Empty White House briefing room lectern at dusk, U.S. and Chinese flag silhouettes through window, signaling AI policy memo

White House Memo on Chinese AI Theft Lands Three Weeks Before Trump’s Beijing Summit

The White House on April 23, 2026 issued a national security memorandum accusing China of running “deliberate, industrial-scale campaigns” to distill American frontier AI models, three weeks before President Donald Trump is scheduled to land in Beijing for a May 14 to 15 summit with Xi Jinping. The directive, NSTM-4, signed by Office of Science and Technology Policy chief Michael Kratsios, commits federal agencies to share intelligence with U.S. AI labs and develop joint defenses, but stops short of new export controls, sanctions or Justice Department referrals.

That gap between the memo’s rhetoric and its operational teeth is reshaping two prediction markets at once. On Polymarket, the contract on whether a Chinese company will hold the best AI model at month’s end is trading at low single-digit probabilities for Alibaba, even after the Hangzhou firm released Qwen 3.6-Max-Preview on April 20. The separate market on whether Trump will visit China by May 31 sits at roughly 70 cents on the dollar, with roughly $54,000 in daily volume and a five-point depth of about $5,500.

What the White House memo actually orders

NSTM-4 directs federal departments to do four specific things: share threat intelligence with U.S. AI developers about foreign distillation campaigns, coordinate defensive best practices across the private sector, partner with industry on detection and remediation tools, and explore accountability measures against foreign actors. It does not authorize new chip export restrictions, criminal referrals or fines.

In the memo’s own language, distillation occurs when a foreign actor “feeds thousands or millions of carefully constructed queries to a frontier AI model, collects the responses, and uses those responses to train a cheaper rival model.” Kratsios wrote that Chinese entities are using “tens of thousands of proxies and jailbreaking techniques in coordinated campaigns” against leading U.S. systems. The OSTP statement issued alongside the memo framed the practice as a national security threat, not a commercial dispute.

Enforcement, though, is the part the document leaves vague. Researchers at NYU Shanghai’s RITS program noted that the memo “announces a posture, not a regime,” and that any binding restrictions would require either Commerce Department rulemaking or congressional action. Both are slow. The Beijing summit is fast.

The evidence base: Anthropic’s 24,000 accounts, OpenAI’s House testimony

NSTM-4 rests on a factual record built over the past 90 days by two of America’s largest AI labs. On February 23, 2026, Anthropic published a technical disclosure titled its detection write-up on industrial-scale distillation attacks, naming DeepSeek, Moonshot AI and MiniMax as the operators of an extraction operation that ran approximately 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude.

Anthropic said attackers routed traffic through residential proxies to bypass geofencing, and that one cluster orchestrated tens of thousands of accounts simultaneously while scripting long, training-shaped conversations. The company called the technique a “hydra cluster.”

Eleven days earlier, on February 12, OpenAI sent a memo to the House Select Committee on the Chinese Communist Party alleging that DeepSeek had developed “new, obfuscated methods” to evade access controls, including programmatic scraping by employees and networks of unauthorized resellers. Yusuf Mahmood’s April 16 written testimony to the same committee identified adversarial distillation as the single highest-priority concern raised by industry witnesses, and asked Congress to push the executive branch to act. NSTM-4 followed seven days later.

Why Alibaba’s frontier-model bet is on the line

Alibaba is the test case for whether a Chinese model can crack the global frontier without distilling Western outputs, and the timing is brutal. The Hangzhou firm rolled out Qwen 3.6-Max-Preview on April 20, five days before this article’s publication. The release topped six coding benchmarks, posted a 2.3 percentage-point gain on the SuperGPQA reasoning test, and beat Anthropic’s Claude on a tool-call instruction-following benchmark.

None of that has been enough for Polymarket traders. The contract on which company holds the best AI model at month’s end still prices Alibaba in low single digits, with U.S. labs OpenAI, Google and Anthropic absorbing nearly all the implied probability mass. The narrower contract on the best Chinese AI company by April 30 puts Alibaba near 90 percent, with Z.ai’s GLM-5 trailing around 4.5 percent and DeepSeek near 2.5 percent.

The split tells you what traders are pricing. Alibaba is nearly certain to lead China. It is nearly certain not to lead the world. NSTM-4 widens that second gap by raising the cost of the cheapest path to catch up.

