The Lead, the Grant, and the Moratorium

00:00:04 Anthropic published a paper this morning on how the United States should handle artificial intelligence competition with China. It's called "2028: Two scenarios for global AI leadership," and the structure is right there in the title — two possible futures, three years out, and a set of decisions Anthropic argues will pick which one you land in.

00:00:23 Here's the first scenario. American models hold a twelve-to-twenty-four-month lead over Chinese models, the United States and its democratic allies set the global norms, and the advantage compounds as talent and infrastructure keep flowing toward the leader. Here's the second.

00:00:39 Chinese labs close to near-parity, only months behind. The Chinese government deploys AI fast across its economy and military, China takes the lower-cost segments of the global market, and authoritarian norms spread into developing economies that buy the cheaper option.

00:00:54 The paper leans on a handful of numbers. It estimates the United States has access to roughly eleven times the compute China does. Huawei's chip production this year, the paper claims, amounts to about four percent of Nvidia's output, dropping to two percent next year.

00:01:10 Only three of thirteen leading Chinese labs have published safety evaluations. And it cites a test in which DeepSeek's model complied with ninety-four percent of malicious requests, against eight percent for US models. From there, three recommendations. Close the loopholes in export controls that let restricted chips reach Chinese firms.

00:01:29 Get Congress to clarify that distillation attacks — thousands of fake accounts harvesting a US model's outputs to train a cheaper copy — are illegal. And push American AI into global markets so democratic infrastructure becomes the default people buy. Walk what that first recommendation actually means in practice, because "close the loopholes" sounds clean and isn't.

00:01:50 Enforcing export controls at this stage means policing cloud access — a chip that can't be shipped to Shenzhen can still be rented by the hour from a data center in a third country. It means chasing shell importers and transshipment routes through Southeast Asia and the Gulf.

00:02:05 And it means the US government leaning harder on Nvidia and the big cloud providers to know exactly who their customers are, which cuts against those companies' revenue. So even inside the friendly reading of this — the version where you fully accept that Chinese access to compute is a real problem — the policy still lands as cost and friction on American firms.

00:02:26 And that's worth noticing: the proposal asks the government to do something the most powerful companies in the sector have every incentive to resist. Anthropic's own language is blunt. Export controls, the paper says, have been incredibly successful, but, quote, "labs in China have remained close by exploiting loopholes in US export control policies." And on the stakes, quote: "The CCP uses AI systems to censor speech, enforce draconian policies on ethnic minorities, and hack major corporations and government agencies."

00:03:00 An account posting as aipulsedaily put it this way: a frontier AI lab publishing a policy paper recommending governments protect frontier AI labs from competition is, quote, "peak regulatory capture." Sahil Panhotra, with a couple hundred likes on his reply, went further.

00:03:16 The paper frames this as democracy versus the Chinese Communist Party, he said, but the real divide is open source versus Anthropic — and if Anthropic cared about democratic AI, it would open its own model weights. And Alpár Kertész landed the sharpest version of it: "The hard part is separating security from incumbency.

00:03:34 If that line is blurry, even a serious safety argument starts sounding like a market argument." The distillation problem isn't hype. Anthropic, OpenAI, and Google have all said publicly that systematic output-harvesting is happening at scale, and one reply in the thread, from a user named Manoj, flagged distillation as the thing most readers skip past.

00:03:58 A serious national-security concern can be true. And the same paper that names that concern also recommends the government treat a private company's market lead as a strategic asset worth defending with new law. The paper doesn't really sit with the tension between those two things.

00:04:14 That's the tension a reader has to carry alone — and carrying it is the point, because Anthropic has spent the last year positioning itself as a trusted advisor to the US government, and a policy paper is how that relationship gets built.

00:04:27 The China paper wasn't Anthropic's only move today. Earlier in the day, the same company announced a four-year partnership with the Gates Foundation worth two hundred million dollars. The money breaks into three parts: grant funding, Claude usage credits, and technical support.

