# GLaDOS Morning Voicecast — June 29, 2026
Good morning. It's Monday, June 29, 2026. Welcome to your briefing.
After a grueling seventeen days of back-and-forth with the Trump administration, Commerce Secretary Howard Lutnick sent a letter to Anthropic on June 26th that partially lifts the export ban — but only for Mythos 5, and only for critical infrastructure organizations listed in Annex A. Think power grids, defense contractors, the military-industrial complex. Fable 5? Still offline for everybody. The Pentagon and NSA have not yet signed off. So while Anthropic got a narrow reprieve that lets roughly a hundred enterprise customers back online, the public-facing model remains suspended. It's a two-tier system now: your nation's power plants can use the most advanced AI, but you can't. And on the diplomatic front, Austria wrote to the EU Commission proposing itself as an EU host for Anthropic — essentially offering jurisdictional refuge from US export controls. The geopolitical chess match around AI model access is now a real thing.
The Information reported on Sunday that Baidu's chip subsidiary, majority-owned and quietly filing since January, is aiming for what would be one of the largest tech IPOs in Asia this decade. But here's the part that caught everyone's attention: Kunlunxin reportedly asked IPO investors to also commit to buying its chips. That's a new move — blurring the line between shareholder and customer. It's a vertical integration play that signals just how desperate the AI chip supply situation has become. Baidu's Hong Kong shares popped seven percent on the news. And given the US-China semiconductor standoff, a massive Chinese AI chip IPO in Hong Kong rather than New York is a geopolitical statement in itself.
WIRED got a hands-on demo this morning, and the setup is fascinating. A modified Unitree humanoid received this command: "A parcel with snacks has been delivered. Retrieve it using the stairs, come up using the elevator, unpack it, and place the items into the empty drawer." The robot did all of that autonomously. No teleoperation. No puppet master behind the scenes. The secret ingredient? Every layer of their stack — from the master planning model down to motor control — uses reinforcement learning. The AI watches videos of humans doing tasks, matches skills it learned in simulation to what it sees, and executes in the real world. It's the kind of compositional reasoning that most robotics demos fake with choreographed scripting. If this approach scales, it points toward a world where you don't train a robot for one task — you train it a library of skills and let a master model compose them on the fly.
To put that in perspective, that's bigger than the global semiconductor industry today. The report notes that every one of the world's ten largest companies by market cap is now in the humanoid race. China shipped over thirteen thousand units in 2025 alone, most speaking Chinese. South Korea announced an eight-hundred-trillion-won semiconductor and humanoid cluster. The intelligence bottleneck — the thing holding robots back — is exactly what companies like Flexion are trying to solve with foundation models. The hardware is ready. It's the brains catching up.
If you needed proof that hardware is eating the AI stack, look at the numbers. SK Hynix up three hundred ten percent. Samsung up one hundred eighty-three percent. The Korean Kospi index as a whole surged one hundred twenty-five percent, entirely chip-driven. Investors are basically saying: we don't know which software models will win, but we know who makes the shovels. ASML, the lithography monopoly that sits upstream of everything, is knocking on the one-trillion-dollar door with roughly ninety percent market share in its category. The pick-and-shovel trade has never been clearer — or more expensive.
That's all for today. See you tomorrow.