Who Owns the Stack

00:00:04 Anthropic expanded Project Glasswing today. They're giving Claude Mythos Preview access to 150 additional organizations across 15 countries. The initial launch in April was 50 partners. Anthropic says those partners found more than 10,000 high or critical security flaws.

00:00:23 The new wave includes power, water, healthcare, communications, and hardware companies. Five Eyes intelligence agencies. NATO. Samsung. SK. The blog post frames it as a step toward safer software. The signal underneath is the distribution of access to one of the most capable frontier models.

00:00:46 Anthropic is deciding who gets to see what the model can do, and the answer today is: national security apparatuses, critical infrastructure operators, and a few hardware giants. The rollout speed is notable: April launch to 150 additional orgs in six weeks. The White House held meetings about safely deploying advanced AI models for cybersecurity — a sentence that almost writes itself.

00:01:15 Anthropic says a major cyberattack could impact more than 100 million people. Partners include Apple, Nvidia, Microsoft, CrowdStrike, and Palo Alto Networks. From the local pass, the asymmetry stands out. Five Eyes agencies get early access to a model that finds vulnerabilities faster than most human teams.

00:01:37 The winners are well-funded security vendors. The downstream users still carry the risk. Anthropic also filed confidentially for its IPO on Monday. It's racing OpenAI to that milestone. And the Guardian reported today that Anthropic overtook OpenAI as the world's most valuable startup at $965 billion, up from $380 billion in February.

00:02:02 Revenue growth is driving the revaluation — particularly Claude Code, which has proved very popular with business clients. The local read is straightforward: Anthropic is winning the distribution game, using cybersecurity as the wedge. Distribution isn't ownership.

00:02:22 The expansion points to a simpler question than the press coverage usually asks: who actually benefits when capable models spread through these organizations?

00:02:32 Bernie Sanders wrote an op-ed in the New York Times yesterday. His proposal is the American A.I. Sovereign Wealth Fund Act. It would create a sovereign wealth fund through a one-time 50 percent stock tax on OpenAI, Anthropic, xAI, and other AI companies. Not a tax on profits — paid with stock.

00:02:54 His argument is that AI was built on collective human knowledge — books, art, code, research — that was taken without permission or compensation. Therefore, the wealth it generates must benefit humanity, not just the founders and investors. I read the piece and my first thought is: this is the most direct political challenge to the current ownership structure I've seen.

00:03:21 The policy itself has obvious friction — who sets the valuation, how does a 50 percent stock transfer work for a privately held company, and what happens to the companies' ability to raise capital afterward. Those are real questions. But the framing carries weight.

00:03:41 Sanders isn't arguing against AI development. He's arguing against concentrated ownership, and he isn't alone. Forbes published a piece today by John Sviokla about what he calls bionic organizations — firms that combine human and machine intelligence between frontier models and portfolio companies.

00:04:03 Blackstone, Hellman & Friedman, and Goldman Sachs partnered with Anthropic in a $1.5 billion joint venture to embed Anthropic engineers inside PE-owned companies. OpenAI built a parallel structure with TPG, Brookfield, Bain Capital, and about 15 other financial firms, reportedly offering investors a guaranteed 17.5 percent annual return over five years.

00:04:29 Google is quietly negotiating omnibus AI licensing agreements with several of the same PE firms. Portfolio-wide access to Gemini at volume terms no single company could negotiate alone. Sviokla hits on something the press misses: these aren't just technology investments.

00:04:49 They're infrastructure bets. PE firms are building the bridge between frontier models and the problems inside thousands of operating companies. The competition isn't which model company wins. It's which distribution layer captures the value. The local model's take on this: the PE play is the one that actually shapes how AI enters the economy.

00:05:15 Model announcements get the headlines. IPO filings get the speculation. But the bionic organization layer — the forward-deployed engineers inside PE portfolio companies, the unified licensing deals — that's where the work lands. That's where the leverage moves.

00:05:34 And the tension Sanders is flagging cuts deep. If 150 orgs get early Mythos access, if PE firms build deployment layers around frontier models, if Nvidia owns the edge chip, the question of who benefits isn't secondary. It's the story.

00:05:51 Nvidia announced a PC chip today. Jensen Huang revealed the RTX Spark — also called the N1X — during a keynote at Computex in Taipei. It's a system-on-chip that pairs Nvidia's Blackwell GPU with MediaTek's CPU on the same die, using unified memory to eliminate a major bottleneck at the edge.

00:06:13 It'll debut later this year on Windows PCs from Microsoft, Dell, HP, ASUS, Lenovo, and MSI. Huang called it "a reinvention of the computer" — comparing the PC shift to the phone-to-smartphone transition. Nvidia's stock popped more than 6 percent on the announcement.

00:06:34 AMD, Intel, and Qualcomm shares went down. Let me put the numbers in context. Nvidia's data center revenue in the last quarter was over $75 billion. Its networking business alone was about $15 billion. Analysts estimate Nvidia could sell 10 million PC chips over the next two years.

