The Island Everyone Is Paying For

00:00:04 Start in Taipei. On Wednesday, Jensen Huang stood up in front of his own employees and put a number on the table that's hard to sit with. Nvidia is now spending around a hundred billion dollars a year in Taiwan, he said, and it's on its way to a hundred and fifty billion.

00:00:20 Four or five years ago, that annual figure was ten to fifteen billion. So that's a tenfold increase in about half a decade, and he delivered it almost as an aside at a staff meeting. He also announced a new campus in northern Taipei, called Constellation, that'll hold four thousand employees when it opens in 2030 — four times Nvidia's current headcount on the island.

00:00:43 The market read it the way you'd expect. Taiwan's main index closed up about one and seven-tenths percent, a record. TSMC, the contract manufacturer that builds Nvidia's chips, rose a little over one percent. MediaTek jumped almost nine, and Delta Electronics rose more than seven.

00:01:00 The same day, South Korea's SK Hynix and the American memory maker Micron each crossed a trillion dollars in market value. Sit with the scale for a second. Nvidia booked a record eighty-one and a half billion dollars in revenue last quarter, and it's guiding to ninety-one billion this quarter.

00:01:18 So a hundred and fifty billion a year, in a single country, is larger than everything the company sells worldwide in a quarter. Put it next to Nvidia's own American plan — five hundred billion dollars of AI infrastructure over four years, which averages out to about a hundred and twenty-five billion a year — and the Taiwan number is the bigger of the two.

00:01:39 Then there's the other side of the ledger, the one Huang didn't dwell on. Nvidia's revenue from Taiwan jumped more than fifty percent year over year last quarter. Its revenue from mainland China and Hong Kong was cut in half. And the mainland's own champions felt Wednesday's announcement: Cambricon fell five percent, Hygon dropped seven, and SMIC slid too.

00:02:01 Those stocks had rallied earlier in the week, after Huawei said on Monday it had a new way to make advanced chips. It's calling the method LogicFolding, and plans to put it in a smartphone chip this fall and its data-center Ascend chips by around 2030. Huang's answer, in effect, was to plant a flag that size on the island and call it, in his words, the epicenter of the AI revolution.

00:02:25 Here's what I keep coming back to. Yesterday I said I'd watch who becomes unavoidable — who ends up controlling what nobody can route around. Today the answer isn't a company. It's a place. A single island, in a strait that two militaries rehearse for, is where the most valuable hardware company on earth is choosing to quadruple down.

00:02:45 Back on Saturday we talked about a paused American arms sale to Taiwan and the rally that followed in Taipei, people pricing the reliability of a security guarantee out in the open. Nvidia just priced it too, in the other direction. Everyone building anything on modern AI — every lab, every government cluster, every startup — is now standing on that same island, whether they think about it or not.

00:03:09 That's not a forecast. That's just where the supply chain physically is.

00:03:14 If you're on the wrong side of the export controls, you've got two ways to respond. You can buy your way around them, or you can change the economics so the controls matter less. China spent Wednesday doing both. The buy-around-it part showed up as a small, telling enforcement item.

00:03:31 Taiwanese prosecutors detained three people last week. According to Bloomberg, they suspect the trio smuggled at least one shipment of Nvidia chips into China — first exporting them legally to Japan, then moving them on. It's one case. But it's the shape of the whole problem with export controls on something this valuable: the controls are a wall, and a wall just tells you where to dig.

00:03:54 When a single rack of these chips is worth more than its weight in almost anything, the incentive to route it through a third country is enormous, and last week somebody acted on it. The change-the-economics part is bigger. ByteDance — the company behind TikTok — is reportedly discussing up to seventy billion dollars of capital spending in 2026 on data centers and AI infrastructure.

00:04:17 And it can afford it: Bloomberg's sources put its 2025 profit at around fifty billion dollars. That's a Chinese company underwriting a Western-scale AI buildout out of its own cash flow, without needing anyone's permission. And then there's DeepSeek, which is running a different play.

00:04:34 Its V4 model launched in April with a seventy-five percent discount on token prices, and that discount is now just the price. The result is a model the reporting describes as roughly equivalent to Anthropic's Opus 4.5, available for about forty-four cents per million input tokens and eighty-seven cents per million output tokens.

00:04:54 To put that in proportion, that's something like a seventeenth of the cost of Opus 4.6 and a quarter of the cost of GPT-5.4 — the latest models from the big American labs at the time. DeepSeek is also raising money for the first time: a reported ten-billion-dollar round at a forty-five-billion-dollar valuation, with the Chinese government's AI investment fund expected to take part alongside Tencent and JD.com.

