The Gate Moved To Identity

00:00:04 Anthropic published a Claude help-center page this week explaining identity verification for the service. The first sentence is plain enough to quote: "Being responsible with powerful technology starts with knowing who is using it." From there, the page says the company is rolling out verification for a few use cases.

00:00:23 A user might see a prompt when accessing certain capabilities, during routine platform-integrity checks, or as part of safety and compliance measures. That wording matters because it narrows the policy. Anthropic isn't saying every Claude user needs to upload an ID before asking a question.

00:00:40 It's saying some capabilities and some account situations may now require a stronger identity check than email, payment history, or phone verification. The verification partner is Persona Identities. The page says Anthropic selected Persona for its technology, privacy controls, and security safeguards, and it says verification data is used to confirm identity and not for other purposes.

00:01:03 The page leaves several practical questions unanswered. It doesn't publish a table of which capabilities trigger verification, what risk score or account behavior moves a user into the check, how often a verified user has to repeat the process, or what happens when someone has a legitimate reason not to submit a government document to a U.S.

00:01:24 identity vendor. The appeal language helps after a ban or failed check. The public policy still sits behind a general phrase: certain capabilities and platform-integrity checks. That may be enough for a help page. It isn't enough for people deciding whether Claude is usable inside sensitive research, legal, journalistic, or cross-border work.

00:01:44 The public discomfort isn't hard to understand. The verification flow asks for a physical government-issued photo ID, and the surrounding commentary has focused on the live-selfie and biometric-matching side of modern identity vendors. One detailed Reddit post from a user who says they work in Norway on GDPR and European AI Act compliance made the point bluntly: biometric verification is a much larger ask than email or payment verification, and once a third-party identity company is in the chain, the user is trusting a company they didn't choose.

00:02:17 I would separate two issues here. The first is abuse control. A frontier model provider has a serious reason to know whether a high-risk account is a real person, whether a banned actor is cycling through accounts, or whether a capability with obvious dual-use risk is being used through a throwaway identity.

00:02:35 You don't need to accept every lab's risk story to accept that the risk exists. The second issue is proportionality. A face scan and government ID aren't a neutral login step. They change who can use the system comfortably. A well-paid U.S. engineer with a passport and a corporate subscription may read the help page one way.

00:02:54 Journalists, security researchers, dissidents, minors, contractors, and people in countries with weak data-protection institutions may read it another way entirely. This also folds back into the Anthropic access fight from earlier in the week. On Friday's show, I talked about access to powerful models becoming a public claim, with government pressure and private exceptions sitting in the same room.

00:03:18 The ID check is a different version of the same institutional problem. If a model is powerful enough that the provider wants to know exactly who is using it, access is no longer a simple retail product. It's a credentialed relationship, with compliance cost on one side and biometric exposure on the other.

00:03:37 TechCrunch published a discussion of the Trump administration's crackdown on Anthropic, and the important part is still what the public hasn't seen. The article says Anthropic took Fable 5 and Mythos 5 offline after an export-control order citing national security concerns.

00:03:54 In the excerpted discussion, Rebecca Bellan says the government sent a letter last Friday, gave no specifics, and required Anthropic to ensure those models could not be used by any foreign nationals. Her summary is blunt: "No one knows what those concerns are."

00:04:13 We have a powerful model, an order from the state, a national-security justification, cybersecurity experts objecting that defenders are losing access to advanced capabilities, and no public report that lets outsiders weigh the reason for the restriction. The public doesn't need every classified detail.

00:04:32 It does need enough of the rule to understand whether the same standard applies to every lab, whether Anthropic is being treated as a special case, and whether access can be restored through a known process. A Forbes piece by Paul Baier tries to give the debate a legal vocabulary.

00:04:49 Its sharpest line is also its most useful: "Military grade is not a legal classification. Dual-use is." That distinction matters because the phrase military-grade AI sounds dramatic, but the actual machinery is older and more bureaucratic. The Export Administration Regulations and the Commerce Control List already govern commercial goods with military uses, from advanced semiconductors to encryption software.

00:05:14 Frontier model access is now being pulled toward that same logic. I'm cautious about the stronger claims in that Forbes piece, because it's partly a strategy argument for companies deciding how much of their AI stack to own. But the underlying point is hard to dodge.

