The short version
- The thesis. Capability got cheaper and easier to download. Permission to power it got more expensive. Voters favored data center limits in 40 of 41 local measures, governors-elect in Texas, Ohio and Georgia promised pauses, and Texas's outgoing governor had already frozen permits.
- The market noticed. The Nasdaq 100 fell 6.4 percent on Thursday. The election was one stated cause. A weak Qualcomm forecast and pre-earnings nerves were others.
- The marks. OpenAI got its first $11 billion at $1.25 trillion, below its $1.4 trillion ask. Anthropic's unpriced IPO is reportedly being discussed 35 to 45 percent below early October. Money still flowed to defense, physical AI and software with revenue.
- Takeoff. METR found AI is speeding up AI research at Anthropic, by about 1.9 times with a wide range. It makes no takeoff forecast. My confidence in a fast takeoff within one to two years is low.
- What to do. Founders: know where your capacity sits and who carries consent risk. LPs: ask your managers how they mark late-stage AI after this week and what their distributions assume about the IPO window. Investors: add local approvals and grid access to diligence, and stop lumping chip vendors with single-site developers.
- Dates. Cerebras reports November 12, Nvidia November 17, and the Texas grid audit should finish in December.
Two weeks before the election, Gina Hinojosa posted a short video on X with one promise in it. If she became governor of Texas, she said, she would issue a moratorium on all data center approvals until the legislature passed laws to protect Texans. A moratorium is a legal pause. The paperwork stops moving.
She won, and the man she beat was the sitting governor.
Wall Street took two days to work out what it had heard. On Thursday, November 5, chip makers sold off hard. The Nasdaq 100, an index of the exchange's biggest companies, fell about 6.4 percent, and Nvidia dropped roughly 9 percent at one point in the session.
On Friday, OpenAI received the first $11 billion of a funding round at a $1.25 trillion valuation, according to Bloomberg's reporting. Those terms had leaked a week earlier, below the $1.4 trillion it hoped for, so the vote did not set that price. It is the baseline the rest of the week got measured against. People preparing to buy Anthropic stock in its IPO were reportedly talking about $1.1 to $1.3 trillion, down from about $2 trillion a month earlier.
Now look at the same week from the other side. A Chinese lab, Moonshot, put the weights of a 2.8 trillion parameter model on Hugging Face. Weights are the numbers that make up a trained model, and anyone who has them can run the model without asking its maker. An American lab, Thinking Machines, promised to release a model under the Apache 2.0 license, which lets anyone use and modify it commercially, after it scored 81.7 percent on a reasoning test built to resist memorization, at 27 cents a task.
Put the two stories next to each other. The smartest software in the world got cheaper and easier to get. The permission to plug in the machines that run it got harder to get and more expensive. Money pays attention to which input is scarce, and this week it drifted away from anyone who needs a county board's yes and toward anyone who can ship without one.
We wrote last night about Anthropic's amended filing and METR's report on how fast AI is speeding up AI research, in a longer piece. So this issue picks up only what is new, and starts with the votes.
The votes
On November 3, 41 local ballot measures touched data center development. Voters favored limits in 40 of them.
The exception was Harwood, North Dakota, where voters turned down an initiative that would have required review of certain developments, data centers among them. Harwood will have to remain a mystery. Whoever lives there should write the book.
Twenty-seven measures that limit data centers passed and thirteen that would have helped developers failed, which makes forty. Of the 27, five were full bans (in California, Florida and Ohio), twelve were partial bans (in Kansas, Maine and Ohio) and three were six-month moratoriums, all in Maine. The rest were zoning rules, voter-approval requirements and financial guarantees. Votes in Clarksville, Arkansas, and Mandan, North Dakota, are still pending in December.
But politicians read it that way, and the bigger votes were for governor. According to Fortune's rundown, Hinojosa in Texas promised a moratorium and an end to subsidies and tax breaks for data centers. Acton in Ohio promised a conditional moratorium that lifts if developers cover utility costs, use union labor and meet environmental standards. Bottoms in Georgia promised a temporary statewide pause while the state studies the effects. Three states with a lot of the country's planned data centers, three versions of the same sentence.
