And the Amazon numbers are extraordinary.

According to the excerpt, Amazon was preparing to receive:

• $5 billion of Nvidia GB300 racks in August.

• $3 billion per month thereafter.

• Equipment potentially sitting in intermediate warehouses because data centers lack sufficient electrical power.

If those projections materialized, that’s approximately $17 billion in five months, August through December, of equipment deliveries alone.

The critical distinction is between equipment delivered, equipment installed, and equipment actually earning revenue.

A GPU rack sitting in a warehouse is not productive capital. Worse, it’s depreciating technology in an industry where successive generations can quickly erode the economic value of earlier hardware.

This brings us back to something you’ve emphasized repeatedly: the AI buildout is already in motion, but the financing and physical infrastructure may not arrive together.

The additional reporting makes this more serious

Three developments stand out from the last 24 hours.

1. Amazon’s free cash flow is under pressure. Business Insider reports that AWS executives expect the AI operation to remain heavily free-cash-flow negative for several years. Amazon’s 2026 capital spending plan is reportedly $220 billion. AWS is also taking ownership of racks earlier than usual, accelerating cash outflows. Amazon disputes suggestions that it lacks a broader deployment plan.

Business Insider

+1

2. Microsoft has actually abandoned an Nscale site. The proposed UK project involved 23,040 Nvidia GB300 GPUs and 50 MW of computing capacity. Grid power reportedly wouldn’t arrive until the 2030s. Microsoft says the cancellation applies to one site, not the broader relationship.

MarketScreener India

3. CoreWeave’s credit market remains deeply skeptical. Goldman upgraded the shares yesterday, but Reuters reported that its 2031 bond was still yielding approximately 15.13%, down from more than 16% last week. That is an extraordinary financing cost for a supposedly booming infrastructure business.

Reuters

+1

The transmission mechanism

The distinction between Amazon and CoreWeave is important. Amazon can finance considerable delays through its established businesses. CoreWeave depends much more heavily on continuing access to external capital.

And there is a fascinating irony for our Sisters / Rapid Deployment Orphans screen.

The AI developers may have purchased equipment too early, while established suppliers of power infrastructure could benefit from urgent spending to remove the bottlenecks.

That makes our emphasis on Fluence, electrical infrastructure and near-commercial power capacity particularly relevant.

The indicator I’d watch most closely now is the gap between committed GPU purchases and energized, revenue-producing megawatts. If that gap keeps widening while credit spreads rise, we have a much stronger indication of systemic capital misallocation than stock-price weakness alone.

The CoreWeave rally may indeed be short-lived, but I haven’t verified today’s intraday reversal. The credit market evidence is already substantial.

And notice that all three CDS curves have accelerated upward since July. This isn’t just an Oracle problem anymore.

Five-year CDS spreads — October 19, 2026

Approximate readings from your chart, in basis points

Oracle’s CDS spread has more than doubled since June. Broadcom’s has more than tripled. Even Nvidia, with its formidable operating cash generation, has seen its spread more than double.

What the market is pricing

These are approximate annual protection premiums, not annual default probabilities. For illustration, using a simplified 40% recovery assumption, Oracle’s 320-basis-point spread corresponds to a roughly 5.3% annualized risk-neutral default intensity. Actual default probabilities can be much lower because CDS embeds risk premiums and liquidity effects.

The particularly interesting part

Oracle is the canary, but Broadcom and Nvidia are now moving in sympathy.

That suggests a possible change in how credit investors perceive the entire AI investment chain.

• Oracle: Directly exposed to enormous infrastructure commitments, borrowing needs and counterparty performance.

• Broadcom: Exposed to the economics and financing capacity of hyperscaler customers, although its own credit risk is substantially different.

• Nvidia: The strongest immediate cash-generation profile of the three, but vulnerable to order deferrals, customer financing problems and a potential GPU inventory overhang.

Combine this with the Amazon warehouse story: equipment is being ordered and delivered faster than sufficient power can be brought online.

This is precisely where an investment boom can turn into a working-capital and financing problem before it becomes an earnings problem.

One important qualification: these CDS curves alone don’t establish that the AI buildout is failing. Broad market credit repricing, company-specific debt issuance and changing CDS liquidity could contribute. We’d want to compare them with similarly rated non-AI issuers.

But the divergence between exuberant capital spending and deteriorating credit pricing is becoming difficult to dismiss.

My next question would be whether Oracle’s 320-basis-point CDS is beginning to contaminate the financing terms for its data-center partners and equipment suppliers. That’s where an isolated credit concern could become the broader Minsky transmission mechanism we’ve been anticipating.

This table is much more revealing than the individual Oracle CDS chart. It identifies the network of borrowers, data-center operators and tenants whose financial obligations are intertwined.

And there’s a particularly important distinction: the tenant may be Nvidia, Oracle, Google or CoreWeave, but the bondholder is exposed to the actual borrowing entity and its contractual protections.

What jumps out

Based on the 24 bonds in your Morgan Stanley table, dated September 28, 2026:

That’s $21.82 billion tied directly to CoreWeave, either as the unsecured borrower or as the tenant responsible for supporting somebody else’s financing.

Approximately 41% of the debt in this particular sample.