The Beijing summit math

Trump is scheduled to meet Xi in Beijing on May 14 and 15, after the summit slipped roughly six weeks from its original late-March window. The administration has publicly framed the visit as stability-seeking. “What we are not looking for is massive confrontation or anything like that” with China, U.S. Trade Representative Jamieson Greer told the Hudson Institute on April 7, as reported in Greer’s April 7 remarks on the May meeting.

Greer said the U.S. and China are negotiating two new bilateral structures: a Board of Trade to define what the two countries can sustainably exchange without crossing national security red lines, and a possible Board of Investment to handle company-specific roadblocks. Rare-earth supply, he said, is being addressed through plurilateral deals with price-floor mechanisms to deter future predatory pricing by Beijing.

None of that suggests AI distillation will dominate the public agenda. It does suggest a private back-channel role.

How the prediction markets are pricing the dual question

The two relevant Polymarket contracts have different liquidity profiles, and reading them together produces a sharper picture than either provides alone.

MarketImplied probabilityResolutionDaily volume
Trump visits China by May 31~70 percent YESMay 31, 2026~$54,000 USDC
Trump visits China by April 30~0.4 percent YESApril 30, 2026Thin
Best Chinese AI company by April 30Alibaba ~90 percentApril 30, 2026Active
Best AI model overall by April 30Alibaba low single digitsApril 30, 2026Active

The structure tells a coherent story. Traders are nearly certain Trump will not fly before April 30 and reasonably confident he will fly before May 31, consistent with the publicly announced May 14 to 15 summit. They are also nearly certain Alibaba will not unseat U.S. labs at the global frontier in the next six days, regardless of Qwen 3.6-Max-Preview’s benchmark wins.

At 70 cents, a YES share on the May 31 contract returns 1.43 times the stake. A trader would need to believe summit logistics hold despite the OSTP memo, despite the open Iran war that delayed the original date, and despite any retaliatory leak Beijing might choose to issue between now and boarding.

The enforcement gap nobody is pricing

Here is the detail buried in the memo that competitor coverage has skipped. NSTM-4 does not name a single Chinese company. It references foreign actors and adversarial distillation as a category, but DeepSeek, Moonshot AI, MiniMax and Alibaba appear nowhere in the four-page directive. That is a deliberate choice, and it preserves diplomatic flexibility before Beijing.

By contrast, the Anthropic and OpenAI submissions named names. Mahmood’s House testimony named names. The White House did not. A senior administration official telling reporters the memo “sends a clear signal” is doing rhetorical work the document itself declines to do.

“Adversarial distillation attacks were the number one concern I raised, and the number one ask was for exactly what this memo describes,” Yusuf Mahmood, who testified before the House Select Committee on April 16, wrote on X after NSTM-4’s release.

The memo does open the door to future action. It instructs OSTP to “explore measures to hold foreign actors accountable,” language that maps to potential entity-list designations, Commerce Department licensing tightening or DOJ economic-espionage referrals. None are scheduled. All are possible. The summit comes first.

What to watch in the next six weeks

Three datable events will resolve the ambiguity NSTM-4 leaves on the table.

  • April 30, 2026. Both the best-AI-model and best-Chinese-AI-company Polymarket contracts resolve. Alibaba’s odds in the global market are weak. The Chinese-only market is effectively decided.
  • May 14 to 15, 2026. The Trump-Xi summit in Beijing. Watch the joint statement for any reference to AI, distillation, model access or licensing. Silence is itself a signal.
  • May 31, 2026. The broader Trump visit Polymarket contract resolves. At 70 cents, the market is pricing modest tail risk that the trip slips a second time.

Background on the broader race is captured in our recent reporting on how Chinese AI agents are reshaping consumer commerce ahead of U.S. rivals, and on the concentration of frontier AI capability in nine dominant labs, six American and three Chinese.

https://x.com/AnthropicAI/status/2025997928242811253

The bottom line for traders and policymakers

Kratsios picked his moment. Issuing NSTM-4 three weeks before Beijing gives Trump a live talking point without committing the United States to any specific punishment. Issuing it the same week Alibaba shipped Qwen 3.6-Max-Preview gives U.S. labs cover to claim that Chinese benchmark gains rest on extracted American outputs, a charge Alibaba denies.

For prediction-market traders, the pricing is consistent. Alibaba is locally dominant and globally a long shot. Trump is going to Beijing. The interesting question, the one neither contract resolves, is whether anything Kratsios wrote on April 23 will still be policy by June.

Mahmood, asked on X what success looked like, gave the answer the memo itself avoids: “Name the companies. Then act.”