Half the AI stack you rely on can be revoked overnight

On April 4, Anthropic blocked third-party agentic tools — OpenClaw and the ecosystem it seeded — from Claude Pro and Max flat-rate subscriptions. Downstream, everything routes to per-token billing or nothing. Peter Steinberger, OpenClaw's creator (now at OpenAI), said the quiet part out loud: Anthropic absorbed the popular features into Claude Code, then pulled up the drawbridge. The same week, OpenAI moved Codex to usage-based pricing. Two of the three vendors your team probably depends on changed their cost model inside seven days.

If your CI loops on Claude via a third-party wrapper — reviewing, retrying, self-correcting — your bill just multiplied. Flat-rate made iteration free. Per-token makes it a line item. A 15-pass review agent that cost the same as a one-pass agent last week now costs 15x.

This is the day's story. Everything else is context for what to do about it.

The vendor layer just proved it isn't a platform

The cleanest read on where trust sits: Rainmaker Securities, brokering ~1,000 private stocks, reports $2B in unfilled Anthropic buy orders with zero sellers, against $600M of OpenAI shares nobody wants. OpenAI's COO moved to "special projects." Its revenue lead is on medical leave. Its response to a cold sentiment cycle was to acquire a tech livestream property in the low hundreds of millions and convert the hosts into in-house marketing — a company buying media talent for perception management, not capability.

Meanwhile, Microsoft's own terms describe Copilot as "for entertainment purposes only." Microsoft's legal team looked at the liability of standing behind Copilot for real work and declined. If you're a Copilot enterprise customer, that clause is your problem, not theirs. It's also the sharpest competitive weapon anyone selling against Copilot has been handed in a year.

Yes, but — a fair counter is that platform providers reshape terms constantly and mature buyers price it in. Snowflake did it. AWS does it every quarter. The counter falls apart on timing: Anthropic gave developers days, not quarters, and simultaneously shipped the replacement feature inside its own product. That's not repricing. That's a lockout with a coincident product launch.

Google released Gemma 4 under Apache 2.0 the same week. Read that as the market answering: the developers Anthropic just burned now have a permissively licensed foundation model with real weight. If your abstraction layer swaps providers with a config change, this week is a non-event. If it swaps with a code change, you have work to do this sprint.

The physical layer says none of the above scales anyway

Roughly 50% of planned US data center builds in 2026 face delay or cancellation. High-power transformer lead times: five years, up from two pre-2020. AI workloads want 18-month deployment cycles. The math doesn't reconcile. China supplies over 40% of US battery imports and roughly 30% of transformer and switchgear categories — a geopolitical chokepoint on the physical layer that runs everything above it.

The federal government just made this its problem. Trump's FY2027 budget proposes $1.5T for defense (+42%, largest single-year increase since WWII) with $15B explicitly redirected from clean energy to AI supercomputers. Domestic cuts are negotiating positions; the defense line and the AI redirect are the durable signals. Once AI compute is coded as national security spending, it's politically durable across administrations. The government is now the marginal AI infrastructure buyer at the exact moment private buildout is stalling.

The efficiency research landing this week is the counter-current. KV cache compression in polar coordinates: 2-bit quantization claiming 99% accuracy retention and 8x memory reduction. Self-distillation lifting a 7B model to 60.4% on HumanEval — matching models 10x its size. Mercury Edit 2 claiming 10x code generation speedup via diffusion. Treat the 10x claim as marketing until an independent group reproduces it. Treat the KV cache result as worth a benchmark on your actual workload this sprint, because if the accuracy claim holds up on long-context serving it changes the GPU math directly.

The agent surface is now a security surface

Anthropic shipped Claude desktop control — full user session privileges, inherited credentials, browser cookies, the works. Cursor 3 shipped an agent-first IDE with parallel autonomous code execution on developer workstations. MetaClaw reads Google Calendar during meetings. Default OAuth consent in most Workspace and M365 tenants lets users grant all of this without admin approval. Google DeepMind published a taxonomy of six specific traps that hijack autonomous agents — the first structured red-team framework for the class.

Add today's separate finding that reasoning models decide tool selection in their first few tokens, before the chain of thought begins, and the security picture sharpens: an attacker who controls the first ten tokens of a prompt controls the tool call, regardless of what the visible reasoning trace says afterward. Your EDR was built for malware, not an AI assistant with your employee's credentials clicking through their inbox.

And Jay Edelson — the litigator who made Facebook pay — is filing chatbot-specific lawsuits now, before case law exists and while public sympathy runs against the industry. Every anthropomorphization decision, every guardrail your team shipped or skipped, is discovery material.

What to do this week

One action, specific and time-boxed: audit every path your systems take to a hosted LLM and verify that swapping provider is a config change, not a code change. Pick your two heaviest-traffic AI features and prove the swap works against Gemma 4 or an OpenAI equivalent by Friday. Same prompts, same tools, same evals. If any pipeline fails the swap test, that pipeline is a single-vendor dependency masquerading as an integration — and the last seven days showed exactly what that costs.

Do that first. The transformer supply crisis, the litigation wave, the agent access surface — all of it matters, and none of it matters more than proving your production stack survives the next Tuesday one of your vendors decides to change the rules.