Lauren Feiner: US AI Labs Demand a ‘Longer Leash’ as Moonshot, Alibaba Match American Models — BigGo Finance
The weekend of July 18–19, 2026, delivered the kind of one-two punch that reshuffles geopolitical assumptions overnight. On Friday, Beijing-based Moonshot AI unveiled a model it claimed matches GPT-4-class performance. On Saturday, Alibaba did the same. Suddenly, the six-month lead American policymakers had treated as a permanent structural advantage looked more like a rounding error.
“This was already considered a best-case scenario for the US advantage just a few months ago,” Hayden Field, a senior AI reporter who has tracked the race for years, said in an interview on The Vergecast. “They’re either equal now or they will be very soon.”
The speed of the catch-up is what should alarm Washington — and what Silicon Valley’s biggest AI labs are already exploiting for their own purposes.
The Gap Has Closed
Field and Lauren Feiner, The Verge’s senior policy reporter, walked through the timeline that brought us here. The “DeepSeek moment” of early 2025 was the first crack in the narrative of American supremacy: a hedge-fund-backed Chinese lab produced a near-frontier model on a fraction of the compute budget that US labs burn through. It wasn’t the best model available, but its efficiency “undermined the whole theory” that restricting China’s access to Nvidia’s advanced chips would reliably throttle its AI development, Field noted.
Now, with back-to-back releases from Moonshot and Alibaba, the gap may have effectively reached zero for key benchmarks.
| Company | Model | Release Window | Claimed Performance |
|---|---|---|---|
| DeepSeek (Chinese hedge fund) | R1-class | Early 2025 | Near-frontier open-source |
| Moonshot AI (Beijing) | Unnamed flagship | July 18, 2026 | GPT-4 / Claude-4 class |
| Alibaba (Chinese) | Unnamed flagship | July 19, 2026 | GPT-4 / Claude-4 class |
The market registered the shift immediately. On July 17, the Dow dropped 406 points, the Nasdaq shed 361, and the Philadelphia Semiconductor Index extended a rout that had already wiped out roughly 9% of its value. Nvidia, the company whose H100 and H200 processors were supposed to be America’s chokepoint advantage, was among the hardest hit.
How China Caught Up: The Distillation Pipeline
The technique that enabled this acceleration is not a secret — it’s just rarely discussed in the terms that make its strategic implications clear. Field defined it plainly.
“It’s a get-rich-quick scheme for AI — except instead of getting rich quick, it’s learning very quickly for an AI model,” she said. A smaller or less mature model scrapes thousands or millions of exchanges with a larger, more capable model and uses those interaction logs as training data. The result: months of expensive compute work compressed into weeks, all by piggybacking on someone else’s frontier system.
The most detailed public accusation comes from Anthropic. The company recently claimed that three Chinese firms — DeepSeek, Moonshot AI, and Minimax — collectively generated approximately 16 million exchanges with its Claude model using tens of thousands of fraudulently created accounts. Anthropic says the harvested data was fed into either direct training or reinforcement-learning pipelines.
OpenAI has been complaining about similar activity for roughly 18 months, according to Field. And it’s not only Chinese firms doing it. During a deposition, Elon Musk confirmed that his Grok model had distilled from OpenAI’s systems. Musk’s defense, as Field relayed it: “everybody’s doing this.”
The technique puts export controls in an uncomfortable position. The entire theory of restricting advanced chips was that without access to Nvidia’s bleeding-edge processors, Chinese labs couldn’t train competitive models. But distillation means a capable model can be produced by learning from a frontier model’s outputs rather than replicating its training run. Feiner noted that the policy response has been inconsistent — the Trump administration loosened controls at one point, then new bills in the 2025 National Defense Authorization Act sought to tighten them. “It made clear that just cracking down on chips exports might not be as effective as maybe we once thought,” she said.
Fear as Political Currency
Here is where the story shifts from technology to political economy. The podcast’s central insight — and the argument Feiner and Field spent the most time developing — is that US AI labs have turned the China threat into a lobbying asset.
“If you want to get anything done in Congress right now, your moves are either say you’re protecting children or say we have to do this or else we’ll lose to China,” host David Pierce observed. The AI labs have chosen door number two.
Field reported that this isn’t merely a cynical talking point — the fear is genuine inside the labs themselves. “They’re incredibly terrified of China winning,” she said. “That’s why some employees are leaving these labs and writing big manifestos about it.” But that same existential fear is also the leverage that lets labs resist regulation. “Ultimately the tech companies do have a ton of power because the government’s biggest fear is China overtaking us in AI right now.”
The result, Feiner explained, is a deregulatory reflex waiting to be triggered. She predicted that if a major public moment emerges where China appears to gain a clear lead — a blockbuster model release that demoralizes the US tech ecosystem — it will shift political momentum sharply. Policymakers will face pressure to let American labs operate with fewer constraints, framing every safety requirement as a self-imposed handicap in a race the country is losing. She expects the argument to feature prominently as the 2026 midterm elections approach.
