Posts

A CoreWeave co-founder told the Financial Times this week where the next round of AI-chip financing could come from. Brannin McBee, who leads CoreWeave's financing, said that winning investment-grade ratings for more of its chip-backed debt would let it borrow from insurers, whose capital "is unlimited". The FT's insurance correspondent had already supplied the other half, saying many large traditional insurers are at their limit for exposure to AI companies. Both can be true at once. Insurance money is already in this market, in investment-grade paper and below it, and the filings name the insurers: 58 purchasers of IREN's notes, more than half of them US life and annuity insurers, and insurers among the lenders on CoreWeave's GPU facilities. Where they are named, the amounts are redacted or not given. What is new is the scale being discussed and the ratings it would take. For scale, the NAIC estimates US insurers' exposure to all private credit at $1.2 trillion at the end of 2025, about 13% of cash and invested assets. It does not break out data centers or AI, and no official figure does.

Tessera's piece of 2 September held back on supplier-side residual value guarantees, saying that whether any given arrangement qualified as circular "depends on documents we have not read". The filings have now been read, and they had been on file all along. AMD disclosed a $4.1 billion guarantee of a partner's data center lease obligations in February, Nvidia disclosed guarantees of partners' facility lease obligations "in exchange for warrants" the same month, Broadcom's backstop appeared on 9 June, and Nvidia's largest guarantee in an 8-K on 17 August. Every one was public before the 2 September piece said otherwise. Checking those sources against Tessera's August work turned up errors in three pieces, corrected here and in a note on each page. They cover the date at which anyone first had to price AI hardware, four telecom figures from the Nortel and Lucent era that were wrong or unsupported, and the shape of AMD's investment in Anthropic. One argument is weaker as a result. Staging was offered as what separates today's commitments from the telecom era, and Lucent's were largely undrawn too. Revenue separates them more.

On 28 September Reuters reported that two customers accounted for nearly a quarter of Anthropic's revenue in 2025. The reporting that followed points at something larger. Anthropic's filing is still confidential, and a search of SEC EDGAR on 5 October found no public registration statement, but Reuters, which has seen it, reports the share of revenue sold through Amazon and Google rising from 11% in 2023 to 32% in 2024 and 47% in 2025, about $2.16 billion, through their cloud marketplaces. The reading here is that the larger concentration is the channel rather than the customer list. Nearly half of last year's revenue arrived through two marketplaces, and the companies that own them are, by Reuters' account of the filing, also Anthropic's investors, its compute suppliers, its competitors and its customers. Reuters still does not name the two customers who each generated 12% of revenue in 2025, and this piece does not guess. If either of them is one of those same two companies, the two concentrations fall on the same names and add up to more than either figure shows on its own.

On 21 April, SpaceX ended seven years of financial silence. For the first time the company disclosed actual numbers publicly. The Analyst Day sessions at Starbase, together with the simultaneous release of the IPO prospectus, gave investors a first look at the financial shape of the business. The headline is Starlink. The combined launch and Starlink businesses generated $15 billion to $16 billion in revenue in 2025, with profits of approximately $8 billion. Starlink alone ended 2025 with roughly 9.2 million subscribers and over $10 billion in revenue. That is a business with disclosed unit economics rather than a projection, and it is the first time those figures have been public. Analysts project Starlink revenue reaching $24 billion by the end of 2026, which is a projection rather than a disclosure. Starlink passed 10 million active subscribers globally in early 2026, adding customers faster than any telecom network has. SpaceX acquired xAI in February 2026 in an all-stock transaction valuing the combined entity at $1.25 trillion, with xAI at approximately $250 billion and SpaceX at $1 trillion at the time. The xAI numbers remain the opaque part. Grok's monetization, compute costs and the capital committed to AI infrastructure have not been fully broken out, and that is what the full S-1 still needs to provide. The governance structure is now confirmed. Musk holds approximately 42% of SpaceX's equity and controls roughly 79% of the votes through super-voting shares. Public investors will hold standard voting shares. This is a founder-controlled company at the ballot box regardless of how many shares trade on Nasdaq.

