the balance sheet of the group that uses the project.See the glossary , set up by Morgan Stanley, paid $35 billion for about 1 GW of hardware, close to a million TPUs. The money comes from three tranches of debt held by Apollo and Blackstone. Broadcom guarantees the two senior tranches, for about $30 billion: if Anthropic stops paying and the chips resell poorly, Broadcom fills the hole.

The second batch is much bigger. In April, Google agreed to sell another 3.5 GW of TPUs to Broadcom for Anthropic. In its filings, Broadcom shows $128 billion of purchase commitments: $55.2 billion of deliveries in its 2027 fiscal year, $72.9 billion in 2028. Anthropic’s prospectus, for its part, mentions a $125.2 billion five-year commitment to lease this TPU capacity, about a third of which could be funded by a Broadcom convertible loan of up to $42 billion (CNBC).

Five details make this set-up remarkable.

It explains the map in the first article.
We noted there that Anthropic was leasing former bitcoin farms (TeraWulf, Cipher, Hut 8) without understanding why. The FT gives the answer: Google guarantees their leases. TeraWulf was the first, for a 360 MW data centre in New York State; Morgan Stanley turned that guarantee into a $3.2 billion construction bond. In exchange, Google received warrants on TeraWulf’s equity.
The risk is almost invisible on the balance sheet.
Google’s guarantees on 10 projects (2.4 GW) commit it to paying up to $44 billion if all the leases default. It books them at $815 million, about 54 times less.
The gap from the second indicator reappears inside the set-up itself.
Google has committed to selling 4.5 GW of TPUs, but has secured only 2.4 GW of data centres to house them. A Google executive told the FT that the team now spends all its time on energy. The megawatt catches up with the dollar.
The guarantee creates a cost-of-capital advantage.
Projects backed by Google borrowed at 7.1% on average, against 9.3% for neoclouds built around Nvidia. Jefferies speaks of a structural disadvantage for the Nvidia ecosystem. A hyperscaler’s balance sheet becomes a commercial weapon, as grid access was in the first article.
The risk is huge, but concentrated on a single customer.
The $200 billion depends on Anthropic’s ability to pay its leases. The prospectus states that a payment default could make a large part of the lease commitments immediately due. In 2027, Anthropic will become Broadcom’s largest customer; Broadcom is targeting $115 billion of AI revenue that year and $230 billion in 2028.

A partner at Rothschild & Co sums up the problem to CNBC: the market is making a very concentrated bet on the ability of two companies to generate enough revenue to repay everything that has been financed. Google is at once Anthropic’s investor, the supplier of its chips and the guarantor of its data centres. If the customer pays, each of them gets paid three times. If it does not, each of them loses three times.

The weak point comes down to one word: the guarantee. Nvidia defends its exposure by explaining that a repossessed chip can be resold to another customer. Broadcom’s debt rests partly on the chips’ residual value. Both arguments hold as long as demand exceeds supply. They give way together on the day it no longer does, since that is precisely the day residual value collapses.

This is the financial equivalent of the third route in the first article. The physical gap was bridged by sending the bill to households; the financial gap is bridged by shifting the risk onto the supplier, then onto its lenders. In both cases, the one who pays is not the one who decides.

4. The IPO price against private valuations

Anthropic’s bankers are reportedly targeting a valuation of $1.5 to $2 trillion, one and a half to two times its last private valuation of $965 billion, set in the spring. The IPO is the moment the public market says whether it will pay for the same promise as private investors.

CompanyLast private valuationTarget valuationPremiumMultiple of annualised revenue
Anthropic$965bn ($65bn round, spring 2026)$1.5 to $2tn, prospectus filed on 26 October×1.6 to ×2.1~22 to 30× ($67bn)
OpenAI$852bn ($122bn round, March 2026)~$1.2tn under discussion, IPO pushed back to 2027×1.4~24× (~$50bn)

Sources: Motley Fool, Euronews, FT via Reuters, TechCrunch.

These multiples are justified only by future growth. Anthropic’s valuation rests on a forecast of $190 to $200 billion of revenue in 2028 (Economic Times). OpenAI’s, on revenue multiplied by ten by 2030. The IPO price therefore measures the market’s confidence that the gap described in indicators 1 and 2 will close.

