Cursor tells House panels it may drop Chinese AI models over geopolitical risk

A House report says the tool SpaceX bought for $60 billion is exploring a phase-out of the Chinese open-weight models its Composer 2 was built on. The bill for the substitute lands inside SpaceX's $7.56 billion AI segment, which discloses no split.

Vincent JiangVincent Jiangthree-minute read
Share
Official portrait of Representative John Moolenaar of Michigan, against a blue backdrop with a US flag at left
Representative John Moolenaar of Michigan, chair of the House Select Committee on China, whose report says Cursor may phase out the Chinese open-weight models its Composer 2 was built on.

Cursor told Congress it may drop the models it was built on

Cursor, the AI code editor SpaceX−6.35% — SpaceX, down 6.35 percent today bought at an implied $60 billion equity value in August, told the House Homeland Security Committee and the House Select Committee on China that it plans to "explore phasing out PRC-developed open-weight models because of geopolitical risk," according to their report, shared with CNBC before publication on Friday 1. Cursor's Composer 2 model was built on Kimi K2.5, from the Chinese lab Moonshot AI 1. The letters to Airbnb−1.27% — Airbnb, down 1.27 percent today, Cursor's maker Anysphere$30B — Anysphere, private, latest valuation $30B and DoorDash−2.45% — DoorDash, down 2.45 percent today broke last week; Cursor's answer is the new part 2.

The Cursor code editor open on a dark screen, with a project's file tree at left and Cursor's agent chat panel at right
The Cursor editor with its agent chat open. Cursor's Composer 2 model was built on Kimi K2.5, from the Chinese lab Moonshot AI.Newmcpe, via Wikimedia Commons

The bill lands in a segment that shows no split

Cursor sells inside SpaceX's Super Intelligence segment, $7.56 billion of third-quarter revenue, 52% of the company's $14.575 billion total, blended with compute rentals and Grok subscriptions and disclosed with no split 37. SPCX fell 4.57% to $132.11 on Friday, below its $135 IPO price, in a selloff that left The Inference's US AI 100 index down 3.45% 453.

Cursor sells inside SpaceX's largest revenue segment

The segment holding Cursor is 52% of SpaceX revenue

  • Super Intelligence$7.56B51.9%
  • Connectivity (Starlink)$5.25B36%
  • Space (launch)$1.77B12.1%
Data
SliceValueShare
Super Intelligence$7.56B51.9%
Connectivity (Starlink)$5.25B36%
Space (launch)$1.77B12.1%
SpaceX Q3 2026 revenue by segment, US$ billions, from the company's third-quarter report. Super Intelligence blends compute rentals, Grok subscriptions and Cursor sales with no disclosed split.3,7

SpaceX revenue more than doubled in the year to the September quarter

$0B$5B$10B$15BQ1 '25Q2 '25Q3 '25Q1 '26Q2 '26Q3 '26$14.58B
Data
Revenue
Q1 '25$4.07B
Q2 '25$4.07B
Q3 '25$5.58B
Q1 '26$4.69B
Q2 '26$7.81B
Q3 '26$14.58B
SpaceX revenue by quarter, US$ billions, from the company's SEC filings. The filings retrieved report nothing for Q4 2025, so that quarter is omitted. Q3 2026 revenue of $14.575 billion was up 161% from a year earlier, and the Super Intelligence segment that holds Cursor supplied $7.56 billion of it.7

Washington wants the cheap option gone before the substitute exists

The report wants the dependency broken while conceding the substitute is missing: the United States lacks a competitive open-weight alternative at "the price and performance the market demands" 1. Open-weight models publish their parameters for anyone to download and run. Chinese models were the three most-used on OpenRouter, the marketplace that tracks token usage, in October 2, and the committees trace their share of AI workloads from about 1% in late 2024 to about 30% a year later 6. OpenAI pulled its models from Cursor with a November 12, 2026, cutoff 3.

