The Board Room

What 25% versus 82% is actually negotiating

Structured chip finance has exactly one executed precedent, and it is worth reading closely before anyone treats the next deal as routine. In the $35 billion June securitization, Apollo and Blackstone raised the debt, bought Broadcom-designed Google TPUs, and leased them to Anthropic. Broadcom guaranteed $29 billion of residual value to get that first deal done, roughly 82% of the total. Nvidia's consortium is offering up to 25%, has not specified which assets that support covers, per Dakin Campbell's reporting, and its six financial firms have signed memorandums of understanding while executing nothing.

The spread between those two numbers is the entire negotiation. Nvidia is asking lenders to carry about four times more residual risk than this market has ever accepted, on an asset whose depreciation schedule Nvidia itself sets through its product cadence. Jensen Huang's counter is fungibility: CUDA, networking and a large installed base make GPUs easy to redeploy, and CoreWeave has contracted 2020-vintage A100s out to 2029. Credit the reframe, because it is a clever one. The proprietary stack regulators call lock-in gets repositioned as collateral liquidity. The support percentage on the first funded transaction is the cleanest read you will get on what sophisticated capital thinks a GPU is worth in year five.

Structure, not credit rating, decides who gets chips

Anthropic is private and unrated, and it still secured chips at a scale its own balance sheet could not carry. That inverts an assumption still sitting inside most capacity models, namely that GPU financing requires a contract or backstop from an investment-grade hyperscaler. Rating is no longer the gate; structure is. The planning consequence is that competitors' compute constraints loosen, and differentiation shifts toward data, distribution and efficiency per token rather than raw capacity. The same template also demonstrated that custom ASICs are financeable, which retires the capital-access objection that used to end dual-silicon conversations internally.

Where the reporting disagrees

Read together, these say the cost of serving intelligence is falling while the cost of owning the metal is rising. That is the textbook argument for renting short and repriceable rather than buying long. It is also the opposite of what a five-year lease structure asks anyone to sign.

The failure mode is a refinancing event

Debt in these structures matures in five years or less and is amortized from compute cash flows, which changes what a demand slowdown looks like from the inside. CreditSights' Andy Li describes an industry building "on the back of debt financing." If demand plateaus, a capacity provider does not have a soft quarter. It has a refinancing event, and contracted capacity travels with it. The reported $30 billion junior tranche in Broadcom's second deal carries no vendor backstop, which makes junior-tranche pricing the earliest visible indicator of AI capex sentiment on offer. The provider's lender is part of the risk surface now, whether or not anyone in the org can name it.

A public futures curve is also a public mechanism for shorting GPU rental prices — transparency the vendor cannot manage.

The clause nobody wrote into the last two years of AI contracts

PJM's proposal would classify some new large-load customers, AI data centers named explicitly, as interruptible when electricity supply tightens. The commercial consequence lands one layer down from the grid. An AI service commitment is only as strong as the provider's power contract, and almost no enterprise AI agreement signed in the last two years contains a curtailment clause, per TLDR IT's reading of the market. That is an unpriced option already written to customers.

The tradeoff worth naming is that interruptibility stops being a procurement footnote and becomes an architecture requirement. Checkpointing, tiering of deferrable training against non-deferrable inference, and failover across independent grid regions rather than cloud availability zones are engineering work. Two to four quarters of it, and none of it can be bought in a renewal cycle.

The water asymmetry is already legally operative

Morning Brew reports the federal 10-year Colorado River plan requires Arizona to cut 27%, Nevada 16% and California 10%, while Colorado, Utah, New Mexico and Wyoming cut nothing, with reservoirs at record lows, limits revisited every two years, and litigation openly expected. A reasonable skeptic says the asymmetry cannot hold politically. The skeptic is probably right about the eventual deal and wrong about the timing, because the asymmetry governs siting decisions now. Phoenix versus an exempt upstream site is worth real money in permitting speed, community consent and biennial reallocation risk. Agriculture consumes the majority of river water, nearly half of it for animal-feed crops, which is where the political fight goes. Industrial users are the convenient secondary target once the easy cuts are spent.

