House Report Targets Chinese Open-Weight Models as Trump Threatens a Tax on Overseas Data Centers and Oracle Says Its OpenAI Build Is Fully Funded
Two House chairmen released a report detailing how Airbnb, Anysphere and DoorDash use Chinese open-weight models and urged keeping those models out of federal systems and high-risk critical infrastructure. President Trump threatened a “tremendous tax” on U.S. companies that build data centers abroad, and Oracle said its OpenAI contract is fully funded. Plus: OpenAI’s defenses against adversarial distillation, Thinking Machines’ Inkling 2 Mini, MedChemBench and new valuation talks at Cognition and Databricks.
House Chairmen Urge Removing Chinese Open-Weight Models From Federal Systems and Critical Infrastructure
Two House chairmen released a joint investigative report on Friday arguing that American companies are quietly building core products on AI models developed in China. According to the House Homeland Security Committee’s announcement of the joint report, Homeland Security Chairman Andrew Garbarino and Select Committee on China Chairman John Moolenaar opened the investigation in April 2026 and corresponded with three companies as case studies: Airbnb, Anysphere and DoorDash. The chairmen make 12 policy recommendations, including a national American open-weight AI initiative, a model provenance standard with an AI bill of materials, a ban on PRC-developed models in federal systems and federal contract work, and removal of those models from high-risk functions of critical infrastructure.
The full joint report on PRC-developed AI models gives unusually specific detail on the three companies. It says Airbnb runs customized versions of Alibaba’s Qwen models on U.S.-based infrastructure for customer support, including an assistant that handles roughly 80,000 guest and host interactions a day. It says Anysphere built Composer 2, the flagship model in its Cursor coding platform, on a Moonshot AI base model. And it says DoorDash reported 25 PRC model families authorized and deployed in a non-trivial manner, with one embedded model that could take up to six months to replace. The report also finds that PRC models accounted for 41 percent of downloads on the leading open-model repository over the preceding year, and that the highest-ranked American open-weight model placed tenth in the most recent independent ranking the committees reviewed, behind six PRC models.
“While America is investing billions and racing to build ‘Super Intelligence,’ Beijing is using stolen intelligence to close the gap.”
As The Record’s coverage of the House investigation notes, the report compares the situation to Huawei and ZTE equipment and warns that removing deeply embedded Chinese models later could take years and cost taxpayers billions. None of the three companies immediately responded to The Record’s request for comment.
The report is not law, but provenance questions are about to show up in procurement questionnaires and security reviews long before any bill passes. Any company that sells to government, utilities or regulated industries should inventory which base models sit under its fine-tunes, vendors and coding tools now, including models it never chose directly, so it can answer quickly when asked.
Trump Threatens a “Tremendous Tax” on U.S. Companies That Build Data Centers Abroad
President Trump said at a White House press conference on Friday that U.S. companies will be barred from building data centers in other countries and will face a heavy penalty if they do. CNBC’s report on Trump’s data center remarks says the announcement apparently came as news to his own staff, and quotes the president acknowledging as much.
“I don’t know if that’s a breaking story,” Trump said. “I haven’t even said that to my people. But I’m telling them now they’re here.”
CNBC notes that Alphabet, Amazon, Microsoft, Meta and Oracle have already built data centers in other countries, often to serve nearby customers or meet local regulatory requirements. Trump also said companies that want to keep building abroad “can, but we’re going to put a tremendous tax on them.” The Associated Press account of the tax threat says Trump did not provide details and had not yet told his own officials about the plan. With no rate, scope or legal mechanism announced, it is a statement of intent for now rather than an enacted policy. The remarks also land awkwardly against public opinion: CNBC’s own survey found 64% of registered voters opposed an AI data center within 10 miles of their home.
Do not rewrite cloud region plans on a press-conference remark, but do note the direction of travel. Between new chip export licensing rules and this threat, where AI compute physically sits is becoming a political variable. Teams with data residency obligations in Europe or Asia should ask their providers how they would handle new U.S. restrictions on overseas capacity.
Oracle Says Its OpenAI Contract Is Fully Funded
Oracle pushed back on Friday against investor worries about how it will pay for its OpenAI buildout. In a short Oracle statement on its OpenAI funding, the company said its final capital raise earlier this month made it the first hyperscaler fully funded for the current investment cycle, and that the financing it outlined at its investor day, including vendor financing, is in place.
“Any further private discussions with vendors are just additional options. Our ability to deliver on our OpenAI contract does not depend on those discussions. Any media reports that will come out and say the contrary are incorrect and wrong.”
The raise itself is documented in Oracle’s November 4 Form 8-K: $10 billion of common stock and $10 billion of notes, with coupons ranging from 7.000% on notes due 2029 to 8.650% on notes due 2046. Oracle also said its Abilene campus is nearly complete, its next gigawatt-scale campus in Shackelford is fully permitted, and the handover of Nvidia Vera Rubin racks there is on track for the coming weeks. Reuters’ report on the Oracle statement says the shares rose as much as 3.3% in after-hours trading, and that Oracle and OpenAI plan 4.5 gigawatts of additional U.S. capacity on top of the initial Stargate site in Abilene.
A preemptive denial of reports that have not been published yet is an unusual move, and it tells you how sensitive markets have become to AI financing. For buyers, the useful signal is that capacity timelines at Oracle’s Texas campuses look on track. For investors, those high-7% and 8% coupons on fresh debt are the real price of the AI buildout.
