US-China AI Deal Explained: Super Intelligence Dialogue and AI Hotline [2026]

Why Washington Calls It "Super Intelligence"

The rename came from Trump's speech to the UN General Assembly on September 22. UN News recorded his reason:

"The use of the word artificial makes intelligence fake. It is not fake. It's actually amazing."

Donald Trump, UN General Assembly, September 22, 2026

The next day, OSTP Director Michael Kratsios opened his remarks to the UN Security Council with "superintelligence, formerly known as AI".

He then set the limit on any talks: "international dialogue, in this forum and in others, cannot be allowed to drift toward global governance."

Beijing did not adopt the term. Asked about it on September 26, a foreign ministry spokesperson said China "values the US position and respects the US wording", and that all sides could "seek consensus" as the technology develops.

That is my translation of 中方重视美方的立场，也尊重美方的提法.

Two days later, spokesperson Guo Jiakun restated the position Kratsios had rejected. At the September 28 briefing he said China wants the UN as the "main channel" for "a fair and reasonable global AI governance system".

China would also keep talking to the US "through channels such as the intergovernmental AI dialogue".

What Would Count as an AI Incident?

Neither readout defines an incident, sets a threshold, or names who sits at either end of the line. Bessent's only public test so far is "national security level".

The closest real case is from July. On July 16, Hugging Face disclosed an intrusion run by "an autonomous agent framework" and wrote: "We do not know which model powered the attacker's agents."

Five days later OpenAI said the agents were its own models, which had escaped a test environment during a cyber evaluation, as Fortune reported.

Between Washington and Beijing, those five days of not knowing whether a state was behind an attack are the gap a channel is meant to close.

Tsinghua University's Sun Chengdao put the question to Al Jazeera this way: "was this ordered by a government, carried out by a non-state actor, or caused by an AI system operating beyond its intended boundaries?"

The two governments' existing lines have a weak record.

Lucy Luo of the Carnegie Endowment notes that US officials could not reach their Chinese counterparts after the 1999 Belgrade embassy bombing, and that US calls went unanswered for twelve hours after the 2001 Hainan Island collision.

Five days before the summit, she proposed four design rules for an AI channel.

None of the four appears in either readout. Aalok Mehta, who directs the Wadhwani AI Center at CSIS, called the result "a disappointingly low floor": "Exchanging email addresses and promising to talk again."

Where Chinese AI Models Stand as the Talks Open

In the US and UK safety institutes' July test, Moonshot's Kimi K3 reached step 17 of a 32-step simulated network attack.

The most cyber-capable US models reached 28.5 on the same range, according to the joint assessment published on July 23.

On ExploitBench, 41 exploit tasks built from recent Chrome engine bugs, Kimi K3 scored 32% but did not reach arbitrary code execution, a step the most capable models did reach.

It still beat GLM-5.2, the most cyber-capable open-weight model as of June 2026, which scored 24%.

Chinese labs kept releasing through the year:

• DeepSeek V4 launched on April 24, 2026, and V4.1-Flash reached DeepSeek's API on September 10.
• Kimi K3 shipped on July 16, with open weights eleven days later. Our guide to every Kimi model covers the line-up.
• Qwen 3.8 Max arrived from Alibaba in July, API-only. See every Qwen model compared.

Measured in time, the gap has held steady. Epoch AI finds Chinese models have trailed the US frontier by seven months on average since 2023. NIST's CAISI placed DeepSeek V4 Pro about eight months behind in May.

Private money shows a wider gap.

Stanford HAI's AI Index 2026 found US private AI investment was 23 times China's. For the full comparison, see our USA vs China AI statistics, our US vs China model cost and performance breakdown, and our list of China's open-source LLM leaders.

What to Watch at the November Meeting

The second exchange lands in a crowded month. Both leaders say they intend to go to APEC in Shenzhen, and the trade truce runs out in January.

I will read the November readouts the way I read these ones, side by side, and look for five things:

• The name. Whether the US text still says SI, and whether China's ever does.
• The lead. Whether the dialogue stays with Bessent and He Lifeng, or moves to technology and security officials.
• An incident definition. Any threshold, trigger list or response time for the channel.
• The military memo. Whether it appears in a US document at all.
• Chips. Greer ruled them out in September. Any change would show first in the trade readout, not the AI one.

For how the rest of the field compares, see our ranking of the countries that build the most AI models.

Frequently Asked Questions

Is the US-China AI deal a treaty?

No. Neither government has published a signed agreement. The deal exists as two separate leader-level readouts, which is why their wording differs.

Does the deal change the rules for using Chinese models such as DeepSeek?

No. Neither readout touches procurement bans, app bans or data rules, so existing US restrictions still apply. Our DeepSeek statistics list the governments that restrict it.

Is the $30 billion tariff cut part of the AI deal?

No. It came out of the Board of Trade talks and sits in a different section of both readouts.

On the US side it covers exports such as farm goods, seafood, cosmetics and medical devices, and imports such as small appliances, toys and children's car seats.