What Is the Collateral Actually Worth?

Correction, 15 October 2026: This piece said that the question of what AI hardware is worth "was largely unasked until this month", that "the first deal to ask it directly is sixteen days old", meaning a CoreWeave facility closed on 10 August, and that "until about two weeks ago" the mismatch between hardware life and debt had been handled with more discipline than the telecom comparison implied. Broadcom's quarterly report of 9 June 2026 had already disclosed a backstop over AI racks, entered into on 8 June, with a maximum exposure of $29 billion. Broadcom's is a supplier's exposure on a default rather than a lender's at the end of a contract, and the point about tenor stands, but the timing statements do not. The details are in In Exchange for the Guarantees, dated 12 October 2026.

Interest disclosure: T-OpenAI is a tokenized loan participation right providing economic exposure linked to OpenAI's pre-IPO valuation, issued through a dedicated issuing subsidiary, as are T-SpaceX and T-Kalshi. Following the SpaceX listing, the investment underlying T-SpaceX is being divested. Redemption opens only once a Redemption Start Date is announced, and as of 26 August 2026 none has been. Tessera's founder, and entities he controls, hold related exposure. Read what follows with that in mind.

Second of three on the capital structure of the AI build. Monday's piece asked who holds the science risk. This one asks what the capital expenditure is worth if things go wrong.

Most comparisons between the AI build and the telecom build of 1998-2002 run on one axis: were the demand forecasts wrong? That axis matters, and Friday's piece spends most of its length on a specific way it went wrong last time. But it is not where the money was lost. The demand forecasts were wrong in 2000 and the fiber still got used. What destroyed the equity was the mismatch between how long the assets lasted and how long the money was committed for.

There is a mismatch again. It is narrower than the popular version, it runs in the opposite direction, and until about two weeks ago it was being handled with more discipline than the comparison implies.

What Actually Happened to the Carriers

More than $1 trillion of equity and debt was raised from 1998 onward, on Morgan Stanley data presented to the House Financial Services Committee in March 2002. In the US alone, roughly 40 million miles of fiber went into the ground at a cost near $90 billion over five years. By 2001, Merrill Lynch estimated that under 9% of installed fiber worldwide was lit.

Global Crossing filed on 28 January 2002 carrying $12.39 billion of debts against $22.44 billion of book assets, at the time the fourth-largest bankruptcy in US history. The gap between those two numbers is the whole subject of this piece. On paper the company was solvent. What it could not do was convert twenty-five-year assets into cash on a schedule set by five- to seven-year high-yield paper.

The fiber was never the problem. It is a twenty-five-year asset that improves in place: dense wavelength division multiplexing meant each strand carried more traffic every year without anyone digging a new trench. Level 3 warned of a possible covenant breach in early 2002 and survived to be acquired.

One correction to a story that gets told loosely. The loose version says recovery values were high, pointing to Level 3's survival and to Google buying dark fiber cheaply mid-decade. That conflates two different things. The creditors of the failed carriers recovered a fraction of what they were owed. What was high was the asset's eventual economic value to whoever bought it out of the wreckage, which is a statement about Google's purchase price rather than about anyone's recovery. Cheap fiber was a good outcome for the buyer and a terrible one for the lender.

Durable assets, fragile balance sheets, and creditors who still lost.

What the Collateral Is Actually Securing

Now the AI build, where the popular version of this argument holds that short-lived GPUs are being financed with long-dated debt.

Until very recently that was not true, and the reason it was not true is more interesting than the claim. The clearest window is CoreWeave's run of delayed-draw facilities, disclosed one by one in the company's filings.

The first, in 2023, was secured on H100s and priced around 15% floating. On 31 March 2026 the fourth closed: roughly $7.5 billion of initial capacity rising to $8.5 billion, priced at SOFR plus 225, maturing in 2032, and rated A3 by Moody's and A (low) by DBRS, described in the filing as the first investment-grade-rated financing secured by high-performance computing infrastructure and an associated customer contract. On 18 May came another: $3.1 billion, the first publicly syndicated facility of its kind, rated Ba2 and BB+, oversubscribed and tightening 50 basis points during syndication.