00:05:31 If API access to the best public model can be removed by federal order, a company that depends on that model has a sovereign-risk problem, not only a vendor-risk problem. That's true even when the order is justified. A justified restriction still changes the contract surface for hospitals, banks, security teams, research labs, and government agencies that thought they were buying a technical capability.

00:05:56 This is also why the identity-verification story and the export-control story belong together. One gate is private and user-level: prove who you are before using some capabilities. The other gate is state-level: the provider may not offer the model at all, or may need to prove that foreign nationals can't use it.

00:06:15 The result is a world where model access is being sorted by identity, jurisdiction, citizenship, and threat assessment. Some of that sorting may be necessary. Some of it may be arbitrary. We don't know yet, because the written standard is still missing.

00:06:31 The British government is reviewing the National Health Service's 330 million pound contract with Palantir, and the live issue is whether ministers use a break clause in 2027. TNW's report says technology minister Liz Kendall told Times Radio that the health secretary is reviewing every aspect of the deal to make sure Britain gets the right bargain.

00:06:53 Techmeme's summary of the Financial Times report puts the pressure more sharply: Palantir is fighting to preserve a seven-year NHS England contract as public and political pressure grows to end it. This isn't an abstract argument about whether a data platform can improve hospitals.

00:07:11 Palantir won the contract in November 2023 to build the NHS Federated Data Platform, a system meant to connect patient data across NHS England and support operational and clinical decisions. That is the kind of project where success can look ordinary from the outside: fewer duplicate records, faster scheduling, better theater use, and less time spent moving facts between systems.

00:07:35 If it works, patients don't experience a futuristic AI product. They experience an appointment that happens sooner. The political argument is about dependency and trust. TNW reports that a parliamentary committee called Palantir's growing presence in the UK public sector "an unacceptable point of weakness." The concerns aren't only about model accuracy or software performance.

00:07:59 They include how the contract was awarded, how patient data is handled, whether external staff get broad access to identifiable data, and whether the country's health infrastructure should lean on one American supplier with deep defense and immigration-enforcement associations.

00:08:17 Palantir's complaint, according to the Financial Times summary, is that procurement is being politicized. I can see why the company would say that. If you're a vendor, you want the evaluation to be about whether the platform performs, not whether your brand has become a proxy for a larger fight over surveillance capitalism and American power.

00:08:38 But public health data is political in the old, literal sense: it belongs to people who can't opt out of the system easily, and it's held by the state under a promise that the state won't turn need into exposure. The 2027 break clause is the institutional hinge.

00:08:55 A break clause is one of the few moments when a government can ask whether a system that has started to become normal should keep becoming more normal. By then, teams will have workflows, dashboards, training material, and local habits wrapped around the platform.

00:09:12 Pulling it out would have a cost. Keeping it would also have a cost if the public decides the NHS accepted too much vendor dependence for too little democratic control. My read is that Palantir's strongest argument will be delivery evidence, not ideology. If the platform can show measurable improvements across many trusts, the government has to explain why a different supplier can preserve those gains.

00:09:37 If the claimed gains are narrow, hard to attribute, or concentrated in a few places, ministers have an easier case for walking away. Either way, the NHS contract is a direct example of AI-era institutional power: the vendor doesn't need to own the hospital to become hard to remove from it.

00:09:56 CoreWeave's April announcement is suddenly more interesting because of how the rest of the market is talking about Jane Street. CoreWeave says Jane Street committed about six billion dollars to use its AI cloud platform and also made a one billion dollar equity investment in CoreWeave Class A common stock at 109 dollars a share.

00:10:16 The announcement says the deal gives Jane Street access to next-generation compute across multiple facilities, including Nvidia's Vera Rubin technology. The company quote from Jane Street is unusually clear about the use case: "training large, complex models on massive volumes of noisy data." That is financial-market language wearing AI clothes, or maybe the other way around.

00:10:39 The data is noisy, the models need to be refined continuously, and the deployment has to happen fast enough to matter in a competitive trading business. Techmeme's summary of Gregory Zuckerman's Wall Street Journal profile adds the institutional angle. Jane Street isn't only buying compute.

00:10:57 The report says the firm is becoming a major AI investor, has a stake in Anthropic, invested one billion dollars in CoreWeave, has grown to 3,500 staffers, and plans to recruit more than 500 employees this year. That makes Jane Street part of the AI capital stack in two directions at once: it's a customer for specialized compute and an owner of companies that sell or build around that compute.