Now the part that surprised me. Greg Abbott, the Republican she beat, had already done much of this. By the Texas Tribune's timeline, in June he ordered data centers to fund the grid infrastructure that serves them. On August 3 he ordered an audit by ERCOT, the operator of most of the Texas grid, and said projects that fail get no connection. On September 14 he told the water board to penalize data centers that don't report their water use. On September 21 he ordered TCEQ, the environmental permitting agency, to halt data center permits until the audits finish, with no end date.
So six weeks before the election, a Republican governor had frozen permits in the state with the most planned data center capacity. The freeze has leaks. In October, the Texas Tribune reported, advocates found some data center permits had been approved anyway, which forced the environmental agency to shut down its online filing system. That is a bipartisan story in the way that matters. Both parties reached the same conclusion from different directions, and they now disagree only about the terms.
The Moonshots podcast titled its episode this week "The Great Collision Between Politics and AI." For once the title undersells it. Collisions are accidents, and this one was announced on ballots.
The most interesting response came from OpenAI. It published a community compact, a document that puts specific promises to host communities in writing. The company signed the Ratepayer Protection Pledge, which commits it to pay the full cost of the electricity its campuses use plus the new infrastructure needed to serve them. It also promised:
- No nondisclosure agreements for public officials, with narrow exceptions for security and trade secrets.
- Financial assurances before major spending, so a town isn't left holding a stranded asset (a half-built grid connection to a campus that never opens).
- Public reporting of its water and electricity use.
- Up to 1,000 megawatts of flexible demand response at its Effingham, Georgia site, meaning it agrees to draw less power when the grid is stressed.
Then came the dollar figures. For Effingham County, $80 million, including $4.8 million for free school meals for 14,000 students and $2 million for a park. For Pike County, Ohio, $40 million, including a $6 million childcare commitment made with SB Energy. The school lunch line works out to about $340 a student.
I haven't seen a lab put prices on local consent, site by site, before. Read it as an opening offer in a negotiation that the ballots just made more expensive. And read the fine print. The compact covers only campuses OpenAI owns and operates. It says nothing about capacity the company leases, and plenty of the industry's capacity is built by someone else and rented out.
Once consent has a price list, a market has something to work with. It started on Thursday.
How a county board reached the Nasdaq
A vote in Harwood cannot move the Nasdaq. A pause in Texas can, and the chain between them is short.
Ballots pass limits. Governors promise pauses. Texas, the state with the most planned data center capacity, elects a governor who promised a moratorium. Wells Fargo puts Texas at about 27 percent of planned U.S. data center gigawatts, which is a lot of the building that Big Tech's spending plans turn into. Those plans are what chip companies sell into. So a vote in Texas shows up in Nvidia's future orders, and the people who price Nvidia know it.
They priced it on Thursday. The Nasdaq 100 fell 6.36 percent, its worst day since March 2020. The broader Nasdaq and the S&P 500 had their worst days since April 2025, and about $2.5 trillion of market value disappeared. Reuters put the cause plainly: chip makers sank on fears of tougher AI regulation after the Democratic sweep. Nvidia dropped about 9 percent at one point, and the Philadelphia semiconductor index, which tracks the whole industry, reportedly fell around 11 percent.
Wells Fargo's sensitivity, as CNN relayed it, is that every 10 percent cut to planned Texas capacity trims Big Tech's AI infrastructure spending by about 1.6 percent. A 30 percent cut in Texas would therefore take roughly 5 percent off spending. The Nasdaq 100 fell more than that in one afternoon. If the model is even close to right, the market was pricing more than Texas.
My reading is that it priced a template. Ohio and Georgia have governors-elect promising their own versions of a pause, and Goldman has reportedly estimated that short-term moratoria could touch about 45 percent of planned U.S. capacity. Once investors can imagine the Texas move repeating, the number to fear is the bigger one.
