What Michael Milken Would See In AI Chip Financing

Authored by Patrick Feeley via Substack,

Very few people in the AI financing debate are reading the deals like a credit trader would. We have had a week of headlines about who is raising what. Amazon is moving $8 billion of chips off its balance sheet. Broadcom is lending $42 billion to its own customer. Nvidia is offering to backstop $125 billion of the debt people use to buy Nvidia. The coverage treats each number as a scoreboard. Hardly anyone asks the question that matters once the paper is sold, which is who is holding the risk and whether they know it.

There is one person who would have asked. He went to federal prison, and he was also right about a version of this problem forty years before almost anyone else.

Berkeley, 1965.

In 1965 Michael Milken was an undergraduate at Berkeley when he came across an NBER monograph by W. Braddock Hickman, Corporate Bond Quality and Investor Experience, published in 1958. Hickman had gone through decades of corporate bond records, from 1900 to 1943, and reached a conclusion that should have been scandalous. Investors usually overestimated the risk in higher-yielding bonds. They treated the rating as the risk, and the rating was wrong often enough that the paper everyone feared tended to do better than its reputation.

Milken later wrote that he “was struck by the disparity between theory and reality.” That is a modest way to describe what he saw, which is that a whole profession had confused a letter grade with a probability. Most people who find something like that write a thesis. Milken built an industry. He went to Wharton to study capital structure, joined the firm that became Drexel Burnham Lambert in 1969, and moved the high-yield desk to Beverly Hills in 1978.

The machine.

The numbers still look fake to me. Public junk bonds outstanding went from $8.5 billion in mid-1977 to $59.1 billion by mid-1985 and roughly $200 billion by early 1990. Between 1978 and 1985 Drexel underwrote 57 percent of every new-issue dollar in the market. One desk, and largely one man. The money built MCI, Ted Turner’s empire, Steve Wynn’s casinos and McCaw Cellular, and it funded T. Boone Pickens’s raids on the oil majors, which readers of my last piece will recognize as an early chapter of the constructivist trade. By 1989 the Predators’ Ball at the Beverly Hilton drew more than 3,000 people. Pickens was there.

Most people, including some who worked there, get the ending wrong. Drexel did not die of bad bonds. Drexel Burnham Lambert filed for bankruptcy on February 13, 1990, and the Los Angeles Times described the cause as “a staggering liquidity problem.” The firm defaulted on roughly $100 million of loans, a small number against the market it had created, and the market lost its main dealer overnight. Most of the issuers were fine and most of the paper was fine. What disappeared was the one institution willing to take the other side.

Two months later Milken pleaded guilty to six felony counts and paid $600 million. He was sentenced to ten years and served 22 months. Weeks after his release he was diagnosed with prostate cancer, and he spent the next three decades building the Prostate Cancer Foundation and FasterCures before his pardon in February 2020. As a former biotech person I would guess his second career did more good than his first did harm. That is a different essay.

The detail that belongs in this essay comes from 1987. That year, according to the GAO, insurers held more than 30 percent of all junk bonds outstanding. In April 1991 California regulators seized Executive Life, which had 62.7 percent of its general account in junk. The institution that failed was the marginal buyer of the paper, with long liabilities, a capital charge set by the rating rather than the asset, and a board that believed it owned bonds.

The skeptic.

The popular memory of Milken is a man who sold leverage to anyone who would take it. His own writing says close to the opposite. “Debt isn’t good. Debt isn’t bad. For some companies, close to zero debt is too much leverage.” In the Wall Street Journal in 2009 he laid out six factors and warned that when they point toward rising business risk, “even a dollar of debt may be too much for some companies.” In Bloomberg Businessweek in 2014 he named the industry. Stable, predictable revenue can carry debt. Technology, he wrote, “should finance growth primarily with equity.”

Read that again and then look at what is being sold this week. I suspect the man who built the high-yield market would hesitate to lend against a GPU.

He also kept a short list of principles about credit. The first one is four words long. Rating is not credit. Most of what follows is an application of those four words.

The tape.

On October 2 the Financial Times reported that Amazon is pitching investors on a special purpose vehicle (a separate company set up to hold specific assets and borrow against them) that would own roughly $8 billion of already-installed Nvidia Grace Blackwell chips, raise debt against them, and lease the chips back to Amazon. The stated goal is a lighter balance sheet. The other effect is a rating. Investors expect the paper to come out investment grade on the strength of Amazon’s double-A, which is what opens it to insurers and pension funds. Amazon says each series of chips will last at least five years. In the same article the FT noted that Grace Blackwell will soon be superseded by Vera Rubin. Both facts ran in one story, and I have not seen anyone comment on it.

A day earlier Reuters pulled a number out of Anthropic’s IPO filing. Broadcom will lend Anthropic up to $42 billion in convertible notes, roughly a third of a $125.2 billion, five-year commitment to lease Google TPUs that Broadcom helps design. Anthropic’s own filing flags the potential conflict of interest and warns that certain defaults could make “a substantial portion of its lease obligations immediately due.” Bloomberg reported that banks are separately assembling $60 billion for Broadcom’s chip financing, $42 billion senior and $18 billion junior, with Blackstone leading the junior piece. In August Nvidia set up a platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to bring in more than $500 billion of outside money, and Jensen Huang said Nvidia holds an option to backstop $125 billion of it.

CoreWeave remains the live experiment. In August it closed a $2.6 billion loan with a roughly five-year maturity against customer contracts that average about three years. Meta’s Hyperion data center was financed with $27.3 billion of bonds through Blue Owl’s Beignet vehicle, priced at 6.581 percent, rated A+, one notch below Meta, and kept off Meta’s balance sheet. Pimco reportedly took about $18 billion of it, which is one firm holding two thirds of one deal.

