A March 7, 2025 report by Gregory C. Allen (Director, Wadhwani AI Center, CSIS) for the Center for Strategic and International Studies, published shortly after the release of DeepSeek-R1. Allen argues that the post-DeepSeek narrative that U.S. export controls have "failed" misreads the evidence: the controls have design flaws and enforcement gaps, but their core premise — that constraining Chinese access to advanced compute buys strategically meaningful time — remains sound, conditional on enforcement.
Author: Gregory C. Allen (Director, Wadhwani AI Center, CSIS; export-controls analyst) Publisher: Center for Strategic and International Studies Date: March 7, 2025 URL: https://www.csis.org/analysis/deepseek-huawei-export-controls-and-future-us-china-ai-race
Summary of argument
Allen offers a primary-voice analytic defense of the U.S. export-control regime in the weeks following DeepSeek-R1. The report's throughline is that controls had design flaws (the October 2022 A800 loophole) and have enforcement gaps (BIS under-resourcing, smuggling networks, the Huawei TSMC die bank), but that these are correctable problems rather than evidence the underlying strategy has failed. Allen's reading is that constraining advanced compute can "slow and disrupt, not stop" Chinese AI development, and that slowing has strategic value only if enforcement keeps pace.
The report pairs with SemiAnalysis hardware reporting on the same period: where SemiAnalysis documents the hardware reality, Allen interprets the policy meaning of that reality (Source: Raw Sources/SemiAnalysis - Huawei Ascend Production Ramp.md) (Source: Raw Sources/SemiAnalysis - CoWoS and HBM Supply Chain.md).
Key claims
DeepSeek
Allen argues that DeepSeek's innovations (MoE scaling, MLA, FP8, DualPipe) are genuine, validated by Western replication rather than propaganda. He notes that the technical team was trained almost entirely in China, which he reads as a signal of indigenous research capacity. He cites DeepSeek CEO Liang Wenfeng's public statement that export controls are "the greatest challenge," with Chinese firms needing "two to four times" the compute of unrestricted access — which Allen presents as the strongest direct evidence that controls materially bind.
Export controls
The October 2022 package contained a design flaw that Nvidia exploited via the A800 (a blown-fuse modification), letting roughly $9 billion of advanced chips reach China legally during a 12-month window. Allen documents H100 smuggling networks, each handling $100M+ transactions, citing at least eight networks (drawing on government and intelligence-community sources). He describes BIS as operating with fewer than 600 employees and an approximately $200M budget, which he calls "drastically under-resourced" for the scope of enforcement. His net judgment is that controls can "slow and disrupt, not stop," and that slowing has strategic value only if enforcement keeps pace.
Huawei
Allen reports that the Ascend 910C delivers roughly 60% of H100 inference performance per DeepSeek's internal evaluation, a figure he frames as increasingly relevant as inference heads toward 70% of total AI compute by 2026 (drawn from a single internal evaluation). He states that TSMC manufactured 2M+ Ascend 910B logic dies for Huawei via shell companies, a stockpile that supports roughly 1M 910C units at about 75% packaging yield; he labels this a "strategically significant stockpile" and an export-control violation.
Against the hardware picture, Allen argues that Huawei's binding constraint has migrated from silicon to software: the CANN software ecosystem is immature, with Huawei's own employees describing it as "difficult and unstable" as of September 2024 and DeepSeek's evaluation reported as "very negative," requiring years of further development. He flags as a forward-looking risk that DeepSeek's open-source community could accelerate CANN maturation and partially close the software gap. As a proxy lower bound on the time to mature an alternative framework ecosystem, he cites Google's TensorFlow-to-JAX migration, which took 2–3 years and large investment.
The U.S.-China gap
Allen argues that the gap is primarily in hardware infrastructure rather than algorithms, since open research, distillation, and informal knowledge transfer diffuse algorithms rapidly. He invokes the Jevons Paradox: efficiency gains do not reduce total chip demand, noting that U.S. firms announced roughly 50% spending increases after DeepSeek. His top-line forecast is that the U.S. lead is unlikely to exceed "more than a year or two, even with extremely aggressive export controls."
Policy recommendations
Allen recommends restoring BIS and intelligence-community capacity to Cold-War-era resource levels for adversary-technology controls; aggressively enforcing the January 2025 Foundry Rule to prevent Huawei and SMIC from reaching viable scaled AI-chip production; and closing logic, packaging, and HBM chokepoints in coordination rather than piecemeal.
Relation to other sources
Allen functions as the primary-voice analytic complement to SemiAnalysis hardware reporting cited on the export-controls, Huawei Ascend, and DeepSeek pages: where SemiAnalysis supplies the numbers, Allen supplies the policy interpretation, naming specific political-economy judgments (BIS under-resourcing, Foundry Rule enforcement, Jevons-paradox demand, a one-to-two-year ceiling on the U.S. lead). The report is also a direct source for Liang Wenfeng's statement that export controls are "the greatest challenge." Allen's one-to-two-year lead framing gives the AI Race Dynamics page a Beltway-consensus figure, distinct from Aschenbrenner's existentially framed timeline.
On the TSMC die bank there is close agreement with SemiAnalysis, which puts the figure at roughly 2.9M dies against Allen's "over 2 million"; Allen's is the more conservative, government-sourced number, and there is no direct contradiction (Source: Raw Sources/SemiAnalysis - Huawei Ascend Production Ramp.md). Allen's short-to-medium-term framing of a one-to-two-year U.S. lead is compatible with the long-timeline backfire argument in Ord, but Allen does not address the 2035+ scenario; Ord's concern that controls incentivize Chinese build-out over 13 years is the risk that Allen's "slowing is valuable" frame defers. Allen treats export controls as a bounded national-security tool, whereas Ball documents the Trump administration's willingness to extend supply-chain designations to domestic firms; Allen's recommendations assume a legitimately operated regime, which Ball's evidence challenges.
Relationships
- supports: Export Controls (AI) (analytical framing), Huawei — Ascend AI Accelerators (policy interpretation), DeepSeek-V3 Technical Report, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, AI Race Dynamics (1-2 year ceiling)
- supports: (Source: Raw Sources/SemiAnalysis - Huawei Ascend Production Ramp.md) (Allen's policy reading of same hardware picture), (Source: Raw Sources/SemiAnalysis - CoWoS and HBM Supply Chain.md)
- related: Bureau of Industry and Security (BIS), BIS Framework for AI Diffusion — Interim Final Rule (Jan 13, 2025), Compute Governance, Distillation, Fast-Follow Problem
- contradicts: none directly; tension with long-timeline export-control skepticism (Broad Timelines) and with Ball's view that supply-chain tools are being weaponized domestically
- depends-on: DeepSeek, Huawei — Ascend AI Accelerators