Wall Street is lending against the future value of AI chips before there is a price for them. Private credit already finances GPUs and disagrees about what they will be worth in five to ten years. The first exchange benchmarks for the price of compute were announced in July. BlackRock plans to turn the loans into securities for pension funds, and says the data and ratings that market needs will take time. The banks funding the build-out say they assess “the value of the collateral”, and none has said on a call what that value is.
There are those who have made big bets on the future value of compute five and six and 10 years from now. I have no doubt that some of those bets are going to pay off really handsomely, and some of those bets are going to be disastrous.
Asked whether compute could become a standalone investable asset class
: I do believe the components of compute, because of the shortage of compute, will increase in value.
Asked whether chip financing is next: We already do in our lending business participate in chip finance.
And: It has to be done structurally right.
We're also partnering with Silicon Data to launch a pioneering Compute Futures market later this year.
His colleague Derek Sammann: We are going to be the first daily benchmark tracking the spot hourly rental cost of NVIDIA H100 GPUs.
And on what it is for: this allows customers to effectively hedge a data center.
GPU Compute Futures, developed with Ornn and NativX, that bring price discovery and hedging to the fast-growing AI-driven compute market.
The rating agencies are working on the buildings around the chip. Moody's, October 2025: our corporate ABS team rated a more than $1 billion data center securitization, also the first transaction of its kind.
S&P Global, asked in July about commercializing compute benchmarks
, pointed to its datacenterHawk acquisition and its very specific proprietary data center intelligence, including pricing, supply chain, site selection, and details like that
.
a 1 GW data center in the U.S. is about a $50 billion enterprise cost. 40% of that is the cost of GPUs and chips. The other 60% is power, grid, core and shell.
What he wants for the chips: an asset-based financing related to the chips through data, through analytics.
The playbook: In the early 1980s, mortgages were loans sitting on banks' balance sheets, illiquid and hard to price.
What changed it: What made them tradable was data, years of performance history, and an understanding in how the performance would move with interest rates and with the changes of the broader economy.
Then: Compute capacity is now following a similar path, and we intend at BlackRock to be the center of it
.
On timing: It will take time to get the rating agencies to opine on this new market and come up with metrics.
His CFO, Martin Small, on the buyers: it's also going to require opening up pension plans and traditional fixed income allocations to these exposures.
The mortgage market took years of performance history before it could trade. For compute, the first exchange benchmarks were announced three months ago, and the loans are already on the books.
The count measures talk, not lending. The Q3 2026 bar covers the 14 bank calls transcribed by 16 October, which skew toward the largest banks; the same 14 banks had no such statement a year earlier. Most bank statements describe limits, and on the calls the limits are never sized. Fifth Third's is the only dollar figure any bank has given on a call for its data center lending. Tim Spence, CEO, April 2026: we have less than $100 million of funded exposure to data centers.
On what lenders have to assess: where you sit in the capital structure, the value of the collateral.
On the theme: Because the AI theme is so broadly diffused into the marketplace. It's not just traditional construction lending.
It's kind of everywhere.
And: Currently, our exposures are still quite modest in the context of our overall wholesale exposure, but that's not the same as saying that we would be immune to a major change in sentiment in the AI space.
On the financing related to chips and other things that go inside the data centers
: we have very hard limits in terms of what we are willing to do and what we are not willing to do and are sticking to that.
Asked whether financing will be the choke point for AI capex: I don't know, and nobody knows.
And: There is going to be a lot of financing to do, and we're going to be in the middle of that.
Asked whether the firm is allocating more or less to the AI opportunity: I would say that we were overweight, and in light of it being something that matters not just to our clients, but to us, I would say we are now double overweight.
On bridge loans that turn into equity: I imagine there are going to be cases where we are going to make mistakes.
we're not large in the data center finance space
, and later: If the CapEx broadly slows down related to AI, do we think even though we don't lend directly to that it would slower loan growth? Yes, I do.
We have well-defined buckets as to how much we want to hold. We're also hedging some of those exposures as well.
And: does the pin go in the bubble someday?
His colleague Ted Swimmer added that the lessees are vast majority of them are investment grade.
Across 936 calls, “data center” appears in 411 paragraphs and “GPU” in 12. When a bank mentions the chip, it is Wells Fargo saying it underwrites chips differently, a passing remark, or the bank's own servers. PNC on its own: remembering that we put GPU chips in our data centers.
The vocabulary for pricing and offloading a depreciating asset appears on bank calls only about auto leases, capital relief on loan books and capital rules.
| Phrase | Paragraphs | Of which banks | What the bank paragraphs are about |
|---|---|---|---|
| data center | 411 | 84 | Read in full for the bank series below |
| GPU | 12 | 3 | BNY's own Nvidia hardware, PNC's own depreciation, a Dimon list of infrastructure |
| chip, AI sense | 24 | 6 | Wells Fargo on underwriting chips and chip financing (3), Morgan Stanley (2) and Goldman Sachs (1) in passing |
| H100 | 1 | 0 | None. The one mention is CME's rental benchmark |
| compute futures | 9 | 0 | None. CME, ICE, and an analyst on the S&P Global call |
| residual value | 11 | 9 | Auto leases at Ally, a leveraged lease at Regions |
| risk transfer | 191 | 25 | Capital relief on auto and commercial loan books |
| SRT | 13 | 3 | Capital relief, none on AI |
| backstop | 8 | 5 | Leverage and capital rules, none on AI |
Universe: 60 US-listed financial institutions (21 banks, 15 asset managers, 16 insurers, 8 exchanges and data firms). The corpus holds 936 earnings calls from 13 January 2023 to 16 October 2026, including 23 calls from two firms since replaced in the universe. Q3 2026 is in progress, with 16 calls transcribed so far, 14 of them banks.
Phrase counts are paragraphs containing the phrase in vendor transcripts; the "chip, AI sense" row is hand-checked to drop blue chips, the CHIPS Act and car-part shortages. The bank series is verified against each paragraph: a management statement counts when it describes the firm's own business of lending to, financing, investing in, insuring, underwriting securities for or advising on data centers, compute, chips or power built for AI. Q2 2026 includes two calls the classifier missed and a hand check found; Q3 2026 was read by hand. Every quote on this page is an exact substring of its transcript paragraph with the speaker checked. Quotes have not been checked against the call audio or the firms' own transcripts. Everything here is about what was said on earnings calls; filings and press reports are outside the count.
Companion pages: The Token Tape (what the same firms say about the cost of AI) and the census of realized AI returns. Research and writing: Milos Maricic. The passages behind every number here are available on request.
MAXI