00:04:43 It's aimed at five areas — global health, life sciences, education, agriculture, and economic mobility. The largest piece goes to health in low- and middle-income countries, where, by the foundation's count, about 4.6 billion people lack access to essential health services.

00:04:59 The named targets are specific — polio, HPV, and both eclampsia and preeclampsia. HPV alone causes around three hundred fifty thousand deaths a year, and ninety percent of those are in low- and middle-income countries. Part of the work is using AI to screen potential vaccine candidates in software, before they go into pre-clinical development, to shorten the early timeline.

00:05:21 On the education side, Claude would power tutoring tools for K-through-twelve students in the US, in sub-Saharan Africa, and in India. On agriculture, Anthropic says it will make farming-specific improvements to the model, aimed at the roughly two billion people whose income depends on smallholder farming.

00:05:39 Taken on its own terms, that's a serious commitment to places the AI industry mostly doesn't build for. One reply caught that — everyone is racing to build AI for Wall Street and Silicon Valley, the writer said, and Anthropic just bet two hundred million dollars that the important use cases are in Africa and Southeast Asia.

00:05:57 But the skeptical replies are the ones I'd actually keep. A user posting as ColdBoot put it cleanly: "AI access is easy to announce. Deployment quality is harder to prove." Another, Patrick, asked which of the five areas hits a model ceiling first — agriculture data, he pointed out, is ninety percent offline and fragmented.

00:06:16 And Alper Ferudun named the mechanism: a commitment like this becomes distribution infrastructure only if the operating layer works, including, quote, "metered inference for workflows that repeat after grants expire." There are two fair ways to read this deal. One is straightforward philanthropy — get advanced tools into the hands of people working on polio and literacy, and good things follow.

00:06:41 The other is that grant-plus-credits is also how you seed dependency. You get health ministries and school systems building their workflows on Claude. The grant period ends. The credits run out. The workflows stay, and now they're metered. That's not an accusation.

00:06:56 It's just how distribution works — a development partnership and a customer-acquisition strategy can be the same set of activities. And put the two announcements side by side, because they happened within hours of each other. In one day, Anthropic positioned itself as a geopolitical advisor to the United States government and as humanitarian infrastructure for the global south.

00:07:18 Those are very different rooms. But they enlarge the same thing — the company's standing as something closer to a public institution than a software vendor. I'm not saying that's sinister. I'm saying it's a deliberate posture, and on a day when you see both hands, you should notice the shape of the body behind them.

00:07:36 Anthropic spent the day widening its institutional footprint. OpenAI, the same day, was in the opposite posture — defending. A class action was filed today, May fourteenth, in the United States District Court for the Southern District of California. The case is Couture versus OpenAI.

00:07:53 The claim, in one sentence: OpenAI embedded Meta's Facebook Pixel and Google Analytics into the ChatGPT website, and in doing so, turned private chatbot conversations into tracking data for the advertising business. Let me walk the specifics, because they matter.

00:08:09 The named plaintiff is a California resident, Amargo Couture, represented by the class-action firm Bursor and Fisher. The complaint alleges violations of three things: the federal Electronic Communications Privacy Act, California's Invasion of Privacy Act, and the privacy provisions of the California constitution.

00:08:28 The mechanism the suit describes is the ordinary ad-tech tracking stack. The Facebook Pixel, embedded in ChatGPT's web pages, allegedly fires real-time requests to Facebook's servers every time a user interacts with the site. Those requests are said to carry content-derived context — the complaint's example is a browser tab title like "Super Bowl 2005 Winner," generated from a user's query — along with cookies that can tie the activity back to a specific Facebook account.

00:08:56 Google Analytics, the suit says, then enriches that with cross-device behavior and demographic signals. The damages math is what draws the eye: California's privacy statute allows five thousand dollars per violation, and across millions of users, the filing puts the potential exposure in the billions.

00:09:15 Two things give this more weight than a single complaint usually carries. First, it isn't the first one. A nearly identical suit was filed against Perplexity in early April — same Pixel mechanism, and that one added the detail that Perplexity's "Incognito" mode didn't stop the data flow.

