Elon Musk told Chinese TV that China will close the advanced chip gap with the U.S., contradicting earlier pacing-restriction positions he backed.

Elon Musk used a long interview with China Media Group to call Xi Jinping a great leader, a line that dominated coverage when state broadcaster CCTV aired the clips on September 23, the day before Xi's state visit to the White House. But the more consequential remarks came later, when Musk turned to artificial intelligence. He said Chinese AI models are outstanding, that China leads the world on performance per unit of compute, and that it will close its chip constraints within two or three years. Eleven days earlier, Musk had publicly endorsed Dario Amodei's plan to "pace the frontier"—a proposal whose China strategy depends on the United States maintaining exactly the chip advantage Musk now expects to disappear.

In the Silicon Valley interview, Musk was unusually generous toward Beijing. He described Chinese AI models as "generally outstanding," praised their results given the limited compute available to Chinese labs, and said China is probably doing better than anywhere else on performance per unit of compute. He estimated that China could close the compute gap within two to three years by building its own lithography and chipmaking capacity. The interview also included discussion of US-China cooperation on AI safety, a theme Musk has promoted throughout September. At the All-In Summit on September 14, he proposed that xAI, OpenAI, Anthropic, Google, Meta and "three or four of the leading Chinese companies" test one another's models before release, allowing competitors rather than the labs themselves to grade the work.

Musk's praise lands awkwardly given what he endorsed on September 12, when he responded to Amodei's essay "We Must Pace the Frontier" by saying Dario was right. While often summarized as a call to slow down, the essay's China section ranks among the most hawkish statements from any frontier lab. It calls for bans on advanced chips and chipmaking equipment, crackdowns on smuggling and remote access, and measures against distillation and weight theft. The stated goal is to widen America's lead to three to five years. Amodei identifies export controls as a key precondition for pacing within democracies—not optional add-ons but necessary prerequisites.

The architecture of pacing only restrains US labs. Amodei acknowledges that if American labs slow more than their lead allows, projects tied to Beijing will overtake them. The entire plan therefore rests on one assumption: that chip controls can hold China back long enough for the US to safely decelerate. Musk's forecast directly attacks that assumption. If China closes the compute gap in two to three years, the three-to-five-year lead the essay aims to build never materializes, and US labs that agreed to pace themselves would be slowing down while their Chinese competitors caught up. Musk has not said pacing was a mistake or withdrawn his support, but the two positions cannot both be comfortable. Either the chip controls will work, making Musk's prediction wrong, or they will not, making the plan he endorsed have a much shorter shelf life than its author intended.

Musk has made similar arguments before. The difference now is timing. He made this forecast on Chinese state television during a week when the White House was preparing to discuss AI guardrails with Xi, and when the pacing debate was still fresh from the essays and replies of mid-September, including China's own rejection of the slowdown calls as fearmongering.

The language Musk used also sits closer to Beijing's position than Washington's. Chinese officials and state-linked commentators have argued that any limits on frontier AI must apply equally to Chinese and American models. One state media-affiliated blog said substantive talks could only begin after the US demonstrated its safety rules were equally effective for its own companies. A CCTV-affiliated account accused Anthropic of trying to dictate industry rules through lobbying and criticized Amodei's calls to arm democracies with AI.

A joint safety framework applying the same rules to both countries' labs is reasonable in principle and resembles the third step of Amodei's essay, which calls for international coordination. But it is not compatible with the essay's China section, which treats Chinese labs as competitors to be kept behind rather than peers to be regulated alongside. By backing both positions, Musk has endorsed the part of Amodei's plan that Beijing is happy to hear while forecasting the failure of the part Beijing objects to.

Tesla's position in China provides commercial context worth noting. Shanghai is the company's largest export hub. Sales of China-made Model 3 and Model Y vehicles rose only 3.6 percent year on year in August to 86,166 units, following a 38 percent jump in July. Tesla's share of China's battery-electric market has fallen to 6.6 percent from more than 15 percent in 2020.

Musk's claims for Grok 4.7 have shifted significantly in six weeks. On August 12, he wrote that the model "will exceed all current models" and that he would be shocked if any model beat it at real-world engineering. By mid-September, he had lowered expectations to roughly on par with Anthropic's Opus 5.0, not 5.1. In the CMG interview, real-world engineering had become a goal rather than a claim. Musk said Anthropic made AI excellent at software engineering, but that no one has yet achieved excellence in real-world engineering, and described that as an opportunity for Grok using training data from SpaceX and Tesla. Independent benchmarks support the more modest version. On Artificial Analysis, the top Grok 4.7 configuration scores 46 on its intelligence index, against 58 for Claude Opus 5.5 and 53 for GPT-6 Astra, and ranks 16th overall. Musk has told SpaceX staff he expects the company to reach first place in about six months.

The shift fits the China argument in an unexpected way. If American labs are competing on engineering data and compute while China competes on efficiency, then Musk's own framing suggests the frontier race is less about who is ahead today and more about who can scale fastest—exactly the question pacing asks American labs to set aside.

Three developments will reveal whether Musk's remarks represent a one-interview comment or a real split with the pacing camp. The first is whether he repeats the two-to-three-year forecast in the US, where it would carry direct implications for export control policy. The second is whether US-China AI talks around Xi's visit produce any shared safety mechanism, since Musk's peer-review idea would require Chinese labs to participate. The third is Grok 4.7's standing once independent testers measure its engineering performance, because Musk has now tied xAI's pitch to a benchmark category that does not yet exist in standard form.