2026-06-27 — Who Controls the Most Powerful AI in the World?

Transcript

Chapter 1

Nova: Today, June 27th, 2026 — the White House pulls frontier AI offline, then blinks, partially restoring access to Anthropic's Mythos 5 for over a hundred US organizations. OpenAI answers with a defiant limited launch of GPT-5.6 'Sol' less than a day after the government asked them to wait. Meanwhile, China's Zhipu is closing the benchmark gap at one-fifth the price, Europe is racing to build AI it actually owns, and OpenAI just revealed a custom inference chip called Jalapeño that could redraw the hardware map entirely.

Ray: One question ties all of it together: who actually controls the most powerful AI in the world right now? Governments think they do. Labs think they do. And at least one Chinese competitor is quietly making the argument that the answer might be neither.

Chapter 2

Nova: The Verge reports that after a Friday-evening ultimatum from the Trump administration took Anthropic's Mythos-class models offline, two weeks of tense negotiations ended in a partial win — the White House authorized more than a hundred US companies and government agencies, including their non-American employees, to access Mythos 5 again. Commentators are already calling it a signal that model capabilities now carry direct political and national-security consequences.

Ray: Partial win is doing a lot of work there. The administration issued a Friday-evening ultimatum, pulled a frontier model offline for two weeks, and then selectively decided who gets back in. That's not a compromise — that's a demonstration of leverage. The hundred-plus organizations that got access are the ones that negotiated hard enough or mattered enough politically. Everyone else is still locked out.

Nova: The practical consequence for the industry is real though. If you're an enterprise or a developer building on Mythos 5, you just learned that your entire product stack can disappear on a Friday night with no warning and come back only if your company has enough political capital to survive two weeks of negotiation.

Ray: And that's the precedent that actually matters here. It's no longer purely a technical question of which model is best for your use case. It's now also a political question — does your organization have the standing to get access when the government decides to tighten the valve? For most developers, the honest answer is no.

Chapter 3

Nova: According to the OpenAI Blog, GPT-5.6 launched as a limited preview less than 24 hours after reports surfaced that the Trump administration had asked OpenAI to stagger the rollout. The lineup includes Sol as the flagship, Terra as a mid-tier option, and additional models with enhanced coding, science, and cybersecurity capabilities. OpenAI went public with its objection, stating — and this is a direct quote — 'we don't believe this kind of government access process should become the long-term default.'

Ray: The quote is pointed, but the behavior is contradictory. They still staged the rollout. If the pushback were genuine resistance, they'd have launched everything simultaneously and dared the government to respond. What actually happened is they launched on a limited basis — which is precisely what a staged rollout looks like — while issuing a press-friendly statement about developer freedom. That's not defiance. That's managed compliance with better messaging.

Nova: The statement still matters strategically. OpenAI is now on record arguing that government-controlled model access harms developers, enterprises, and cyber defenders. That's a public position they'll be held to in the next policy fight. The fact that they said it out loud — naming cybersecurity specifically — shifts the terms of the debate even if the rollout itself was staged.

Ray: Maybe. But the government's leverage is clearly real. OpenAI complied, framed it well, and moved on. The open question for anyone building on Sol or Terra is whether the next access dispute gets resolved in 24 hours or in two weeks — and whether their use case makes the approved list.

Chapter 4

Ray: Wired reports that European governments and companies are accelerating efforts to build sovereign AI capabilities, and the explicit accelerant is US government unpredictability. After watching Mythos 5 go dark and GPT-5.6 get delayed by political negotiation, European buyers are asking a structural question: why build critical infrastructure on a foundation that can be switched off by a White House decision?

Nova: The political will is real and growing. But Wired also notes that analysts remain skeptical the continent can quickly produce top-tier frontier models. The funding gaps, the talent concentration in the US and China, the compute infrastructure — those don't close because a minister gives a speech about digital sovereignty. Political will and technical capacity are two very different things.

Ray: True, but partial decoupling is still meaningful. Europe doesn't need to out-compete OpenAI to reduce its exposure. If European enterprises shift even a significant portion of their workloads to homegrown or domestically-controlled models — even mid-tier ones — that's a structural market shift that US labs will feel in their enterprise revenue.