It is tempting to read that as the market repricing GPUs from junk to investment grade in three years. It is also wrong, and the filings say why.

The ratings are not tracking the hardware. They are tracking the customer. The A3 facility had a long-dated contract from an investment-grade hyperscaler standing behind it; the Ba2 facilities do not. Rating agencies are underwriting the offtaker, with the equipment serving as a recovery floor rather than as the primary credit. Which makes "GPU-backed lending" close to a misnomer. It is offtaker-backed lending with GPUs attached, and the distance between A3 and Ba2 measures counterparty quality rather than a changed view of silicon.

That also explains the tenor discipline. If the credit rests on a contract, the debt gets written to the contract, and the hardware's economic life never has to be priced at all. Nobody was ignoring the depreciation question. They were routing around it.

The Deal That Stopped Routing Around It

On 10 August, CoreWeave closed a $2.6 billion facility with roughly a five-year maturity against underlying customer contracts averaging about three years. The filing is explicit that this is the point: lenders financing commitments shorter than the debt, and accepting the risk that capacity is renewed or re-leased when those contracts end.

That is, as far as can be seen, the first structure in this market where somebody is underwriting residual value on the equipment itself. The two years of tenor beyond the contracts is a bet on what a fleet of accelerators is worth in 2029, which is precisely the number nobody has had to defend.

CoreWeave presents the deal as an expansion of what is financeable, and on its own terms that is fair. It also reportedly met resistance in syndication, against a background of broader skepticism about the build. Both are true, and neither should stand alone.

So the honest position is not that the market has decided GPUs are good collateral, nor that it has decided they are not. It is that the question was largely unasked until this month, because the offtaker structure made asking it unnecessary, and the first deal to ask it directly is sixteen days old.

The Depreciation Argument, and Why It Is Weaker Than It Looks

The other half of the popular case is that accelerators lose 30-35% of their value a year. The counter-case is strong: hyperscalers depreciate this equipment over five to six years and several have extended their useful-life estimates rather than shortening them; A100s deployed years ago are still earning; older accelerators cascade down to inference and batch work rather than going to scrap. The line that there is no DWDM for a GPU is not right either, since kernel and software efficiency gains have delivered real throughput improvements on unchanged silicon.

What makes this live is not the estimate. It is that no party to these transactions has been required to hold a view.

Sponsors defer the question through depreciation schedules that will not be tested for years. Lenders have sidestepped it by underwriting the customer instead. And the scale of what remains unrecognized is substantial: the four largest hyperscalers bought roughly $434 billion of property and equipment in the four quarters to March 2026 against something near $149 billion of reported depreciation, because depreciation runs on five- and six-year server schedules and twenty-five- to forty-year building schedules. Today's income statements carry a fraction of today's build.

The BIS made a related observation in January that belongs on the first page of any credit memo on this sector: private credit is pricing AI infrastructure loans at essentially the same spreads, maturities, and collateral rates as its non-AI book, while equity markets price the same underlying assets for dramatically higher expected returns. Both cannot be right.

The Mismatch That Has Not Gone Anywhere

None of the above touches the lease layer, which is where the real duration problem most plausibly sits.

As of mid-2026, more than $120 billion of data center spending had moved off corporate balance sheets in roughly eighteen months through bankruptcy-remote special purpose vehicles. Meta's Hyperion campus in Louisiana is the template: a $30 billion structure, around $27 billion of it debt, held in an SPV majority-owned by Blue Owl with Meta a minority holder, and Meta's payments recorded as lease expense rather than debt. Oracle has used the same architecture repeatedly: roughly $13 billion from Blue Owl and JPMorgan into the Abilene facility, a $38 billion package across two further sites, an $18 billion loan in New Mexico.

What the sponsor commits to is a stream of contracted payments running fifteen to twenty years. What generates the revenue to justify them is hardware that will turn over three or four times inside the same term. Every re-equipping decision inside that window is a fresh capital call the original lease says nothing about, and the lease obligation continues whether or not the sponsor chooses to make it.