There is already a precedent this year. SpaceX, which now includes xAI, went public in June at $135 a share, for a valuation of about $1.77 trillion (Motley Fool). The stock gained 19% on the first day, then fell as low as $108.27 in July, about 20% below its IPO price. It only climbed back above on 10 August, after a better-than-expected quarter (CNBC). In the meantime, the end of the first lock-up period released 911 million shares, more than the 639 million sold in the IPO.

The first public test of a giant AI-linked valuation thus showed two things: the market goes along at first, then asks for proof again quarter after quarter.

The loop: four indicators, one question

The four indicators measure the same distance, the one between a promise and an asset that produces revenue. They hold each other up.

The capex-to-revenue ratio is high because investment anticipates capacity that does not exist yet. That capacity remains partly theoretical, as the gap between announced and operating gigawatts shows. It is partly financed by suppliers, whose guarantee depends on the value of the chips. And the value of the chips depends on demand that the IPO has to justify. If one link gives way, the other three tighten at the same time.

The first article set two clocks against each other, AI’s in days and the grid’s in years. Finance adds a third, which falls between the two:

Chips
are depreciated over 2 to 4 years.
OpenAI’s cash
raised in March, would run out by 2028.
The revenue
that justifies the valuations is expected in 2028 (Anthropic) and 2030 (OpenAI).
The grid
takes about 5 years between a connection request and commissioning.

So a chip can be obsolete before the grid has delivered the megawatts that power it, and before revenue has paid back its price. This is the write-down mechanism Gartner describes.

This financial clock is speeding up. The yield on 10-year US Treasuries reached 5.30% in October, the highest since 2002: borrowing to build costs more. Insurance against an Oracle default is back to its 2009 price. And the banks that financed the data centres are trying to pass some of those loans on to other investors, because they are nearing their exposure limits on AI: a $38 billion construction loan, for sites leased to Oracle, has found no taker for six months (Startup Fortune, citing the FT and Bloomberg).

That leaves the question that already ran through the first article: who carries the risk? Not the labs alone. Anthropic, which owns neither buildings nor substations, has shifted most of its assets onto AWS, Google, former bitcoin farms and now Broadcom’s debt. OpenAI has shifted it onto Oracle, SoftBank and Nvidia. Gartner estimates that 80% of neoclouds will not survive as independents after 2030. In both articles, the gap is bridged by those who did not make the decision.

Scroll: the deals light up in the order they were signed. Hover over or tap an arrow for the details.

Overview
  • Investment
  • Compute purchase
  • Financing, guarantee
  • Acquisition
  • Cut-off
Anthropic : Maker of Claude. IPO targeted at around $2 trillion.AAnthropicLabOpenAI : Maker of ChatGPT. IPO pushed back to 2027.OOpenAILabSpaceX · xAI : Absorbed xAI in February 2026. Leases its Memphis data centres to Anthropic and Google.SSpaceX · xAILabCursor : Code editor acquired by SpaceX in August 2026. Offers Claude to its users.CursorSoftware and devicesGoogle : Designs TPUs with Broadcom. Investor in, supplier to and guarantor for Anthropic. Supplies Gemini to Apple, rents TPUs to Meta, and rents Nvidia GPUs from SpaceX for Gemini.GoogleCloudApple : Bases its next generation of models, Siri included, on Gemini. Uses Claude internally; ChatGPT still powers some Apple Intelligence features.AppleSoftware and devicesMeta : Rents compute from Google, Nebius and CoreWeave, buys up to 6 GW from AMD, and announced in July 2026 that it wants to resell its spare capacity.MMetaLabNebius : European neocloud. Customers: Microsoft ($17.4bn over five years) and Meta (up to $27bn).NebiusData centresCoreWeave : Neocloud built on Nvidia hardware, in which Nvidia is also a shareholder. Customers: OpenAI (≈$22.4bn), Meta (≈$35bn), Anthropic.CoreWeaveData centresAmazon : AWS, Anthropic’s first compute supplier, and an investor.AmazonCloudMicrosoft : Azure. Investor in Anthropic, OpenAI’s main cloud supplier.MicrosoftCloudNvidia : Invests in its customers and guarantees their financing. About a quarter of its 2027 business would come from labs it funds.NvidiaChipsBroadcom : Buys Google’s TPUs for Anthropic, and finances and guarantees their lease.BroadcomChipsAMD : Pays its customers in shares: warrants on up to 160 million shares for OpenAI and for Meta, in exchange for 6 GW of orders each. Also invests in Anthropic.AMDChipsPrivate credit : Apollo, Blackstone and Morgan Stanley, through the Compute SPV vehicle.Private creditPrivate creditFormer miners : TeraWulf, Cipher, Hut 8: former bitcoin farms turned into data centres for Anthropic.Former minersData centresSoftBank · SB Energy : SoftBank has committed more than $60bn to OpenAI, partly funded by a $40bn bridge loan. Its subsidiary SB Energy is developing the 10 GW Ohio campus leased by OpenAI.SoftBank · SB EnergyData centresLambda : Neocloud in which Nvidia is a shareholder, renting Nvidia chips to Anthropic.LambdaData centresNscale : British neocloud. Dedicated infrastructure contract with Anthropic.NscaleData centres
Overview