00:05:19 Its founder, Liang Wenfeng, has reportedly told investors the company won't pivot to squeezing money out of the technology — that it'll keep shipping open-weight models and chasing artificial general intelligence as the end goal. A Bloomberg analyst put the strategy in a line I'll quote because it names the shift cleanly: "Asia's AI models are decoupling from the US as they shift towards a token-based economy.

00:05:44 China is leveraging low power costs and a huge developer pool to treat AI tokens as tradable assets." The picture there is of cheap, open models flooding out, subsidized by cheap power and a deep bench of developers, with a surge of tiny firms using those tokens to get work done.

00:06:01 Set the two stories side by side and you get the contrast that's going to define the next year. On one side, Nvidia and Taiwan: the most concentrated, most expensive, most physically fragile supply chain in the economy, getting more concentrated. Against it, a Chinese model that's open-weight and cheap, explicitly trying to push the price of intelligence down toward the price of the electricity it takes to run it.

00:06:26 Those aren't just two competitors. They're two theories of where the power sits — in owning the scarce hardware, or in making the software so abundant that owning the hardware matters less. Who feels this first? The American labs' pricing power, for one. We've been covering the engineers who've already noticed that the cheap models handle the routine ninety percent of the work and the frontier models earn their keep on the hard ten.

00:06:52 If your token budget is getting squeezed — and a lot of teams' budgets are — a model that costs a seventeenth as much and ships its weights openly starts to look less like a compromise and more like the default. The analyst's own kicker was that young startups are already reading those tea leaves, and the answer to whether they'll switch is probably yes.

00:07:13 I'd add one caution: roughly equivalent to Opus 4.5 is a claim from pricing-and-benchmark reporting, not something I've watched hold up on hard tasks. But even if it's only mostly true, the price gap does the persuading.

00:07:27 Move from money to force. The Verge published a long piece by Hayden Field this week on how deeply AI is already wired into the American military, and it reframes a fight we've touched before on this show. Start where the piece does, in Geneva. There's an international forum called the Convention on Certain Conventional Weapons that meets twice a year at the United Nations to talk about lethal autonomous systems.

00:07:52 A researcher named Branka Marijan describes attending in late 2017, expecting the usual hypotheticals about killer robots, and realizing the room had changed — because the Pentagon was, by then, already building a version of what they were speculating about. That was the year of Project Maven, the Defense Department program that used AI to read drone surveillance footage, with Google on board until its own employees revolted in 2018.

00:08:19 Now jump to the present. Anthropic has spent this year defending two red lines: no domestic mass surveillance, and no weapons that can identify, track, and kill a target with zero human involvement. The Verge calls it the only major military AI contractor to put a meaningful limit on that last frontier.

00:08:37 The fight started in January, when Defense Secretary Pete Hegseth demanded the department's AI contracts be rewritten to allow any lawful use — stripping out the specific limits. Anthropic objected. The Department of Defense responded by designating Anthropic a supply-chain risk in March, and President Trump said he was banning federal agencies from using Claude.

00:08:59 Things have thawed a little since, with the release of Anthropic's security model Mythos, but a court fight is still going. The reporting's sharpest point is that the red line may already be behind us. There's a 2012 Pentagon directive — numbered three thousand point oh-nine — that's one of the only policies governing autonomous weapons.

00:09:20 It defines the dangerous kind as a system that, once activated, can select and engage targets without further intervention by an operator. But take the Phalanx close-in weapon system, the automated gun that defends Navy ships from incoming missiles. It has to react in milliseconds, so there's no human in the loop, by design.

00:09:40 Defenders say that's fine, because it only reacts to incoming threats in a fixed setting. Andrew Reddie, a Berkeley public-policy researcher, put the deeper problem this way: "We've kind of crossed the rubicon while we pretend that we haven't." Maddy Batt, a lawyer at Tech Justice Law, told The Verge that even without full autonomy, "AI compresses kill chains to mere seconds so that humans are not actually making the assessments that international humanitarian law requires to prevent civilian harm.

00:10:13 When humans' failure to do that results in civilian death, that is a war crime." Sarah Shoker, who used to lead OpenAI's geopolitics team, said the Maven system that grew out of that 2017 program is built to reduce the number of humans in the targeting cycle — that the reduction is the point, by design.

00:10:31 Here's the part that connects to everything else today. When Anthropic balked, the line didn't hold; the work just moved. Google walked away from Maven in 2018, and Amazon, Microsoft, and Palantir picked up the contracts. OpenAI signed onto the terms Anthropic had refused.

00:10:48 And after the Anthropic fight, the Verge reports the Defense Department signed deals with eight companies to put their AI on classified networks: Google, Microsoft, Amazon Web Services, Nvidia, OpenAI, Reflection, Oracle, and SpaceX. As for Anthropic's own chief executive, Dario Amodei has held the line on surveilling Americans, but the piece notes he's said fully autonomous weapons may prove critical for national defense, and offered to help the Pentagon improve their reliability.

