The Investor

The term that matters is revenue share, not price

A per-token price increase is a cost problem, and cost problems have engineering answers: quantize the model, cache aggressively, shrink the tool surface, route the easy turns to something cheaper. A revenue-share demand is a different instrument entirely. Alibaba is reportedly weighing exactly that ask, per Benedict Evans, which converts an application company's gross margin from a number its engineers control into a number its supplier reopens at will. Anyone who underwrote a 70% steady-state margin behind a vendor holding revenue-share optionality underwrote a counterparty's forbearance and filed it as a cost curve.

Evans reads the market as a supply crunch in which labs can name their price. The counterweight he reports from the buy side is that large enterprises now presume dual-sourcing, because, in his phrase, you cannot rely on anyone right now. Demand-side behaviour commoditises the model layer in the same quarter that supply-side scarcity lets suppliers raise price. A poor place to own equity and an unreliable place to buy inputs, at the same time.

What the free tier now covers

Meta's Muse Glimmer is the specific fact that resets the COGS line: 30B parameters under Apache 2.0, running on a single 24GB RTX 4090 or a 32GB Apple Silicon Mac, sub-20GB at roughly 4-bit quantization for 0.2-1% accuracy loss, and about 233 tokens/second on an RTX 5090 using DFlash speculative decoding for roughly 3x throughput. It was trained around the agent loop itself, planning, tool calls, self-checking, failure recovery. Nous Research came at the same line from software, collapsing twelve Hermes browser tools into one and cutting token consumption 48-66% with no accuracy drop. Databricks has published its own token-cost optimisation methodology, which is the tell: cost-per-token engineering is table stakes, not differentiation.

The honest caveat: a 30B model does not replace a frontier model on genuinely hard problems. Most agent turns are not hard problems. They are tool calls, retries and the self-checks between them, which is what this class of model was tuned for, and also where the token volume in every agent forecast was supposed to come from.

Where the two readings diverge

One reading says suppliers gained pricing power. The other says harness efficiency cuts revenue per task even as agent volumes climb. Both hold at once, price per token up, tokens per task down. A model vendor whose top line rests on agentic usage growth now needs a token-efficiency discount applied to the per-task assumption. An application company loses predictability in the variable-cost line in either direction, and it is the unpredictability rather than the level that breaks a five-year model.

The mirror image is the trade almost nobody is pricing. This is probably too clean, but zero-egress, zero-per-token local agents under a permissive license look like a structural advantage in healthcare, legal, defense and EU data-residency accounts, where cloud-API competitors cannot follow on price or on data handling. Two ways that goes wrong before the license does: the frontier gap widens faster than the harness improves, or enterprises decide dual-sourcing means two clouds rather than one cloud and one laptop. The unhedged tail risk is license durability: a business whose entire cost structure assumes permanent unrestricted access to one open-weight lineage is carrying regulatory risk it has not named.

If a holding's margin case improves only because tokens got cheaper, you own a market condition, not a business.

The packaging decision is the TAM claim

The price is the least interesting term here. What to underwrite is where GitHub put the SKU: Code Quality shipped outside Advanced Security, which is a monetization judgment dressed as a packaging detail. Microsoft is asserting that maintainability carries standalone willingness-to-pay, separate from the security budget, and that is budget-line creation rather than a feature reshuffle. It also hands challengers a clean total-cost wedge, though Microsoft keeps the option to collapse that wedge by re-bundling the moment anyone gets traction. Both things are true at once, which is usually how these end.

The legitimacy half arrived in parallel, and it removes the last procurement objection. Linux 7.2-rc7 landed on August 9 with more than 400 fixes Torvalds credits directly to automated scanning across nearly every subsystem, including an eight-year-old use-after-free race in ptdump that Syzbot flagged in June, with Claude Opus 4.8 credited in tracing the root cause. Human triage and sign-off remain mandatory on every AI-assisted contribution. When objections surfaced in July, Torvalds told dissenters they could fork or leave. The most scrutinized codebase in software now runs machine-assisted review as ordinary process, under a policy that survived open dissent.

Three positions, one published price

Sources diverge on durability, and that disagreement is the whole decision. One reading says $10 is an introductory number that rises once the category stops being contested. The competing evidence is that incumbents keep giving away last year's wedge: Spotify's Xirp orchestrates across Claude, Gemini CLI and OpenAI Codex, was validated by 1,300 internal engineers, and is now free to anyone. Underwrite the floor, not the recovery. Forty-two reviewers on a single public benchmark is competitive density, not greenfield.

