The AI stack got a landlord this week, and it's not you

Blackstone-led consortiums committed $10B to deploy OpenAI and $1.5B to deploy Anthropic across their portfolio companies. Same week, Anthropic doubled Claude Code enterprise pricing. Same week, Uber's CTO confirmed Claude Code running $500–$2,000 per engineer per month — enough to burn a twelve-month AI budget in four. And same week, three credible escape valves shipped: DeepClaude proxying to DeepSeek V4 Pro at a claimed 17× lower cost, Mistral Medium 3.5 hitting 77.6% on SWE-Bench with open weights on 4 GPUs, and IBM Granite 4.1 at 512K context under Apache 2.0.

Read those as one story. The frontier labs just bought a distribution channel that routes around your enterprise sales motion, priced their premium tier for the customers that channel will deliver, and watched the floor collapse out from under everyone else — all inside a single news cycle.

The PE move is the load-bearing piece. A general partner with committed capital in a deployment JV does not behave like a sponsor with a preferred-vendor list. When Blackstone's operating partner tells 250+ portfolio CEOs to standardize on Claude for back-office automation, the vendor evaluation is over before your AE gets on a plane. If your pipeline touches PE-owned mid-market and you don't know which sponsors sit in which consortium, you're forecasting off a map that was redrawn in April.

Yes, but — the counter-reading has teeth. PE-mandated software selection has historically been looser than the org chart suggests, contracts renew in 18 months, and JVs bought distribution rather than loyalty. Fair. Except the JVs are also funding integration consultants, which is the part that compounds. Once Claude is wired into a portfolio company's operations by an Anthropic-paid consultant, ripping it out in eighteen months is a services engagement nobody signs up for. The switching cost is the consultant, not the model.

The $221 problem is a category, not an outlier

GitHub Copilot burned $221 of inference on a single 15-message agentic session against a $40 subscription. That number is not the outlier. It is the mean once users unlock agentic workflows. The top 5% of users generate roughly half the inference spend on every product I've seen data for, and agentic loops are a 10–50× compute multiplier over chat. Flat-rate pricing on agentic AI is arithmetic that only works if your users don't discover the agent.

Of the coding assistants that matter, exactly one — Replit, at ~$1B run rate with 300% NRR — claims profitability. Codex is heavily subsidized. Cursor runs margin-negative. Claude Pro sits at roughly a 10× per-token premium over the alternatives. The subsidies end when the capital patience ends, and Anthropic doubling enterprise pricing this week says the capital patience is thinning.

The piece most operators are missing: the moat isn't the model. Mason Drxy's controlled ablation moved gpt-5.2-codex from 52.8% to 66.5% on Terminal-Bench 2.0 by changing only the prompts and middleware. Same weights, 13.7-point swing. That is larger than most model-generation upgrades. If your staff engineers are debating Claude versus GPT this quarter, they're optimizing the weekend-swappable layer while the six-month layer — the context pipeline, the tool schemas, the retry policy, the truncation rules — goes unowned.

The 17× DeepClaude number deserves a caveat because agent loops eat margins that token-price ratios don't measure. Retries, longer trajectories, tool-call schema drift. Real end-to-end cost delta lands closer to 4–8× on mixed production workloads. Still worth the migration for most teams. Not worth waiting a quarter for.

The compliance clock started, quietly

Underneath the economic story, the NSA and four Five Eyes partners published joint guidance formally naming autonomous AI agents as a tier-one cybersecurity concern — excessive privileges, cascading agent-network failures, prompt injection, weak auditability, and agent identity. They mapped it onto existing zero-trust and least-privilege frameworks, which is a deliberate choice to ship imperfect controls fast. The uncomfortable part is that most agent stacks already score badly against those existing frameworks. No new rubric. The old rubric was the one you were failing.

Historical pattern on Five Eyes advisories: 12 to 24 months from voluntary guidance to procurement language, then binding audit requirements. The clock is running. Every agent with production access needs cryptographic identity, short-lived credentials, and human-in-the-loop gates on high-impact actions before an auditor asks how you'd revoke a single compromised agent in under a minute. Long-lived shared service account keys fail that test today.

And the PromptMink malware traced to a Claude Opus commit in npm this quarter is not a curiosity. It's the first documented case where a frontier coding agent's signature appears on a software supply-chain compromise. "AI wrote it, I skimmed it" is now an unreviewed path into production. npm 11.10 ships min-release-age natively. Two lines of config buy you a 7-day cooldown that would have blocked every major supply-chain wave this year.

What to do this week

One action, sized to fit before Friday. Pick the three coding agents your team actually uses, pin the harness version, and run a 50-task benchmark from your own last 90 days of merged PRs — not SWE-Bench, not Terminal-Bench, your PRs. Measure cost per successfully merged PR, not cost per token. Log the exact system prompt, tool schemas, retry policy, and truncation rule on every run. Then rerun with the model swapped to Mistral Medium 3.5 or DeepSeek V4 Pro on the same harness.

One of two things happens. The open-weight substitution holds within 15% on your workload, and you have a negotiating position for the next Anthropic renewal that isn't "please." Or the delta is real and you now know exactly which tasks justify the frontier premium and which don't. Either outcome is worth more than another quarter of watching the meter.