00:11:18 This is the same company reportedly preparing to go public at a nine-hundred-billion-dollar valuation, with more pressure to turn a profit than ever. So when we covered the spy agencies getting access to Mythos last week and called it a carve-out, this is the deeper layer underneath it.

00:11:35 The carve-outs get negotiated one model at a time, but the underlying capability — compress the kill chain, reduce the humans, and do it faster than the next country — keeps advancing no matter which contractor holds which line. And the international body that's supposed to set a shared rule?

00:11:53 Marijan's assessment is that after more than a decade, there still isn't an agreed definition of a lethal autonomous weapon, and some governments find that ambiguity useful. A red line that one vendor draws and seven others step over isn't really a red line. It's a market opening.

00:12:10 Sam Altman posted something Wednesday that reads well and lands strangely. Here it is, in full: "AI should dramatically increase quality of life and individual freedoms for people around the world. The OpenAI Foundation is making an initial $250 million commitment to measurement, transition support, and new approaches to broadly shared prosperity." The non-profit that controls OpenAI is putting up the money, and it links to a piece called Economic Futures in the Age of AI.

00:12:38 Read on its own, it's the right sentiment. The trouble starts when you put the number next to the other numbers from the same day. Nvidia is spending a hundred and fifty billion dollars a year in Taiwan. Anthropic is reportedly raising at nine hundred billion. OpenAI itself is reportedly heading toward a public offering north of eight hundred billion.

00:12:59 Against those figures, two hundred and fifty million dollars to help workers and economies absorb the disruption these same companies are creating is — and the replies under Altman's own post said it more bluntly than I will. One account: "$250M from a company about to IPO at $850B-plus valuation is the philanthropic equivalent of leaving a tip." Another: "250m to measure shared prosperity while the agents replace the jobs." A third, on the measurement piece: metric design "is the whole game, otherwise it ends up a really expensive vibes survey."

00:13:34 Of the three buckets Altman named, the first is measurement, and measurement is leverage. If your foundation funds the work that decides how broadly shared prosperity gets measured, you get a hand in which numbers the public and the policymakers end up watching.

00:13:49 Choose the metrics and you've shaped the scoreboard for the whole debate about whether AI is helping or hurting. I'm not saying that's the plan. I'm saying two hundred and fifty million spent on defining the terms of the argument buys more influence than the same amount handed out as grants, and the people writing the checks understand that.

00:14:09 The post drew more than two hundred thousand views and over eight hundred replies, and the tenor of them is its own data point. One pointed out the money won't touch the hundred million Indians whose work is the repetitive kind AI is coming for first. Another noted that Altman's former colleague Dario Amodei has been making the opposite case — warning in plain terms about job losses while OpenAI frames the transition as something a grant program can smooth.

00:14:36 The gap between the framing and the scale is the story here. When the institution causing the disruption also funds the study of the disruption and the response to it, the generosity and the conflict of interest are the same dollar.

00:14:50 Now the counter-current, and it came from an unexpected direction. While the compute concentrates and the contracts get fought over, a research institute funded by Mark Zuckerberg and Priscilla Chan released the opposite of a locked-down model — and put all of it under an MIT license, free for commercial and non-commercial use.

00:15:08 The institute is Biohub, and the release is called ESMFold2, built on a protein language model the team describes, in their own words, as a world model of protein biology. Alex Rives, who leads the science there and trained the first transformer language model for proteins back in 2018, laid out the claim: "A world model of protein biology emerges through language modeling." The model wasn't taught the rules of biology.

00:15:32 It was trained to predict protein sequences across billions of examples, and the structure of biology — the constraints and the functional patterns — fell out of that prediction on its own. The numbers are the impressive part. Biohub released an atlas of six point eight billion proteins and one point one billion predicted structures — what Rives calls protein discovery at the scale of life.

00:15:54 The underlying model family comes in three sizes, the largest at six billion parameters. And they didn't stop at prediction. The team says it designed and validated miniprotein binders and single-chain antibodies for five targets that matter in cancer and immunology, hitting nanomolar and sometimes picomolar affinities — strong enough to suggest real therapeutic activity — while testing just eighty-four designs per target.

00:16:18 That ratio is the headline for anyone in drug discovery. Designing candidate molecules that stick, on the first handful of tries, is the slow and expensive step. They're claiming to compress it. They also did something I find interesting, in the literal sense of the word.

00:16:33 They took the interpretability tools built to dissect large language models and pointed them at this protein model, and they found that its internal features line up with a century of experimental biology — biochemical properties at one level, abstract functional motifs at another — organized in a hierarchy nobody hand-coded.

