The Board Room

Where the line actually falls

The question is not cheap model versus expensive model. It is whether a given call has a finite output space. Published comparisons put the typed-decision model at 85% on page classification against 56% for the generative baseline, then at 27% against 93% on choosing a crawl root, a task that requires a view of the whole site. Adjacent tasks, inverted results. That inversion makes decision-boundary classification an architecture call, made once per task, and almost no engineering organization has a taxonomy for making it.

The second lever is the harness, and the vendor default is the inefficient baseline

NVIDIA's SoL-Pi ran an automated search over agent harness configurations instead of models, the harness being the wrapper that manages context, actions and observations around a coding agent. Four techniques survived the search: Action Fusion, Online Context Compact, ObservationPack and an Evidence-Preserving Reducer. Together they cut recorded token traffic 44.7–49% and roughly a third off the hourly API bill at approximate performance parity, measured against native Codex and Claude Code defaults. The caveat is material: 51 tasks on EdgeBench. Directionally important, evidentially thin. Replicate it on a local task distribution before booking the saving.

Why this matters

OpenAI published the demand side of the same equation. Non-engineering teams went from roughly 0% to 90% Codex adoption in four months, and rising PR volume pushed a 10x load surge through CI in six months. In Gergely Orosz's synthesis of the nine-stage loop, a human defines the outcome and the agent does "pretty much everything after that." OpenAI explicitly declines to judge the output: "whether this is good quality or not, well, we'll see." A skeptic reads the adoption curve as a productivity story, and on its own that reading holds. Set beside the CI surge it becomes a cost-relocation story: the cost moved to CI compute, token traffic and human verification. None of those scale with headcount, and headcount is what capacity models still measure. OpenAI's own incident agent diagnoses but cannot mitigate, so on-call stays staffed.

The layer commoditized before it matured

Within days of the typed-decision launch the pattern was cloned: kev (Apache-2.0, built on a 0.5B base model, runs on Apple Silicon), LocalJev on local inference, Cua's 2.8MB form-filling model, and an open-sourced reinforcement-learning decision family spanning 100+ languages. The leader's residual advantage is reported at roughly 19 points out-of-domain, measured against clones that shipped in days. The adjacent structured-output vendor drew six clones in two days, and its "output tokens free" line appears in a launch table, not a price list. Vercel's 13% of paid teams inside 24 hours measures trial during a free promotion, not retention. Incumbents should be expected to answer with native cheap-decision endpoints within a quarter or two. Teams that abstracted the layer get price competition. Teams that hard-wired a specialist get a migration project.

Cost savings here are table stakes competitors also capture. The defensible part is a product experience that could not exist above 300 milliseconds.

What the two marks have in common

Both marks are priced off the building and the ground, not the model demand inside them. The Jane Street facility was structured so that tenant quality would carry the deal, and a profitable, cash-rich trading firm as anchor lessee is exactly what makes a GPU building read as infrastructure. Credit investors looked past the tenant anyway and underwrote the residual value of the building, an asset whose usefulness depends on hardware cycles that may turn faster than the lease or the debt. The Doña Ana mark sits further still from technology: 1,400 acres, county-level permitting backlash, and syndicate banks repricing paper that trades near par when deals of this type are healthy.

The visibility matters as much as the move. This repricing is observable in FINRA trade reporting, public and near real-time, on financing structures that used to be opaque. Credit has quietly become the highest-frequency indicator of the AI capital cycle, and it leads equities. A monthly readout costs days of work and buys two to three quarters of lead time.

The supplier set is about to sort itself

Where the sources disagree — and why both are right

The short-seller read is a late-cycle signature: infrastructure and cyber software at their most expensive since 2021, visible credit stress, accounting-focused shorts rotating in. The prescription follows from that reading, which is to pull capital events forward and, for buyers, hold dry powder for a credit-driven reset. The Information reaches the opposite operating conclusion from the same marks. Higher hurdle rates kill marginal projects. That tightens supply while demand broadens beyond labs and hyperscalers, with quantitative finance now buying purpose-built leased capacity and taking neocloud stakes. Both hold simultaneously. Assets get cheaper while capacity gets scarcer, because the binding physical input is power, and megawatts arrive on their own schedule while a loan reprices in an afternoon. The squeeze lands in 12 to 18 months.

Equity-adjacent capital is underwriting a different picture. CoreWeave priced $3.7bn in convertible bonds, more than it targeted; converts get upsized because buyers want the equity option, not because anyone underwrote the cash flows. SoftBank raised its Arm-backed margin loan by $5bn, to $25bn, to fund AI bets. Nvidia joined Crusoe's $3.9bn Series F at a reported $30.9bn post-money alongside three sovereign funds. A supplier is financing its own customer. Vendor-concentration math now has to account for who funds whom, and the Crusoe cap table is where that question first showed up.