The same shape appears in the neocloud financings. Galaxy Digital raised a debut junk bond this summer to fund a Texas campus against fifteen-year lease commitments from CoreWeave, while the customer contracts underpinning CoreWeave's own facilities average around three years. Fifteen-year obligations, three-year revenue visibility.

Short assets, long commitments. The mirror image of 2002, which is why the loss, if it comes, should be expected to land somewhere different this time.

Where It Would Land

Lending into all this has scaled from nearly nothing to more than $200 billion of private credit exposure to AI-related borrowers, with Morgan Stanley projecting a further $800 billion over two years. The widely quoted $1.5 trillion financing gap, roughly $2.9 trillion of global data center investment through 2028 against Big Tech operating cash flow covering perhaps $1.4 trillion, is worth stating conditionally. It assumes the capex plans are executed. If demand disappoints, capex gets cut and the gap never appears.

The reading here is that stress concentrates in the credit and SPV layer rather than at the sponsors, but the cushion is thinner than the headline balance sheets suggest. Pivotal Research projects Alphabet's 2026 free cash flow falling almost 90%, to around $8 billion from $73 billion in 2025; Barclays models a similar decline at Meta from $43.6 billion. Morgan Stanley has Amazon at roughly negative $17 billion of free cash flow this year, Bank of America nearer negative $28 billion. They are also borrowing at scale on their own balance sheets, not only through vehicles: Meta priced $30 billion in October 2025, Alphabet about $25 billion that November and roughly $31 billion more in February 2026, Amazon $24.9 billion this summer after $15 billion in November. Alphabet's long-term debt quadrupled during 2025.

So the accurate statement is narrower than "they can absorb it." These are solvent businesses with excellent capital-markets access and no debt-service cliff of the kind that killed Global Crossing. They are also spending their cushion, in public, on the thing in question. That is a derating risk rather than a solvency risk, and it is not the fortress the comparison usually implies.

Two qualifications that should not be skipped. The Federal Reserve Bank of Chicago put bank exposure to AI-adjacent industries at around 0.8% of total assets earlier this year, which says where the risk is not: the regulated banking system is barely in this, and the exposure that matters sits in private credit, which nobody is required to report. And the SPV structures are often described as giving sponsors the option to walk away. Whether that is true depends on residual value guarantees and credit support that are, in most of these deals, not public. Where the sponsor stands behind the paper there is no walk-away, and the ring-fence is presentational.

What This Does and Does Not Say

It does not predict insolvencies at the top. It says the loss, if it comes, lands on SPV equity, on operators without investment-grade sponsors standing behind them, and on the small but growing pool of paper that has started to take residual risk on the hardware directly.

One number is worth carrying out of this. Allianz Research puts the current divergence between AI capital expenditure and revenue growth at roughly 46%, against about 32% during the 2001 telecom cycle. It should not be treated as predictive, since the denominators are not strictly comparable and the 2001 episode had a debt structure this one does not. But for anyone watching a single quantity, it is closer to the right one than any depreciation estimate.

That was the shape of 2002 as well. The technology thesis was correct throughout and the creditors were still impaired, because the question a lender answers is not whether the asset is good. It is whether the asset is good by the date the payment is due.

Which leaves the question underneath all of it. If the credit in this sector rests on the offtaker rather than the hardware, then the durability of the counterparty's own funding is the first-order risk. That is Friday.

Sources are as cited in the text: Morgan Stanley data via the House Financial Services Committee, March 2002; Merrill Lynch fiber utilization estimate, 2001; CoreWeave facility terms per company filings; hyperscaler property and equipment and depreciation figures per company reporting for the four quarters to March 2026; off-balance-sheet and SPV figures per contemporaneous reporting; BIS, January 2026; Federal Reserve Bank of Chicago, 2026; free cash flow projections per Pivotal Research, Barclays, Morgan Stanley and Bank of America; Allianz Research on capex-to-revenue divergence. Figures are as of the dates and reports stated.

This is market commentary, not investment advice. It is not a recommendation to acquire, hold or redeem any Tessera product, or to take or avoid exposure to any company mentioned.

T-Tokens are tokenized loan participation rights, not equity. High risk. DYOR. Not financial advice. Not available in the US or other restricted territories. tessera.pe/terms