Consent moved further than cost did

Techpresso reports opposition to a nearby data center rising from 42% to 75% in twelve months across surveys using identical wording, with 64% strongly opposed against 15% in favor, and net negative with every political constituency: 43 points among Republicans, 65 among independents, 75 among Democrats. A separate Redfin poll in July put opposition at 53%, so the level is contested and the direction is settled. MIT Technology Review's read is that permission, not capability, is the binding constraint. That makes it an election-cycle issue rather than a permitting-desk one.

Where the sources pull apart

Nvidia, the most information-advantaged actor in this market, took a minority stake in site developer Cloverleaf Infrastructure and partnered on gigawatt-scale US "AI factory" projects. When the chip company starts buying access to megawatts downstream rather than supply upstream, it has reclassified the scarce asset. Against that, Sam Altman argues society may absorb AI more slowly than model progress suggests. Both cannot be equally right. The shape that survives either outcome is contracted optionality: power-secured capacity callable without owning the substation.

The ownership gap is the real finding

Power sits with a grid operator. Water sits with a federal plan revisited every two years. Consent sits with local politics ahead of a midterm. None of the three appears in a quarterly objective anywhere in most technology companies, which is why the interaction between them goes unmanaged until a delivery date slips. The facts are public. The gap is accountability.

A firm that cannot name who is able to switch off its inference does not own its roadmap.

The cheapest enforcement action of the year

The Los Angeles Police Department's Inspector General never sued Flock Safety. It read the signed contract out loud instead, quoting language that lets the vendor "retain the right to use the foregoing for any purpose in Flock's sole discretion," plus a separate grant covering anonymized footage for "training of machine learning algorithms," per Term Sheet's reporting. LAPD was renegotiating within weeks, and it is not the only city doing so. Two privacy attorneys independently found the same clause pattern in different jurisdictions, which turns a customer-specific dispute into a template problem across an installed base spanning 5,000-plus communities in 49 states.

The clock speed is the lesson, not the clause. A legislature takes years. A procurement desk with an auditor takes weeks, and the discoverable artifact moves from the published privacy policy to the master services agreement nobody re-reads after signature.

The market marked it before the lawyers finished

Flock went from $3.5B in February 2022 to $7.5B in March 2025, a 114% step-up. The latest round moved roughly 11% to $8.3B while recurring revenue doubled past $500M, landing near 16.6x ARR. That is not a growth discount. That is late-stage capital pricing legitimacy risk into a business with no product-market-fit problem. a16z's earlier "effectively the only game in town" thesis got tested in the same cycle by cities that cancelled signed competitors, several of them less transparent about data collection. Switching costs were lower than the monopoly narrative implied.

A reasonable skeptic would file the Ring episode as one consumer news cycle. The sequence argues otherwise. A viral consumer ad preceded Amazon pulling Ring's Flock integration, officially because it "would require more resources than anticipated," and municipal cancellations in Boston, Flagstaff and Santa Clara followed. Partner concentration is reputational concentration, and the exit arrives in a neutral sentence the smaller party does not get to edit.

Two more readers of the same drawer

Techpresso reports a class action against Twitch and Amazon alleging creators were treated as "free training stock," resting on two facts the company supplied itself: an AI-training opt-out that ships on by default, and product chief Mike Minton saying on a company stream, "if it was opt-in, nobody would opt in." That is a discoverable statement of intent, and a lesson in executive communications discipline as much as in product defaults.

The third reader is statutory. Illinois SB 315 layers onto California's transparency framework less than eight months after that framework took effect, the first state mandate for independent third-party frontier model audits, including auditor access to "all unredacted versions" of published materials, per a16z's policy tracking. Red-team memos and draft catastrophic-risk assessments become externally visible. The behavioral risk exceeds the legal one: teams that know their notes are producible write less useful notes.

The move, and the trap beside it

Flock's own remedy shows the trap. It now advocates a seven-day retention default, down from 30 days, on the claim that this still covers "over 90% of searches," while the LAPD contract guarantees five years. Publishing the technical argument against one's own contracts, without making the guardrail contractual, manufactures the privacy-washing accusation for the other side. A policy that is not a default is evidence.

The tradeoff was named at signature and gets priced later. The number that belongs beside net revenue retention is a retention metric too, of rights rather than revenue: what share of the training corpus survives a hostile contract review at 25%, 50% and 75% revocation. Most leadership teams cannot answer that today.

The question is what percentage of the training corpus survives a hostile contract review. If nobody owns the answer, a customer's auditor will produce it first.