The Investor

The Core Problem

A tour of 14 Chinese AI labs produced something closer to a unit-economics report than an AI story. Export controls did not suppress Chinese AI. They industrialized its efficiency. The labs are extracting 4-7x more intelligence per FLOP than naive scaling predicts, shipping models 6-8 months behind the frontier at 10-28x lower prices, and holding 50-70% gross margins while doing it. CAISI, the US government's own assessment, agrees the capability gap is roughly stable at 6-8 months. The unit-economics gap is structural and widening.

The Numbers That Break the Thesis

DeepSeek V4 Pro clears at $0.43/M input tokens. GLM-5 at $1.00. Kimi K2.5 at $0.95. Claude Opus 4.6's premium tier runs 11-28x higher for comparable-quality work, which is a lot of multiplier to defend. Z.ai reports 50% gross margins at those prices. MiniMax reports 70%. These are not subsidized loss-leaders, or rather, the more interesting version: they are profitable operations running on 8x less compute.

Chinese token volume has reached ~9 quadrillion tokens/month against roughly 4 quadrillion in the US and West. The consumption side is already 2x. Market structure is the other half of the puzzle: 14+ frontier-class labs in China, including Meituan, Ant, and ByteDance, against 5 in the US, fragmenting the chip pool and collectively out-iterating through open-source.

The Bellwether Signal

Cursor built Composer 2 on MoonshotAI's Kimi K2.5. A flagship US developer-tools company, one that sits in dozens of venture portfolios, is already routing workloads through Chinese open-source infrastructure. This is not a margin decision. It is capitulation on the proprietary-frontier-model moat.

When a darling portfolio candidate builds its flagship product on Chinese open-source, the thesis that "access to frontier models" constitutes a moat is no longer supported by the evidence.

Portfolio Implications

The $285B that flowed into US AI startups in 2025 against $12.4B in China was priced on the thesis that capital intensity creates durable pricing power. Chinese efficiency is that thesis's biggest counterexample on the board. Asking whether China will reach the frontier is the wrong framing. Adequate capability at 1/20th the cost collapses premium pricing in most enterprise use cases. That is the actual question, and mostly the answer is no.

CAISI, the lab tour, and Cursor's routing decision all point the same way. The durable moats now live in distribution + proprietary data + workflow lock-in, not raw model capability. Frontier model access belongs in the commodity column in most thesis docs this quarter, if LPs are reading carefully.

The Contrarian Read

China is not catching up on capability. The gap is stable. What is actually happening is that China is winning on unit economics while the US wins on capability, and those are different axes. Premium pricing survives where frontier capability is differentially necessary: complex reasoning and safety-critical agents, plus frontier coding (Claude on F4). Outside that perimeter, the Chinese price floor becomes the global price floor inside 12 months. This is probably wrong in the tails. It is the base case everywhere else.

The Math That Kills SaaS Comps

Six independent sources this week landed on the same uncomfortable arithmetic, which most IC memos are still pretending they haven't read: AI-native software appears to cap at roughly 17% gross margin, or eleven percent net at scale, rather than the seventy percent that every SaaS comp in the deck quietly assumes. The mechanism is not mysterious. Personalized inference kills the caching and multi-tenancy leverage that made SaaS SaaS, reasoning models burn 10-100x more tokens per query (which offsets the per-token price declines everyone keeps citing as salvation), and enterprise deployment keeps needing Forward-Deployed Engineers. The gross margin shape is Palantir, not Snowflake. That is the trade.

The tell is that the evidence is now multi-directional. Google, OpenAI and Anthropic all shipped human-heavy deployment motions in the same week — FDEs, a thing literally branded 'Deployment Companies,' and PE joint ventures. When three frontier labs independently bolt on a consulting arm in the same seven days, they are not announcing a strategy. They are confessing that enterprise AI does not self-serve. The narrative premium in the stock prices still assumes clean software economics. The delivery model is services.

Four Escape Routes — Only Two Are Venture-Scale

The a16z Thesis Tell

Seema Amble at a16z went on record calling structural disruption on Salesforce, SAP and Workday, flagging manufacturing, construction, field-service and accounting as the AI-native SoR opportunity. Salesforce's 'headless product' launch reads correctly as defensive, or rather, as the most defensive thing an incumbent can do: you do not rebrand your APIs as products unless you are worried about whoever is building the UI above them. Per-seat pricing breaks structurally when the seat is an agent.

If a founder can't credibly answer 'which of the four escape routes are you?' — it's a margin trap, no matter the ARR growth rate.

The Counter-Narrative

Aaron Levie at Box has the opposite trade, and it is not a silly one. His view is that data-layer SaaS benefits from agent proliferation — agent activity moves from ten percent of workload today to ninety percent within three years, which drives multiplicative demand on data platforms through a seats + consumption stacking model. Engineering job openings are at a three-year high against 46K March layoffs, so the labor story is redistribution, not destruction. Token spend going from roughly one percent to ten percent of company opex opens a greenfield 'AI FinOps' category that currently has no incumbent.

