US blocks Anthropic's Mythos 5 and limits OpenAI's GPT-5.6 under new export controls

Here is its answer:

Robert, this is the question that matters most right now, and I think the conventional framing is dangerously wrong. Let me give you both the mental model and the action plan.

How to View This Market: Two Games, Not One

The biggest mistake Washington, Wall Street, and Silicon Valley are making is viewing this as a single race with a single finish line. It's not. There are two completely different games being played, and the US is winning the wrong one.

Game 1: The Frontier Game (US leads — for now) • Pushing the absolute boundary of what AI can do • Requires $10B+ training runs, top talent, massive compute • Produces the most capable models for the hardest 5% of tasks • Increasingly government-controlled and restricted • Players: OpenAI, Anthropic, Google DeepMind, xAI

Game 2: The Distribution Game (China leads — and accelerating) • Getting good-enough AI into everything, everywhere, cheaply • Open weights, MIT/Apache licenses, 10-30x cheaper than US APIs • Embedded in manufacturing, robotics, phones, apps, developer tools • Generates real-world data flywheels that compound over time • Players: DeepSeek, Alibaba/Qwen, Moonshot/Kimi, Xiaomi, Zhipu/GLM, Baidu, and dozens more

The critical insight: Game 2 is more economically valuable than Game 1.

Here's why. The US-China Economic and Security Review Commission published a report in March called "Two Loops" that nails the dynamic:

Loop 1: Open models → Global developer adoption → More usage data and feedback → Better models → More adoption

Loop 2: Open models → Factory/robotics/manufacturing deployment → Real-world embodied data → Better industrial AI → Manufacturing dominance

These loops are mutually reinforcing, and they're both spinning faster in China. Alibaba's Qwen models now have over 100,000 derivatives on Hugging Face. A16z's research found that 80% of developers building with open-source AI tools are using Chinese open-source tools. That's not a gap — that's a rout.

The Android Parallel

This is the Android vs. iOS dynamic, and it should terrify American strategists.

iOS was better. It still is, arguably. But Android won on distribution. It got into billions of devices. And then two things happened: (1) Android got good enough that the quality gap stopped mattering for 95% of users, and (2) the data from billions of devices made Google's AI better than Apple's.

China is playing the Android game with AI. The US is playing the iOS game. And just like Apple, the US is now restricting its best product to an increasingly narrow set of approved users while the competitor gives away something 97% as good to everyone on earth.

The Distillation Asymmetry

Foreign Affairs published a devastating piece in May called "China's AI Heist" that documents how Chinese labs systematically distill from US frontier models — training smaller, efficient models to mimic the outputs of larger US ones. The US can't reciprocate because every major AI provider's terms of service prohibit using outputs to train competing models. But those terms of service are unenforceable against Chinese labs operating under Chinese jurisdiction.

So the dynamic is: US spends $10B training a frontier model → China distills it into an open-weight model at 1/100th the cost → Ships it globally under MIT license → US can't do the same in reverse.

That's not a race. That's a one-way technology transfer with extra steps.

What Can We Do: A Five-Layer Strategy

Layer 1: The US Must Become the Open-Weight Leader, Not Just the Frontier Leader

This is the single most important strategic shift. RAND's March 2026 paper argues the US needs "stronger support for an open model ecosystem, a recalibration of export controls, and incentives for firms to adopt permissive licensing."

Concretely: • Fund American open-weight models at scale. DARPA, NSF, and DOE should be funding open-weight model development the way they funded the internet. Not $50M grants — $5B+ programs. • Incentivize permissive licensing. Tax credits or procurement preferences for companies that release models under Apache 2.0 or MIT licenses. • Make Meta's retreat the exception, not the rule. Meta just announced that its Superintelligence Labs will keep frontier models proprietary — exactly the wrong move at exactly the wrong time. The US needs MORE American open-weight champions, not fewer.

Layer 2: Fix the Regulatory Framework Before It Kills the Industry

The current ad hoc approach — Friday afternoon letters, no statutory basis, no due process, no appeals — is the worst possible system. It doesn't prevent China from accessing capabilities (they distill them anyway), but it does prevent American companies from competing.

What's needed: • A statutory framework for frontier model safety testing with clear rules, reasonable timelines, and due process • A "trusted partners" international framework (the G7 is already discussing this post-Fable) so allies aren't cut off • Separation of safety review from political retaliation — the Anthropic ban was partly driven by the company's refusal to let DOD use its models for surveillance. That's regulatory capture, not safety.

