Trump vs Harvard, Nvidia export controls, how DEI killed Hollywood with Tim Dillon

By November 30, 2025, there is substantial evidence both that U.S. export controls have constrained China’s AI hardware and that China’s frontier AI capabilities have nonetheless reached near parity with the West, making the prediction impossible to score cleanly as right or wrong.

1. The condition of continued restricted access is largely met
Since 2022–2023, the U.S. has barred or tightly licensed exports of Nvidia’s leading data‑center GPUs (A100, H100, A800, H800, H20, L40/L40S, etc.) and expanded controls on advanced semiconductor manufacturing equipment to China. (stblaw.com) These controls remain broadly in force through 2025, with only prospective discussion of loosening some sales (e.g., H200) still under consideration. (reuters.com) So the conditional premise—China lacking straightforward access to Nvidia’s top-tier chips and U.S. leading-edge fab tools for multiple years—is essentially true.

2. Evidence that China has been materially “slowed down”

• Compute share collapse: One 2025 analysis using Epoch AI data finds that, as of March 2025, the U.S. controls about 75% of global AI compute capacity, while China’s share fell from 37.3% in March 2022 to 14.1%—a sharp relative decline aligned in time with U.S. export controls. (research.contrary.com)
• Bottlenecks acknowledged by Chinese and U.S. officials: The same report notes Chinese leaders and firms openly describe bans on advanced chips as their key constraint, and that deployment (inference) at scale is hampered—e.g., DeepSeek limiting API access soon after launching its R1 model, and Tencent warning that chip shortages could severely limit nationwide AI adoption or drive up costs. (research.contrary.com) Separately, a U.S. export‑control official estimated Huawei would be able to produce no more than about 200,000 advanced AI chips in 2025—insufficient for domestic demand—underscoring a real hardware shortfall despite Huawei’s Ascend line. (reuters.com)
  These points support Chamath’s intuition that the missing U.S. technology “slows them down quite a bit,” at least in terms of available compute and ease of scaling.

3. Evidence that frontier AI capabilities have not been strongly held back
On the other hand, multiple independent indicators show Chinese models closing the capability gap quickly:

• A detailed 2025 export‑controls review finds that despite the U.S. holding roughly 5× more compute, the performance gap between top U.S. and Chinese LLMs shrank from double‑digit margins in 2023 to near parity in 2024. (research.contrary.com)
• Chinese LLMs have reached or exceeded top global benchmarks:
  • Alibaba’s open‑source Qwen 2.5‑72B topped the OpenCompass LLM leaderboard, surpassing closed‑source state‑of‑the‑art models such as GPT‑4o and Claude 3.5 on many metrics. (alibabacloud.com)
  • Qwen2.5‑Max entered the Chatbot Arena global top‑10 and ranked #1 in math and coding and #2 on hard prompts in early 2025. (en.people.cn)
  • On Hugging Face’s Open LLM leaderboard, all top‑10 open-source models were derivatives of Alibaba’s Qwen series; Qwen2.5‑1.5B‑Instruct became the single most downloaded open‑source model globally, highlighting massive adoption and competitive quality. (aibase.com)
• Journalistic and policy analyses (e.g., Time and Foreign Policy) stress that China has advanced in AI despite chip controls, citing workarounds such as stockpiling and smuggling Nvidia GPUs, offshore training on Nvidia hardware in Southeast Asian data centers, and rapid improvement of domestic chips like Huawei’s Ascend series. (time.com) One 2025 Foreign Policy piece concludes that export controls "have neither failed nor worked as well as hoped," since Chinese frontier models like DeepSeek V3 and R1 were only months behind U.S. models when released. (foreignpolicy.com)
• Nvidia‑linked commentary similarly argues that specific bans (e.g., on the H20) did not stop China’s AI progress, contending instead that they mainly hurt U.S. industry, even while some experts still defend them as strategically useful. (tomshardware.com)

These results imply that, while China’s scale and cost of deployment are hurt by hardware constraints, its frontier model capabilities have continued to improve rapidly and, on some benchmarks, match or beat Western systems.

4. Why this is ultimately ambiguous rather than clearly right or wrong
Chamath’s claim is qualitative: that lack of access to leading‑edge U.S. manufacturing tech and Nvidia’s top chips would "slow [China] down quite a bit" over multiple years. By late 2025:

• There is strong evidence of a significant compute and scaling slowdown (loss of compute share, acknowledged chip bottlenecks, deployment constraints), which supports his view. (research.contrary.com)
• There is equally strong evidence that headline AI capabilities have not been slowed to the same degree, with Chinese frontier models arriving within months of U.S. releases and achieving top rankings on major leaderboards, despite the hardware gap and controls. (alibabacloud.com)

Because “slowed down quite a bit” could reasonably be judged true if one focuses on compute and deployment, or false if one focuses on frontier model quality and timing, the available data support both interpretations. The subjective, non‑quantitative nature of the prediction and the mixed empirical record make it impossible to classify it as clearly right or clearly wrong. Therefore the fairest score is ambiguous.

Sacks’ prediction is conditional and long‑term: if Harvard keeps DEI‑style, race‑conscious policies, then every 1–2 years there will be new whistleblower‑driven litigation revealing ongoing racial discrimination in admissions, with repeated legal findings against Harvard. The key testable elements are (a) a recurring pattern over multiple 1–2‑year cycles and (b) court or formal legal findings that Harvard is again liable for race‑based discrimination in admissions, post‑2023. The Supreme Court’s 2023 Students for Fair Admissions v. Harvard decision did find Harvard’s prior race‑conscious admissions unlawful, but that is the baseline event Sacks is referring to, not a post‑prediction data point. (en.wikipedia.org) Since the podcast aired on April 19, 2025, only about seven months have elapsed (to November 30, 2025), which is less than the first 1–2‑year window he specified. In that interval, there has been a federal Office for Civil Rights compliance review and a denial‑of‑access letter over Harvard’s refusal to turn over admissions data, but that review has not yet produced a legal finding that Harvard is again violating Title VI in admissions. (ed.gov) There has also been a civil‑rights complaint by America First Legal alleging unlawful DEI programs and DEI‑linked use of federal funds, which calls for investigation and potential sanctions but is not itself a whistleblower admissions‑discrimination lawsuit with a liability finding. (nypost.com) Other prominent Harvard cases since 2024 concern hostile‑environment and antisemitism claims under Title VI and have led to settlements and policy changes, not new judicial findings that Harvard is again discriminating by race in its undergraduate admissions decisions. (en.wikipedia.org) Because (1) the first 1–2‑year period following the prediction has not yet elapsed, and (2) the pattern of repeated whistleblower‑driven admissions cases with fresh liability findings has not had time either to materialize or be ruled out, the available evidence neither clearly confirms nor clearly falsifies Sacks’ forecast. It is therefore too early to judge, so the prediction is best classified as inconclusive at this time.