Who pays, and who is lined up to gain

If the Kimi line goes, SpaceX either absorbs a higher inference bill inside a segment built to hide it, or passes the cost to seats where third-party models already bill at list, and the price edge over rivals Claude Code and Codex goes with it 3.

The named winners are the American open weights Meta's−4.82% — Meta, down 4.82 percent today Muse Glimmer, Nvidia's−10.86% — Nvidia, down 10.86 percent today Nemotron and Reflection AI's$2.1B — Reflection AI, private, latest valuation $2.1B Beam 2, plus a recommended "national American open-weight initiative" with public compute and lawful access to US frontier models for training 1.

Washington is stacking tools: bar PRC models from federal systems and contracts 1, order inventories on a 180-day Commerce clock 6, and make companies post their Chinese AI use on their websites, which Treasury and the White House science office have discussed 2.

Airbnb says its China use is limited, and Beijing objects

Airbnb, also named in the letters, says its AI "runs overwhelmingly on US-origin models," with limited China-origin use on approved US providers 8. Beijing says politicizing technology will "only stifle global AI advances," and committee chair John Moolenaar calls disclosure "an important step" 8. Nothing is decided: Cursor gave no timeline, and the disclosure idea may never reach formal rulemaking 12.

Watch for a date, a rule and the fourth-quarter print

Watch whether the phase-out gets a date, whether disclosure becomes a rule, and the fourth-quarter report in early 2027, the first read of the blended segment after the OpenAI$1.18T — OpenAI, private, latest valuation $1.18T cutoff and any retreat from Kimi 3. The committees are demanding an American model at Chinese prices, and until one exists, the difference is SpaceX's bill.

Get Policy & Labor by email, free

The rules governments write for AI, and what AI does to jobs. In our daily email, every weekday.

Loading SPCX

Reader comments

0 comments

    More in Policy & Labor

    All in this topic →

    Policy & Labor

    Accenture to cut 950 of 1,400 Meta moderation jobs in Portugal as unions strike Friday

    Accenture is cutting 950 of the 1,400 Meta content-moderation seats in Lisbon, announced on November 18, 2026, ahead of union strikes on November 21. The cut is a live read on what happens to outsourcers' revenue when their biggest clients automate the per-head work those outsourcers bill for.

    Policy & Labor

    House Democrats set a Dec. 1 deadline for Nvidia-Megaspeed records as $1.4 trillion of AI value is gone

    House Democrats have put the Nvidia-Megaspeed inquiry on a dated calendar with answers due December 1, and the approval data released alongside it freeze the China chip pipeline; the $1.4 trillion of AI value already gone is the market pricing a political overhang that outlasts December.

    Policy & Labor

    xAI asked the Ninth Circuit to block California's training-data law as judges pressed both sides

    The Ninth Circuit heard xAI's bid on Nov. 18; a ruling would set how much a frontier lab that rents compute from its own parent, SpaceX, may keep hidden, just as its biggest rival, Anthropic, banked $107.3 billion.

    Policy & Labor

    Trump's accounts bought Palantir in September, weeks before his agencies awarded it $42.5 million

    The November 19 disclosure puts the president long a stock his own agencies pay, weeks before a Democratic Congress arrives promising scrutiny of that book. The risk to Palantir holders is scrutiny of every award, not steering of the trades.

    Policy & Labor

    Roblox CEO skips Senate child-safety hearing with 196 suits pending and a Jan. 13 arbitration ruling

    Two bipartisan bills would void the click-through terms that keep 196 grooming lawsuits in private arbitration, and a January 13 bellwether ruling may open the courthouse door first. Shareholders hold the tail risk: Roblox guides Q4 bookings down 10–15% into the GTA VI launch, with the February print three weeks behind the ruling.

    Policy & Labor

    Alibaba asked for 948,950 US AI chips; 75,000 were approved, none shipped

    License records released on November 19 shut Alibaba's overseas chip route, and a New York Times account exposed the Johor workaround it used instead. A securities suit landed the same day, and the 53% AI-cloud machine now rides on Alibaba's own Zhenwu V900.

    Subscribe to TheInference for free