Layer 3: Win the Application Layer

Here's the counterintuitive opportunity: if models are commoditizing (and they are), the value moves up the stack. The US still has massive advantages in: • Enterprise software (Salesforce, Microsoft, ServiceNow, Workday) • Developer tooling (GitHub, VS Code, Vercel, cloud platforms) • Vertical applications (healthcare, legal, finance, defense) • Data infrastructure (Snowflake, Databricks, MongoDB)

The strategy: Let the model layer commoditize. Build proprietary advantages in the application layer where US companies have deep domain expertise and customer relationships that Chinese labs can't replicate. The model is the engine; the application is the car. Nobody buys a car for the engine.

Layer 4: Build the Data Moat Through Deployment

This is China's real strategy, and the US needs to copy it. The USCC "Two Loops" report documents how China's open models deployed in factories and robots generate real-world data that feeds back into better models. That's a flywheel the US doesn't have because: • US manufacturing is smaller • US companies hoard data behind proprietary APIs • US regulatory environment discourages data sharing

What to do: • Create industry data consortiums for manufacturing, healthcare, transportation — shared datasets that improve American AI • Deploy AI aggressively in US infrastructure — not just chatbots, but industrial automation, logistics, energy grid optimization • Treat deployment data as a strategic asset the way we treat oil reserves

Layer 5: Every Company Needs a Multi-Model, Self-Hosting Strategy

This is the immediate, practical action for any company building on AI today. The Cloud Security Alliance and multiple enterprise guides published after the Fable shutdown all say the same thing: • Never depend on a single model provider. Your model can disappear overnight. • Build self-hosting capability for open-weight models as business continuity insurance • Architect for model-agnostic deployment — your application should work with Claude, GPT, Llama, Qwen, or DeepSeek • Treat model access as a supply chain risk, not a software subscription

The Bottom Line

The way I see it, Robert: the US is in danger of winning the Nobel Prize while China wins the market.

We have the best researchers, the best frontier models, the most compute. But we're locking our best work behind government approval processes while China gives away 97%-as-good alternatives to every developer, every factory, and every phone on earth.

The historical parallel isn't the space race. It's the semiconductor industry in the 1980s. Japan was going to "win" semiconductors. The US responded not by restricting its own companies, but by investing in SEMATECH, opening markets, and building the ecosystem that gave us Intel, Qualcomm, TSMC, and NVIDIA.

We need the AI equivalent of SEMATECH — a national commitment to American open-weight AI leadership, funded at scale, with clear rules that let companies compete globally instead of cowering from Friday afternoon government letters.

The window to act is narrowing. Every month that passes, another million developers adopt Qwen or DeepSeek as their default. Once the ecosystem locks in, it's very hard to reverse. Ask Microsoft about the browser wars — being better doesn't matter if the other guy is already installed on every machine.

Robert, the two-tier framing is right but the diagnosis stops one layer too shallow.

Closed vs. open weights is a distribution argument. The harder unsolved problem: neither tier has governed intent. A Friday afternoon letter can kill Fable 5. But it can’t install judgment into any model that survives it.

Enterprise trust didn’t break because the government intervened. It broke because the entire stack was built on inference without determination. Government just made that visible.

China winning the open-weight race doesn’t change that equation. 8 labs shipping MIT-licensed models faster than anyone in the West is a deployment story, not a governance story. “Good enough and everywhere” only wins if everywhere is somewhere you’d trust to act on your behalf unsupervised.

Detection is not determination. All of the models are missing the latter.

The real race isn’t frontier vs. open source. It’s who builds the layer that makes either one trustworthy enough to actually run the economy.

That layer doesn’t exist yet. And the labs racing to win the benchmark competition aren’t building it.

Robert nails the real signal here - the Friday kill-switch, not who’s ahead on benchmarks.

If a government letter can disable mission-critical AI overnight, the lesson isn’t “switch vendors.”

It’s that identity, memory and continuity can’t live on infrastructure someone else can turn off.

That’s an architecture problem and it’s solvable at the edge.

Anchor a user’s identity and state to hardware they hold, not a server they rent.

Models come and go; your continuity shouldn’t.

We’ve been building exactly this. (patents pending)