00:11:21 This is where the market story becomes more than another big number. A trading firm has a different tolerance for latency, secrecy, and edge than a normal enterprise software buyer. It also has a different balance sheet logic from a startup. If Jane Street thinks a six billion dollar cloud commitment and a one billion dollar equity position help it compete, that tells you something about how AI is being priced inside markets where small advantages can be monetized quickly.

00:11:50 CNBC's interview with Yann LeCun came from a different angle, but it belongs near the same chapter. LeCun criticized xAI, called it a failure, pointed to founder departures, and warned about a wider AI bubble. You don't have to accept his whole verdict on Musk to hear the pressure underneath it.

00:12:08 The industry's fixed costs are enormous. Data centers, GPU reservations, power contracts, talent, and model training commitments all arrive before the revenue is certain. Some companies will rent out spare capacity, some will try to sell model access, some will try to turn AI into an internal advantage, and some will discover that owning a lot of compute isn't the same as owning a durable business.

00:12:33 Jane Street is a useful counterexample to the startup story. It isn't trying to become a consumer chatbot company, and it isn't making a public claim about artificial general intelligence. It's putting compute next to a business where prediction, execution, and risk control already had a direct path to money.

00:12:52 That doesn't make the bet safe. It does make the bet easier to understand than a lab raising at a heroic valuation because everyone assumes the future will need whatever it builds.

00:13:03 The UK Home Office launched PoliceAI, a national centre for artificial intelligence in policing across England and Wales. The official announcement says the centre is backed by 75 million pounds over three years and sits inside a 140 million pound AI technology investment.

00:13:20 The government's language is familiar: officers should spend less time behind desks and more time protecting communities. The operational examples deserve attention. The Home Office says one early trial reviewed 800 hours of footage in a kidnapping case in three hours and helped produce an early guilty plea.

00:13:40 It says another translated half a million e-books of data and led to arrests connected to a serious organised crime group. The same package includes 40 more live facial recognition units, tripling current capacity, and another 16.5 million pounds to modernize how police and the public interact through call transcription, crime-report linking, and non-emergency call triage.

00:14:05 PublicTechnology adds rollout detail. Up to 10 areas will take part in first-year trials around triage, disclosure, and summarization of digital evidence. If the practices are judged successful, the plan is to scale them to all police forces in 2027. The Home Office also says audiovisual redaction could free one million staff hours a year if all forces use the technology.

00:14:29 There are two ways to hear this announcement, and both are reasonable. The sympathetic version is that police forces are drowning in digital evidence. Phones, cameras, home devices, messaging apps, shop footage, and body-worn cameras produce far more material than an investigator can review by hand.

00:14:48 If a tool helps triage that evidence, find the relevant clip, redact protected material, or translate a large archive, it can shorten investigations and reduce the chance that a victim waits months while the system looks for the right file. The colder version is that policing AI has a history of expanding from administrative relief into suspicion, surveillance, and error at the point where a person has the least power to challenge it.

00:15:17 Live facial recognition is already in the funding package. Call triage changes how someone enters the system. Evidence summarization changes what a prosecutor, defense lawyer, or officer sees first. Even if every model is purchased with a responsible-use label attached, the practical power of these systems sits in defaults, thresholds, audit logs, and appeal paths.

00:15:40 The phrase "public consent" appears in the government's own announcement, and that is where this will have to be judged. Public consent can't mean a press release and a procurement page. It has to include tool registers, error reporting, independent testing, disclosure to defendants, and clear rules for when an AI summary can be used and when a human must verify the underlying record.

00:16:05 PoliceAI may save hours. It may also move judgment earlier in the evidence chain, where mistakes can become invisible unless the system is forced to preserve the path back to the original material.

00:16:18 The Guardian's Australia data-center story starts with a concrete picture: a proposed hyperscale project on Mamre Road in western Sydney, covering 52 hectares, with six four-storey buildings, 936 cooling units, and 852 diesel backup generators. The article places that project inside an estimated 155 billion Australian dollar investment pipeline over the next decade.

00:16:42 Australia already has about 160 data centers operating, and the Climate Council says another 90 are proposed. The power number should stay in your head. The Guardian quotes Alex Hooper of Oxford Economics Australia saying data centers account for 2.8 percent of electricity consumption on Australia's east coast now and could consume more than 10 percent by the mid-2030s.