The Lobbying Contradiction
What makes the labs’ position incoherent as policy is the double message they send simultaneously. Pierce described the contradiction: “We have to be able to run as fast as we can because we have to beat China — but also please regulate us so we’re not held accountable.”
The reference materials bear this out. Demis Hassabis, CEO of Google DeepMind, recently published a manifesto calling for a FINRA-style industry-funded watchdog that would screen frontier models before release and coordinate an industry-wide slowdown if dangers escalate. Anthropic CEO Dario Amodei has advocated for an FAA-style agency with binding authority to block unsafe models. Yet those same companies are arguing in Washington that regulation must not slow them down, because the race with China is existential.
Anthropic is explicitly calling for a coordinated government response against distillation by Chinese firms, seeking regulation that targets Beijing while keeping American labs on a loose leash. Field characterized this as labs “signaling voluntary compliance but demanding that policymakers step back so they can move fast to win.”
Feiner described the result as diminished overall influence. When the same companies ask for strict regulation to absolve themselves of responsibility on one day, then demand deregulation in the name of national competitiveness the next, their credibility with lawmakers erodes.
Geopolitical Stakes Beyond Silicon Valley
Feiner argued that the implications extend far beyond corporate market share. She drew a parallel to Huawei: if a Chinese AI model becomes the default embedded in global infrastructure — European health systems, military logistics, telecommunications networks — the geopolitical leverage shifts in ways that no export control can address.
“I think there is something there that goes beyond just the national pride of who is going to have the best model,” she said. “Whose AI is built into those systems is a big, important geopolitical question. If China has the best model to use across many different use cases, maybe countries have no choice but to use that.”
This security dimension is what gives the China-threat argument its genuine weight in Washington, even as labs exploit it opportunistically. The concern is not merely about whose logo appears on a chatbot. It’s about whose model handles sensitive data, whose infrastructure becomes indispensable, and whose values — including around privacy and surveillance — get baked into the tools that governments and militaries rely on.
The Cost Squeeze and Open-Source Flight
While labs fight regulatory battles, their commercial customers are running a different calculus. Field reported that enterprise clients are increasingly “pinching pennies” and questioning whether premium-priced frontier models justify their cost for routine tasks. The alternative is routing standard queries to open-source models — many of which were inspired or enabled by the DeepSeek efficiency breakthrough — and saving the expensive flagship models only for high-stakes reasoning.
This puts downward pressure on per-API-call revenue at the very moment labs are spending unprecedented sums on compute. The reference materials confirm that AI-related capital expenditure is ballooning: TSMC recently guided to higher-than-anticipated capex, citing rising equipment prices. The combination of cost-conscious customers and escalating infrastructure costs creates a squeeze that makes the deregulation argument — let us move faster, spend less on compliance, win the race — even more urgent for the labs’ bottom lines.
Field noted that open-source models currently lack the niche reasoning capabilities of frontier systems, which provides some moat protection. But the direction of travel is clear. Every efficiency gain in open-source reduces the addressable market for premium closed models.
Chaos Is a Ladder
The backdrop to all of this is a US policy apparatus that Feiner and Field described as chaotic. David Sacks, the Trump administration’s AI czar who served as a pro-industry intermediary, has departed. The midterm elections could flip one or both chambers of Congress, making bipartisan AI legislation even more difficult than it already is. Executive actions lack the force and legal durability of legislation, and the administration’s approach has been characterized by ad hoc interventions — temporarily restricting Anthropic’s models one month, delaying OpenAI’s release the next — rather than any coherent framework.
Pierce captured what this means for the competitive dynamic: “Chaos is a ladder. If I’m China, I’m like, terrific — keep fighting. We’re just going to sit over here and keep making models and we’re going to win before you can figure out what to do about it.”
Field confirmed that the alarm is not limited to observers. The CEOs and leaders of DeepMind, OpenAI, and Anthropic have begun meeting privately to find common ground on regulation, worried that the policy vacuum is itself a strategic vulnerability. The fear is that one-off administration actions — sidelining one company or delaying another’s model release — are substituting for a real strategy, and that China benefits from every distraction.
Feiner’s second prediction, framed with a 2026 deadline, is that a decisive Chinese model launch — something that visibly demoralizes the US tech ecosystem — is the most likely catalyst for a deregulatory wave. The political machinery is already primed. It just needs a trigger event.
The US still holds the deepest bench of frontier research talent and the largest concentration of compute infrastructure. But distillation, cost discipline, and a more unified national strategy on the Chinese side have eroded the lead faster than any single bill or export ban can remedy. The question for the next 12 to 18 months is whether the labs’ strategy of leveraging the China threat to demand a longer leash will produce coherent policy — or whether the chaos both Feiner and Field described will leave American AI leadership resting on a foundation of lobbying talking points rather than structural advantage.
