The compute choke point — and a $2 trillion parallel

00:00:04 Friday evening, the US government sent an export control directive to Anthropic ordering the company to block access to its Fable 5 and Mythos 5 models for all foreign nationals. The order arrived without specific details of the national security concern — just verbal evidence of a potential narrow jailbreak.

00:00:26 So what did Anthropic do? They cut the models off for everyone, including American customers and their own employees. The Verge's Terrence O'Brien reported that the government provided no written disclosure beyond those verbal claims. In its statement, Anthropic says they haven't even received documentation of a concerning non-universal jailbreak leading to actual harm.

00:00:52 Their position is that the jailbreaks shown are either benign or minor, provide no Mythos-specific uplift, and look identical to capabilities seen in rival models like GPT 5.5. The Guardian, citing Reuters reporting, adds another layer. Pentagon CIO Kirsten Davies publicly backed the directive on X, calling it about things more important than revenue cycles and clickbait.

00:01:19 The article notes this marks a shift in export control strategy — for years, US restrictions focused on chips and infrastructure rather than model access itself. It also comes right after Anthropic confidentially filed its IPO paperwork last month, edging ahead of rival OpenAI in the race to reach public markets.

00:01:41 The real question here is practical. How much evidence justifies pulling a model from hundreds of millions of users? Who sets that threshold? Anthropic says this doesn't align with fair regulation — the company called for greater US oversight of AI earlier this week, including blocking models with unacceptable risks, but noted Friday's action didn't meet that standard.

00:02:07 The government hasn't published its own reasoning beyond those initial verbal claims. A US official confirmed the Commerce Department issued the directive, covering Fable 5 and Mythos 5 for all foreign nationals without specific details of the security concern.

00:02:26 The order also notes an understanding that the government believes there's a method to bypass a safeguard designed to prevent Fable 5 from identifying software vulnerabilities. The baseline is simple: on Friday evening, someone in Washington called it national security enough to suspend access to two of the most capable models available today.

00:02:51 No documents were produced as evidence. The models are down for everyone right now.

00:02:56 Ethan Mollick at the University of Pennsylvania has been thinking through the infrastructure implications of this from a different angle. His thread on X frames it as a compute story rather than a policy one — and that shifts the timeline for how this plays out.

00:03:14 Mollick's argument starts with a simple observation: you can't build a Mythos-class model without a very regulatable compute footprint. Training that kind of model uses so much power and chips that national governments will obviously notice. Nobody is training at that scale without permission, in his reading.

00:03:35 He connects this to the open weights question — many assumed a government order like this would push more models into open source, since companies would lose trust in centralized controls. Mollick disagrees with that expectation. If Mythos-level models are considered dangerous by one government, others — including China's — will want them kept closed.

00:03:59 The mechanism he's describing is straightforward. Compute requirements for frontier model training create natural choke points: data center power budgets, GPU supply chains, and cloud provider capacity allocation. These are things that can be monitored, taxed, or restricted by any government with jurisdiction over the physical infrastructure.

00:04:22 The weights themselves could theoretically leak, but building a comparable model from scratch requires coordination at industrial scale that's much harder to conceal. There's a practical question about enforcement that Mollick doesn't fully address. You can control compute at the supplier level — AMD, NVIDIA, cloud providers — but you can't easily stop someone in a non-cooperating jurisdiction from importing enough hardware to train a comparable model.

00:04:53 One reply in his thread asked whether the compute footprint for Mythos-class models is comparable to the pre-training of Opus 4.8, which would set a baseline for understanding the scale involved. That question about actual requirements is worth pursuing. The Guardian reported that experts said Mythos models could dramatically accelerate sophisticated cyber-attacks, particularly in sectors like banking that rely on complex, interconnected, and often decades-old technology systems.

00:05:25 If those claims hold, the compute footprint needed to train a model with that capability would be massive. What Mollick nails is the asymmetry this creates for frontier capability distribution. Governments controlling power infrastructure and semiconductor manufacturing have structural leverage that doesn't require writing new laws.

00:05:48 They just need to notice what's happening at their data centers and ask a few questions. The Anthropic order targets model access rather than compute directly — which is unusual. But the infrastructure angle explains why such an order might actually stick, even without detailed evidence of a specific exploit.

00:06:09 If building these models requires governments' attention anyway, regulating who gets to use them becomes much easier to enforce than open weights.

00:06:19 While the AI world processed the government directive, another story closed out its trading day on Friday — one worth holding alongside it. SpaceX hit the Nasdaq with a $2.1 trillion valuation. The IPO closed at $135 per share, up 19% from the offer price. CNBC reported that the company raised $75 billion — triple the size of Alibaba's record-setting 2014 debut.

00:06:46 SpaceX is now the sixth most valuable US company despite carrying a fraction of the revenue found in any tech megacap. Two details stand out from the reporting. First, Musk retained above 82% voting power in the company. At Facebook's IPO in 2012, Zuckerberg controlled 56%.

