AI labs' record lobbying quarter is your vendor risk

Q2 federal lobbying disclosures landed this week, and the AI labs set records. Anthropic spent $1.97 million. OpenAI spent $1.2 million. Both are the largest quarters either company has posted. That's a Washington story on its face, but it's a procurement story underneath: the model you build on now has a regulatory dependency, and you don't get a vote in it.

What actually happened

Per CNBC, the two labs combined to spend $3.17 million from April through June. Anthropic's $1.97 million was up 26% quarter over quarter; OpenAI's $1.2 million was up roughly 18%. Anthropic has now spent more than $3.5 million in the first half of 2026 alone — more than the $3.1 million it reported for all of 2025 — and per Bloomberg Government, it's running an expanding in-house team plus nine outside K Street firms.

The filed issues are the tell: cybersecurity, copyright, cloud computing, and defense contracts. Axios reports Anthropic also worked export controls and AI safety standards, meeting with both chambers of Congress and with the White House, Commerce, and Treasury. Note the timing — this ramp follows the Commerce Department forcing Anthropic's flagship models offline over national security concerns, a decision that hit paying API customers, not policy staff.

Why AI policy risk is vendor risk for your business

Here's the part operators skip. You already run vendor risk on the boring dimensions — uptime, pricing, whether the company survives the year. Add a fourth: can a government make this model stop existing for you? In 2026 the answer has been yes, more than once, with days of notice.

That's not an argument to avoid frontier models. We use them daily and they're worth it. It's an argument about how you wire them in. Concretely: put every model call behind one internal interface, not scattered SDK imports across forty files. Keep prompts and evals in your repo, not in a vendor's console. Know the second-best model for each job and roughly what it costs, so switching is a config change and an afternoon of eval runs — not a rebuild.

The labs are spending millions to shape the rules because the rules move their revenue. Your version of that leverage is cheaper: build so that when a rule moves, you change a string.

Sources: CNBC, Axios, Bloomberg Government.