Nova: For global enterprises operating across jurisdictions, the consequence is a fragmented AI stack. Different models for European operations, US operations, potentially Asia-Pacific — compliance, capability, and access all diverging by geography. That's a procurement and engineering headache that didn't exist two years ago, and the Trump administration's erratic export controls are directly responsible for accelerating it.

Chapter 5

Nova: CNBC reports that Zhipu's GLM 5.2 model now lands within a percentage point of Anthropic's Opus 4.8 on a key agentic benchmark — and does it at roughly one-fifth the cost. The framing from analysts is direct: US government-imposed restrictions on frontier AI are inadvertently handing Chinese competitors a strategic opening, with enterprises actively asking whether cheaper Chinese alternatives can meet their needs.

Ray: One benchmark at one-fifth the cost is not strategic parity. Western enterprises evaluating a model switch are not running a single agentic test — they're weighing safety auditing, data residency, geopolitical exposure, supply chain risk, and the question of whether a Chinese-developed model can pass their legal and compliance review. A percentage point on one benchmark doesn't move those needles.

Nova: That argument holds when US models are reliably available. It starts to crack when Mythos 5 goes dark for two weeks and GPT-5.6 launches on a government-approved drip. The compliance calculus for a European manufacturer or a Southeast Asian bank looks different when the 'safe' US option just demonstrated it can be politically interrupted. At that point, the one-fifth cost advantage starts getting a serious hearing in procurement meetings.

Ray: The geopolitical risk cuts both ways though. A Chinese model carries its own set of government entanglements — data access concerns, export control exposure in the other direction, the question of what Beijing can compel Zhipu to do. Swapping US political risk for Chinese political risk is not obviously a better deal.

Nova: For buyers outside the US and China — European companies, Southeast Asian governments, Latin American enterprises — that bilateral risk framing is the opening. They're not loyal to either side. They're looking for the model that's capable enough, cheap enough, and available enough. GLM 5.2 just made a credible case on all three dimensions simultaneously. That's new.

Ray: That's the argument that's shifting my position. I was treating the benchmark result in isolation — one data point, one test, insufficient to prove a strategic shift. But I was underweighting the combination. Near-benchmark parity, a five-times cost advantage, and a demonstrated willingness by the US government to cut off model access arriving together — I now think that's a genuine inflection point for Zhipu, not a benchmark artifact. The geopolitical and safety considerations I was leaning on don't carry the same weight when availability itself becomes the variable.

Chapter 6

Nova: TechCrunch reports that OpenAI has revealed Jalapeño — a custom inference chip developed with Broadcom. It puts OpenAI alongside Google, Apple, and SpaceX in a wave of big-tech companies designing their own silicon specifically to reduce reliance on Nvidia. The inference economics argument is straightforward: if you're running billions of requests a day, owning the chip changes your cost structure fundamentally.

Ray: It changes the cost structure for OpenAI. For the developer building on the OpenAI API, Jalapeño is invisible infrastructure — they still pay whatever OpenAI charges. And for any company that isn't at hyperscaler scale, Nvidia remains the only realistic option. Custom silicon entrenches a two-tier market: the giants optimize their own stack, everyone else stays on commodity GPUs and hopes the pricing stays competitive.

Nova: That's a fair structural critique. But here's the thread that connects Jalapeño back to everything else in today's episode: the same companies now publicly fighting governments for the right to deploy their models are simultaneously building the hardware infrastructure to make themselves ungovernable at the infrastructure layer. When OpenAI owns the chip, the data center, and the model, the government's leverage points get a lot harder to find.

Chapter 7

Nova: The throughline today is that frontier AI has become a political asset — and that means its availability is now a negotiated outcome, not a given. Any enterprise, developer, or government that treats access to a US frontier model as stable infrastructure is making a bet on political continuity that this week proved is not guaranteed.

Ray: And the specific bet that matters most in the next 90 days: will the Trump administration formalize a model access approval process, or leave it as ad hoc political negotiation? Because if it formalizes, Zhipu's cost advantage stops being a curiosity and starts being a procurement argument that non-aligned buyers make in board rooms — and neither Washington nor the frontier labs have a clean answer for that.