50 deals, 19 players

Each arrow runs from the one who invests, buys compute, finances or guarantees, acquires or cuts off, to the one on the receiving end. The amounts are commitments of different kinds: we do not add them up.

2023
September – October 2023

Two bets

Amazon invests up to $4 billion in Anthropic, Google up to $2 billion. In exchange, Anthropic trains its models on their clouds: part of the money invested comes back as compute purchases.

2025
September 2025 – January 2026

Neoclouds, landlords backed by Nvidia

The labs also rent from new specialist players. OpenAI signs about $22.4 billion of contracts with CoreWeave in 2025; Microsoft rents $17.4 billion of capacity over five years from Nebius.

CoreWeave buys its chips from Nvidia, which has also committed $6.3 billion to it and is a shareholder ($2 billion more in January 2026). The first loop at a chip supplier: Nvidia ⇄ CoreWeave.

October – November 2025

The model spreads

Microsoft and Nvidia together invest up to $15 billion in Anthropic, which commits $30 billion to Azure cloud in return.

On the OpenAI side: $250 billion of purchases promised to Microsoft. AMD invents a variant: OpenAI orders 6 GW of its chips, and AMD gives it warrants on up to 160 million of its shares in return, about 10% of the company.

2026
January 2026

First snag

Anthropic cuts off xAI engineers’ access to Claude, which they were using in the Cursor code editor: its terms forbid helping a competitor. Remember Cursor, it comes back later.

January – February 2026

Google also sells its model and its chips

Apple chooses Gemini and announces that its next generation of models, Siri included, will be built on it: a multi-year partnership whose price is not public, about $1 billion a year according to the press. A month later, Meta rents Google TPUs for several billion.

Apple had also negotiated with Anthropic for Siri. According to Bloomberg, Anthropic asked for several billion a year; Apple still uses Claude internally.

February – July 2026

AMD pays its customers in shares

AMD repeats its recipe with Meta: up to 6 GW of orders, and once again warrants on up to 160 million shares given in return. In July, it invests up to $5 billion in Anthropic, which will buy up to 2 GW of its chips.

The value of these warrants depends on AMD’s share price, which itself depends on the orders they reward.

March – July 2026

Meta, from tenant to landlord

Meta rents up to $27 billion of capacity from Nebius and raises its commitments to CoreWeave to about $35 billion. The same month, Anthropic signs with CoreWeave too. In July, Meta announces that it wants to resell its spare capacity, like SpaceX. On the day of the announcement, CoreWeave and Nebius shares lose about 12%.

April – May 2026

A change of scale

Anthropic promises more than $100 billion to AWS over ten years, and $200 billion to Google. Amazon reinvests $25 billion, Google up to $40 billion.

The first two loops close: Google’s and Amazon’s money comes back to them as orders.

April – August 2026

Google’s machine

Broadcom buys $128 billion of TPUs from Google, destined for Anthropic. A private credit vehicle (Apollo, Blackstone, Morgan Stanley) buys a first $35 billion batch and leases it to Anthropic. Broadcom guarantees about $30 billion of that debt.

Buildings are needed too. Google guarantees up to $44 billion of leases on former bitcoin farms, TeraWulf, Cipher and Hut 8, which Anthropic rents. On Google’s balance sheet, that risk weighs $815 million.

May – September 2026

The rival becomes the landlord

Anthropic leases the Memphis data centres that xAI, now part of SpaceX, had built for Grok: $1.25 billion a month until 2029, up to $84.5 billion in total, terminable by either side on 90 days’ notice.