The trade-off has a name. Locking long-dated capacity in a market that is pricing obsolescence means taking the residual-value bet credit investors just declined. That bet holds up where workload volume is stable and hardware-refresh substitution rights sit in the contract rather than in the relationship.

A multi-year capacity commitment to a thinly capitalized provider is a credit position you took without a credit process.

Two revocation mechanisms

Routing telemetry is public and near real-time, and it showed traffic moving at a speed no procurement cycle contemplates. That is the competitive mechanism, and it is the less dangerous of the two. The governance mechanism ran on a different track: three sophisticated buyers pulled a frontier model from use on data concerns, not capability. A buyer controls the first path by choosing when to switch. The second path is triggered by the vendor's conduct and arrives with no notice period.

The share move is leverage, and leverage decays. A vendor that shed thirty points wants anchor logos back and will trade on price protection and capacity guarantees. A vendor that gained them wants reference accounts. Both positions are exploitable only by a buyer who can actually move the workload, which has a concrete test: half of inference shifted to an alternate provider inside ten business days, with under 5% eval regression. Dual-sourcing counts when the second provider is already serving production traffic at volume.

What survives a vendor change

The durable assets sit below the model layer, and the same material makes the case. OpenAI's legal research product moved correctness from 38.7% to 54% with the base model held constant, buying the gain with a 230M+ URL curated index plus instruction and thoroughness scaffolding. Stanford's Paper2Agent beat Claude-with-repository-access, 98.7% and 100% against roughly 83% and 79%, on validated tool scaffolding alone, auto-generating and validating 593 of 599 tools. Model integration is the one layer a vendor can take back at renewal. The curated corpus, the retrieval quality and the eval harness that proves any of it works stay on the buyer's side of the contract.

The verification problem transfers with the deal

The 54% correctness figure comes from a private, 200-question validation set, and vertical AI is migrating steadily toward non-public benchmarks. Capability claims repeated to downstream customers are therefore claims no one downstream can independently check. At roughly half autonomous correctness, someone absorbs the verification cost and the liability that comes with it, and a firm selling outcomes without human review priced into the contract has volunteered for the job. The response that holds up is to make verification a product surface: citation grounding, because it lets a reviewer test the claim against the source, and confidence-based routing, because it decides which answers a human sees before a client does. Audit trails settle the argument afterward. Incumbents in regulated verticals hold that counter-position while the new entrant's access stays gated, and API access is already announced. The window is measured in quarters.

Thirty points of share moved in eight weeks. That vendor will negotiate, and only with customers who can move the workload in ten business days.

Price the prescriptions, not the provocation

The work here is not agreeing or disagreeing with the diagnosis. Four prescriptions are now circulating with wildly different evidence quality and wildly different costs if you adopt them and they turn out wrong. Treat them as hypotheses with kill criteria attached, because that is what they are.

The sleeper item is a revenue ceiling, not an org chart

Sellis attributes Snap's advertising business never reaching its potential to auction mechanism design, invoking Vickrey–Clarke–Groves — the pricing rule that determines what a winning bidder actually pays — explicitly. That reframes a decade of monetization underperformance from a demand problem into an architecture problem. Generalize it past Snap: if you own an auction, ranking or dynamic-pricing surface and nobody on your staff can defend its mechanism from first principles, you have an unmeasured revenue ceiling and you are funding a sales organization to push against it. That review is one senior hire and a quarter of work, and it is cheaper than the demand-generation spend it would replace.

The credibility discount, stated plainly

The track-record claims underwriting every prescription are unquantified self-attributions — first PM at Snapchat, the fastest growth the platform had seen, no numbers anywhere. The substantive material sits behind a paywall, and the host discloses he may be an investor in companies discussed. None of that makes the ideas wrong. It makes them untested, which is a different thing from false and demands a different response. The leaders who keep their product organization intact are the ones who walk into the conversation with a ratio thesis already written: PM-to-engineering-to-design ratios, what PMs exclusively own, and the marginal output of the Nth product manager expressed in shipped outcomes rather than alignment artifacts. Leaders asked cold lose headcount arbitrarily.

One adjacent read that is immediately actionable: Discord has lost its head of product, and the layer beneath a departed product leader is the loosest it will be for twelve months. That organization was credited with the platform's fastest growth period. If you compete anywhere near community, social or communications, consumer-scale community growth experience is a capability you hire or go without.

If you cannot name the marginal output of your Nth product manager, someone above you will name it for you.