Where the money goes instead

Spotify's own precedent is the template, or rather the more useful version of it: Backstage became the standard and the money accrued one layer up, in audit logging, SSO, cost attribution and SLAs. The equivalent layer for agents is already evidenced as a gap rather than a preference. A survey of roughly 336 papers found GUI and computer-use agents systematically lack error recovery, safety checks and auditability, which are architectural gaps rather than model-capability gaps. Separately, hidden-activation correctness probes proved extraction-method-dependent and unusable as a substitute for running tests. The cheap shortcut to agent verification is closed.

The tooling wave has names attached already: Prefactor on real-time run scoring and drift, Cekura on voice and chat agent QA inside CI/CD, FetchSandbox on deterministic replay across 60-plus APIs. This is probably wrong in the particulars, but the one worth funding is whichever becomes the audit system of record for agent actions, because regulated buyers pay for provenance and do not pay for a dashboard. Assume a 12-24 month window before platforms bundle the rest.

Underneath sits the demand driver nobody markets: agents solve each assigned task correctly and never plan for future features, silently compounding technical debt that slows both engineers and subsequent agents. Every dollar of agent-generated code manufactures future demand for refactoring, and for the debt visibility that makes refactoring schedulable. Discipline note: PR-AF's 0.706 recall, its #2-of-42 placement and the 3x-findings-at-10x-lower-cost claims are vendor-supplied and unverified — annotate them before they anchor a valuation conversation.

In devtools the money just moved from shipping agent output to proving and cleaning it.

The 25% rule

The heuristic travels further than the financials. Optiver runs 30-40% of its roughly 950 engineers on platform work, against 15-20% typical at large tech companies, per The Pragmatic Engineer's reporting. Take about 25% platform engineering as the dividing line: above it accounts build, below it they buy. That line is probably wrong at the edges, and it still sorts accounts better than revenue does. Optiver forked asyncpg, contributed a nanosecond-precision timestamp type to Postgres, built PG Feed on the Postgres write-ahead log specifically to avoid Kafka's extra disk I/O and latency, and shipped its own AI gateway and MCP hosting platform internally this year.

That last item closes a beachhead. Selling agent gateways or MCP infrastructure into elite financial firms means pitching accounts that already built the product, which sends that go-to-market down into tier-2 financial services and mid-market enterprise, where nobody staffs 300 platform engineers. Then again, Databricks won the entire data platform at a firm operating at nanosecond precision, alongside Kafka and Postgres. Where these buyers buy, they buy the incumbent lakehouse, which scopes challenger theses to research and backtest.

The layer the budget moved to

Alex Itkin, Optiver's CTO US, says what the industry used to keep indoors: ultra-low latency is no longer a moat in itself. The retreat system went from seconds to nanoseconds over a decade, and that decade is now the floor rather than the edge. Budget migrated one layer up, into research compute, where serious prop shops spend hundreds of millions each. Hence NVIDIA, Groq and Cerebras courting trading firms, Hudson River Trading on a GTC stage discussing Blackwell, and Jump among the first to deploy Vera Rubin.

The counter-thesis deserves a hearing: chip vendors court prop shops for the logo, and the revenue never scales past pilot. Nine-figure research budgets argue otherwise, as does a buyer profile enterprise software rarely sees. No external customers, so no procurement committee and no external deadlines. Own capital only, so budget authority sits with people who can attribute a GPU purchase to profit and loss within weeks. Partnership structure, so nothing waits on a board cycle.

Two second-order repricings

1. Talent costs more than the models assume. AI labs now recruit out of prop shops, not just Big Tech, because the binding constraint at frontier labs is low-level systems and hardware talent rather than ML researchers. Kernel, HPC and FPGA hiring lines are underfunded against a market clearing at prop-shop levels and above.
2. Agentic coding turns headcount into compute. AI coding tools have started raising builds per engineer per day, against bare-metal CI clusters that require production-grade capacity planning. GitHub Actions exposes no queue times or utilization, so Optiver built bespoke observability on webhooks. A product wedge sitting on a demand model, and the CI category is not priced on it yet.

Two things belong in the documents before anything ships. Knight Capital nearly went bankrupt on a single bug that produced a $440M loss, so anything touching the live execution path carries reputational and legal exposure far beyond its ACV, which is one more argument for the research, backtest and CI paths where budget is growing anyway. And 2025's record profit reflects a favourable volatility environment; profitable strategies decay quickly. Prop-shop revenue belongs at under a quarter of any portfolio company's forecast: a proof point for adjacent high-performance verticals, not the terminal market.

The dozen firms that can fund nine-figure research clusters are customers, not targets, and they have no procurement committee to slow you down.