I should also give the other explanations their day in court. Qualcomm fell about 7 percent on Wednesday after a weak forecast it tied to a memory shortage. Nvidia reports earnings on November 17, and people sell ahead of reports. The Nasdaq had run hard for months. Two causes were stated out loud on the same news cycle, and cutting off one of them to tell a cleaner story would be bad analysis.
Who is hurt most when a pause arrives? Not the biggest buyers, since a company with data centers in a dozen states can move spending to the ones that still say yes. This is my inference rather than anything reported, but the exposed holders look like three layers: developers with one big site stuck in a Texas queue, the lenders who financed them, and GPU lessors, the companies that rent out Nvidia chips by the hour and now hold chips that need somewhere to plug in.
The macro backdrop did not help. October payrolls rose by 165,000, more than three times the 48,000 monthly average of the prior twelve months, with unemployment about 4.1 percent. A hot labor market makes rate cuts harder to justify, and growth stocks hate rates that stay up. The Fed held at 3.75 to 4.00 percent on October 28 and meets again December 8 and 9.
The marks
A public stock reprices every second. A private company reprices when someone writes a check, which makes private prices slower, stickier and more revealing. This week several of them moved at once.
OpenAI. The company received an $11 billion first close at a $1.25 trillion valuation, according to Bloomberg's reporting on Friday. A first close means the first investors have funded while the round stays open. The target had been $30 billion or more at $1.4 trillion, cut back amid an AI stock sell-off, and the round may still reach $20 billion. The Wall Street Journal reported these terms on October 30, before the vote, so treat this mark as a baseline and not a result of the election. Against the old ask, the price is about 11 percent lower and the check is well under half the size. The size is the quieter signal, because it says how many buyers showed up at that price.
Anthropic. Handle this one with the most care. Its IPO is not priced, and the amended filing we covered last night still leaves the price range blank. Bloomberg reported on November 6 that people preparing to buy are discussing $1.2 to $1.3 trillion, some as low as $1.1 trillion, down from the $1.4 to $1.5 trillion reportedly given to the banks and about $2 trillion in early October. That is 35 to 45 percent below the October figure. Reuters reported that investors doubted a roadshow could start on November 9. Chatter before a price range is chatter.
Cerebras. The chip company's remaining lock-up releases on Monday, November 9. A lock-up is the period after an IPO when insiders agree not to sell. Third-quarter results follow after the close on Thursday, November 12, according to the company's investor relations page. The last filed numbers are second-quarter core revenue of $210 million against $25.4 billion of remaining performance obligations, meaning contracted revenue not yet delivered. That is about 121 quarters of revenue in backlog, and delivering it needs power and permits.
If you want a one-sentence version, here it is. The week did not mark down artificial intelligence. It marked down the part of AI that has to ask a county for permission.
Where the money still said yes
Checks kept clearing all week, and the pattern in who got them is the interesting part. Everything below comes from trade press reports of the deals, not term sheets, so read the figures as reported.
Defense got the biggest check. Shield (formerly Shield AI) is reportedly buying Echodyne, a radar maker, for $5.3 billion. It raised $2.5 billion to pay for it: a $1 billion Series H at a $28.8 billion post-money valuation, a $1 billion structured-equity sleeve from Blackstone funds, and $500 million of debt led by JPMorgan. Structured equity sits between debt and ordinary stock, usually with downside protection in exchange for less upside. Closing is expected in early 2027, and the financing falls away if the deal does. Shield's March round priced it at $12.7 billion, so the valuation more than doubled in about eight months.
Physical AI kept its audience. Zipline, the drone delivery company, reportedly raised $1.5 billion at $21.5 billion. CesiumAstro raised a $730 million Series D at more than $4.5 billion, Northwood Space a $375 million Series C and Harbinger a $240 million Series D. Construction robotics has reportedly taken in about $1.6 billion in 2026 against $389 million in all of 2025, mostly because of one company, Bedrock, so I wouldn't build a thesis on that total.