All of this is being priced with the 10-year Treasury above 5.2 percent. It touched 5.34 percent on October 1, the highest since 2002. JPMorgan expects $4.1 trillion of AI-related debt to be issued through 2030. For scale, the entire junk bond market that Milken built peaked at around $200 billion.

How a trader reads it.

First, the rating is borrowed. A vehicle that owns depreciating silicon is being rated off Amazon, not off the silicon. That works for as long as Amazon pays the lease, and the lease is most of the trade. Strip the structure away and a buyer of that paper is long Amazon credit at some spread over Amazon’s own bonds, plus a residual bet on what a 2026 GPU fetches in 2030. The first leg is easy to price. The second is hard to hedge, because I am not aware of any real forward market in used Blackwells. If the spread does not pay for the second leg, the buyer is being paid for Amazon risk and taking the chip risk for nothing. Milken would have asked about it in four words.

Second, the debt can outlive the asset. Five-year money on chips with three-year contracts and a successor generation already announced is a bet on renewal. The CoreWeave lenders said as much and described themselves as comfortable underwriting renewal risk. That may be a reasonable bet, but it is closer to a venture bet than a term loan, and I would expect it to be priced like one. Amazon trimmed the useful life of some servers from six years to five in 2025. Michael Burry spent his October 1 Substack arguing the real economic life is shorter still. I do not know the right number, and I doubt anyone does. What I do know is that the coupon is fixed and the collateral value is not, and if the collateral falls short, the lender holds the difference.

Third, the marginal buyer looks familiar. In 1987 it was the insurers. In 2026 it is insurers and pension money again, reached through investment-grade structures that look like corporate bonds and behave partly like equipment leases. Cliff Asness put it better than I can in March. “The lion doesn’t care if the ostrich is first loss or higher up in the capital structure.” Tranching sorts out who takes the first loss, and the size of any loss stays the same. To be fair to Asness, he has been careful to say he did not predict the timing of trouble in private markets, and I am not predicting it either. I am describing a structure.

There is a real counterpoint, and Milken would make it himself. Anyone with a contract from Anthropic or Amazon has absorbed every rate increase so far. One executive told CNBC last month, “If you have a deal with Anthropic, will 50 basis points really stop you?” Probably not. My guess is that trouble, if it comes, starts two tiers down, with the smaller neoclouds lenders have started turning away and the second-string chip lessors who priced their books when the 10-year was in the 4s. Credit cycles have a habit of starting where few people are looking.

The club.

There is one more pattern Milken would recognize, and it is the one I think about most as a public-market investor. In 1989 the Los Angeles Times described the Drexel network as a club whose members were lending to and borrowing from each other. Deals cleared because the same handful of balance sheets sat on both sides of the trade.

In 2026, Broadcom lends to its largest customer so that customer can lease Broadcom’s chips. Nvidia offers to backstop the debt its buyers use to buy Nvidia. Several of the sponsors in the Nvidia platform also appear in the Broadcom financing. Each deal is defensible on its own. Taken together, the suppliers have become a meaningful source of their own demand, and the public is about to be invited into the arrangement through an Anthropic IPO that people close to it have put at up to $2 trillion.

To be precise, I think AI demand is real, perhaps the most real thing in the economy right now. My concern is narrower. When the junk market seized in 1990, the companies were mostly sound. The dealer that made the market walked away, and nobody else would take the other side. For AI credit, the chips will almost certainly get used. What I would want to know is who makes a market in GPU-backed paper in the quarter one of the large sponsors decides to sit out. Today, as far as I can tell, it is largely the sponsors themselves.

People.

The part of Milken I like most is the optimist. “People are the scarce resource,” he said. “It’s not buildings, it’s not printing presses, and it’s not factories.” I did not expect to find him on chips, but he got there too. “For today’s computer chips, it’s less than two percent of the cost. Brainpower has become the ‘raw material’ for building companies.”

Hold that up against this week. Borrowed money is going into the factory at a 5 percent risk-free rate, and very little is going to the people who have to make the factory pay. I would guess the return on AI accrues mostly to the operating companies that take the compute and turn it into margin, and many of them are small and mid caps that have not done it yet. They are sitting on the adoption gap.

Anthropic seems to see it the same way. It just committed $100 million to train 10,000 engineers to deploy its models inside customers. Read that as a signal. A model company is paying out of pocket to close the gap between buying compute and using it, which suggests the chips alone do not close it. The lenders are financing the supply side of a market whose bottleneck, I think, sits on the demand side, with people.

Six questions Milken would ask.

1. Does the debt pay down before the chips stop earning?
2. Should this business be financed with debt at all, or, as he put it, primarily with equity?
3. Whose credit is actually being rated, the chip’s or the guarantor’s?
4. Who owns the residual if a 2026 GPU is worth a fraction of book in 2029?
5. Is the seller financing its own customer, and who makes the market if the dealer steps back?
6. Who turns the compute into margin?

The sixth is the one I spend my days on, and it is the only one where the answer is a company rather than a structure. Milken liked to say the past is always triple-A and the future is always single-B. He meant it as encouragement, and I take it that way. The future is worth financing, as long as the people doing the financing know what they own. This week a good number of them will own Amazon credit with some used-chip risk attached, and I hope they are being paid for both.

This note is for research and discussion only. It is not an offer to sell, or a solicitation to buy, any security. Sargasso Capital Management may hold positions discussed or related instruments and may change those positions without notice.