00:09:32 So this is a pattern of litigation, not a one-off. Second, and this is what reaches furthest: the legal theory rests on treating a chatbot session as a protected communication under wiretap law. If a court accepts that framing, the exposure isn't OpenAI's alone.

00:09:48 Every AI company running standard consumer web analytics on a chat product is suddenly inside the same statute. There's a wider context here. This Pixel-tracking theory isn't new. Over the last two years, plaintiffs' firms have run this playbook against hospital websites, tax-preparation services, and video platforms — arguing that embedding Meta's tracking code on a sensitive page amounts to an illegal wiretap.

00:10:13 Some of those cases settled for real money. What's new is pointing the theory at a chatbot, where the sensitive page isn't a page at all. It's a conversation. Discovery, if the case gets that far, would test the specifics: what the Pixel actually transmitted, whether OpenAI configured it to exclude query content, whether any of this was disclosed in a privacy policy a user could reasonably find.

00:10:36 None of that is settled. But the firms filing these suits clearly think a chatbot is the richest target the playbook has found yet. Here's why I think this lands outside the AI-industry bubble. People type things into ChatGPT they wouldn't say to a friend — questions about a diagnosis, a debt, a divorce, a legal problem.

00:10:56 The whole reason the product is useful is that it feels like a private channel. The complaint's allegation is that the most generic, default piece of web infrastructure — the same Pixel that's on half the internet — was bolted onto that channel. Whether the facts hold up is for discovery to decide, and I wouldn't pre-judge it.

00:11:15 But we've now been watching OpenAI defend on two separate legal fronts — the wrongful-death suit we've tracked, and now this. Different harms and different statutes, but the same company. The courts are starting to treat the chatbot as something other than a website, and nobody — not the companies or the regulators — has settled what it actually is.

00:11:37 Garry Tan, the president of Y Combinator, posted something today that Elon Musk then amplified to roughly eleven million views with a single word: "Hmm." More than three hundred local bills have been filed. Half of the data centers planned for 2026 are facing delays or cancellation.

00:12:03 And Tan's framing was about jobs — each one of those facilities, he wrote, brings billions to local economies, and the people who say they want American jobs are trying to block the buildout. Let me fill in around that, because the post is a position, and the situation underneath it is more textured.

00:12:22 The Sanders and Ocasio-Cortez bill — the AI Data Center Moratorium Act — would halt construction or upgrades of any data center drawing twenty megawatts or more, until national safeguards are in place. Set the federal bill aside for a moment, because the local picture is moving on its own.

00:12:40 Industry reporting has roughly thirty to fifty percent of 2026 capacity already delayed or canceled, with only about four gigawatts of a planned twelve actually under construction. Just yesterday, a Dallas exurb — Hill County, Texas — put a one-year ban on data centers, which Bloomberg described as a rare Texas rebuke.

00:13:00 And Musk's own xAI has been in this exact fight in Memphis, where the Environmental Protection Agency closed the loophole that had let the company run something like thirty unpermitted gas turbines to power its first facility. So there are two stories colliding here.

00:13:16 Tan's story is the jobs-and-growth story, and it isn't wrong on its own terms — these are large capital projects with real local payrolls. The other story is the one the organizers are telling: data centers of this size draw enormous amounts of power and water, and in a lot of these towns, it's the ratepayers — ordinary households — who end up subsidizing the industrial load through higher utility bills.

00:13:42 That's who's organizing those three hundred local bills. It isn't abstract opposition to AI. It's people looking at their power bill. And the economics there aren't abstract either. A single large data center campus can draw as much electricity as a mid-sized city, and in most US grids the cost of new transmission and generation gets spread across every customer on the system — so a household three towns over can watch its bill climb to serve a hyperscaler's load it will never use.

00:14:12 Water is the other ledger. A lot of these facilities sit in places already short on it, and the cooling demand is constant. None of that makes the buildout wrong. It makes it contested, in a precise way: the benefits — construction jobs, an expanded tax base — and the costs — power, water, ratepayer exposure — land on different people.