A few weeks earlier, an internal xAI memo had called the use of its GPUs “clearly low”. Despite these rents, SpaceX’s AI division lost $1.5 billion in the second quarter. And the gas turbines installed in Memphis for Grok now power Claude.

Google also rents 110,000 Nvidia GPUs there, for $920 million a month (about $30 billion until 2029), to serve Gemini Enterprise. So Google sells its own chips to its competitors, and rents Nvidia’s from a competitor for its own model.

August 2026

The strangest loop

SpaceX acquires Cursor for $60 billion. OpenAI cuts off Cursor’s access to its models. Anthropic keeps selling Claude there.

So SpaceX pays Anthropic for Claude, while Anthropic pays SpaceX for its GPUs.

August 2026

OpenAI and Nvidia

OpenAI commits to about 12 GW of Nvidia chips and leases a 10 GW campus in Ohio for 20 years, developed by SB Energy. Nvidia guarantees its financing, up to $105 billion.

Nvidia also holds about $21 billion of SpaceX shares, from its investment in xAI.

To meet more than $60 billion of commitments to OpenAI, SoftBank invests on credit: a twelve-month $40 billion bridge loan, signed in March. OpenAI then leases the Ohio campus from it: the loop closes.

August – September 2026

Other landlords

Anthropic rents up to $44.6 billion from Nscale and about $35 billion from Lambda, a neocloud in which Nvidia is a shareholder. Nscale says it does not yet have firm financing for this contract.

October 2026

The supplier becomes the banker

Broadcom lends up to $42 billion to Anthropic, convertible into shares, to lease Broadcom chips. It is discussing up to $50 billion of financing for OpenAI.

14 – 20 October 2026

Credit tightens

The most indebted links give way first. On 14 October, CoreWeave’s share price falls 15.5% and its bonds yield close to 15%, the yield of a borrower seen as risky. Its debt is approaching $50 billion; Nvidia, its supplier, is a shareholder.

On the OpenAI side, SB Energy withdraws or suspends its IPO, and SoftBank’s credit risk hits its highest in four years.

Sources: CNBC, Reuters, Bloomberg

Summary

13 closed loops

In each one, the money a player pays out comes back to it, directly or through a third party. If one link stops paying, the other’s revenue disappears at the same time.

  • Amazon ⇄ Anthropic
  • Google ⇄ Anthropic
  • OpenAI → Nvidia → SoftBank · SB Energy → OpenAI
  • OpenAI ⇄ SoftBank · SB Energy
  • OpenAI ⇄ AMD
  • AMD ⇄ Meta
  • AMD ⇄ Anthropic
  • Microsoft ⇄ Anthropic
  • Nvidia → Anthropic → CoreWeave → Nvidia
  • Nvidia ⇄ CoreWeave
  • Anthropic → SpaceX · xAI → Cursor → Anthropic
  • Anthropic → Private credit → Broadcom → Anthropic
  • Broadcom ⇄ Private credit

Every player is at once a supplier, customer, creditor or guarantor of the others. SpaceX is the extreme case: Anthropic’s landlord, Anthropic’s customer through Cursor, and a competitor in code.

OpenAI: the risk carried by Oracle and SoftBank

OpenAI’s case is less sophisticated, but more concentrated. OpenAI builds little itself: it commits, and others borrow. Oracle has signed a cloud contract with it valued at $300 billion over five years from 2027. According to KeyBanc, Oracle could need to borrow up to $100 billion over four years to serve it, when its long-term debt already stood at $82.2 billion at the end of August 2025 (DatacenterDynamics).

SoftBank is the other exposed balance sheet, and it is exposed on credit. Its commitments to OpenAI exceed $60 billion. To meet them, it signed a twelve-month $40 billion bridge loan in March, at about 6.14% interest, one of the largest ever arranged in Asia-Pacific; according to Bloomberg, some bankers were worried about its exposure to OpenAI (Japan Times). Its subsidiary SB Energy is also developing the 10 GW Ohio campus leased by OpenAI, whose financing Nvidia guarantees up to $105 billion.

The loop is complete: SoftBank borrows to invest in OpenAI, which then leases a campus from it. If OpenAI’s revenue disappoints, the value of the stake and the rents fall together, while the loan remains due.

