What six months of filings add up to
The reform group Issue One went through the newly filed federal lobbying disclosures for the first half of 2026 and totalled them. Eleven of the largest technology, social media and artificial intelligence companies, together with their two leading trade associations, spent 41 million dollars between January and June. That is more than 226,000 dollars a day, and it is up around 3 million dollars, or 8 percent, on the 38 million the same group spent in the first half of 2025.
The second-quarter breakdown is where the shape of the year shows. Meta reported close to 6 million dollars for April to June, Alphabet 5.3 million, Microsoft around 3 million. Anthropic booked 1.97 million, its largest quarter on record, and OpenAI 1.2 million. Nvidia reported 1.25 million and ByteDance 1.14 million. The trade bodies were smaller: TechNet 440,000 dollars, NetChoice 60,000. Across the six biggest names, 324 registered lobbyists were at work in that quarter alone, which Issue One puts at roughly one lobbyist for every one and a half members of Congress.
Alix Fraser, the group's vice president of advocacy, described the result as an influence campaign that has burrowed deep into Washington. The underlying records are public: every registrant files with the Senate lobbying disclosure system, and the filings name the individual lobbyists and the issue areas. Anthropic's own registration, for instance, lists its lobbying subject in three words, artificial intelligence policy, which tells you how much latitude the disclosure format allows.
The ask is a waiver, not a rulebook
Why it matters: the money is not being spent to write a detailed federal standard. A large part of it sits behind the SANDBOX Act, which would let AI developers apply for exemptions from federal regulations they consider an obstacle to experimentation, provided they agree measures to mitigate consumer and safety risk. Waivers would run two years and could be renewed up to a ten-year total. The White House Office of Science and Technology Policy would administer the scheme and report annually to Congress, and Congress could make any successful waiver permanent.
The second vehicle is blunter. The American Artificial Intelligence Leadership and Uniformity Act would write preemption into statute, imposing a temporary moratorium on state laws that restrict AI models and systems engaged in interstate commerce. The argument behind both is the same one the OSTP director has made in public, that a patchwork of state regulations is anti-innovation and falls hardest on companies too small to navigate fifty regimes. Whatever you make of that case, note what it produces if it wins: fewer things a developer is obliged to measure, document and publish.
Why a US exemption fight lands on a European desk
The bottom line: vendors do not build one model for Washington and another for Brussels. They build one, and the obligations they carry at home shape what exists inside it to be reported on. That is the connection most coverage of these filings misses, because it treats lobbying totals as a story about American politics rather than about the supply chain that ends at your desk.
Your side of that supply chain has a date on it. From 2 August the Commission's enforcement powers over providers of general-purpose AI models come into application. The AI Office moves from persuasion to compulsion: it can request documentation, evaluate models directly, order corrective measures, restrict or withdraw a model from the EU market, and levy fines up to 15 million euros or 3 percent of global turnover. The Article 50 transparency duties land on the same day, covering disclosure that a user is dealing with an AI system, and the marking of synthetic audio, image, video and text.
Read those two things together. Much of what you will be expected to show about a model you did not build has to come from the party that did build it. If that party succeeds in reducing what it is required to generate, the shortfall does not stay in Washington. It arrives as a gap in your file, at the point where a regulator asks you to evidence a claim about a system you licensed.
The clause to add before August
The practical response is contractual and it is small. At the next renewal, name the artefacts you need rather than referring to compliance in general terms: the model documentation, the description of training data sources at the level the Act requires, the notice of substantial modification, and the marking behaviour for generated output. Attach a date to each and a duty to notify you when any of them changes. A vendor that intends to supply the EU market will already be producing most of it; one that hesitates at the clause has told you something useful before you have paid for it.
Do the same thing on the way in. Procurement questions about AI vendors still tend to stop at security and hosting location, which were the right questions three years ago. The question that matters from 2 August is narrower and more awkward: show me the documentation you would give a market surveillance authority. Ask it before signature, when you still have the leverage, and keep the answer with your record of processing rather than in the deal folder.
Read next: Anthropic Wants Every Model Tested, Open or Closed | 25 AI Firms Ask Washington to Back Open Models