Software with revenue got paid. Applied Compute reportedly raised $350 million at $3.25 billion from Elad Gil and Nvidia and says its annual recurring revenue passed $100 million after 18 months, which is the company's own claim. fal, which runs infrastructure for image and video models, reportedly raised a $1.3 billion Series F at $16 billion, double its April valuation.
Line them up and a pattern shows. The money went to companies that sell a finished thing to a customer who already exists: a government, a warehouse, an enterprise with a security problem. In the rounds I could find, it went less eagerly to anyone whose business first requires building a large facility and waiting for a regulator.
A reasonable person can read the same list as a bubble finding new hiding places, since several of these valuations jumped fast. That reading deserves its hearing. My answer is that bubbles tend to be indiscriminate, and this week's money was making distinctions.
Capability you can download
While one half of the industry was arguing with county boards, the other half kept giving away the thing the boards were arguing about.
On Hugging Face, the site where models get published, Moonshot posted Kimi K3.1 Pro. It is a mixture-of-experts model, meaning a very large network split into specialist sub-networks of which only a few activate for each piece of text. It has 2.8 trillion parameters in total and 104 billion active at a time, so it costs roughly what a 104 billion parameter model costs to run. The 4-bit weights come to about 1.5 terabytes by my arithmetic, so this is not a weekend download on hotel Wi-Fi.
The benchmark scores on the model card are Moonshot's own, so grade them accordingly. Artificial Analysis lists the model as open weights with commercial use allowed under restrictions, and its license is custom. Open weights means you can download and run the model, not that you can do anything you like with it, so read the license before the leaderboard.
ARC-AGI is a puzzle test built to resist memorization, since each task is a small novel reasoning problem. DeepSeek's V4.1 Flash scored 78.1 percent on ARC-AGI-2 on November 2. On November 6, Thinking Machines' Inkling 2 Small scored 81.7 percent on the same test at $0.27 per task. Its weights are only promised, under Apache 2.0, "next week," so hold that loosely until the files exist. Artificial Analysis also found it uses about 1.7 times the median output tokens of its peers, so a cheap price per task can hide a talkative model.
Now the other column of the ledger. On the newest and hardest version, ARC-AGI-3, DeepSeek's model scored 5.8 percent. Same lab, same week, 78 percent on one test and under 6 on the next. That gap is the most useful number in this section, and I'll use it in a few pages.
What does downloadable capability do to the thesis? It moves the scarce input. When a frontier-quality model can be copied, the interesting question is no longer who trained it. It becomes who has the power, the chips and the permission to run it, and who can run it in a place that still says yes. Weights can cross a border in an afternoon. A 500 megawatt campus cannot.
The takeoff question
One theory going around this week is that we are a year or two from a fast takeoff, the point where AI improves itself faster than anyone can plan around. METR's new report is the evidence people cite. Last night's post covers its numbers, so this section adds only what changes the conclusion.
The verdict, up front. The report is serious evidence that AI is speeding up AI research. It does not support a dated takeoff, and its authors make no timeline or takeoff forecast. My confidence in a fast takeoff within one to two years is low.
What METR found. In Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, METR tracked AECI, Anthropic's version of the Epoch Capabilities Index, a composite score of model capability. Since October 2025 it has grown about 1.9 times faster than before, with a 95 percent confidence interval of 1.5 to 2.8 times. Robustness checks give 1.1 to 2.9 times, and the same measure across all frontier developers gives 1.4 times. METR attributes about 80 percent of the acceleration to AI, with substantial uncertainty, and says growth in training compute is not the driver. It saw no convincing bottlenecks over the next twelve months.
The finding I trust most is the least exciting. Staff polls on Slack produced a self-reported speedup of about 12 times, while interviews centered near 2 times. When people feel twelve times faster and a careful measurement says two, believe the measurement.
What limits it. Anthropic redacted most of the most informative evidence. A check on how the company allocated compute was canceled. METR had no query access to unreleased models, and much of the analysis relied on Claude, so the instrument under study was also an instrument of the study. The strongest evidence is the evidence outsiders can't inspect.
