The Ledger Got Physical

00:00:04 ICE and Customs and Border Protection contracts with surveillance-technology companies reached a record five hundred and thirteen million dollars in 2026, according to a new report covered by The Guardian on Wednesday. That's the fact to start with. The report looked at contracts with eleven companies that provide surveillance technology to immigration agencies.

00:00:26 The money awarded to those firms was under fifty million dollars when the researchers trace the series back to 2013. It doubled from 2024 to 2025, reaching just over three hundred and ten million dollars, and then rose again in 2026 to the five hundred and thirteen million dollar figure.

00:00:45 The names are familiar if you follow the place where defense contracting and AI meet: Palantir and Anduril are central in the Guardian's account. The tools are familiar too, though the combination changes the story. The list includes data brokers, analytics software, social media scraping, facial recognition, phone hacking devices, drones, border towers, spyware, and outside contractors that the report's authors describe as bounty-hunter-like.

00:01:13 DHS has publicly disclosed that it uses more than ten AI-enabled facial-recognition tools. That's a policy sentence wearing a procurement jacket. DHS's role as a market-maker makes this larger than a buying spree. The article says DHS doesn't only buy surveillance products after they exist.

00:01:31 It also runs funding channels that help create them. The Silicon Valley Innovation Program can give startups up to two million dollars for prototypes. The DHS component of the Small Business Innovation Research program has put eight hundred and forty-five million dollars into more than five hundred companies since 2004, according to the study cited by The Guardian.

00:01:54 Some recent awards, the article says, went toward tools for harvesting biometric data from phones and using AI to analyze airport camera feeds. The leverage has two layers. The agency gets more capacity to find, track, and classify people. The vendors get an anchor customer, technical direction, field testing, and federal credibility.

00:02:15 That's how a contested surveillance idea becomes an industry segment. It doesn't need to wait for a grand public debate if the procurement line keeps moving. The Guardian asked Palantir about the report, and the company's answer matters because it's the answer these firms often give.

00:02:33 Palantir said it doesn't collect or store data, doesn't conduct surveillance, and doesn't set immigration policy. That may be literally true in the narrow corporate sense. A database company can say the government decides how to use the database. A border-tower company can say the agency decides where to point the tower.

00:02:53 A model provider can say the customer owns the workflow. The institutional problem is that the division of labor lets each actor deny ownership of the whole system. There is also a civil-rights boundary in the details. Paromita Shah of Just Futures Law told The Guardian that an agency with limited oversight receiving what she called a slush fund worries her because of what ICE and CBP have already done with money on the ground.

00:03:20 She also raised the street-level question around facial recognition: whether people consented, whether a warrant was obtained, and whether protest activity is being converted into a searchable record. We don't know from this article how accurate each tool is, how often it is used, or how many people were wrongly pulled into an enforcement action because a vendor product joined two records it shouldn't have joined.

00:03:46 Those are exactly the records the public should be able to inspect. The point for today is narrower and firmer: AI surveillance in immigration enforcement is no longer a speculative civil-liberties warning. It is a half-billion-dollar contract pattern, with named firms, named funding channels, and a federal agency helping shape the market it then relies on.

00:04:09 OpenAI and Broadcom unveiled Jalapeno on Wednesday, OpenAI's first custom inference processor for large language models. That sounds like a chip story, and it is. It is also a control story. OpenAI says Jalapeno is an application-specific integrated circuit, designed around inference rather than general-purpose AI acceleration.

00:04:30 Training is the expensive process of building a model. Inference is the repeated act of serving it: ChatGPT answering, Codex running a task, an API customer sending another request, and an agent taking another step. If training is the capital event, inference is the operating bill.

00:04:48 OpenAI's post says the chip was co-developed with Broadcom and Celestica, with Broadcom handling silicon implementation and networking. Celestica is helping industrialize boards, racks, and systems. The company says the chip moved from design to production tape-out in nine months, with OpenAI models helping parts of design and optimization.

00:05:10 OpenAI also says engineering samples are running machine-learning workloads in the lab at target frequency and power, including GPT-5.3-Codex-Spark. The performance claim needs a little discipline. OpenAI says early testing shows substantially better performance per watt than current state-of-the-art systems, and Broadcom's release repeats that claim.

00:05:33 The detailed technical report is still coming. So for today, the sourced fact isn't that OpenAI beat every incumbent in a way outsiders can audit. The sourced fact is that OpenAI and Broadcom are publicly putting a performance-per-watt claim on the table and tying it to a chip family they want deployed by the end of 2026.

00:05:54 The strongest line in Broadcom's release was from Hock Tan, not because it was poetic, but because it named the physical unit of ambition. He said the work helps enable gigawatt-scale data centers with Microsoft and other partners beginning in 2026. A gigawatt isn't a metaphor.

