The Investor

How a bank markdown travels to your software holdings

The channel is back-leverage: loans that banks make to private-credit funds, secured by the loans those funds hold. When a bank writes down the software loans inside that collateral, it can lend less against them. The fund then has less room to refinance or extend credit to the same borrowers. In our view, bank marks lead fund marks, which makes the software values in your private credit statements the lagging number.

Semafor reports the Fed is examining collateral and risk controls at the four banks. A supervisor asking about collateral quality tends to come before tighter lending terms, not looser ones. That is our inference; the reporting describes no change in terms so far.

Why AI disruption becomes a credit problem

Much sponsor-backed software is priced per seat. Its revenue tracks customers' headcount, while its debt service does not move. a16z data cited by The Bear Cave shows the pressure: call center jobs grew about 4% a year for 15 years and are now shrinking about 4% a year. Every lost agent seat is lost revenue for seat-priced customer-service software, and the interest bill stays the same.

The substitute also gets cheaper each year a loan is outstanding. Alejandro Saucedo's ML roundup puts the inference cost of reaching a fixed benchmark score as falling roughly 13x a year. A five-year loan sized on today's seat count is exposed to automation that will cost a small fraction of today's price well before the loan matures.

What the bank tape and the take-private fights add

Markets appear to reward early recognition. Among ten large US banks in correction, JPMorgan, whose markdowns helped prompt the review, is down least at 11%. Capital One, the most consumer-credit-heavy name, is down most at 25%. Those figures come from a single widely shared social post with no stated measurement window, so read the ranking as directional.

Borrowers are also testing lender protections. EA is relying on an investment-grade rating from Egan-Jones, which The Bear Cave calls 'ever-controversial', plus Treasuries set aside for future payments, to avoid repaying bondholders at a premium after its take-private. Creditors allege a default. For anyone holding take-private debt, change-of-control protection is now being litigated rather than assumed.

What to discount, and where the opening is

The Bear Cave is published by Hunterbrook Media, whose affiliate Hunterbrook Capital is short several names the publication covers, and some of its market figures come from social posts. The Fed review itself rests on Semafor's reporting, not on a short thesis, which is why we treat it as the anchor. The weaker figures affect how big the problem is, not whether it exists.

For patient capital, the same pressure creates demand. Software that AI is disrupting but that still produces cash will need structured equity, preferred stock or holding-company financing as conventional refinancing gets harder. Whoever arrives with dry powder sets those terms.

When banks start marking down AI-disrupted software loans, the private credit fund's statement is the last place the loss shows up.

Why the backlog behaves like a credit exposure

Lambda's pre-money valuation comes to about 0.29x its backlog, a multiple that looks cheap only on the assumption that the backlog turns into cash. Most of it is owed by one customer. Lambda's lenders and shareholders are underwriting Anthropic's ability to keep paying for the life of the contracts.

Anthropic's capacity to pay rests on revenue, which is usage times price, and Anthropic has been cutting its own prices hard. The commitment holds while usage outgrows the cuts. Lambda added $1B of debt days before the equity raise. New shareholders rank behind those lenders. The take-or-pay obligations (payment owed whether or not the capacity is used) were not disclosed, and neither were contract duration or termination rights.

The same dependency at market scale

Figures cited by The Bear Cave put the one-day bond wave at about a quarter of 2026 data center issuance, implying roughly $480B for the year. Lazard CEO Peter Orszag states the dependency plainly: if AI token spending drops, the debt-financed buildout has a problem. The bond figures trace to social posts rather than filings, so treat them as indicative.

Firmus is the small-cap version of the same trade, with $51M of revenue and a $68M loss carried against $30B of debt and 46 MW installed. We think the $30B is mostly committed facilities for capacity not yet built. The reports also disagree on the valuation Firmus sought. The Bear Cave describes a $44B IPO, while The Information reports a ~$5B raise at more than $30B. Both versions were raised against megawatts that do not exist yet.

Where public comps enter the chain

The Information's sum-of-the-parts values AWS by reference to CoreWeave and Nebius, and Lambda's planned 2027 IPO would price against the same two peers. If Amazon's Q3 shows AI demand lagging $220B of 2026 capex, the neocloud benchmark moves first and every private GPU-cloud mark pegged to it follows. Our back-of-envelope on The Information's growth figures says Amazon's blended forecast needs the AWS-and-other block compounding in the low-to-mid 20s.

The reports agree on the dependency and differ in emphasis. Lambda coverage stresses the IPO upside if Anthropic keeps compounding. Bond-wave coverage stresses a demand curve that has never been tested through a slowdown. A third outcome, in which token spending grows but more slowly than the debt schedules assume, gets less attention than either. Our view, which we are prepared to be wrong about, is that Amazon's Q3 tests the thesis before Lambda files, and the first place it shows up is the CoreWeave and Nebius multiples.

A GPU cloud's backlog is only as strong as its largest customer's ability to grow usage faster than it cuts prices.

The workload under the valuation is getting cheaper

TypeSafe's product, Jev, returns typed decisions with confidence estimates for the small judgments inside business processes: classify a request, pick a tool, escalate an exception. Those high-volume calls are exactly the workload Anthropic's cheapest model now serves at $0.10 per million input tokens. Microsoft Research Asia's Agent Lightning lifted a 9B open model 14.6 points on SWE-bench Verified with about 6,000 training samples. Alejandro Saucedo's roundup adds a production case: Pinterest replaced a multi-stage clustering pipeline with a single fine-tuned Qwen3 4B model.

In our view, $7.5B holds only if calibration is the real product, meaning confidence scores reliable enough for an enterprise to automate on. That would make TypeSafe a verification company priced like a model company. Its Fortune 500 penetration figure is a company claim with no revenue attached.

Nous offers the clearest test date

Nous's multiple falls to about 15x if it clears its $100M ARR target by year-end, which would make it the cheapest forward multiple among the week's rounds. The usage claims behind the forecast, 24M+ Hermes clones and about 2.5% of global AI token usage, are self-reported. Year-end ARR is the checkpoint that shows whether open-source agent distribution turns into revenue.

Nvidia is setting marks across the stack

Nvidia plans to invest in d-Matrix, which sought a $5B valuation in July after a $2B Series C. Nvidia is also opening NVLink Fusion and MGX, its chip-interconnect and rack designs, to rival inference chips, and it joined the Nous and Mecka AI rounds. A strategic investor pays for ecosystem position, so its price need not reflect the return a financial investor needs. We read the interconnect move as a hedge: Nvidia accepts that inference silicon will be mixed and keeps control of the connection and the rack. Inference-chip startups now sort into those inside its ecosystem, which gain distribution, and those outside, which face pressure. d-Matrix's first NVLink Fusion systems are not due until Q4 2027, a long runway for slippage.

Where these marks get tested

Open weights arrive first. Reflection's Beam (501B total parameters, 23B active, Apache 2.0) and Mistral Large 4 (1T total, 49B active) are due to ship weights by the end of October, the first chance to measure their costs independently. Beam's claimed 3-4x inference-compute advantage over GLM-5.2 is an estimate, not a measured deployment cost. Public tests follow in 2027 with Lambda's planned IPO and DeepSeek's possible STAR Market listing.

A mark set by a strategic investor tells you what the ecosystem is worth to that investor, not what the company is worth to you.