The companion review. METR also reviewed the R&D section of Anthropic's August Risk Report, and the details are sharper. The public report said the speedup was "less than a factor of 2," and 2 is the company's early-warning threshold. The private numbers imply about 1.97, which METR calls around 2. A redacted leading-indicator analysis gave 1.3 to 2.8. METR says the public report overstated its case that the threshold had not been crossed, and calls it plausible that it was crossed by the July 15 model date.
The case for a short timeline has three parts.
- If METR is right that the 2x tripwire was plausibly crossed in July, the speedup is already here.
- Cheap capability is spreading. Open models are near 80 percent on ARC-AGI-2 at cents per task.
- Linux 7.4's merge window closed with a record 19,671 commits, about 21 percent above the previous record, and one tally has "Assisted-by" commits up 2.4 times. Maintainers report being overwhelmed by AI-generated patches. A commit count measures activity, not value.
The case against has four.
- The open model that scores 78 percent on ARC-AGI-2 scores 5.8 percent on ARC-AGI-3.
- A SemiAnalysis index called GDPx reportedly put the best model at 12 percent. I got that second-hand through a newsletter, so weight it lightly.
- The most important evidence is redacted.
- Acceleration is a rate, and takeoff is a rate that keeps rising. METR compares two periods and, as far as I can tell, doesn't show the rate climbing quarter after quarter, which a takeoff would need.
A wrinkle for my own thesis. METR says training compute isn't driving the acceleration. A faster research loop runs on software engineering, which a permit pause can't touch. The thesis of this issue applies to the next generation of training runs, not to the speedup METR measured. A fast takeoff would be a software story paid for with hardware, and the hardware just got harder to site.
Where I land, with confidence levels:
- High: capability on well-specified tasks is getting cheaper and more portable.
- Moderate: AI is speeding up AI research at the leading lab by somewhere around 1.5 to 2.8 times. The interval is wide and part of the evidence is hidden.
- Unknown: whether that rate is still climbing.
- Low: a fast takeoff within one to two years.
What would move me: an independent measurement above 3x, the rate rising across consecutive quarters, an open model moving ARC-AGI-3 off the floor, or access to the redacted data. Any one would shift my view, and two or more would shift it a lot.
Three combinations the market is missing
Each item above got covered as its own story. The interesting things happen in pairs. Here are three, each my inference, each with a way to be wrong.
1. Downloadable models plus immovable campuses. Kimi's weights can be copied anywhere in an afternoon. A campus waiting behind a Texas grid audit can't be moved. So the scarce asset is a site that is already permitted and energized, and operating campuses should trade at a premium to announced ones. I'd be wrong if Nvidia's November 17 guidance shows buyers unfazed, or if the ERCOT audit, which Houston Public Media reports should finish in December, clears most projects quickly.
2. OpenAI's compact plus leased capacity. The compact prices consent, but only for campuses OpenAI owns and operates. My guess is that the cost of consent flows down the contract chain, so leases between developers and tenants start carrying ratepayer and community terms, and owner-operators look politically safer than renters. I'd be wrong if nobody else publishes anything similar within a few months.
3. A backlog plus a siting risk. Cerebras reported $25.4 billion of remaining performance obligations against $210 million of quarterly core revenue. Delivering that backlog depends on customers having sites that are powered and permitted, which this week made harder to assume, so backlog quality now has a permit component the accounting doesn't show. Results on November 12 are the first check. If management says delivery timing is unchanged, I'm wrong, at least for now.
For founders
If you sell a finished thing to a customer who already exists, the week was mostly a tailwind. Don't relax anyway. Investors who just watched a $2.5 trillion afternoon will ask harder questions about anything with a physical dependency.
The questions I'd be ready to answer in your next investor meeting:
- Where does your capacity physically sit? If you rent compute, which states are your provider's sites in, and what happens to your price if one pauses permits for six months?