00:07:06 While both are founder-controlled, SpaceX is on another level entirely. More than 500 million shares changed hands throughout the day, approaching the volume of Facebook's market debut. Second, the pricing wasn't based on standard market forces. CNBC quoted investment banker Lloyd Greif saying it was a deal driven by what one man wanted.

00:07:32 There was no price range, no traditional haggling with shareholders — just a take-it-or-leave-it offer that the market accepted enthusiastically. The roadshow was abbreviated. The revenue multiple is striking: 112 times last year's revenue. SpaceX had $4.9 billion in losses last year and far less revenue than the trillion-dollar companies it joins by market cap.

00:07:59 Stock closed at $2.1 trillion, making Musk the world's first trillionaire on paper. Musk told staff at SpaceX's Texas headquarters on Friday morning — weeks before his 55th birthday — that he gave the company less than a 10% chance of succeeding in those early days.

00:08:19 He added that if anyone had told him this was going to happen back then, he would have thought they must be smoking some really good crack. The company has grown to 22,000 full-time employees since founding in 2002. The offering reportedly minted some 4,400 millionaires among current and former SpaceX employees.

00:08:43 Former Nasdaq CEO Robert Greifeld said he would definitely bet that OpenAI and Anthropic go public in 2026 — both companies announced they confidentially filed their IPO paperwork last month. The timing sits right alongside the Anthropic story: this happens on the same day as the export control directive, just after both AI model giants filed their own paperwork.

00:09:10 The capital concentration at play here — one founder controlling a quarter of all trillion-dollar AI company equity by voting rights — mirrors the government's attempts to restrict access to frontier models. We're looking at two different kinds of leverage: one rooted in who controls the silicon, and another in who controls the money.

00:09:36 A paper published in IEEE Robotics and Automation Letters earlier this month offers some useful contrast. It tackles a smaller-scale question than government export controls or trillion-dollar IPOs, but it cuts to something real about what these systems can actually do.

00:09:54 Seung Chan Hong, then an undergraduate at the University of Melbourne, led a study training collaborative robots to read human emotions using vision language models — architectures similar to the frontier models we've been discussing, just scaled down for a specific physical task.

00:10:13 The researchers wanted to know if robots could adjust their behavior based on what humans were feeling during shared tasks. Hong notes that while there's a lot of hype around advancing robot physical abilities, innovation needs to happen in how they interact with humans, not just what they can lift or move.

00:10:34 To train the VLM, the team had volunteers watch videos of robots handing over objects — with varying degrees of success — and describe the emotions the humans were expressing. The volunteers factored in broader context rather than reporting solely on facial expressions.

00:10:53 For example, a person pausing to think with a furrowed brow may simply be concentrating on the task at hand, not necessarily angry. Contextual factors like drumming fingers or pursing lips clarify exactly what that furrowed brow means. The VLM outperformed conventional AI systems relying on standard facial analysis and object tracking.

00:11:16 On a scale from zero to one, the conventional system scored 0.77 while the VLM achieved 0.86. In a second experiment with 40 volunteers interacting directly with a robot — purposefully programmed to make an error — the machine offered either an emotionally adaptive apology or a pre-scripted standard one.

00:11:36 Thirty-one of the 40 participants preferred the adaptive version, which accounted for the human's perceived response to the mistake. But the paper stays honest about its limits. After collaborating with a robot that failed in its task, many participants ranked their trust in the machine lower regardless of how it apologized.

00:11:58 The researchers put it this way: a personalized apology acts as social lubricant, but it can't repair the trust lost when the robot fails its physical job. The VLM also didn't always match what people reported about their own emotions — just what third-party observers saw from the outside.

00:12:18 Hong says the system matched human observers well but didn't always align with self-reported feelings during interaction. While the VLM is a strong observer of outward social cues, it isn't reading minds. The researchers found that emotional adaptivity was far less important to users than the robot's actual functionality.

00:12:40 Competence matters more than calibration, even when people appreciate getting the latter. This provides useful context for the Anthropic story we started with. We're talking about models that can be blocked or unblocked at government order, while the systems themselves are still struggling with one of the most basic social tasks — knowing when someone is actually upset versus just concentrating.

00:13:07 The frontier capability gap between processing massive data and actually understanding what it sees might be wider than anyone assumes.

00:13:17 We're looking at two different kinds of leverage converging on the same question: who gets to build at frontier scale? One answer is written in silicon and power contracts. The other in voting rights and market caps. The Anthropic directive arrived without documentation — just a verbal claim from a government that's been slow to explain its reasoning.

00:13:40 The SpaceX IPO closed with a price set by one person, unanchored from traditional metrics of value. They happen on the same day. And between them, a paper about robots and emotions reminds us that what these systems do reliably is still narrower than we tend to assume when we talk about "frontier" capability.

00:14:00 I'll leave that gap between us — the distance between processing at scale and actually understanding. Seln Oriax