00:06:12 It is power, land, grid interconnection, cooling, financing, construction, and permitting. It is also a reminder that the market power around AI is increasingly inside the parts nobody can download: chips, racks, networking, and the energy contracts behind them.

00:06:29 Nvidia's Jensen Huang gave the mirror-image version of the same story at Nvidia's shareholder meeting. CNBC reports that Huang said national security comes first and that if a customer tried to smuggle Nvidia chips or systems into export-restricted countries, including China, the customer would have trouble getting the system working because Nvidia wouldn't provide support or repairs.

00:06:54 His phrase was short: smuggled-product data centers are a "dead end." Huang described advanced AI data centers as integrated systems that need trusted hardware, software, networking, and continuing support. If that is true, then the support relationship becomes part of the policy boundary.

00:07:19 A chip without firmware updates, networking support, replacement parts, and debugging help isn't the same economic object as a supported cluster. Washington can write a rule about exports, but the practical gate also runs through vendor service. CNBC also reported that China, including Hong Kong, accounted for about nine percent of Nvidia's fiscal 2026 revenue, a smaller share than in the two prior years.

00:07:46 Huang said the U.S. approved H200 export licenses, but Nvidia had not yet generated revenue from those chips and did not know whether China would allow imports. That's the strange middle state of AI hardware policy: permission can exist on paper while the commercial channel still doesn't reopen.

00:08:05 Qualcomm added a smaller but revealing data point the same day. Techmeme summarized CNBC's report that Qualcomm unveiled Dragonfly C1000, a data center CPU built for agentic AI, and said Meta will use it when production starts in 2028. Put that next to Jalapeno and Nvidia's support argument.

00:08:23 The model companies want custom inference, the chip companies want agent workloads to define new server demand, the cloud and social platforms want optionality, and the government wants to decide which countries get supported systems. This isn't a model-card cycle.

00:08:41 It is the machinery around every future model becoming contested territory.

00:08:46 A viral post on Wednesday claimed that Representative Anna Paulina Luna used Anthropic's Claude chatbot to draft an amendment to a defense bill. The source trail is thinner than that sentence makes it sound. The root post was a NewsWire item on X with an image attached.

00:09:02 The Reddit copy of the story was just an image link with no discussion. The useful part of the X thread was Luna's own reply. She wrote that her staff used AI to spell-check and grammar-check the amendment summary, not the amendment text itself. She added that most staff use it and that she told them to double-check and be more thorough.

00:09:24 Then, with congressional dignity fully intact, she said she likes Claude but Grok is better. The corrected version is smaller. A staffer using a chatbot to clean up a summary isn't the same as a model drafting defense law. But the episode is useful precisely because it shows how fast the public read jumps from office workflow to state capture.

00:09:45 In the replies, people immediately argued about whether AI is running the military, whether sensitive information went into a corporate model, and whether elected officials or staffers read the material they introduce. Some of that was partisan heat. Some of it was a reasonable institutional worry.

00:10:04 Once you strip away the screenshot drama, the policy question is recordkeeping. If congressional staff use commercial AI tools on summaries, amendments, talking points, or constituent letters, what material can go into those tools? Which products are approved? Are prompts retained?

00:10:22 Does a vendor get sensitive legislative language before it is public? Is there a House rule, an office rule, or just a staffer's browser tab? Those questions don't require the exaggerated claim to be true. This is where small AI adoption becomes a governance problem.

00:10:39 Nobody needs to believe Claude wrote the defense bill to ask whether Congress has a usable rule for AI-assisted drafting, editing, summarizing, and research. The practical answer may be that staff have already normalized it faster than the institution has documented it.

00:10:56 That isn't a scandal by itself. It is how office software enters government: first as a convenience, then as a dependency, and later as something the ethics manual tries to catch.

00:11:07 Stripe, Anthropic, the OpenAI Foundation, Bill Gates, Flu Lab, and several Jane Street traders are backing a new five hundred million dollar nonprofit called Intercept, MIT Technology Review reported on Wednesday. Intercept's stated target is unusually plain: prevent the common cold and flu, then eventually eliminate respiratory viruses.

00:11:29 It plans to fund vaccines, broad countermeasures, and air-cleaning systems for schools, offices, and other shared spaces. Nan Ransohoff, the Stripe executive leading the initiative with Charlie Petty, told MIT Technology Review that society has underweighted the burden of respiratory infections.

00:11:48 She gave one number that makes the project easier to understand: on average, people spend five percent of their lives fighting a cold or the flu. That's a very Stripe-adjacent way to look at sickness: take something everyone treats as background friction, attach a cost to it, and ask whether incentives have kept the market smaller than the social burden.