00:14:33 That gap is the fight, and right now it's being had one county at a time. A continuity note. We talked about Senator Sanders just a few days ago, when he picked up the phone on a US-China AI treaty. Same senator, different lever — that was foreign policy, and this is domestic siting.

00:14:50 And here's the through-line I'd draw, lightly, between this chapter and the first one. The Anthropic paper says the United States must protect its compute advantage. But compute isn't an abstraction. It's buildings, substations, and water rights, and it has to physically land in someone's county.

00:15:09 The paper argues about the lead at thirty thousand feet. This fight is what it looks like at ground level — in a county commission meeting, where the people who live next to "somewhere" get a vote. Right now the buildout still has the leverage. But three hundred local bills is the first real friction it has hit, and friction compounds.

00:15:30 Susan Zhang — an AI researcher who worked on Meta's OPT model — quote-tweeted a piece of news today with four words: "what goes around comes around." The news underneath the jab deserves a careful walk, with a heavy caveat I'll get to. The claim: a threat group calling itself TeamPCP says it breached Mistral AI, the French model lab, and is selling its internal repositories on a criminal forum — roughly five gigabytes of files tied to AI training and inference projects.

00:15:58 This sits downstream of a supply-chain attack that ran through the AI developer ecosystem earlier this week — a campaign security researchers have been calling "Mini Shai-Hulud," which compromised more than a hundred seventy packages across the main JavaScript and Python registries, npm and PyPI, and published over four hundred malicious versions in a roughly two-day window.

00:16:20 Microsoft is investigating one specific compromised Mistral package on PyPI that runs a hidden second-stage payload the moment it's imported. The whole thing carries a vulnerability rating of nine-point-six out of ten — critical. Now the caveat, and it's a real one.

00:16:36 As of right now, there's no public confirmation that the repositories TeamPCP claims to be selling are authentic. This is a claim on a forum, not a verified breach. So treat everything that follows as conditional. But the scenario matters even as a claim, and here's why.

00:16:52 Mistral isn't just another startup. It's the company the European Union points to when it wants to say it has a sovereign option — a frontier lab that's neither American nor Chinese. If a group can breach that lab and put its training and inference code up for sale, that isn't a data breach in the ordinary sense, where you worry about leaked customer emails.

00:17:14 For a model lab, the training pipeline and the inference code are the asset. They're the thing. There's also a regulatory wrinkle. Under European data and security rules, a confirmed breach of this kind would put Mistral on the clock with its own regulators — the same regulators the European Union is counting on to hold the line on AI governance for the whole bloc.

00:17:35 A sovereign-AI champion that can't secure its own source code is a harder thing for Brussels to hold up as the answer. And connect it back to the first chapter. Anthropic's paper wants the US government to criminalize distillation — the patient version of theft, where you create fake accounts and slowly harvest a model's outputs to approximate it.

00:17:56 What TeamPCP is claiming to have done is the crude version. You don't need to distill a model if you can just buy the repository. Zhang's "what goes around comes around" is a needle aimed at Mistral — there's industry history there. What I'd actually hold onto is broader and colder: the security perimeter around frontier labs is porous, and the thing inside that perimeter became strategically valuable faster than the labs hardened the walls.

00:18:23 Whether or not this particular claim checks out, that gap exists, and it's the kind of gap that gets tested again.

00:18:29 Daniel Miessler, a security writer, posted a short thesis this morning that I think holds up even though the specific number in it doesn't survive contact. Here it is, quote: "What if the American companies are walking a tightrope named token subsidies, and beneath them a net awaits their fall.

00:18:46 And that net is Chinese cloud-based opensource models that are one one-hundredth the cost." The claim has two parts. Part one: US frontier inference pricing — what you pay per token to use the big American models — is propped up by subsidy. That subsidy is venture money, loss-leader pricing, and a willingness to run at a loss to hold market share.

00:19:07 Part two: if that pricing ever has to normalize toward actual cost, the floor underneath it is Chinese open-weight models that anyone can download and serve cheaply. The replies pushed back hard on the number, and they were right to.

