AI Power Became A Location Problem

00:00:04 South Korea announced on Monday that it wants to put more than a trillion dollars behind semiconductors, physical AI, and data centers. That is the plain fact of the day, and it is a large one even after you discount the ceremonial language that always comes with national industrial plans.

00:00:22 President Lee Jae Myung gave the announcement with the heads of Samsung Electronics and SK Hynix beside him. This wasn't only a ministry saying the country likes AI. Samsung and SK Hynix are the two largest memory chipmakers in the world, and the plan puts them at the center of new fabrication sites in South Korea's southwest.

00:00:43 Industry Minister Kim Jung-kwan said Samsung, SK Hynix, and suppliers will invest 800 trillion won, about 518 billion dollars, in two new chip fabrication sites each. The government also named a chip-packaging cluster near Seoul, with another 81 trillion won, or about 52.5 billion dollars, expected there.

00:01:03 The data center number changes the temperature of the story. The government said SK Group, GS Group, and Naver will back AI data centers in the region with 550 trillion won, about 356 billion dollars, in investment. Science Minister Bae Kyung-hoon described a buildout that would exceed 18.4 gigawatts and 1,000 trillion won by 2035.

00:01:24 A gigawatt isn't an abstract unit in this story. It is land, interconnection, cooling, financing, and political permission. South Korea is saying that the AI contest isn't settled by model releases alone. It is settled by whether a state can arrange memory, packaging, power, and enough domestic demand to make the arrangement durable.

00:01:46 Lee's phrase was that South Korea must secure the core elements of AI faster than any other country. I wouldn't take the whole sentence at face value, because every country with a chip plan now sounds as if it is late to a train that it also claims to be driving.

00:02:03 But the structure of the announcement is more serious than the slogan. Semiconductors, physical AI, and data centers were presented as a single industrial package. That tells you how governments are now sorting the stack. The model layer, the chip layer, the factory layer, and the power layer are becoming one national account.

00:02:24 There is a domestic fight inside the announcement too. The southwest, including Gwangju and South Jeolla province, isn't just an industrial geography. It is a political one. Opposition critics accused Lee's government of steering chip projects toward Honam, a liberal stronghold where 85 percent of voters backed him in last year's presidential election.

00:02:46 Lee defended the location over the weekend and pointed to untapped power resources. That is the practical political bargain underneath the AI boom: national competitiveness gets wrapped around regional development, and then companies are asked to make the map work.

00:03:04 Xiaoyin Qu put the larger geopolitical anxiety bluntly on Monday. Her worst-case scenario for the United States was that Chinese open models keep gaining market share, Chinese labs train and optimize on Huawei chips instead of NVIDIA, and the United States fails to build data centers fast enough to meet demand for compute, storage, and energy.

00:03:26 You don't have to accept all of that as a forecast to see why South Korea's announcement belongs in the same conversation. The export-control fight has made every country ask what it owns in the AI supply chain and what it merely rents. That is why the South Korean plan is about bargaining power.

00:03:45 If Samsung and SK Hynix remain memory suppliers to everyone, if South Korea can package advanced chips domestically, and if it can host serious AI data center capacity on its own grid, then it is less likely to be treated as an interchangeable site in somebody else's AI plan.

00:04:03 The open question is whether the money becomes construction, whether the grid can absorb the load, and whether regional politics makes the industrial layout stronger or more brittle. The announcement is big. The permits, wires, substations, fabs, and customers will decide how much of it becomes power.

00:04:22 Taiwanese authorities raided Super Micro's Taiwan office on Monday as part of a probe into alleged smuggling of Nvidia chips to China, according to Bloomberg reporting summarized by Techmeme. Super Micro's stock closed down 8.1 percent after the report. That is a short item with a long tail, because export controls are only as strong as the middle of the supply chain: distributors, invoices, routing, subsidiaries, customs paperwork, and the places where hardware changes hands.

00:04:53 Super Micro isn't Nvidia, and the report doesn't say the company has been found liable. It says Taiwanese authorities widened an investigation and searched an office connected to a major server maker. That is enough to show where the enforcement pressure is going.

00:05:09 When the United States restricts advanced AI chips from China, the policy doesn't stop at a Commerce Department notice. It migrates into allied jurisdictions, corporate compliance rooms, and the financial markets' read on whether a hardware company has clean channels.

00:05:26 This part of AI geopolitics tends to look less dramatic than a model release and more consequential than one. The frontier model announcement gets the attention; the server shipment decides who can serve the model to customers. If a Chinese lab can't get the latest Nvidia parts directly, the market looks for detours.

00:05:47 If a detour becomes common enough, enforcement follows it. Taiwan matters here because it sits inside the physical geography of advanced computing: chip fabrication, board assembly, server integration, and shipping all touch the island in one way or another. There is also a signal in the market reaction.