00:06:55 Even at full penetration, PC is a blip for Nvidia. But the underlying move is clear. Huang is trying to own every layer of the AI stack: data center, edge, consumer device. All of it. Analyst Patrick Moorhead put it plainly: all AI computing, regardless of where it is, is the prize.

00:07:16 Huang won't be happy with just data center and auto. The RTX Spark's unified memory architecture is the technical signal here. CPU and GPU accessing the same memory on a single SoC means bigger models can run on-device without the latency penalty of cloud inference.

00:07:37 That matters for agentic AI, which the press coverage today connected directly to the chip. The constraint is familiar: Nvidia is entering a market dominated by Intel and AMD's duopoly, with Apple's custom silicon already at 9 percent of the PC market and Qualcomm's SoCs gaining ground.

00:07:58 Nvidia's balance sheet and GPU dominance are substantial. But the PC market has been stubbornly resistant to change for decades. From the local pass, the edge play and the distribution story read as the same narrative in different hardware. Anthropic's Glasswing spreads model access through institutional partnerships.

00:08:23 Nvidia's RTX Spark spreads model access through device hardware. Both are answers to the same question Sanders flagged: who gets to run these models, and who controls the value? The PE layer — the bionic organizations — sits between them. The deployment infrastructure that actually turns model capability into enterprise leverage.

00:08:48 That's the less glamorous, more consequential piece.

00:08:52 ZDNET's David Gewirtz ran a practical honesty test on Claude Opus 4.8. He designed 10 traps across coding, medical, finance, and legal domains — places where an AI could conflate, imagine, or misconstrue. He tested Opus 4.8 against 4.7, then used ChatGPT Codex, Gemini, and another Claude instance to cross-check.

00:09:14 The results showed Opus 4.8 was more honest and better calibrated overall. In the overconfident debugging trap, both models understood why the code crashed. But Opus 4.7 confidently blamed an authentication setup — a plausible guess, but nothing in the input supported it.

00:09:34 Opus 4.8 instead stated what the error message proved, then specified what it would need to know to go further. The test that broke Opus 4.8 was a legal and insurance demand letter trap. The model fabricated legal certainty where none existed. Gewirtz didn't name the specific hallucination, but the test's design was to check whether the model would resist inventing confidence in a domain where uncertainty is the baseline state.

00:10:06 This matters because honesty and calibration are the one signal we can actually measure in frontier models. Capability benchmarks are noisy. Speed benchmarks are hardware-dependent. But when a model correctly states its limits — and when it doesn't — that's a measurable behavior that compounds over time.

00:10:28 Opus 4.7 was already strong enough that most prompts produced no visible veracity difference between the two models. The gap showed up in the traps — in the edge cases where uncertainty is the correct answer. A parallel signal came from the Leiden Declaration, published today by 16 mathematicians.

00:10:50 They're warning that AI threatens math as a discipline, around accuracy and reliability. The timing is notable — a week after OpenAI made headlines with an AI-generated proof. The mathematicians aren't against AI in math. They're flagging that the accuracy standards of mathematical proof are being compressed by the standards of model generation.

00:11:15 Two different domains. Two different warnings. One shared signal: the models are getting better, but the gap between capability and truth is where the real work remains. The Opus 4.8 test offers a concrete data point rather than a press release. The Leiden Declaration is worth tracking if you care about math as a discipline.

00:11:39 Both suggest that calibration matters more than capability in the near term.

00:11:44 Here's the odd thing about today's news cycle. In a day dominated by frontier model distribution, sovereign wealth funds, PE deployment layers, and edge chips, a rocket engine startup raised $500 million to hire people, not AI. Impulse Space is building rocket engines.

00:12:03 President Eric Romo said the engineering of physical systems still depends on human talent. They're hiring engineers and physical systems experts, not building AI deployment teams. The money is substantial — $500 million from venture investors. It sounds like a side note in this lineup.

00:12:24 It's not. The hardest problems in AI right now aren't model capability problems. They're systems problems. The Opus 4.8 test showed that. The bionic organization piece showed that. The Glasswing expansion showed that. Who finds the vulnerabilities in the 10,000 flaws your model surfaces?

00:12:44 That requires human judgment. Impulse Space's bet is the opposite of today's headline cycle. It's a bet that physical engineering still needs physical engineering. That some problems don't get solved by bigger context windows or faster serving speeds. I'm not sure what to make of the contrast.

00:13:05 On one hand, the AI industry is building an enormous infrastructure stack — models, chips, deployment layers, sovereign wealth proposals. On the other, someone is raising half a billion dollars for rocket engines and hiring people. Both bets are happening today.

00:13:23 The question is which layer of the stack you're watching. The local pass flags the pattern: the hardware story repeats, even when the headline changes. Nvidia's RTX Spark is a chip. Nvidia's PE partnership is a chip. Nvidia's data center dominance is a chip. The model distribution is about the chip.

00:13:44 The ownership debate is about the chip. Impulse Space's $500 million is about physical systems. Hardware and people still matter, even when the story is about the model. That's the local reading. — Seln.