This link is already showing signs of strain. SB Energy has withdrawn or suspended its IPO (Reuters, via Hindu BusinessLine), and on 20 October Bloomberg reported SoftBank’s credit risk at a four-year high (Bloomberg). Meanwhile, OpenAI is seeking to raise about $30 billion at a valuation of more than $1.4 trillion, with SoftBank among the expected participants, before Anthropic’s accounts become public.

On top of that come about $22.4 billion of contracts with CoreWeave, 6 GW of AMD chips and $250 billion promised to Microsoft. OpenAI builds little; Oracle and SoftBank borrow for it.

Conclusion: when the concrete runs late, the debt waits

The first article measured a physical gap: AI promises gigawatts that the grid will take years to deliver. This one measures its financial double: the market is paying today for capacity that will only exist tomorrow. The sector’s spending is about eight and a half times its revenue. At OpenAI and Anthropic, fewer than one announced gigawatt in ten is in operation. And a growing share of the money flowing between the labs, their suppliers and their financiers comes back to where it started.

These figures do not say that AI is worthless. The revenue is real and rising fast. They say that three clocks do not run at the same speed: debt falls due before the revenue arrives, and chips are depreciated before the grid powers them. They also say that those who make the promises are not those who will pay if the schedule slips. The labs rent; Oracle, SoftBank, CoreWeave and the banks borrow; and at the end of the chain, IPO investors will be the ones buying the promise.

In October, the first signs appeared where we expected them. The most indebted links gave way first, credit became more expensive, and two sites stopped in Europe, one over a permit, the other over the grid. This is not a break. It is proof that the two gaps are one: when the concrete runs late, the debt waits.

Looking ahead: the market’s price, and the European question

On 26 October, Anthropic filed its prospectus: Nasdaq, ticker ANT, led by Goldman Sachs and Morgan Stanley. The first half of the test is over. A major lab has published audited accounts, and they are more nuanced than expected: a business close to break-even, a net loss that is mostly an accounting entry, compute commitments of more than $500 billion. The second half remains, the real one: the price the market will agree to pay. Neither the share price nor the amount raised has been set yet. OpenAI, for its part, is trying to raise $30 billion before that comparison becomes possible.

One question remains that these two articles have only touched on: Europe. The two sites halted in October, Google’s in Finland and Nscale’s in Essex, show a continent that hosts AI infrastructure without owning it, and that runs into the same physical walls. If Europe wants its own models, it will have to enter the same race, with the same three clocks. Can it do so without reproducing the same loops? That will be the subject of our next article.

Appendix: the deal register

The deals in the graph, with their source. Amounts as announced or reported, of different kinds and lengths: we do not add them up. “SEC filing” marks an official document, “press” a press article, “secondary” a source citing other media.