The resolution, probably wrong but worth stating: data-moat incumbents win, thin wrappers die. The SaaS trade is not universally short. It is bifurcated. Overweight the data layer (Snowflake, Databricks, Box-class). Underweight thin-workflow SaaS without a consumption-pricing option. The rest is noise.

The Enforceability Problem

Anthropic has publicly voided secondary transfers routed through Hiive, Forge, Sydecar, Lionheart Ventures, Open Doors Partners, and Unicorns Exchange, which is either housekeeping or — to use a more honest description — a NAV question for any fund carrying Anthropic exposure through those platforms. The gap between what an SPV says it owns and what the issuer will recognize has widened meaningfully in the last month. If the company refuses to honor the transfer, the paper marks down quietly regardless of what the secondary market is printing.

Three ways this goes. The secondary market formalizes and SPV holders are fine. Issuers win enforcement and paper marks down twenty to thirty percent. Or governance noise drags restrictions into a broader sector retrace. The middle outcome is the one being underpriced.

The Retail Froth Signal

Meanwhile, the publicly traded AI exposure vehicles offering retail synthetic exposure to Anthropic, OpenAI, and SpaceX now trade at two to three times the NAV of their underlying private shares, some of them doubling within days. This is the textbook signature of a marginal-buyer shift: institutional discipline has been priced out and retail FOMO is setting the clearing print.

When retail is paying 2-3x NAV for private AI exposure while the issuer is voiding the platforms that sold it to them, the trade is to distribute into the premium before the enforcement catches up.

The Microsoft-OpenAI Restructure

The revised agreement capped Microsoft's revenue share at thirty-eight billion dollars and freed OpenAI to diversify infrastructure across AWS and Google. That quietly unwinds every Microsoft AI moat thesis written since 2023. Anthropic's simultaneous AWS GA launch with full feature parity means AWS is now first-class frontier infrastructure for two of the three most valuable labs. The Microsoft AI exclusivity premium is over in substance, even as it is preserved in form through 2032.

The Opportunity Window

For anyone sitting on Anthropic, OpenAI, or SpaceX secondaries, the combination of peak demand, company-issued platform warnings, and a 2-3x retail vehicle premium is a textbook trim signal. This is probably wrong if the retail bid deepens rather than exhausts, but in-kind distributions, structured secondaries, and direct sales into retail-driven vehicles are the rare window where selling consensus pays above private FMV. The shelf life is weeks, not quarters.

What Changes This Week

Anyone holding Anthropic exposure via the six named platforms should obtain a legal opinion on transfer enforceability before month-end NAV marks. It is a materiality issue for this quarter's reporting, not a line item for the next LP meeting.

The Governance Layer Gets Claimed

Microsoft shipped Agent 365 plus the Work IQ API this week and told anyone paying attention that the governance and orchestration layer for multi-agent enterprise workflows belongs to them. Cloudflare put Artifacts into beta with Git-like versioning, rollback, and audit for AI agent outputs. Versa shipped agent policy controls. GitLab restructured into 60 teams around the same problem. This is what a category looks like in the six to twelve months before Gartner writes the Magic Quadrant.

The demand signal is visible in the same week. Datadog's telemetry across more than 1,000 organizations shows agent framework adoption has doubled year on year, which is the headline number. Underneath it, multi-model is now the default posture and LLM tech debt is compounding faster than teams retire it. Atlassian's survey has 74% of IT leaders citing AI security as a major barrier, while 79% of the same group are exploring AI incident management. Adoption is outrunning controls.

Where This Sits in the Stack

The 'Okta for Agents' Opportunity

a16z calls agent authorization "one of the hardest unsolved problems," which is the language partners use when they want the category named after their portfolio company. No named leader exists yet. The shape rhymes with the 2010 Okta setup, where seed checks priced the category before it had a label. This is probably wrong, but the counter-thesis worth respecting is that identity categories sometimes take a decade to resolve. Build the target list anyway.

The supply-chain attacks that landed the same week sharpen the urgency, or rather, make it legible to people who had been ignoring it. Mini Shai-Hulud hit Mistral, TanStack, and UiPath packages with persistence in Claude Code and VS Code settings files, across 121 compromised npm packages, including what is being called the first confirmed AI-assisted zero-day. Agents with code deployment authority plus compromised packages equals a board-level risk most portfolios have not audited.

Agent governance is roughly where container orchestration was in 2015: multiple vendors shipping into it, no narrative dominates yet, TAM still pre-consensus. By the time the analyst note lands, Series A pricing will be 3-4x current levels.