# The AI build-out has reached the credit committee

Global AI-related debt issuance fell to $23bn in September from a $113bn June peak, even as the industry has raised an estimated $466bn this year. The constraint is shifting from access to chips towards whether lenders can see and price the web of leases, guarantees, customers and physical dependencies behind them.

**AyEye Today · 2026-10-11**

LEAD SUMMARY — ANALYSIS. The AI build-out has spent much of 2026 behaving as if the principal question were where to find the next chip, transformer or power connection. A new number adds an awkward extra appointment to the calendar: the credit committee.

Global AI-related debt issuance fell to $23bn in September, less than half August’s total and far below the $113bn raised at June’s peak, according to Morgan Stanley data reported by the Financial Times on 10 October. Yet the industry has still raised an estimated $466bn of AI-linked debt this year, compared with $101bn in the same period of 2025. This is not a shortage of ambition. It is the first clear pause to ask who, exactly, is carrying it.

## 1. The new constraint is not capital alone. It is confidence in the chain.

REPORTED — INDEPENDENT JOURNALISM. The [Financial Times reports](https://www.ft.com/content/9c13d40e-d2b2-45d5-921c-a68cd4b308f9) that global AI-related debt issuance declined steadily after June and that US investment-grade AI bond issuance stopped in September, after blue-chip technology groups had borrowed about $306bn between January and August. Morgan Stanley attributed much of the slowdown to the exceptional volume already raised, while investors are also examining leverage, returns and opposition to data-centre projects more closely.

The verb matters. Issuance has slowed; it has not vanished from the pipeline. Banks are still assembling financing measured in tens of billions for chips and compute. The [Wall Street Journal reported](https://www.wsj.com/tech/ai/oracle-broadcom-and-spacex-seek-blockbuster-debt-deals-to-pay-for-ai-chips-848e8032) this week that Broadcom was seeking more than $50bn for custom OpenAI chips, while Oracle and SpaceX were also exploring large hardware financings.

ANALYSIS. The meaningful change is that capacity plans can no longer be read as engineering schedules with money attached. They are credit propositions. A lender must believe that future demand will fill the machines, that the customer will keep paying, that the guarantor can absorb a failure and that the hardware will still have value when the contract ends.

## 2. The balance sheet no longer shows the whole machine

CONFIRMED — INVESTOR ANALYSIS. [UBS Asset Management describes](https://www.ubs.com/ie/en/assetmanagement/insights/investment-outlook/the-red-thread/trt-end-year-2026/articles/webs-of-ai-debt.html) a roughly $35bn private-credit structure in which a special-purpose vehicle bought about a million Google custom chips and leased them to Anthropic. Anthropic’s rental payments service the debt, Broadcom supports the senior notes and Google backstops site leases. UBS says the arrangement gave Anthropic about a gigawatt of compute without putting the debt itself on Anthropic’s balance sheet.

That is financially ingenious and operationally consequential. The same network can contain supplier, customer, lessor, guarantor and creditor roles at once. Reported corporate debt may therefore be the easiest part to see rather than the complete exposure. UBS estimates that aggregate hyperscaler obligations rise from about $800bn to roughly $3.5tn when leases, service commitments, production commitments and backstops are included, before deducting cash.

IMPORTANT LIMIT. These figures combine different kinds of obligations and do not mean $3.5tn is immediately due, equally risky or hidden in a single place. Nor does an off-balance-sheet structure automatically imply misconduct. The point is narrower: the ability to build compute increasingly depends on contractual support distributed across multiple entities, and the operating company’s reported debt cannot describe that dependency by itself.

## Concept to learn today: the compute credit stack

THE COMPUTE CREDIT STACK is the chain of promises that turns a proposed AI system into usable capacity. It sits beside the technical stack, but it can now decide whether that stack exists on time.

### Demand promise

Who commits to use the capacity, for how long and at what price? A forecast is not the same as a contracted payment.

### Asset owner

Who buys the chips, land, cooling and power equipment, and who carries them if the intended user no longer wants them?

### Credit wrapper

Which stronger company, insurer or reserve makes a weaker customer acceptable to lenders? What, precisely, is guaranteed?

### Physical dependencies

Which grid connection, water supply, permit or local agreement must arrive before the asset can earn revenue?

### Reset and recovery

What happens when delivery slips, a model changes hardware strategy, compute prices fall or the customer’s economics weaken?

ORIGINAL SYNTHESIS. The compute credit stack is the financial control plane beneath the AI control plane. It allocates not only capital but time, scale and bargaining power. When it tightens, a model laboratory may still possess the technical recipe for a larger system while losing the schedule, price or autonomy needed to build it.

