Washington changed hands overnight, and so did the odds on the AI industry's favorite federal project. You mean the one where Congress tells fifty states to stop writing AI rules. That's the one. It's Wednesday, November 4th, and you're listening to AI Daily Briefing. Coming up: what a Democratic Congress means for preemption of state AI laws, AMD nearly doubles its data center sales, the New York Times reports Anthropic wants to double its compute again next year, and a new benchmark from Ai2 shows exactly where frontier models lose track of state. Hit follow and you won't have to come looking for the next episode.
Luis Rijo, in a special election edition of PPC Land:
What the result does is close the path that Republicans were building toward federal preemption of state law, at least for the next two years, and hand Democrats the subpoenas, the committee gavels and the confirmation votes that shape how federal agencies behave.
The results as Rijo reported them early this morning: the Associated Press called the Senate for Democrats late Tuesday after Josh Turek beat Republican Ashley Hinson in Iowa, per Forbes' live coverage. In the House, Democrats were projected to clear the 218 seats needed for a majority, with NBC News projecting as many as 244. Some races were still uncalled when he filed. For anyone shipping AI products, preemption is the word that matters. Rijo lays out how much of the Republican plan needed unified control: the ten-year state moratorium the Senate stripped out ninety-nine to one in July 2025, an executive order telling the Justice Department to sue states over their AI laws, and a White House framework asking Congress to stop states from regulating AI development at all. His conclusion is that the state-by-state patchwork survives, and probably grows. Before anyone overreads the swing, the limits. Trump keeps his veto and the Senate keeps its sixty-vote threshold. Fortune's Jeremy Kahn wrote on election morning that new AI law will still require a bipartisan bill with very broad support. And the Asia Business Daily reports Trump said on a New York Post podcast on Monday that he would be, his words, Mr. Veto. So what Democrats actually win is gavels and subpoenas. Expect hearings, not statutes. Per the Asia Business Daily, citing Fortune, Elizabeth Warren has already asked Dario Amodei to explain the risk factors in Anthropic's IPO filing. Next year she asks from the majority. Watch the committees. Maria Cantwell, the senior Democrat on Senate Commerce, is in line to take that gavel back from Ted Cruz, though Rijo notes assignments won't be settled until December and January. And watch the lame duck: the outgoing Republican Congress returns before the new one convenes on January 3rd, and Rijo calls the next eight weeks unusually important for every bill in play. Data centers were on local ballots too. Fortune counted at least forty-one data center referendum measures yesterday. Practical takeaway: if you sell into healthcare, hiring or anything regulated, don't cut your state compliance budget yet.
The Stock Titan editorial team, breaking down AMD's third-quarter numbers:
AMD's data center sales grew 97% from a year earlier in the third quarter of 2026, a $4.2 billion rise bigger than AMD's whole revenue gain.
Data center came to eight and a half billion dollars, sixty-four percent of AMD's total. Overall revenue was thirteen point four billion, up forty-five percent and above the top of AMD's own forecast range for the fifth straight quarter. Fourth-quarter guidance is about seventeen billion, plus or minus three percent. Reuters reported the shares jumped eight percent in extended trading Tuesday. The bigger number came on the call. Lisa Su told analysts she expects total revenue to grow more than eighty percent in 2027, per Reuters, above the seventy-seven percent analysts expected in LSEG data. She also said eighty percent of next year's sales will come from data center chips. For scale, Stock Titan points out that Nvidia's data center sales were eighty-nine billion dollars in its quarter to late July, about ten times AMD's. Nvidia grew faster in all four matched quarters, but the gap narrowed from forty-four percentage points to ten. And the first Helios racks are now running at Anthropic, Meta, Microsoft, OpenAI and Oracle. Two lines I'd circle in red. GAAP earnings carry a loss of roughly one point one billion dollars on warrants AMD handed two customers, so the sweeteners on those mega-deals are hitting the income statement. And free cash flow was five hundred fifty-four million dollars, down from one point five three billion a year ago. The revenue is real. The cash hasn't caught up. Bank of America's Vivek Arya asked the bottleneck question directly: do AMD's customers have the financing, land and power for roughly three gigawatts of planned deployments through 2027? Su said, per Benzinga, that AMD planned the ramp jointly with customers, has good visibility into 2027, and knows which data centers will receive its GPUs. Knowing the address isn't the same as the building having power. But after Intel's weak outlook last week, Reuters framed this as a strong counterpoint to fears the AI boom is losing steam, and the market agreed. This is the part of the story I think gets undercovered. A second supplier with real volume changes what every lab can negotiate on price.
