The Board Room

The Contradiction in One Frame

Anthropic's valuation has crossed $1-1.2 trillion on 10x annual revenue growth. In the same news cycle, Fleet swapped Anthropic's Sonnet 4.6 for Kimi K2.6 with zero quality degradation at one-fifth the cost. Capital markets are pricing monopoly rents while the underlying technology commoditizes; both are true today, and only one survives the decade.

The capital markets are paying for the layer they believe compounds. The open-source community is proving that layer is reproducible at 5x lower cost. Both bets are currently correct. The question is which one a three-year vendor commitment is exposed to when they diverge.

The Evidence Is No Longer Anecdotal

Several convergent signals say open-source has crossed a threshold that matters for procurement decisions this quarter:

• Kimi K2.6: Opus-level quality at 20% of API cost, production-validated
• Aurora and ZAYA1: Frontier parity achieved with 100x fewer training tokens
• vLLM: 72% throughput improvement on H20 hardware
• SGLang: Processing 57 billion tokens daily at scale
• Zyphra: Training competitive models on AMD, breaking the NVIDIA dependency

If inference costs fall 5-10x over the next twelve months, which this evidence trajectory supports, any business whose margin depends on API markup is building on sand. Firms whose value sits in orchestration, domain knowledge, or workflow get the opposite outcome: their cost per task falls with each model release without a single renegotiation.

The Funding Paradox Makes This Urgent

The infrastructure financing the frontier labs requires those labs to hold pricing power. Big Tech's collective free cash flow has collapsed 91% — from $45 billion to $4 billion per quarter — under AI capex weight. SoftBank just cut its OpenAI-backed loan from $10B to $6B, the first serious note of capital impatience. If returns arrive 2-3 quarters late, the correction lands on cloud pricing, API costs, and startup funding at the same time.

The scenario most enterprises have not modeled: a 20-40% cloud price increase by Q1 2027 as hyperscalers attempt to recover margins. Any strategy predicated on flat or declining compute costs deserves a stress test against that scenario now, rather than when the price increase arrives.

The Decision This Forces

Buying locks in a cost basis set by a vendor whose pricing power is under active assault from free alternatives at one-fifth the cost. Building on open weights accepts a modest engineering tax in exchange for owning the curve. Neither choice is obviously correct this quarter. One of them will look obviously correct in eight quarters.

The honest framing for any enterprise with more than 60% of AI spend concentrated in a single frontier provider is this: a contingency plan has to cover a 30% price increase when that provider moves to satisfy its investors, and it has to cover the open-source alternative a competitor adopted six months ago reaching the same quality at zero marginal cost.

The Coverage Gap Nobody Is Pricing

Berkshire Hathaway and Chubb are removing AI-related damages from standard commercial policies, and they have received 80% regulatory approval to do so. The total specialty AI insurance market today is $40 million, smaller than a single Series B round. Projections put it at $5B by 2032, which is another way of saying adequate coverage at reasonable prices will not exist for several years.

Companies are deploying AI aggressively while their insurance quietly exits coverage of the thing being deployed. The first major uncovered AI loss is still hypothetical. Every one of them is, right up until it isn't.

Why This Matters This Quarter, Not Next Year

The incentive structure is economically irrational and organizationally predictable. The team deploying AI reports to the CTO. The team managing insurance reports to the CFO. In most organizations, these two decisions are made by different people who do not coordinate. Deployment accelerates. Coverage narrows. The gap widens quarterly.

The window matters because bespoke AI liability coverage is still being written at reasonable rates today, on the strength of a minimal loss history. That window closes the moment a headline-making uncovered incident forces repricing across the specialty market. The Mythos result of 423 Firefox vulnerabilities in a single month shows the attack surface. The Canvas breach affecting 275M people during finals week shows the exposure. Combine the two with a coverage exclusion and the balance-sheet risk is material.

Insurance Posture as Deployment Ceiling

A reasonable skeptic would call this a risk-management question for the general counsel and move on. The skeptic is half right. It is also a deployment-ceiling question. An AI application that creates $50M in value but carries $200M in uninsured liability is not a net positive. It is a contingent liability that belongs in the 10-K. Firms that lock in coverage now get a roadmap set by their engineers. Firms that wait get a roadmap set by their underwriter.

The PE Angle Amplifies This

TPG, Blackstone, and Brookfield committed $10B alongside OpenAI, and private equity firms are now mandating AI adoption across portfolio companies. The liability profile splits along an awkward seam: the operating partner at the fund writes the deployment mandate, the portfolio company's balance sheet absorbs the uninsured exposure. That mismatch produces the first litigation cycle, and it arrives before the insurance market scales to meet it.

The Existence Proof That Moved

A non-programmer — a newsletter publisher — built a feature-complete email client in seven days that replaced a $30/month Superhuman subscription. Split inboxes, command palettes, undo-send, email rendering, unsubscribe flows. Every feature that took Superhuman's engineering team months is now a one-week project for someone who cannot write code. This is not a prototype. It is a working production inbox.

When implementation cost approaches zero, the entire surface of competitive advantage moves to problem selection and structural positioning. The winning organizations will be the ones that can answer, in one sentence, why their product cannot be rebuilt in a week by a motivated user with a $100/month coding agent subscription.

The Pattern Is Now Weekly, Not Quarterly

Three data points make this structural rather than anecdotal:

1. Airbnb: 60% of new code is AI-written — a company operating at this ratio has a fundamentally different cost curve than a competitor at 10%
2. Stripe's projects.dev: Combines Vercel + Stripe + Supabase into single-click deployment, collapsing days of setup into minutes
3. Factory: Positioning as a "coding harness for non-coders" — the institutional version of the one-week build

The uncomfortable companion number: current LLMs corrupt 25% of document content in long editing workflows. Both things are true. The firms that chase the 60% without building verification scaffolding will pay for it in production incidents.

Which Moats Survive and Which Don't

The Agent-Native Architecture Shift

The builder deliberately added hidden selectors and debug endpoints to make his app operable by AI agents. This is the mobile-responsive moment for the agent era. Applications that expose agent-friendly interfaces become composable in automated workflows. Applications that do not become dead ends. Mandating agent-operability in product architecture today is a decision that looks cosmetic for two quarters and structural for the next ten.