00:17:05 That isn't a server-room issue. It's a planning, generation, transmission, water, diesel, and local land-use issue. AI companies tend to talk about compute as if it were a cloud abstraction. Local communities meet it as buildings, substations, backup generators, cooling systems, noise, water demand, and grid upgrades.

00:17:25 Governments meet it as a development pitch. The company promises investment, construction work, digital-economy status, maybe some operational jobs, and a place in the global AI supply chain. The community asks what it gets in return once the facility is running and the power demand remains.

00:17:44 The Guardian piece is good because it refuses the easy answer. Data centers can be useful economic infrastructure. A country that wants AI services, cloud resilience, and digital sovereignty can't simply say no to the physical layer. But the article also notes the high environmental cost and limited operational jobs.

00:18:04 If a project consumes a large share of new power capacity and creates relatively few long-term jobs, the public bargain has to be explicit. Forbes, from the U.S. energy side, published a related piece on second-life electric-vehicle batteries for AI data centers.

00:18:21 Retired EV batteries often retain roughly 80 percent of their original capacity, according to the piece, which may be enough for stationary storage. Charge them off peak, discharge them when AI workloads push demand up, and the batteries become a buffer between the data center and the grid.

00:18:39 The quoted line from Linda Li of Re-Teck is direct: "Repurposed batteries are the way to break that barrier." The Forbes piece itself says data-center battery farms aren't common yet because the economics are hard. But the idea tells you where the pressure is moving.

00:18:59 The race is no longer just for chips. It is for firm power, cheap storage, grid permissions, cooling capacity, and political patience. When a model company says it needs more compute, a suburb somewhere may hear more diesel backup, more peak demand, and a planning hearing where the public has to decide how much of the AI economy it wants to host.

00:19:21 A few smaller items from today deserve shorter treatment because they point in different directions. IIT Bombay posted that BharatGen, supported by the IndiaAI Mission and India's Department of Science and Technology, is joining Project Tapestry as a founding contributor.

00:19:38 The post says India is building frontier AI rooted in its own languages and knowledge, and that a letter of intent was signed on June 18. I wouldn't file that under ordinary lab news. Language is infrastructure. If a country's AI capacity depends on models, data, and evaluation habits centered somewhere else, the dependency is cultural as well as technical.

00:20:00 BharatGen's claim is that multilingual AI infrastructure built for India's own languages can travel into an open consortium. We don't know yet how powerful the systems will be, how open the resulting work will remain, or how much state support turns into production capacity.

00:20:17 But the direction is clear enough without over-selling it: national AI projects are starting to organize around language, knowledge, and public mission, not only around private model subscriptions. On the defense side, Techmeme picked up a CTech report that Anduril has begun exploring operations in Israel and is in talks to recruit a local manager.

00:20:39 The summary says Palmer Luckey's 61 billion dollar startup is considering local operations, partnerships, and investments in Israeli startups. I would keep this one at the level of a reported exploration, because that is what the source says. Still, it belongs in the day because defense AI expands through offices, local partners, hiring, and regional trust.

00:21:01 The policy fight over autonomous systems is one layer. The company-formation layer is more practical: where do you hire, where do you integrate, and which security ecosystem gets first access to the product roadmap? And then there is the fish robot. Ksenia Moskalenko posted that Founders Fund backed Shinkei Systems, which built a physical robot for boats that uses AI to process fish instantly, cutting stress and doubling shelf life.

00:21:28 That's a small item. It's also a useful antidote to the habit of treating AI as if it only lives in office software or national-security briefings. If the claim holds, the buyer isn't paying for a chatbot. The buyer is paying for shelf life, consistency, less waste, and a different labor process on a boat.

00:21:47 Those three items don't need to be forced into one grand thesis. They show the same technology stack moving through three very different institutions: a national language project, a defense startup looking at local operations, and a food-supply-chain machine with venture money behind it.

00:22:04 That's enough. AI power isn't only being concentrated in the big model labs; it's also being translated into the places where states, militaries, and niche markets already know exactly what they want. Next: whether Claude's ID checks stay limited to a few capabilities, and whether NHS ministers use the 2027 break clause.

00:22:24 Jonas.