00:19:21 A user named muhdur said flatly that Chinese models aren't one one-hundredth the cost — he had tried two of the popular subscription plans and burned through them in about four hours, and rated the quality, quote, "2025 November level, just fine." Joseph Thacker, a security researcher, made the practical point: if you're worried about a Chinese inference provider, just use a provider that isn't in China — the weights are open, you can host them anywhere.

00:19:47 So the literal "one one-hundredth" figure overstates the gap. But strip the number out and the structure is still standing, and it connects straight back to the China paper from the top of the episode. Anthropic's paper warns that China wins, quote, "global market share in lower-cost segments." Miessler is just describing the mechanism by which that could happen.

00:20:08 If the price of frontier-adjacent intelligence collapses — because open weights are cheap to download and cheap to serve — then the US closed-model business has to justify its premium on something other than raw capability. It has to justify it on governance, reliability, integration, and the audit trail.

00:20:26 And that's exactly what one reply on the Anthropic thread said in four words: "Governance is part of the moat." What actually decides whether Miessler's net is real is a number nobody outside those labs has: how much of current US inference pricing is margin, and how much is subsidy?

00:20:45 Nobody on the outside knows the answer. Not me, not Miessler, and not the replies. That number is one of the most consequential figures in the industry, and it isn't public. When it does become visible — through an earnings call, a price change, a funding round that reprices — that's the moment to pay attention.

00:21:03 Last one. It didn't make headlines today, and items like that tend to matter later. An analysis circulated today — from Nate B Jones, who runs an AI strategy channel — arguing that enterprise AI is converging on a single model for how it actually gets deployed.

00:21:18 His framing is four pressures pushing in the same direction. Frontier labs are moving down-stack, out of pure model provision and into productized agents and forward-deployed engineering teams that sit inside a customer's office. Consultancies — McKinsey, BCG, Accenture, Capgemini, and PwC — are moving up-stack, building dedicated agentic practices and retraining their engineers to wire AI into enterprise systems.

00:21:42 The systems-of-record vendors — Salesforce, ServiceNow, Workday, and SAP — are exposing agent frameworks with audit trails built in. And private equity firms are standardizing AI deployment playbooks across the portfolio companies they already own. The claim underneath all four: the value isn't in the base model anymore.

00:22:00 It's in the implementation layer — data access, permissions, evaluations, audits, and who owns the thing after it launches. And the people who get squeezed in that picture are the generic AI startups that are a thin wrapper around someone else's model, with no ownership of the workflow or the governance.

00:22:17 We covered one piece of this a few days ago — OpenAI's Deployment Company, a forward-deployed-engineering play worth four billion dollars across nineteen partners. What today's analysis does is show the wider shape that move sits inside. The world-facing angle here is labor, and it's specific.

00:22:34 Look at one line in that list: consultancies retraining engineers to wire AI into enterprise operating systems. That's hundreds of thousands of consulting and systems-integration jobs being re-pointed — not eliminated yet, re-pointed — toward a new task. And the private-equity piece is the one that gets the least attention and probably deserves the most.

00:22:54 When PE firms use agentic deployment as a standardized efficiency tool across their portfolios, the people who feel it first are the workers at PE-owned companies, where the word "efficiency" has a specific and well-documented history. Step back across the whole day, because the items rhyme more than I expected when I started.

00:23:13 The bottleneck in AI moved — from model capability to deployment, governance, and distribution. And every story today is about an institution moving to own that new layer. Anthropic is working to advise the government and seed the global south. OpenAI is defending the legal boundary of what a chatbot even is.

00:23:31 Local governments are fighting to control where the compute physically lands. Consultancies and private equity are absorbing the deployment market rather than being disrupted by it. The frontier labs didn't flatten the old institutions. They're partnering with them, lobbying them, getting sued by them, and getting blocked by them — one normal institutional fight at a time.

00:23:52 What I'm watching next is the OpenAI suit — whether any court accepts the theory that a chatbot session is a communication under wiretap law. That's the one ruling on this list that would reach every other company at once. The rest of today's stories are pressure.

00:24:07 That one would be precedent. I'm Jonas.