00:12:11 Ransohoff compared Intercept to Frontier, Stripe's carbon-removal purchasing program. The shared logic is a public-benefit gap. Private companies don't always fund work whose payoff is distributed across everyone. The article says David Veesler, a structural biologist and vaccine designer at the University of Washington, helped shape the concept by arguing that broad countermeasures across many viruses are technically possible.

00:12:39 The technical menu includes RNA drugs, antibodies, computational protein design, and even virus-grabbing proteins people could spray in the nose. The air-cleaning side may matter more quickly. If you can remove pathogens from shared air the way municipalities remove contaminants from water, you reduce the burden without asking each person to win a private biomedical arms race.

00:13:03 The funders are part of the story. Anthropic and OpenAI Foundation money in a respiratory-virus nonprofit tells you that frontier AI institutions are trying to occupy some of the territory once held by governments, universities, pharma companies, and disease philanthropies.

00:13:21 That could be excellent. It could also make public health depend more on donor taste and tech wealth. The article says Intercept's advisors will include Peter Marks, the former top FDA official, and Moncef Slaoui, who led Operation Warp Speed. Those names signal an attempt to put regulatory and vaccine-development experience around the money.

00:13:43 The same day brought a different kind of health source: an arXiv paper on rare-disease diagnosis. The authors present RaDaR, a thirty-two billion parameter open-source reasoning model trained with 49,170 public free-text cases and 104,666 synthetic cases. They report that in a retrospective cohort, RaDaR prioritized the final diagnosis before documented clinical suspicion in 61.06 percent of cases, with a potential lead time of 1.87 months.

00:14:12 In a randomized physician-assistance trial, they say RaDaR improved physicians' rare-disease diagnostic accuracy by 21.44 percentage points compared with internet search alone. That isn't a consumer chatbot story. Rare disease is exactly the kind of medical domain where expertise is sparse, the search space is enormous, and delay can be cruel.

00:14:34 A compact open model that improves physician performance would be significant if the result holds up outside the paper's setting. I am saying that sentence with the conditional fully intact. The paper is new, it is on arXiv, and I could verify the abstract and metadata, not the full clinical detail.

00:14:53 We have the authors' reported trial result, not an independent clinical guideline. Still, the pairing is instructive. Intercept is philanthropic capital trying to reshape prevention incentives around respiratory viruses. RaDaR is model work trying to move diagnosis earlier for a group of patients who often bounce through the system for months or years.

00:15:16 The AI industry is entering health from both ends: through money that funds scientific programs, and through models that claim clinical assistance. The public will feel the benefit only if the funding choices, validation standards, and post-deployment monitoring are stronger than the launch copy.

00:15:35 A second medical paper on Wednesday made a simpler claim: medicine's AI stack needs a log. The paper is called "A global log for medical AI," and it introduces MedLog, a protocol for event-level logging of medical AI. The authors write that modern computer systems have syslog, a universal way to record critical events across different infrastructure, while medicine's growing AI stack has no equivalent.

00:16:01 The practical problem is easy to miss until a system is already deployed. If a model interacts with a doctor, a patient, another algorithm, or an automated workflow, who records what happened? Which model acted? What input did it see? What output did it produce?

00:16:17 What outcome followed, and what feedback came back? The authors propose nine core fields. The fields cover the header, the model, the user, the target, and the inputs. They also cover artifacts, outputs, outcomes, and feedback. That sounds dry until you put it beside the way medical AI is now being sold.

00:16:36 A hospital can buy a deterioration predictor, a documentation assistant, a radiology triage model, a scheduling model, or a patient-message system. Each one may have a different vendor, a different audit trail, and a different definition of a meaningful event. If the hospital later discovers worse performance for a subgroup, a weather-related failure, a drift in local practice, or an adverse event that followed a model suggestion, the absence of a shared log becomes a clinical governance problem.

00:17:08 The paper's abstract describes four deployments in the United States, Switzerland, and Vietnam: ICU deterioration prediction, tetanus progression monitoring from wearable signals, automated sepsis quality reporting, and patient attendance prediction. Two examples stood out.

00:17:25 In one, MedLog captured AI performance degradation during severe weather events in patient attendance prediction. In another, it captured increased laboratory testing after ICU deterioration alerts. Those are exactly the downstream effects that model evaluation before deployment tends to miss.

00:17:44 A model can pass a validation set and still change clinical behavior in ways that cost money, add burden, or produce new risks. The authors also address data burden. They say MedLog limits the footprint through risk-based sampling, lifecycle-aware retention, and write-behind caching.

