The AI market just bifurcated — pick the survivors before the repricing

Anthropic's annualized revenue ran from roughly $9B to $20B in a single quarter. Menlo Ventures' enterprise survey shows the company now captures 40% of LLM spend versus OpenAI's 27% — a complete inversion from the 12%/50% split two years ago. In coding specifically, the gap is 54% to 21%. ChatGPT's US mobile share fell 24 points in twelve months to 45.3%. Max Schwarzer, OpenAI's VP of Research and head of post-training, walked across the street to Anthropic during the Pentagon backlash.

In the same week, Josh Wolfe at Lux Capital sent a letter to portfolio founders telling them to extend runway, then said publicly what nobody in his seat says publicly: fewer than ten AI startups actually matter, and the rest are riding a bubble the industry is afraid to name. Lux had just closed $1.5B in January. The firm is hoarding ammunition while telling its companies to dig trenches.

These aren't contradictory data points. They're the same story told from two ends.

The math under the warning

AI infrastructure burned $443B in 2025 against $51B in traceable revenue. That's a 10.3:1 spend-to-revenue ratio — roughly four times worse than cloud computing at the equivalent stage. Barclays' back-of-envelope says you'd need twelve thousand ChatGPT-scale products to justify current capex. MIT found 95% of enterprise AI initiatives delivered zero measurable P&L return.

Meanwhile the public market is already pricing the correction. Salesforce announced a $50B buyback. ServiceNow added $5B. Pinterest took $2B with Elliott Management leading a fresh $1B position. That's $57B in defensive buybacks announced in weeks — management teams collectively betting their own stock is mispriced because AI-disruption fear overshot fundamentals. SaaS has shed roughly $2T in market cap this cycle.

And a new structural tell: AI startups are now selling the same equity at two prices in the same round. Aaru priced at both $450M and $1B simultaneously. Redpoint led the lower tranche; other VCs joined at the headline number. The stated purpose, per Primary Ventures' Jason Shuman, is to scare other investors away from backing the number-two and number-three players. If your diligence relies on headline valuations, you're systematically overpaying.

The correction isn't a 2000-style broad washout. It's a power-law consolidation. The winners are pulling away at historic velocity — Anthropic's ramp is faster than Salesforce ever managed — while the long tail returns capital. Both sides of that sentence need to be in your model.

What's actually load-bearing for the survivors

Two data points this week point at the real moat, and it isn't the model.

Stripe's eleven-task agent benchmark — full-stack payment integrations evaluated end-to-end — clocked Claude Opus 4.5 at 92% versus GPT-5.2 at 73%. Useful, but the more interesting number sits inside the methodology: identical model, identical task, scaffold-only variation, and the score swings from 42% to 78%. Thirty-six percentage points from orchestration design alone. Boris Cherny, who built Claude Code at Anthropic, ships 20-30 PRs a day running five parallel agents in separate worktrees with a plan-first, one-shot-implementation pattern. His team tried vector DBs, recursive model indexing, and RAG for code search. They landed on glob and grep. The operational overhead of embeddings didn't pay for itself on a structured corpus.

If you're A/B testing foundation models without controlling for harness, you're confounding two variables with very different effect sizes — and the smaller one is the model.

The other load-bearing surface is commerce. Mastercard Agent Pay went live for all US cardholders this week. Stripe and OpenAI's Agentic Commerce Protocol has Etsy in production and a million-plus Shopify merchants onboarding. Visa shipped Trusted Agent Protocol with cryptographic agent identity. Visa and Mastercard both announced stablecoin settlement in the same news cycle. Six months ago this was a whitepaper category. It's now production infrastructure, and your checkout flow has a new user that isn't human. CAPTCHAs, visual confirmations, multi-step forms with dynamic layouts — those are agent-blockers. Being invisible to agents in 2026 is what not being indexed by Google was in 2005.

What to do this week

Three moves, in order of how much they hurt to delay.

First, audit every AI vendor in your stack against a 50% funding-reduction scenario. The cost of doing this in March is an afternoon. The cost of doing it in October, after the first marquee AI startup misses payroll, is a forced migration on someone else's timeline. Any product critical-path locked to a single foundation model needs a documented multi-vendor contingency presented at the next board meeting. Anthropic's Import Memory tool — copy-paste your ChatGPT history into Claude — has explicitly weaponized the fact that switching costs at the model layer are near zero. Your customers will do the same to you if you don't get ahead of it.

Second, push iOS 17.3 across your managed fleet by end of business. Coruna is a 23-vulnerability zero-click chain, suspected US-government origin, now confirmed in the hands of Chinese cybercriminals and Russian state actors. Forty-two thousand devices confirmed compromised, watering-hole delivery, no user interaction required. This is EternalBlue for mobile, and any iPhone below 17.3 in your environment is known-compromisable today. While you're in there, patch CVE-2026-22719 in VMware Aria Operations — CISA added it to the KEV catalog March 3, which means actively exploited, not theoretical. Aria has god-view of your virtualization infrastructure.

Third, instrument your scaffold the way you instrument your model. Version it in your experiment tracking. Run ablations. If a harness change is worth 36 points on Stripe's benchmark, it deserves the same rigor as a model swap. Pick one workflow this sprint, hold the model fixed, and benchmark three orchestration variants on tasks from your actual codebase.

The organizations that triage AI vendor survivability, mobile security posture, and orchestration discipline this month will navigate what's coming. The ones waiting for clarity are the ones who generate it — for their competitors.