OpenAI Details How It Fights Adversarial Distillation
The House report’s charge that Chinese labs extracted capabilities from American models got a technical counterpart on the same day. In a post on how OpenAI is combatting adversarial distillation, the company described a defense stack that combines adversarial testing, model training, input and output classifiers, API-level controls and investigations into coordinated extraction campaigns. It puts special weight on protecting the model’s raw chain of thought, which it calls protected reasoning.
“The average attack success rate fell from 54.0% against GPT‑6 Astra with its original safeguards to 0.2% against GPT‑6.2 Astra with its updated safeguards.”
OpenAI also reported an internal study on how cyber capabilities transfer. A student model trained on a teacher’s exact reasoning traces generally outperformed one trained only on reasoning summaries after supervised fine-tuning. After reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested, while the summary-trained student did better at lower budgets. OpenAI’s conclusion is that protecting reasoning traces can limit the transfer of cyber capability, even when the task itself is permitted under its usage policies.
Expect hidden reasoning to stay hidden. If your product depends on reading or logging full chains of thought from frontier APIs, plan around summaries instead. The numbers come from OpenAI’s own red-teaming, so treat them as a vendor claim, but the policy direction across labs is clear: reasoning traces are now treated as protected assets.
Thinking Machines Releases Inkling 2 Mini, a Looped Model That Runs on a Laptop
Thinking Machines Lab released Inkling 2 Mini, its open-weight laptop-sized model, on Friday. It is a looped transformer: the middle layers run three times for every token, which the lab says lets the model reason more without writing out extra tokens. The Inkling 2 Mini model card lists 26.4 billion total parameters, 3.4 billion active (5.4 billion effective active with looping), text and image input, a context window of up to 1 million tokens and an Apache 2.0 license. The checkpoint is about 20GB in mixed NVFP4 and MXFP8 precision, and the card lists support for Apple silicon devices with 32GB or more of memory and a single Nvidia DGX Spark.
“Inkling 2 Mini leads every open-weights model with less than 100B total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%).”
The lab is frank about the limits. Looping improves reasoning but does not help the weights store more facts, so the model’s factual recall is in line with others its size. The model card warns that it can get stuck in long reasoning loops and may take destructive actions, such as deleting or overwriting files, without confirmation. Hugging Face’s launch post for Inkling 2 Mini covers day-one availability for local runtimes.
This is an American open-weight model released on the same day Congress complained there are not enough of them, and it is small enough to run on a well-equipped MacBook. It is worth testing for offline or private coding agents. Take the card’s warning about destructive file actions seriously: sandbox it, and cap tokens and agent turns.
OpenAI Introduces MedChemBench for Drug-Discovery Reasoning
OpenAI also published MedChemBench, a medicinal chemistry reasoning benchmark. It contains 99 problems drawn from 117 papers and patents, covering 761 distinct molecular structures, and grades models on four things: reconstructing molecular structures, explaining experimental evidence, ranking and selecting compounds, and predicting numerical properties. At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of available credit, compared with 83.3% for GPT-6 Astra, 82.0% for GPT-6.1 Sol and 58.2% for GPT-5.6 Sol. OpenAI says selecting compounds and predicting numerical properties remained the hardest parts.
The company is careful about what the results mean. It says the benchmark does not establish whether models can discover new drugs, that prior exposure to the published sources cannot be ruled out, and that each answer was a single attempt without tools or new experiments.
The jump from GPT-5.6 Sol to the GPT-6 family is the headline, but the honest caveats matter more for anyone in life sciences. A benchmark built from published papers measures how well a model reads the literature, not whether it improves lab decisions. Use results like these to choose models for literature triage, then validate against your own unpublished assay data.
AI Valuations Keep Climbing: Cognition Talks at $75 Billion, Databricks at $225 Billion
Two reported funding talks show that private AI valuations are still rising even as public chip and AI stocks wobble. Reuters, citing Bloomberg News on Cognition’s funding talks, says the maker of the Devin coding agent is considering raising $2 billion or more at a $75 billion valuation, about two months after a Series E at $48 billion. Separately, Reuters, citing the Wall Street Journal on Databricks, says the data and AI software company is in early talks to raise several billion dollars at about $225 billion, up from the $190 billion valuation it announced in August. Both reports say terms could change, and neither company immediately responded to Reuters.
Coding agents and the data platforms that feed AI applications are where investors see durable revenue right now. For customers, rich valuations bring aggressive sales pushes and bundling. That is a good moment to negotiate multi-year pricing, while keeping exit clauses in case the market cools and vendors consolidate.
Quick Hit: Gemini Now Makes GIFs
Google added GIF creation to the Gemini app. In the official Gemini app announcement on X, Google said users can create looping photo effects and memes with its latest video model, Gemini Omni 1.2 Flash, which it says delivers smooth motion and consistent quality across frames. It is a small feature, but it shows video models moving into everyday chat use.
What We Left Out
Several leads did not make the cut. Reports about Nvidia’s leadership and a market-moving rumor that Jensen Huang would back a rival chip venture were either unconfirmed or rested on claims we could not verify from primary sources. We did not restate Anthropic’s first-day trading or weekly Nasdaq-100 figures without a primary market source. Social posts about a Cursor rebrand and an AI-assisted Riemann hypothesis proof are unconfirmed. OpenAI’s neurodegenerative-disease work with PerturbAI and Scripps Research was covered earlier this week, and ChatGPT’s always-on “dots” assistants still lack a primary release note.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →