The Board Room

The savings are gated on a capability, not a purchase

Every dollar of routing savings depends on proving that a cheaper model did not degrade a business outcome, which means the money is downstream of measurement rather than procurement. The router is a purchase. The evaluation harness is not. Routing software of the LiteLLM class is thin, replicable and easy to in-source, so the durable position belongs to whoever owns the eval policy behind the router. An organization that cannot segment its workloads by task complexity has no baseline to negotiate against, and no way to bank the arbitrage even if a vendor concedes it.

Two details in the AT&T account travel further than the headline. Inference is partially repatriating from cloud to owned silicon: open models now run on AT&T's own Nvidia and AMD hardware, because owning beats renting at that volume. That puts hosted-frontier-only products at a disqualifying disadvantage in a growing share of large regulated deals. The second detail is that DeepSeek and Moonshot models are held out of production at governance review, not capability review. Frontier pricing on the low-complexity tier is protected by a compliance moat rather than a technology one, and compliance moats move with policy in both directions.

Where the sources agree, and where the numbers are soft

The price points corroborate the direction of travel rather than the magnitude. Gemini 3.7 Flash posts 84.6% on ARC-AGI-2 at $0.25 per task, GLM-5.3 Max reaches 1597 points at $3.65 per million tokens, and Gemma has passed a billion downloads. The Information sets this against capital moving the opposite way: government bond yields at 20-year highs, with disclosed capital demand from the AI sector above $600 billion. Infrastructure capital is getting scarcer while model access gets cheaper. One discordant note is worth holding: OpenAI is discounting GPT-5.6 Sol 50% through Router while Pro subscribers report exhausting a $200-a-month plan in a single heavy coding day. Discounting on one side and caps on the other is what a supply-constrained vendor under price pressure looks like.

A note of discipline on the inputs, because a skeptic would start here and would be right to. The widely repeated "hundreds of millions annually" spend figure is The Information's extrapolation from public list prices, not disclosed spend, and it ignores enterprise discounts and caching. The 56%/2% result is self-reported on the buyer's own methodology, and a 2% aggregate quality decline can conceal severe regressions inside specific high-stakes workflows. The claim that 83% of organizations need infrastructure upgrades for production agentic AI comes from sponsored Google Cloud research.

The vendor's only rational counter

With account value flat at their largest customers, frontier vendors have one move left, which is up the stack into services and outcomes. Anthropic's implementation joint venture with private-equity firms, and its enterprise arm's purchase of a consultancy, are that move. For anyone selling software or services into the enterprise, the model vendor is quietly becoming a systems-integration competitor. That conflict is far cheaper to find in a partner agreement than to discover during a renewal.

Enterprise AI moved from growth budget to managed cost, and the buyer who can prove quality holds at a lower price now sets that price.

The mechanism is underwriting drift, not bad management

The tempting read is that these were execution failures, and it is the wrong read. Underscore's Lily Lyman describes a whole cohort where "the market's shifted so much in terms of what's possible with Claude" that the value proposition is worth less than it was underwritten at. Monashees' Eric Archer names the structural version: AI has shortened the half-life of a thesis, so a ten-year instrument is being underwritten against an 18-to-24-month reality. That arithmetic does not stay inside the fund. It applies with equal force to the three-year product roadmap sitting in most operating plans.

The four lanes, and what evidence sits in each

The supply-side news closes the argument on the fourth lane. Nvidia committed roughly 54% of a lab's enterprise value for rights that lock out nobody, which prices model IP as a rentable input rather than an ownable asset. Any roadmap line defended by "we have the better model" is renting a moat in a deflating market, and the landlord is free to re-let it to the competition.

The buildable moat is the one most companies are deleting

The most useful counter-example is small and unglamorous. Every, a roughly 30-person company, collected 30,000 historical edits from its editor in chief, built a copy-editing agent from them, and back-tested that agent against her own past work. No fine-tuning, no research team, no meaningful capital. The enabler was the dataset, and the dataset was exhaust: labeled expert judgment that already existed.