AI infrastructure partnerships are often described as giving companies the flexibility to "walk away." But what does that actually cost? Using Meta's Louisiana data center venture as a case study, this article examines one of the few fully disclosed residual value guarantee (RVG) structures in the sector. Meta describes the project as providing "strategic optionality and flexibility." The filings also disclose up to $46.03 billion of maximum exposure to loss, including lease commitments, future funding obligations, and a $28 billion residual value guarantee. The takeaway is that optionality isn't free. The right to walk away has a price, even if that price may never ultimately be paid. As AI infrastructure spending accelerates, understanding what sits behind terms like flexibility and optionality becomes just as important as understanding the assets themselves.

The final piece in this 3-part series asks whether the revenue supporting the AI build is actually measuring what investors think it is. The telecom boom offers a useful warning. One of its most influential statistics claimed internet traffic was doubling every 100 days. The number spread through analyst reports, earnings calls and prospectuses, helping justify enormous infrastructure investment. But researchers later traced the figure back to something different: network capacity was growing at that pace, not internet traffic. Actual traffic was doubling roughly once a year. That distinction matters. Getting a growth forecast wrong is normal. Building a capital cycle around a metric that measures the wrong thing is a different problem entirely. As hundreds of billions flow into AI infrastructure, the question is worth asking again: are the revenue and demand signals financing this build measuring genuine end demand, or are parts of the industry counting activity generated by the build itself?

The AI build is often compared to the telecom boom of 1998–2002. But the biggest lesson from that era may not be about demand forecasts. It is about what happens when long-lived assets are financed with money that comes due much sooner. Telecom companies spent heavily on fiber that ultimately proved enormously valuable. The problem was timing. Global Crossing entered bankruptcy with $22.44 billion in book assets and $12.39 billion in debt. Its fiber could last 25 years, but its financing couldn't wait that long. The eventual value of those networks didn't save the original equity holders or creditors. Much of that value went to whoever bought the assets cheaply after the balance sheets broke. The second piece in this 3-part series looks at today's AI infrastructure build through that lens: if expectations disappoint, what is all that capital expenditure actually worth, who owns the collateral, and does the financing last long enough for the assets to prove their value?

The AI build is usually discussed as a technology story. But underneath it sits a capital-allocation question that may matter just as much: who is actually structured to hold the risks being created? This first piece in a 3-part series starts at the bottom of the stack with basic scientific research. It is long-duration, failure-prone, and difficult for the company funding it to fully monetize. Historically, that has made it an awkward fit for private balance sheets. Yet frontier AI companies are spending heavily on exactly this kind of work. The article examines why basic research behaves like the first-loss equity tranche of technological development, why neither governments nor private companies are obvious natural holders of that risk, and what the current AI investment cycle may be telling us about a decades-old economic assumption.

SpaceX reported $18.37 billion in quarterly capex. Two days later, it announced the first phase of a semiconductor campus that could eventually involve far more spending. The difference between those numbers is more important than the market reaction. Committed spending gets measured. Announced spending keeps its optionality. That matters as AI infrastructure spending accelerates toward extraordinary levels and a growing share is financed with debt. It matters even more as frontier AI companies prepare to enter public markets. The deeper question is whether quarterly reporting is built to price research programs whose costs arrive now but whose payoff may be decades away. SpaceX offers a useful case study in what happens when long-duration technological bets meet the short-duration accountability of public markets.

Since its June 12 IPO, SpaceX has completed a full round trip, pricing at $135, surging above $225 within days, and now trading below its offer price. The debate is no longer about what SpaceX was worth as a private company. The market is trying to determine what kind of public company it is becoming. One increasingly useful way to think about SpaceX is as an emerging hyperscaler. It may not follow the path of Amazon, Microsoft, or Google, but its combination of launch infrastructure, Starlink, and AI compute is creating a business model that looks increasingly familiar, while pointing toward a very different destination.