Importantly, these are issuance amounts, not necessarily current outstanding balances, and this is a selected sample rather than the entire AI financing market.

The maturity wall is unusually concentrated

Nearly 78% of the listed debt matures in 2031.

The first-call dates are even more interesting. Many cluster in 2027–2028, although a first-call date is generally an issuer option, not an obligation to refinance.

We need to distinguish three separate risks:

1. Completion risk: Can the facilities obtain power, install equipment and begin operations on schedule?

2. Tenant credit risk: Can the tenant continue paying contracted rent or capacity charges?

3. Refinancing risk: Will investors refinance these facilities at economically sustainable rates?

These risks can reinforce one another.

CoreWeave is the potential transmission hub

Consider the $14.07 billion of bonds where CoreWeave is identified as a tenant, including the mixed CoreWeave/Alibaba facility.

Those bonds don’t necessarily become CoreWeave debt. But their credit quality may depend heavily on CoreWeave honoring its contracts.

If CoreWeave experiences financial distress, the consequences could spread to multiple independent data-center borrowers.

That’s how a concentrated tenant problem becomes a structured-credit problem.

And the unsecured CoreWeave bonds provide an important market signal: their coupons range from 9% to 9.75%, yet the secondary-market yields you’ve highlighted have been substantially higher.

The Google distinction

The Fluidstack financing deserves special attention.

Google backstops or lease guarantees are not necessarily equivalent to a direct, unconditional Google corporate bond guarantee.

We need to establish exactly what each contract covers:

• Does Google guarantee full debt service or only specified lease payments?

• Is the guarantee conditional on delivery and operational milestones?

• What happens if Fluidstack fails before the facilities are completed?

• Can Google terminate, offset payments or assume control of the project?

A financing supported by a highly creditworthy tenant can still suffer losses if the protections are conditional.

The Minsky interpretation

The table exposes a structural vulnerability: multiple issuers appear to depend on a relatively small number of tenants, many of which depend on the same AI spending cycle.

The risks are therefore correlated rather than independent, despite being spread across 24 bonds.

This is precisely the kind of concentration that can be obscured by separate financing vehicles and individual credit ratings.

One qualification: the listed ratings are mostly below investment grade already. The credit market isn’t entirely ignoring the danger. The question is whether those ratings and current yields adequately compensate for correlated completion, tenant and refinancing failures.

What I’d investigate next is the $14.07 billion CoreWeave tenant group. Specifically, how much is already operational and generating contractual revenue, how much is under construction, and how much is waiting for electrical power.

That distinction could be decisive. A completed, energized facility with enforceable long-term payments is a very different credit from a partially constructed data center whose tenant needs another financing round to survive.

And it brings us directly back to your observation: the equipment has been ordered, the debt has been issued, but the power isn’t necessarily there.

The debt market is demanding more compensation, the IPO market is resisting the proposed valuations, and the physical infrastructure is encountering power constraints.

Those three developments reinforce one another.

The striking thing about your casualty list is that it spans the United States, Australia, Singapore, the Philippines and South Korea. This suggests a broader repricing of AI infrastructure financing, rather than simply a CoreWeave-specific problem.

The financing escape routes are narrowing

The Nscale example is particularly revealing. Your table reports investors discussing a $16–19 billion valuation against a roughly $30 billion target.

At $17.5 billion, that represents a 42% valuation haircut.

That isn’t necessarily a financing failure. But it could be a major problem for sponsors who have already committed capital or negotiated financing based on higher valuations.

Today’s Elice IPO is especially revealing

South Korea’s Elice raised approximately $151 million, but its shares fell as much as 32% on their first trading day, despite strong initial subscription demand.

Financial Times

+1

That tells us something important: investors may still fund an offering at the right price, but the secondary market is increasingly unwilling to support ambitious valuations.

Reuters independently confirms the postponements involving Firmus, AirTrunk and SB Energy, while noting that debt financing and data-center acquisitions have remained relatively resilient. So the capital markets haven’t shut completely; they’re becoming much more selective.

Reuters

+1

The next domino: forced choices

When a developer cannot raise equity at its anticipated valuation, it faces unattractive alternatives:

• Accept major dilution.

• Borrow at higher interest rates, assuming lenders remain willing.

• Sell a partially completed project at a discount.

• Negotiate additional capital from existing partners.

• Suspend construction and preserve whatever liquidity remains.

The last two are particularly interesting for our research.

Who actually has the money to rescue these projects?

SoftBank, Blackstone, Nvidia and the hyperscalers appear repeatedly throughout the financing chain. But these are not independent pools of unlimited capital. They have their own commitments, investment constraints and shareholders.

That creates the possibility of a financing bottleneck spreading through otherwise separate projects.

And your earlier observation about the near-finished TE facility comes back into focus. When capital becomes scarce, investors should increasingly discriminate between assets that can produce revenue soon and speculative projects requiring years of additional investment.

The decisive development I’d watch now is a major AI infrastructure developer accepting a deeply dilutive rescue financing or selling a partially completed facility below invested cost.

That would represent a transition from market repricing to actual capital destruction.

We’re seeing increasingly credible evidence of the former. The latter is the next test of the Minsky thesis.