There’s growing speculation about DeepSeek’s R1 model, U.S. export controls, and China’s AI progress. Let’s separate fact from fiction. Here’s why people are wrong to bet against compute. 🚫 Fiction #1: “DeepSeek can match Western AI without advanced GPUs.” ✅ Fact: DeepSeek has 50,000+ Hopper (H800, H100, H20) GPUs, but U.S. firms (OpenAI, Google, Microsoft, Meta, Anthropic) access larger clusters (H100, H200). DeepSeek’s CEO has cited U.S. export controls as a major bottleneck—highlighted in a January 20, 2025 meeting with Premier Li Qiang. 🚫 Fiction #2: “Memory and bandwidth aren’t critical for next-gen AI.” ✅ Fact: Upgrading from H100 to H200 boosts performance by 43% and reduces costs by 3× due to larger batch sizes and lower KVCache bottlenecks. New “reasoning” models like OpenAI o1/o3 and DeepSeek R1 rely on long-chain inference, requiring high memory capacity. 🚫 Fiction #3: “Chinese AI labs don’t need Nvidia’s latest GPUs, moat of compute is no more.” ✅ Fact: Huawei’s Ascend GPUs exist, but top Chinese AI labs still prefer Nvidia hardware for its mature software ecosystem. China has imported millions of H20 GPUs, proving demand for more advanced hardware. SMIC is stuck at a 7nm production process using DUV and yields are low. 🚫 Fiction #4: “Export controls don’t matter now.” ✅ Fact: U.S. policy was never about stopping Chinese AI—it’s about slowing progress and making it more expensive. While China stockpiled GPUs before 2023, the real impact of restrictions on H200, GB200, and GB300 will emerge in the next 6–12 months. 🚫 Fiction #5: “Inference compute isn’t a major bottleneck.” ✅ Fact: AI inference is now as costly as training. New models like DeepSeek R1 and OpenAI’s o1/o3 use multi-step reasoning, which drastically increases GPU demand. NVL72 and GB300 are designed to handle 100K+ token contexts, reducing per-token costs. 🚫 Fiction #6: “Scaling pre-training further yields minimal benefit.” ✅ Fact: Overtraining by 10–100× can drastically lower inference costs. This strategy requires massive GPU clusters, which China struggles to access due to export controls. 🚫 Fiction #7: “AI hardware advancements are incremental.” ✅ Fact: Nvidia’s GB200/GB300 introduce major leaps, including new interconnects and high-memory architectures. These shifts significantly improve AI performance beyond standard generational upgrades. 🚫 Fiction #8: “China’s AI breakthroughs won’t benefit the PLA.” ✅ Fact: China’s civil-military fusion ensures commercial AI supports the PLA in applications like drone swarms and intelligence analysis. Even off-the-shelf AI tools can enhance China’s military capabilities. 🚫 Fiction #9: “GPU smuggling isn’t a real concern.” ✅ Fact: Smuggling via Malaysia, Singapore, UAE, and secondhand markets undermines U.S. export controls. Nvidia’s H800 and H20 chips skirt restrictions, requiring stronger enforcement. The AI race isn’t just about better models—it’s about who can deploy and scale AI the fastest. 🚀

After having met with many people regarding the TikTok issue, it is apparent that what’s known as a “Chinese golden share structure” is standing in the way of saving TikTok. If a company has a “golden share structure ” it means that one of the shareholders is Xi Jinping, the head of the Chinese Communist Party. He owns you and makes all the decisions. While I want TikTok to keep operating in a manner consistent with our national security, it is long past time to deal with the golden share dilemma more broadly. I will soon introduce legislation with my Senate colleagues that prevents any company that has a “Chinese golden share” from being listed on any American exchange and further, remove any existing company that has a golden share structure from any American exchange.