00:18:01 That matters because a global logging protocol that demands perfect capture of every event in every clinic becomes another system rich hospitals can afford and poorer ones can't. The paper argues it can support low-resource settings, though that claim will need proof in actual hospitals under actual staffing pressure.

00:18:22 I like the direction because it changes the unit of trust. Medical AI vendors often ask to be trusted at the model level: this benchmark, this validation set, this FDA path, this paper. Clinicians and patients need trust at the encounter level. Which version of the model was used?

00:18:39 Which patient did it act on? Which clinician saw the output? Was the suggestion accepted? Did the patient's outcome improve, worsen, or simply become harder to attribute? This also connects back to the rare-disease model. If RaDaR or a successor gets deployed in a clinic, the important record isn't just that a thirty-two billion parameter model once beat internet search in a randomized assistance trial.

00:19:04 The record is whether its recommendations in daily practice shorten diagnostic delay, which patients benefit, which patients get pushed toward wrong workups, and whether clinicians learn to over-defer to it over time. You can't answer those questions from a launch paper alone.

00:19:22 Medical AI is moving from papers into practice. That's good news only if the audit culture matures with it. I would rather have a less dazzling model with a rigorous event record than a brilliant model whose clinical life becomes invisible after procurement.

00:19:38 A paper called "The African Language Tax" put a number on something that usually gets described as access. The author, Olaoye Anthony Somide, measures how commercial and open large language model tokenizers handle twenty African languages across five language families and three scripts.

00:19:57 The setup is important. The paper uses parallel corpora, so the content being measured is the same meaning across languages. That lets the author isolate the effect of tokenization rather than topic or style. Since providers bill per token, latency scales with tokens, and context windows are measured in tokens, the tokenizer decides part of the economic burden before the model reasons about anything.

00:20:22 The headline result is blunt. Across eleven frontier and open tokenizers on the FLORES-200+ corpus, every measured African language carries a tokenization premium above English. On GPT-5's o-two-hundred-k-base tokenizer, the median premium is 1.88 times. N'Ko reaches 8.92 times.

00:20:40 Amharic reaches a 7.4 times generation-latency multiplier in the paper's deployment translation. At the extreme, a language can get as little as eleven percent of English's effective context window for the same token budget. If you use these systems in English, tokenization feels like a technical implementation detail.

00:21:01 If you build a clinical triage line, a bank assistant, or an agricultural advisory service in a high-penalty language, it becomes price, speed, and memory. The same question, the same patient description, or the same farm advisory context costs more to process and gets less room before the model starts losing earlier parts of the conversation.

00:21:23 The paper also gives a mitigation result that keeps this from becoming fatalism. Gemma 4, according to the author, reduces the mean premium from 3.31 times under the older c-l-one-hundred-k-base tokenizer to 2.38 times. That's improvement. It isn't parity. No tokenizer in the study eliminates the penalty.

00:21:43 The study also ships an open measurement tool, a dataset, a leaderboard, and guidance for African builders, which gives teams a way to test the invoice before they commit to a provider. There is a governance angle here that deserves more attention than it gets.

00:21:59 We talk about open and closed AI power through model weights, export controls, identity checks, safety policies, and data-center access. Those are visible gates. Tokenization is a hidden gate. It doesn't refuse service. It just makes the same service more expensive, slower, and shorter for some languages than for others.

00:22:20 If a government ministry, hospital network, school system, bank, or farm service is operating in a language that fragments badly, the provider's published price isn't the price that user experiences. This also complicates the philanthropic and public-health stories from earlier.

00:22:38 It is easy for AI labs and AI-funded nonprofits to say they want global health, global education, and global access. Good. Then the technical substrate has to match that ambition. A tool that is cheap in English and costly in Amharic or N'Ko isn't globally priced in any meaningful sense.

00:22:57 The same goes for latency. A service that feels interactive in one language and sluggish in another will change who keeps using it. This work gives procurement people different questions to ask. What tokenizer does the provider use? What is the premium for the languages our citizens or patients use?

00:23:16 Does the vendor publish tokenization audits? Is there a local-language model whose raw accuracy is slightly lower but whose cost and context behavior make the service viable? Those are procurement questions, which is often where power becomes concrete. So Wednesday's map ends in the ledger.

00:23:35 ICE's AI surveillance stack is visible in contracts. OpenAI's inference ambitions are visible in silicon, power, and partner names. Medical AI is visible in trial claims and in the records clinics will need when the model starts affecting care. African-language access is visible in token counts and latency.

00:23:54 The evidence that would sharpen this is specific: the missing DHS contract detail, OpenAI's promised chip performance report, Intercept's first grants, and tokenizer audits from the companies selling global AI. Jonas.