- Who carries the consent risk in your contracts? Check whether your vendor can pass rising approval costs on to you.
- What is your fallback if a model provider raises prices? Have a second provider tested, not just named in a slide.
- Did you read the license? Apache 2.0 and a custom license with restrictions, like Kimi K3.1 Pro's, are different things.
- Does your financing depend on a date you don't control? Shield's evaporates if its acquisition doesn't close. If yours has similar conditions, say so before an investor finds them.
The founders who will look smart in a year are the ones who treated permission as a line item in their model back in November.
For Team Ignite limited partners
If you invest in venture funds, the week matters to you in two ways, and the first is slower than the second.
The slow way is marks. Funds value private holdings at their last priced round until something forces a change. The OpenAI first close came in about 11 percent under its asking price, and the talk around Anthropic's IPO is reportedly 35 to 45 percent under early October. Neither is a final mark, and the first is a haircut to an ask rather than a drop from a prior valuation. But funds carrying late-stage AI positions will face questions at quarter-end about whether the last round is still the right number.
The faster way is liquidity. The IPO window many portfolios were counting on is a schedule, and schedules slip. Reuters reported that investors doubted an Anthropic roadshow could start on November 9. Cerebras has a lock-up release Monday and results Thursday, and SpaceX's employee lock-up balance comes off on December 8. If those go badly, the window narrows for the next company in line. If they go well, the sell-off will look like a pause. Nobody selling you a confident view knows which.
Questions worth asking any fund manager you back, ourselves included:
- How do you mark late-stage positions after a week like this one, and when will the next mark reflect it?
- How much of the portfolio depends on a single site, single state or single customer for physical capacity? Concentration in developers and lessors is where a permit pause bites hardest.
- What does the plan for distributions assume about the IPO window? A realistic answer includes a delay case.
- Where in the portfolio is your exposure to companies that sell a finished product to existing customers, compared with companies that first need to build? This week's money drew that line itself.
For other investors
Add consent to the diligence list. For anything with a physical footprint, the questions are now local approvals, interconnection queue position (the line to connect to the grid), who pays for grid upgrades, and whether the project sits in a state with a governor-elect who promised a pause. A year ago those questions lived in a footnote. This week they moved a $2.5 trillion index.
Separate the exposures that got lumped together. Chip vendors, hyperscalers (the giant cloud buyers), owner-operators, lessors and single-site developers don't carry the same risk from a permit pause. A company with sites in a dozen states can reroute. A developer with one large Texas project can't. My inference, flagged as such earlier, is that the pain concentrates in the developers, their lenders and the chip lessors, and I'd want to know which of those a given position resembles before reading a headline index move as relevant to it.
Don't over-read the vote. Forty of 41 is striking, but it is a sample of local ballots from ten states, which tend to bring out motivated voters. It shows direction and intensity. It says little about the national median, and nothing about what a legislature will pass. The sell-off likewise had several causes, so be wary of anyone who gives it one.
The calendar
Dates to keep in front of you over the next few weeks. Items marked "reported" come from press accounts I couldn't confirm at the source.
- Monday, November 9: Cerebras's remaining lock-up releases. Also the date the Anthropic roadshow was previously reported to start, which investors reportedly doubted.
- Thursday, November 12: Cerebras reports third-quarter results after the close.
- Tuesday, November 17: Nvidia reports earnings.
- Next week, unconfirmed: Thinking Machines has promised Inkling 2 Small weights under Apache 2.0.
- December, date open: the ERCOT audit of Texas data center projects should finish. Local votes are pending in Clarksville, Arkansas, and Mandan, North Dakota.
- December 8: SpaceX's employee lock-up balance releases, and the Fed meets December 8 and 9.
- January 12 and 19: the Texas legislature convenes, and Hinojosa takes office a week later.
- Early 2027: Shield's Echodyne acquisition is expected to close.
If I had to pick the one date that tells us the most, it's November 17. Nvidia's tone on where its customers plan to build will say whether Thursday's sell-off was a flinch or a verdict.