## 3. Commercial success is already written into technical capacity

CONFIRMED — PRIMARY SOURCE. In an April filing, [Broadcom told the US Securities and Exchange Commission](https://www.sec.gov/Archives/edgar/data/1730168/000119312526144028/d87999d8k.htm) that Anthropic would begin accessing about 3.5 gigawatts of next-generation TPU-based compute in 2027 through an expanded collaboration with Broadcom and Google. The filing states that Anthropic’s consumption of that capacity depends on its continued commercial success.

That sentence is unusually useful. It joins a frontier lab’s revenue trajectory directly to gigawatts of future capability. A benchmark curve may suggest what a laboratory could train next; a credit document may reveal whether the physical capacity will be available to try.

INFERENCE. If lenders start demanding higher yields, shorter commitments, more collateral or stronger guarantees, the effects will not remain inside finance departments. Laboratories may favour models that reach users sooner, architectures that use available chips, products with clearer revenue or deployments that secure an anchor customer. They may defer redundant capacity and expensive assurance work unless those costs are protected deliberately.

This does not mean lenders will choose model architectures directly. It means the cost and conditions of capital can shape the feasible set from which technical leaders choose.

## 4. Local delay has become credit information

REPORTED — INDEPENDENT JOURNALISM. The connection between physical delivery and financial confidence is already visible. [Axios reported](https://www.axios.com/2026/09/25/oracle-debt-data-centers) that Oracle reserved the right to delay lease payments on the New Mexico Project Jupiter data centre if the project did not open as planned. The project had helped developers obtain about $18bn in loans, while permitting and local opposition contributed to delay. Oracle said the project remained on schedule.

ANALYSIS. A planning objection, transmission delay or construction problem is no longer merely a site issue. It can change when lease payments begin, alter the value of the debt, widen spreads on related securities and affect how readily the next project is financed. The local operating compact and the capital-market contract have become part of the same system.

This is why a data-centre announcement should be read along four axes at once: technical capacity, physical deliverability, community permission and financeability. A project can look enormous on the first axis and fragile on the other three.

## 5. Assurance capacity needs its own protection

ORIGINAL SYNTHESIS. A tighter credit environment introduces an easily missed safety question. When a project is under pressure to generate revenue, which parts of the build-out remain non-negotiable?

Monitoring clusters, evaluation environments, incident-response staff, redundant controls and independent testing all consume money and compute without producing the most visible product demo. If those functions compete with training and inference inside a stressed financing plan, they risk becoming variable costs precisely when a more highly leveraged system can least afford operational surprises.

The remedy is not to make bondholders the AI safety board. It is to define assurance capacity as a protected requirement in the operating and financing case: reserved compute, funded control functions, tested incident obligations and clear stop authority that survive a schedule squeeze. A system whose safety case depends on surplus capacity has no safety case once the surplus disappears.

## Noise: this is not evidence that the AI build-out is over

NOISE CHECK. September’s decline follows extraordinary issuance earlier in the year, and Morgan Stanley says the volume already raised was the principal reason for the slowdown. Large new financings remain in preparation. The data does not prove a credit freeze, a collapse in AI demand or an imminent systemic crisis.

Nor should one month be mistaken for a durable trend. Global debt markets are responding to interest rates, geopolitics and government borrowing as well as AI-specific risk. A later surge in issuance could reverse September’s pattern.

The narrower conclusion still matters: the sector has borrowed at historic scale, some of the exposure travels through complex lease and guarantee networks, and investors are now pausing long enough to price those relationships. The AI race has acquired a cost of capital and a memory.

## Mental-model update

Earlier: AI capacity was constrained mainly by chips, power, permits and local consent.

Now add: every gigawatt also rests on a chain of commercial promises. The relevant unit is not only compute delivered, but compute whose customer, asset owner, guarantor and physical dependencies can all survive the life of the financing.

In the illustration, fast red and blue pressure gathers against a broad exposed interval while slower dark strokes hold the opposite side. A small yellow passage continues across, but does not settle which force will move next. That is the financing question now hanging over the build-out: not whether capital exists, but under what conditions it will cross.

## Questions to carry forward

  - Which future compute commitments depend on an unrated or loss-making customer, and what supports the debt if that customer misses its plan?

  - How much AI infrastructure exposure sits in leases, guarantees and special-purpose vehicles rather than reported corporate debt?

  - Which grid, permit and community milestones can pause payments or reprice the financing?

  - Will tighter capital reward efficient architectures and useful products, or simply concentrate frontier development inside the strongest balance sheets?

  - Are evaluation, monitoring and incident-response resources protected when a project’s economics tighten?

Canonical: https://www.lecxie.com/publications/ayeye-today/2026-10-11.html