Anthropic Rented From Its Enemy. Your Stack Just Repriced.

Anthropic leased xAI's entire Colossus 1 cluster this week — 220,000 GPUs including GB200s — from the CEO who publicly called them "misanthropic and evil." Dario Amodei has now admitted the company planned for 10x growth and hit 80x. Renting compute from a declared rival is not strategy. It's what a capacity emergency looks like with a counterparty attached.

On the same calendar, Anthropic converted every Claude subscription into dollar-matched API credits across the Agent SDK, GitHub Actions, and every third-party harness — Cursor, Cline, OpenCode, Zed. The 70-90% implicit discount that made those tools economically viable disappears on June 15. OpenAI answered within hours with two months of free Codex for enterprise switchers, window closing July 13. ServiceNow, sophisticated by any measure, burned its full-year Anthropic budget by May and couldn't tell which users drove it because Anthropic ships no per-user telemetry.

Three events, one story. The compute is scarce enough to bend rival strategy, the pricing is being reshaped for an S-1 that likely lands in October, and the revenue underneath the headline numbers doesn't behave like SaaS revenue — no SLAs, no attribution, low switching cost, reversible in a single procurement cycle.

Yes, but — Ramp's 34.4% vs 32.3% crossover is a card-billing metric, not token volume or workload criticality. The specific ranking is measurement noise. The directional signal — multi-vendor procurement is now the default, and enterprise buyers are unhappy with what they bought — is robust across every source that touched the story.

The margin event is 30 days out

Every portfolio company, internal tool, and coding harness that priced COGS off subsidized Claude tokens lost 20-40% of runway last Friday. Most founders haven't flagged it to their boards yet because the invoice doesn't arrive until July. The teams that catch this in May renegotiate with leverage. The teams that catch it in July explain it to the audit committee.

What to do this week: strip the harness from one representative workload, log raw input/output tokens and tool-call fanout, and compute (current third-party usage − plan credit equivalent) × API rates. That's the new bill. Then run OpenAI Codex on the same tasks during the free window. Worst case, you have fresh comparative data before your Anthropic contract renewal. Best case, you find out your harness was worth less than the arbitrage.

The deeper move is instrumentation. Deploy an LLM gateway — LiteLLM, Portkey, something that gives you per-user, per-feature, per-model tagging with daily budget alerts. ServiceNow's problem wasn't that Claude cost too much. It's that they couldn't attribute the cost until the money was gone. If you're running any Claude-dependent path in production without that layer, you're one usage spike away from the same postmortem.

The other three floors that fell

Vercel published the first production-grade index of AI traffic across 200,000+ teams: 59% of tokens are now agentic. Multi-turn, tool-calling, state between turns. If your eval harness still scores single-turn responses against reference answers, you're benchmarking the minority workload. Anthropic captures 61% of dollar spend on Opus for reasoning; Google captures 38% of volume on Flash for throughput. Two different businesses inside "foundation models," and the routing layer is where cost lives now.

Security took a harder hit. UK AISI confirmed Anthropic's Mythos cleared both simulated attack ranges — full autonomous network takeover, a discrete jump from the prior generation's persistence ceiling. Mozilla wrapped the same model in a custom harness against Firefox and surfaced 271 previously-unknown bugs including sandbox escapes. Daniel Stenberg ran a generic scan against curl and got one low-severity CVE. Same model, 271x delta from scaffolding alone. Harness engineering dominates model choice on this task by at least fifty times.

Meanwhile the endpoint layer quietly gave up. TrustedSec pointed LLMs at five commercial EDR products and found they share one architecture — YARA rules, Lua engines, local ML classifiers — reversible in days. PraisonAI went from disclosure to active exploitation in four hours. NGINX shipped an 18-year-old unauthenticated RCE in the rewrite module. Traefik dropped a CVSS 10.0 auth bypass. Argo CD leaked plaintext Kubernetes secrets. Five critical CVEs at five consecutive stack layers, chainable end-to-end, in one week.

The patch SLA calibrated to human attacker tempo is calibrated to last year's threat model. Ingress first, control plane second, kernel third. Tonight, not the weekend.

The one thing to ship this week

Instrument per-request cost attribution at your LLM gateway with team, feature, and model tags — and pipe it into the same dashboard as your patch SLA and eval trajectory metrics. That's not three projects. It's the same project three times. Every failure mode this week — the pricing cliff, the agentic eval blindspot, the collapsed patch window — is a failure to see what your systems are actually doing until the invoice, the exploit, or the reward-hack arrives. Fix the observability layer first. Everything else is downstream of what you can measure.