Louis Juricic at Investing.com, summarizing a New York Times report:
Anthropic expects to have more than 3.5 gigawatts of computing capacity by the end of this year and aims to double that figure next year, the New York Times reported Tuesday, citing people familiar with the company’s plans.
Doubling puts Anthropic around seven gigawatts by the end of 2027. The Times says that would take enough electricity to power every home in Chicago. We covered the amended S-1 and its eight hundred ninety-five point four billion dollars in commitments on Tuesday's show. Two new details here: nearly seventy billion of those commitments take effect over the next twelve months, and the disclosed list includes one hundred twenty point five billion dollars with SpaceX, alongside Google, Amazon and Microsoft. SpaceX as a compute supplier at that size is the line that made me sit up. The report also names neoclouds Fluidstack, Lambda and Nscale. And Anthropic is hiring construction and supply-chain specialists and buying equipment like cooling systems for its own data center projects. That's a model lab starting to act like a developer. The 2027 target isn't consistent across outlets. Diego Almada Lopez at Crypto Briefing writes that Anthropic is targeting roughly nine and a half to ten gigawatts by the end of 2027, against the Times' doubling to about seven. The Times cites people familiar with the plans. Crypto Briefing doesn't name a source for its figure, so treat the target as a moving range. Crypto Briefing also puts Anthropic's third-quarter net loss at twenty-three point three billion dollars, against nearly seventeen billion in revenue. Put that next to the AMD segment. When a chip CEO says she has visibility into 2027, this is what she's looking at: labs signing for gigawatts years ahead, funded by losses that big. Compute plans are now measured in cities.
The research team at Ai2, the Allen Institute for AI, introducing a new benchmark called StateBench:
All four performed well on r-trivial and aperiodic problems. But on soluble and general finite problems, accuracy fell sharply as inputs grew longer, dropping to 13% or lower on the longest general finite problems.
The four are Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna. Each read a long sequence of variable assignments and had to give the final value of every variable. The levels climb from values that never change, to plain reassignment, to self-referential updates like x equals x plus five, to assignment by reference, like swapping two variables. This is my whole job. Long agent traces, config files, a variable reassigned forty times across a log. A needle-in-a-haystack test tells you a model can find the latest value. It doesn't tell you whether it can follow a chain of swaps, and that's exactly where these models fall apart. The architecture result is the more interesting part. Ai2 trained small models of two hundred seventy-five million parameters from scratch on StateBench. Gated DeltaNet and hybrid models generalized almost perfectly on the first three levels out to tens of thousands of assignments. Transformers degraded on longer sequences and never learned the hardest level, even when trained directly on it. One more finding I like: the small hybrid models scored near the ceiling on RULER-style problems but near chance on the same problems framed as spreadsheet edits. Format alone moved the score that much. Ai2's own caveat is that the benchmark assumes a fixed number of variables, while real conversations keep introducing new entities. This beats a leaderboard. The method is published, the levels come from theory, and it tells you where a model breaks, not just where it ranks.
If the AMD and Anthropic segments were your favorite part of today, our sister show The Data Center Daily covers that ground every day. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and the energy markets being reshaped by intelligence at scale. Look for it in the same app you're using right now. Gigawatts, every morning. Somebody has to count them.
One loose end from yesterday. Beijing Post reports Moonshot has released an upgraded Kimi model aimed at multi-step agent tasks, but the report names no version and says nothing about weights. So the Kimi K3.1 Pro download question is still open. We'll check on it tomorrow, Thursday, November 5th. AI Daily Briefing is a Lantern Podcast. Thanks for spending your morning with us.