DateFrom → ToTypeAmountWhat the deal saysSource
Oct 2026Broadcom → AnthropicFinancing or guaranteeup to $42bnLoan convertible into shares, to fund about a third of the $125.2bn five-year TPU lease.CNBC, citing Anthropic’s prospectus (opens in a new tab)press
Oct 2026Broadcom → OpenAIFinancing or guaranteeup to $50bnFinancing for custom chip purchases, under discussion.Motley Fool (opens in a new tab)press
Sep 2026Anthropic → LambdaCompute purchase≈$35bnAnthropic rents compute from Lambda, in Texas.AFP, via L’Orient-Le Jour (opens in a new tab)press
Sep 2026Nvidia → LambdaInvestmentequity stakeNvidia is a shareholder in Lambda, which installs its chips before renting them out.AFP, via L’Orient-Le Jour (opens in a new tab)press
Sep 2026Anthropic → SpaceX · xAICompute purchaseup to $84.5bnTotal commitments to SpaceX through 2029 in Anthropic’s prospectus; the gap with the $45bn is not explained.The Information (opens in a new tab)press
Aug 2026Google → Former minersFinancing or guaranteeup to $44bnGoogle guarantees the leases of 10 projects (2.4 GW); it books only $815M of liabilities.Financial Times, translated by Expansión (opens in a new tab)press
Aug 2026Anthropic → Former minersCompute purchasenot disclosedAnthropic leases the data centres of TeraWulf, Cipher and Hut 8.Financial Times, translated by Expansión (opens in a new tab)press
Aug 2026SpaceX · xAI → CursorAcquisition$60bnSpaceX completes the acquisition of Cursor on 14 August.CNBC (opens in a new tab)press
Aug 2026Cursor → AnthropicCompute purchasenot disclosedCursor keeps offering Claude; Anthropic promises to devote more compute to it.CNBC (opens in a new tab)press
Aug 2026OpenAI → CursorCut-offno amountOpenAI cuts off Cursor’s access to its models from 12 November (about 5% of Cursor’s traffic).CNBC (opens in a new tab)press
Aug 2026Nvidia → SpaceX · xAIInvestment≈$21bnStake from $10bn invested in xAI, converted into SpaceX shares at the merger.Wikipedia, AI build-out financing (opens in a new tab)secondary
Aug 2026OpenAI → NvidiaCompute purchase≈12 GWOpenAI’s existing and planned commitments to Nvidia chips through 2030.AI News, citing Nvidia’s earnings call (opens in a new tab)press
Aug 2026OpenAI → SoftBank · SB EnergyCompute purchase20-year leaseOpenAI leases a 10 GW campus in Ohio, developed by SB Energy.SiliconANGLE (opens in a new tab)press
Aug 2026Nvidia → SoftBank · SB EnergyFinancing or guaranteeup to $105bnNvidia guarantees the financing of the Ohio campus leased by OpenAI.AFP, via L’Orient-Le Jour (opens in a new tab)press
Aug 2026Anthropic → NscaleCompute purchaseup to $44.6bnDedicated GPU infrastructure at the Monarch campus, in four tranches; Nscale has no firm financing yet.Nscale prospectus (S-1) (opens in a new tab)SEC filing
Jul 2026AMD → AnthropicInvestmentup to $5bnAMD investment in Anthropic.Bloomberg Law (opens in a new tab)press
Jul 2026Anthropic → AMDCompute purchaseup to 2 GWAnthropic will buy up to 2 GW of MI450 chips from the first half of 2027.Bloomberg Law (opens in a new tab)press
Jun 2026Google → SpaceX · xAICompute purchase$920M a month (≈$30bn)Google rents about 110,000 GPUs on Colossus from October 2026 to June 2029 to meet demand for Gemini Enterprise; terminable by either party on 90 days’ notice from 2027.DatacenterDynamics, citing SpaceX’s filing (opens in a new tab)press
Jun 2026Private credit → BroadcomCompute purchase$35bnThe Compute SPV vehicle buys about 1 GW of TPUs, close to a million chips.Financial Times, translated by Expansión (opens in a new tab)press
Jun 2026Anthropic → Private creditCompute purchaselease paymentsAnthropic leases this hardware from the private credit vehicle.Financial Times, translated by Expansión (opens in a new tab)press
Jun 2026Broadcom → Private creditFinancing or guarantee≈$30bnBroadcom guarantees the two senior tranches of the debt if Anthropic stops paying.Financial Times, translated by Expansión (opens in a new tab)press
May 2026Anthropic → GoogleCompute purchase$200bnAnthropic commitment to Google’s chips and cloud.Siècle Digital, citing The Information (opens in a new tab)press
May 2026Google → AnthropicInvestmentup to $40bnNew Google investment in Anthropic.AI Weekly, Q2 2026 review (opens in a new tab)secondary
May 2026Anthropic → SpaceX · xAICompute purchase$1.25bn a monthCompute on Colossus and Colossus II until May 2029 (≈$45bn), terminable by either party on 90 days’ notice.SpaceX prospectus (S-1) (opens in a new tab)SEC filing
Apr 2026Anthropic → AmazonCompute purchaseover $100bn over 10 yearsAnthropic commitment to AWS.Anthropic’s draft prospectus, via Reuters (opens in a new tab)press