Most organizations are destroying their version of it. Code review comments from the best architect, the GC's redline patterns, the deal desk's approval logic, the strongest CSM's escalation calls — expert judgment with outcomes attached, aging out of a chat retention policy. A reasonable skeptic would say that hoarding logs is not a strategy, and the skeptic is right that storage alone builds nothing. The tradeoff is still lopsided: retaining the exhaust costs almost nothing and deleting it is irreversible. Retention is the cheapest strategic option on the table and almost nobody exercises it. Casey Newton reports that builders on frontier models must be willing to throw products out every three to six months; decision exhaust is the only layer that compounds while models churn.

The supply forecast hiding in the concession

Smith's admission is not contrition. It is a supply forecast. Sponsor-owned software with real data and weak AI execution is coming to market, and the natural buyer is whoever has genuine agentification capability. Two calibration notes: the moat framework is investor pattern-matching, and seven of twelve private-equity transactions disclosed no terms, so treat the rotation into regulated and physical assets as directional rather than precisely quantified. The deal flow argues the pattern-matching is holding up anyway: specialty pharmacy, water treatment, railroad emergency response, a defense IPO.

If a model release can write a product's obituary, what the company holds is lead time, not a moat. Lead time should be spent buying data, licenses and liability.

The retention money landed away from the exposure

Every confirmed retainer recipient was IC6 or IC7. No one has documented an IC4 or IC5 receiving one, which leaves the mid-level and senior bench — the people carrying the operating knowledge of specific systems — sitting at standard market bands. That is the arbitrage, and it has an expiry date: January refreshers.

The amounts matter less than the failure pattern. In one confirmed case an engineer showed Google the Meta retainer number, Google raised its bid, and the engineer left. Meta's retention budget funded a competitor's winning offer. The lesson is not that money fails to retain engineers. It is that money fails to retain engineers whose autonomy was taken away, and there is now a public, well-sourced demonstration of the distinction.

The mechanic generalizes, and almost nobody has modeled it

Meta stock sat near $200 in March 2022 and again in March 2023, and trades around $540 today, so those grant cohorts appreciated roughly 3x. As that equity fully vests at year end, total compensation falls unless January refreshers are unusually large, and engineers are openly saying they are staying to see the number. Any company whose equity appreciated meaningfully since 2022 carries the identical exposure, and the fix is a compensation-modeling exercise rather than a culture initiative. The second attrition wave is arithmetic, not sentiment.

What the disruption bought, and what supply looks like now

An honest scorecard is worth writing down, because a board member will eventually ask for one. The reassignments restarted Meta's AI coding model work and shipped Meta Muse Spark, now ranked 7th most capable and tied with Grok 4.5. A reasonable skeptic would point out that the capability gain is real, and the skeptic is correct. The gain is real and the lead is not decisive. It was purchased with reported "almost zero productivity across teams," collapsed confidence in directors and VPs, and the departure of the exact people teams depend on. Coercive reallocation cost more than the capability it purchased when the alternative on the table was a rationale and an opt-in path.

Two supply-side facts from elsewhere change the hiring math. a16z documents a seed-stage Stockholm company that hired six people who had each been CTO at a unicorn — leadership density not purchasable in the Bay Area at any price. And at Every, where AI writes essentially all the code, headcount went from about 15 to about 30 in a year. At the frontier of adoption, automation raises demand for senior judgment instead of lowering it, so an investment case built on FTE avoidance is underwriting an outcome practitioners are not observing.

Underneath all of it, private labs can now match megacap total compensation through secondary share sales, with listings plausibly 3–12 months out. Illiquidity has stopped being the discount it was. If those listings land, senior AI compensation resets again, and this quarter's retainer becomes next quarter's floor.

A million dollars cannot buy back autonomy already taken away, and every engineer being paid to stay is shopping that number to your competitors.