根据彭博社审阅的一份备忘录（日期为 February 13, 2026（6:30 AM GMT+8）），OpenAI 告诉美国立法者，DeepSeek 正在使用「蒸馏」（在美国前沿模型的输出上进行训练）以及「新的、经过混淆的方法」来规避防御。OpenAI 表示，在 DeepSeek 于 2025 发布 R1 之后不久便开始提出担忧，并与 Microsoft 展开调查，以确认资料是否以未经授权的方式取得。
OpenAI 主张经济诱因被扭曲，因为 DeepSeek 与许多其他中国模型没有每月订阅费用，而像 OpenAI 和 Anthropic 这样的美国公司在基础设施上投入了数十亿并对高级服务收费。备忘录指称，与 DeepSeek 有关的帐户尝试透过第三方路由器、程式化存取，以及未经授权的转售商网路来绕过防护栏；OpenAI 也警告，透过蒸馏复制能力可能削弱安全控制，增加在生物与化学等领域的滥用风险，同时也提到在 Taiwan 与 Tiananmen Square 等议题上的审查。
硬体取得仍是第二个压力点：彭博社指出，2025 年末 President Donald Trump 推动放宽晶片限制，以允许销售 Nvidia H200，并称其大约落后 Nvidia 领先的 Blackwell 晶片 18 个月。彭博社引用众议院委员会纪录称，Nvidia 提供技术支援以协助 DeepSeek 改进并共同设计 R1，且 DeepSeek-V3 基础模型完整训练需要 2.8 million 小时的 H800 GPU；H800 在 2023 仅被允许对中国销售数个月，之后于 2023 的规则又停止销售，期间美国也就可能透过 Singapore 规避出口管制展开调查。
OpenAI told US lawmakers that DeepSeek is using “distillation” (training on outputs from US frontier models) and “new, obfuscated methods” to evade defenses, according to a memo reviewed by Bloomberg dated February 13, 2026 (6:30 AM GMT+8). OpenAI said it began raising concerns shortly after DeepSeek’s R1 release in 2025 and opened a probe with Microsoft into whether data was obtained in an unauthorized way.
OpenAI argued the economic incentive is skewed because DeepSeek and many other Chinese models have no monthly subscription cost, while US firms like OpenAI and Anthropic have invested billions in infrastructure and charge for premium services. The memo alleges DeepSeek-linked accounts tried to bypass guardrails via third-party routers, programmatic access, and networks of unauthorized resellers; OpenAI also warned that copying capabilities via distillation can weaken safety controls, increasing misuse risk in areas like biology and chemistry, alongside censorship on issues such as Taiwan and Tiananmen Square.
Hardware access remains a second pressure point: Bloomberg notes that in late 2025 President Donald Trump moved to ease chip restraints to allow Nvidia H200 sales, described as about 18 months behind Nvidia’s leading Blackwell chips. House committee records cited by Bloomberg say Nvidia provided technical support to help DeepSeek improve and co-design R1, and that the DeepSeek-V3 base model required 2.8 million H800 GPU hours to fully train; H800s were permitted for sale to China for only a few months in 2023 before rules later in 2023 halted sales, amid a US probe into possible export-control circumvention via Singapore.