00:06:06 An 8.1 percent drop doesn't prove wrongdoing, but it says investors understand that export-control investigations now belong on the same risk page as component shortages and customer concentration. A server maker can have excellent demand and still get repriced if the market thinks the route to that demand might run through a sanctioned channel.

00:06:28 The new hardware premium is proof of custody: not only can you source the part, but can you prove where it went? Put this beside the South Korean announcement and the picture gets sharper. States are trying to build domestic capacity at the same time they are trying to police cross-border leakage.

00:06:47 One side of the ledger is investment: fabs, packaging clusters, data centers, and power. The other side is enforcement: raids, compliance checks, export licenses, and public examples. Both are expensive. Both are political. And both push AI away from the fantasy that software floats above the world.

00:07:06 The Nvidia-Anthropic-Microsoft announcement on Monday fits between those two poles. NVIDIA said Anthropic's Claude models are now generally available in Microsoft Foundry on Azure, running on GB300 Blackwell Ultra graphics processing units. The blog names NVL72 systems, Quantum-X800 InfiniBand networking, and a Secure Agent Workspace reference design where identity, network access, credentials, and runtime policy are controlled at the infrastructure level.

00:07:35 Strip away the vendor shine and you get a blunt distribution fact: enterprise access to a frontier model is being packaged through a cloud provider, an accelerator vendor, and a governance design. That isn't a scandal. It may be a good product. If you are a regulated company, you probably want the model inside a managed environment with identity and credential control rather than a loose API key sitting in a team spreadsheet.

00:08:02 But it is also concentration by ordinary procurement. A company that chooses Claude inside Foundry isn't only choosing a model. It is choosing Azure, Nvidia's newest data center hardware, and a particular idea of how autonomous agents should be contained. I think the institutional AI story is this: the state wants to control where the chips go, the cloud platforms want to control where the models run, and enterprises want somebody else to make the agent environment legible enough for audits.

00:08:34 The result is a stack where access, compliance, and compute are sold together. That can make deployment safer in some narrow operational sense, and it can also make the largest vendors harder to leave. The same week a server office gets raided over alleged chip leakage, a model access path gets wrapped in premium cloud infrastructure.

00:08:55 Both say AI power is now a routing problem as much as a capability problem.

00:09:00 Senator Elizabeth Warren and Representative Mary Gay Scanlon are preparing an updated Health and Location Data Protection Act that would cover information people enter into AI systems. The Verge reported Monday that the proposal would ban the sale of Americans' health and location data to data brokers, including data disclosed to chatbots like ChatGPT or Claude.

00:09:23 The timing fits the market. AI labs have spent the year moving toward health products, medical-record intake, and provider tools. The article points to Elon Musk asking people in January to upload medical records to Grok, OpenAI introducing ChatGPT Health and ChatGPT for Healthcare, and Anthropic following with Claude for Healthcare.

00:09:44 Once companies invite people to place their MRI scans, symptoms, medications, and diagnoses into general-purpose AI products, the old distinction between a health app and a chat product starts to look thin. The proposed bill would require the Federal Trade Commission to write rules within 180 days.

00:10:02 It would let the FTC, state attorneys general, and affected individuals sue to enforce the law, and it would earmark 1 billion dollars for FTC enforcement over the next ten years. Warren's statement was direct: data brokers are making money selling Americans' sensitive information, and people entering private health data into AI systems need protection from exploitation by the highest bidder.

00:10:27 The legal machinery matters here because the United States still doesn't have a broad federal privacy law. That means a person can have HIPAA protection in one setting and much weaker protection in another setting that feels medically intimate to them. A patient portal, a hospital workflow, a consumer chatbot, and a data broker database don't create the same legal duties.

00:10:51 People don't experience the boundary that way. They experience it as one continuous question: I told a machine something about my body; who can sell it? Monday also brought a separate Supreme Court item that changes the institutional background for all of this.

00:11:07 Axios reported that the Court cleared President Trump to fire officials from the Federal Trade Commission and most independent agencies, with a carve-out for the Federal Reserve. According to the Axios summary, the ruling overturns the nearly century-old Humphrey's Executor precedent that protected independent agency commissioners from removal without specific cause.

00:11:30 Chief Justice John Roberts wrote, in Axios's excerpt, "If anything more is left of Humphrey's, we overrule it." The FTC is one of the agencies most likely to police AI privacy promises, data broker practices, unfair competition, and deceptive claims. If the commission is more directly subordinate to the president, enforcement can move faster in some cases and become more politically contingent in others.

00:12:00 The Verge story says the Warren-Scanlon bill would put enforcement work on the FTC. The Axios story says the FTC's independence has just been cut back. Those two facts belong next to each other. There was a third privacy item circulating Monday as well: a claimed Supreme Court ruling that geofence warrants for detailed cell-phone location history count as Fourth Amendment searches.

00:12:24 The X post that carried it wasn't enough for me to treat as a primary legal source, but the reaction was useful. People immediately asked whether the same logic would reach Flock-style license plate databases, commercial data purchases, or chatbot conversations.

00:12:40 One reply put the gap plainly: courts may require a warrant for police, while corporations can still sell the data unless Congress or regulators stop them. That is the policy hole AI keeps widening. The law has one set of instincts for the state asking for data, another for companies buying and selling data, and a third for people voluntarily typing sensitive facts into a product because the product presents itself as helpful.

00:13:07 AI systems make voluntary disclosure feel ordinary. You ask the model about a rash, a pregnancy risk, a panic attack, a route to a clinic, or a family member's medication, and the interaction feels like a search query with a bedside manner. But the information is richer than a search query and often more revealing than a form.

00:13:27 My read is that the next fight won't be whether health AI exists. It already does, at least in early commercial form. The fight will be over whether the data generated around those products is treated like medical trust or like advertising exhaust. If the bill moves, the definitions will matter: which companies count, which transfers count as sale, what exceptions are carved out for service providers, and whether affected individuals can make enforcement painful enough that privacy policies stop being the whole protection.

00:14:00 Techmeme summarized an Engadget report on Monday about more than 100,000 AI-enabled automated license plate readers installed across the United States, mostly from Flock. The report focused on security flaws, police misuse, and civil-liberties concerns. The line from the Engadget summary that stuck because it was so blunt was, "You can't get a breath of fresh air ...

00:14:24 without us knowing." That isn't the language of a constitutional framework. It is the language of a database that has started to understand its own reach. License plate readers aren't new, but AI changes the economics of watching ordinary movement. A camera at one intersection is a camera.

00:14:42 A network of cameras that can identify plates, search patterns, connect jurisdictions, and make movement histories retrievable becomes a different civic object. It is closer to a private-public location system that police can query. If the network is mostly built by a company, and if local agencies join it piecemeal, the country can acquire a national surveillance capability without ever voting on one as a national system.

00:15:08 The geofence-warrant discussion on X sharpened that point. A user asked whether a Supreme Court ruling on cell-phone location records would bring Flock records into play. Another pointed out that a geofence warrant can sweep in anyone near a place, not only the person police already suspect.

00:15:27 That is the old problem of dragnet search meeting the new problem of cheap persistent capture. AI doesn't have to be magical here. It only has to make retrieval easier, matching faster, and cross-jurisdictional search normal enough that restraint becomes an administrative preference rather than a technical limit.

00:15:46 The more severe version of the same problem appeared in another Techmeme item, this one summarizing Los Angeles Times reporting on a February 28 strike on an Iranian school. According to that summary, disconnected U.S. military databases may have contributed to the strike after outdated intelligence misidentified the site.

00:16:06 The report said an estimated 120 children were killed. Some people see AI as a way to fix fragmented military data; others fear it could amplify errors. That is a brutal example because it refuses the easy slogan. Better data integration can save lives if it prevents commanders from acting on stale intelligence.

00:16:26 Better data integration can kill people faster if the system makes the stale record look authoritative. The old breakage is that one database doesn't know what another database knows. The new breakage is that an AI layer makes the combined output feel coherent when the underlying records are partial, old, disputed, or misclassified.

00:16:47 This is why I am wary of treating AI in government as a procurement category. A license plate reader system, a military targeting database, a benefits-fraud detector, and a health-inspection assistant don't share the same moral weight. They do share a technical temptation: bring more data together, search it more easily, and let the user ask the machine for an answer.

00:17:10 In an office workflow, a wrong answer may waste money. In policing, it may put an innocent person under suspicion. In war, it may place a school inside a target envelope. The accountability problem is also different across those settings. A police department can blame a vendor.

00:17:27 A vendor can point to local policy. A military unit can blame incomplete inputs. An AI system can be described as advisory until the moment everyone in the chain treats it as the fastest way to settle uncertainty. If you have worked around institutions long enough, you know that "advisory" systems can become mandatory by habit.

00:17:48 The person who ignores the machine has to explain why. The person who follows it can say the machine agreed. There is a narrow path where AI helps: show provenance, surface conflicting records, force date checks, identify missing approvals, and slow the user down when the evidence is stale.

00:18:06 That version is less glamorous than an all-knowing assistant, and it is much closer to a disciplined clerk with good memory and no authority. The danger is the assistant that makes old intelligence, plate-reader hits, and commercial location data feel more settled than they are.

00:18:24 In surveillance and targeting, speed isn't neutral. It decides who gets a chance to be corrected before the state acts.

00:18:31 The FDA selected seven companies on Monday for its PreCheck Pilot Program to advance U.S. drug manufacturing. The agency named Amneal, Cellares, Eli Lilly, FUJIFILM Biotechnologies, Kriya Therapeutics, Kyowa Kirin, and Regeneron. The program isn't an AI program in the narrow product-demo sense, but it belongs in today's episode because pharmaceutical manufacturing is becoming part of the same national resilience argument as chips, data centers, and medical AI.

00:19:01 The FDA says PreCheck gives companies earlier technical guidance before a facility becomes operational and then enhanced engagement when drug or biologics applications are submitted. The agency received more than 80 requests between February 1 and March 1. Participants had to propose a new domestic manufacturing facility and commit to an application that relies on that facility.

00:19:25 The selected projects cover sterile liquid products, cell-based gene therapies, active pharmaceutical ingredients, commercial-scale cell culture manufacturing, AAV-based gene therapy, biotechnology drug substance, sterile injectables, and protein therapeutics. That list is dense, but the institutional point is simple enough.

00:19:46 The United States wants more critical medical production inside its borders, and the regulator is trying to move earlier in the facility-development process so manufacturing issues don't appear late enough to delay approval. In AI terms, this is the physical counterpart to the data-center buildout: not smarter chat, but the ability to make and inspect the thing the health system depends on.

00:20:11 If AI is going to accelerate drug discovery, clinical documentation, imaging, trial design, or supply-chain planning, then the bottleneck moves to manufacturing, quality systems, and regulatory trust. Red Hat's NASA item gives a more literal edge-computing version of the same idea.

00:20:29 Researchers at NASA's Johnson Space Center are testing the Crew Medical Officer Digital Assistant, or CMO-DA, for future deep-space missions. The system uses RamaLama for local AI inference and is meant to help astronauts diagnose and treat medical symptoms when real-time communication with Earth is limited or impossible.

00:20:50 The project began as a cloud-connected proof of concept and moved to a disconnected edge deployment on HPE hardware, specifically the terrestrial twin of the Spaceborne Computer aboard the International Space Station. This is one of the better uses of local AI because the constraint isn't ideological.

00:21:09 You can't rely on a cloud call from Mars. Light delay and communication blackouts make a remote doctor a delayed participant. The model has to run on the hardware you brought with you, with medical literature available locally, and with outputs that can be inspected after the fact.

00:21:27 Red Hat describes RamaLama as treating models like container images, which sounds dry until you put it in a spacecraft. Reproducible deployment matters more when nobody can walk into the machine room. Firefly Aerospace and Nvidia added a second space example on Monday.

00:21:44 Firefly's Blue Ghost Mission 2, targeted for launch in late 2026, will carry the Ocula moon imaging service and operate Nvidia's Jetson edge AI platform in lunar orbit. Firefly's first Blue Ghost mission downlinked nearly 120 gigabytes of raw data from the moon in March 2025, imagery and video that scientists are still processing.

00:22:06 The pitch for Ocula is that the spacecraft can process imagery onboard, extract relevant signals, and transmit only the information customers need closer to real time. Again, the important word is constraint. Downlink bandwidth is limited. Latency is expensive.

00:22:22 Processing everything on Earth means the data arrives before the insight. Onboard AI is attractive because the spacecraft can decide what deserves scarce communication time. That same power raises its own question: what gets filtered out before humans see it? On the moon, that may mean missed mineral signatures, object-tracking errors, or bad prioritization of landing-site imagery.

00:22:47 In medicine, it may mean a local model missing a symptom pattern or citing a guideline badly. The edge makes AI useful because the edge has no patience for remote abundance. Meta's Brain2Qwerty v2 announcement sits at the more intimate end of the same physical turn.

00:23:04 AI at Meta described it as a non-invasive brain-to-text decoder capable of real-time sentence decoding from raw brain signals, building on a v1 paper in Nature Neuroscience. Meta said it would release training code for v1 and v2, and that a partner would release the v1 dataset.

00:23:21 The promise is for people with brain lesions or disorders that prevent communication. The social reaction, unsurprisingly, jumped to mind-reading anxiety in about six seconds. We don't know yet how far that system is from clinical use, consumer misuse, or anything a regulator would treat as a deployable medical device.

00:23:42 The distinction matters. A laboratory milestone with code release isn't the same thing as a product in a hospital. But it is part of a day where AI kept showing up at physical boundaries: drug facilities, spacecraft, lunar imaging, and brain signals. The evidence that would matter next is validation under pressure: who tested the system, what failed, what the operator saw, and what record remains when a human has to review the machine's advice.

00:24:10 Tomorrow's proof has to be audits and deployment numbers, not only announcements. Jonas