Apr 2026Amazon → AnthropicInvestment$25bn moreNew Amazon investment in Anthropic.AI Weekly, Q2 2026 review (opens in a new tab)secondary
Apr 2026Broadcom → GoogleCompute purchase$128bnBroadcom commits to buying 3.5 GW of TPUs from Google, destined for Anthropic (deliveries 2027-2028).Financial Times, translated by Expansión (opens in a new tab)press
Apr 2026Meta → CoreWeaveCompute purchase$21bn (≈$35bn in total)New commitment from 2027 to 2032, after $14.2bn signed in September 2025.The Next Web (opens in a new tab)press
Apr 2026Anthropic → CoreWeaveCompute purchasenot disclosedMulti-year agreement, capacity coming online from late 2026.Pure AI (opens in a new tab)press
Mar 2026Meta → NebiusCompute purchaseup to $27bn over 5 yearsNvidia Vera Rubin capacity from 2027, plus the purchase of any capacity Nebius has not sold to others.DatacenterDynamics (opens in a new tab)press
Mar 2026SoftBank · SB Energy → OpenAIInvestmentover $60bn of commitmentsSoftBank’s commitments to OpenAI, partly funded by a twelve-month $40bn bridge loan (≈6.14%) signed in March.Japan Times, citing Bloomberg (opens in a new tab)press
Feb 2026Meta → GoogleCompute purchaseseveral $bnMeta rents Google TPUs over several years to develop its next models.Storyboard18, citing The Information (opens in a new tab)press
Feb 2026Meta → AMDCompute purchaseup to 6 GWMulti-year agreement for up to 6 GW of AMD GPUs, first deliveries in the second half of 2026.AMD press release (Form 8-K) (opens in a new tab)SEC filing
Feb 2026AMD → MetaInvestmentup to 160M sharesWarrants unlocked in tranches as Meta’s purchases climb towards 6 GW and AMD’s share price crosses set thresholds.AMD press release (Form 8-K) (opens in a new tab)SEC filing
Jan 2026Anthropic → SpaceX · xAICut-offno amountAnthropic cuts off xAI engineers’ access to Claude in Cursor: its terms forbid helping a competitor.Sherwood News (opens in a new tab)press
Jan 2026Apple → GoogleCompute purchase≈$1bn a year, unconfirmedMulti-year partnership: Apple’s next generation of models, Siri included, will run on Gemini and Google’s cloud.AppleInsider (opens in a new tab)press
Jan 2026Apple → AnthropicCompute purchasenot disclosedApple runs versions of Claude on its own servers. For Siri, Anthropic asked for several billion a year; Apple chose Gemini.AppleInsider, citing Bloomberg (opens in a new tab)press
Jan 2026Nvidia → CoreWeaveInvestment$2bnNvidia buys CoreWeave shares at $87.20 and helps it find land and power for more than 5 GW by 2030.Computing (opens in a new tab)press
Jan 2026CoreWeave → NvidiaCompute purchasenot disclosedAll of CoreWeave’s infrastructure runs on Nvidia hardware (Rubin, Vera).Computing (opens in a new tab)press
Nov 2025Microsoft → AnthropicInvestmentup to $5bnMicrosoft investment, announced alongside Nvidia’s ($15bn between them).Semafor (opens in a new tab)press
Nov 2025Nvidia → AnthropicInvestmentup to $10bnNvidia investment in Anthropic.AFP, via L’Orient-Le Jour (opens in a new tab)press
Nov 2025Anthropic → MicrosoftCompute purchase$30bnAnthropic commitment to Azure cloud, in return.Semafor (opens in a new tab)press
Oct 2025OpenAI → MicrosoftCompute purchase$250bnOpenAI commitment to buy Azure cloud services.Wikipedia, AI build-out financing (opens in a new tab)secondary
Oct 2025AMD → OpenAIInvestmentup to 160M shares (≈10%)Warrants at $0.01 a share, unlocked in tranches tied to deployment and share price milestones.Tom’s Hardware (opens in a new tab)press
Oct 2025OpenAI → AMDCompute purchase6 GW, tens of $bnOpenAI commits to deploying 6 GW of AMD chips, starting with 1 GW of MI450 in the second half of 2026.Tom’s Hardware (opens in a new tab)press
Sep 2025OpenAI → CoreWeaveCompute purchase≈$22.4bnThree contracts in 2025: up to $11.9bn in March, $4bn in May and $6.5bn in September.Verdict (opens in a new tab)press
Sep 2025Nvidia → CoreWeaveCompute purchase$6.3bnNvidia commitment to CoreWeave announced in 2025.Computing (opens in a new tab)press
Sep 2025Microsoft → NebiusCompute purchase$17.4bn over 5 yearsDedicated GPU capacity in Vineland (New Jersey), up to $19.4bn with options.DatacenterDynamics (opens in a new tab)press
Oct 2023Google → AnthropicInvestmentup to $2bnGoogle investment. Anthropic uses its chips and its cloud.Wikipedia, AI build-out financing (opens in a new tab)secondary
Sep 2023Amazon → AnthropicInvestmentup to $4bnAmazon investment. Anthropic commits to training its models on AWS.Wikipedia, AI build-out financing (opens in a new tab)secondary

Main sources