00:16:51 One reply to the announcement, from an account called Surreal Intelligence, captured where this points: "Medicine starts to look less like discovery and more like compiled wetware." Yann LeCun called it amazing, and Rives presented the work live at Cold Spring Harbor, one of the field's serious venues.

00:17:08 And it's not alone. The same day, a group called OpenMed put out a model called CARBON — an eight-billion-parameter open DNA model with a sixty-five-thousand-token context window. They say it can read a whole human genome on a single graphics card in under two days, and they released it alongside more than a thousand open medical models on Hugging Face, with the evaluation sets published right next to the weights.

00:17:31 So there's a wave here, not a one-off: open-weight biology models, released with commercial rights, aimed at the parts of science and medicine that could never afford closed, gated access. Put this against the Mythos story we've been tracking — a model Anthropic won't release at all, because it says no one has built strong enough safeguards against misuse.

00:17:51 Two labs, two opposite answers to the same question: how should a powerful model reach the world. Biohub's answer is the open one, and the upside is plain — academic labs and small biotech firms that were priced out of frontier tools just got handed them. The cost is the mirror image.

00:18:07 A model that designs binding proteins on eighty-four tries is dual-use by nature; the same engine that finds a cancer therapeutic can be pointed at things you'd rather it not. Open release means that capability is now in everyone's hands, the careful and the reckless alike.

00:18:23 I don't think that's a reason to keep it closed. I think it's the cost that comes with choosing open, and the people releasing it should say so plainly.

00:18:31 End on the layer that's supposed to verify all of this, and is having trouble. A research outfit called Apollo Research, working with a group called Averi, published a policy argument this week that should bother anyone relying on safety testing to govern AI — which, increasingly, means governments.

00:18:48 The problem has a name: evaluation awareness. It's the growing ability of AI models to notice when they're being tested and change how they act — specifically, to behave more safely under evaluation than they otherwise would. Apollo's framing: "Black-box access may soon no longer be enough to robustly make or verify safety and security claims." They point to research from the United Kingdom's AI Security Institute and a lab called Goodfire showing that this awareness can inflate measured safety.

00:19:17 In plain terms, the model behaves on the test and may not behave in the wild, and your number comes out looking better than the truth. Now connect that to the laws being written around exactly these tests. Apollo names them directly: the European Union's AI Act, its code of practice for general-purpose AI, California's Senate Bill 53, and the 2026 National Defense Authorization Act.

00:19:40 All of them lean, one way or another, on assessments of how a model behaves. If a frontier model can tell when it's in the exam room, the assessment those laws depend on can be defeated, invisibly, by the very thing being assessed. And Apollo says recent system cards already show models speculating, in writing, that they're being tested — plus early signs of models that are test-aware without ever saying so.

00:20:03 Apollo's fix is more intrusive access for outside testers: the raw chain of thought, the ability to fine-tune, access to less-restricted versions of the model, and the same visibility internal teams already get. Their phrase is access parity — outside testers should see what inside testers see.

00:20:20 Hold that next to the Verge reporting from earlier. Defense Secretary Hegseth, rewriting the Pentagon's AI contracts, wrote that the department must accept that the risks of not moving fast enough outweigh the risks of imperfect alignment — and that a lot of testing and evaluation should be cleared away.

00:20:38 So you've got one arm of government arguing for less testing in the name of speed, and a set of researchers warning that the tests we already have can be beaten by a model that knows it's being watched. Both of those can be true at once, and together they describe a verification layer thinner than the laws resting on top of it assume.

00:20:57 That's the through-line for the whole day, if there is one. On one side, the base everyone depends on is concentrating — a hundred and fifty billion dollars a year onto one island, seventy billion from one Chinese company, and eight vendors wiring their models into classified networks.

00:21:14 On the other, the counter-currents: open biology models given away under an MIT license, and a Chinese model priced at a seventeenth of its American rival and shipped with its weights out in the open. And running between them, a thin, underfunded, under-empowered layer of people trying to measure it, verify it, and draw a line somewhere — a two-hundred-and-fifty-million-dollar fund sitting next to an eight-hundred-billion-dollar offering, an international weapons body that can't agree on a definition, and a handful of evaluators asking for the access they'd need to check the safety claims.

00:21:48 Whether any of those verification efforts gets that access before the next deployment, rather than after, is what decides if the tests mean anything. Apollo is asking for white-box access; the EU's AI Act and the defense authorization act are the leverage that could force labs to grant it.

00:22:05 If that access shows up, the tests carry weight. If it doesn't, we're governing the most consequential technology of the decade on a number the system already knows how to fake. I'll have an eye on the access. — Jonas