Dynamic precision offers the United States a smarter way to slow China’s AI climb while keeping developers tied to American ecosystems and standards.
DeepSeek’s meteoric rise—and the stumbles that followed—has put in stark relief how US-China artificial intelligence (AI) competition now turns on ecosystems and standards more than any single chip. Earlier this year, DeepSeek stunned the field with an open-weight reasoning model that catalyzed a wave of reproductions and variants across the open-source community. But when the company moved to train its follow-on model, R2, it ran into the hard edge of geopolitics: US export rules and supply constraints slowed the project, and Beijing’s push to steer national champions toward Huawei’s Ascend processors compounded the delay. This episode is not an isolated hiccup. It captures a larger strategic contest over which stack—hardware, software, and developer tools—will become the default not only in China but across the Global South. If Washington responds with crude, on-off bans, it will likely accelerate China’s drive for a sovereign stack. If instead the United States applies dynamic precision—targeted controls that evolve with technical metrics and market realities—it can slow Beijing’s climb, dampen its resolve for full self-sufficiency, and expand global adoption of American-led standards.
The DeepSeek Shock, and What Its Slowdown Really Reveals
In January, DeepSeek’s R1 exploded onto leaderboards and social media, signaling how “open-weight” reasoning models could be trained efficiently and then reproduced by others. Within days, independent teams launched fully open reproductions of R1’s approach, amplifying the model’s impact far beyond its original lab. The point was less the novelty of a single architecture and more the diffusion advantage of releasing usable weights—the opposite of the “closed-weight, Application Programming Interface (API)-only” model favored by many US providers. That diffusion set the stage for the real contest: whose stack would everyone build on?
By summer, however, DeepSeek had delayed R2. Reporting tied the holdup to a combination of US export rules and the difficulty of training at scale on Huawei’s Ascend chips—hardware Chinese authorities have been encouraging domestic AI firms to adopt for reasons of sovereignty and control. In August, The Financial Times summarized that DeepSeek “pushed back” R2 amid those constraints, while industry trackers described the broader pattern: China’s largest labs are being prodded to rely on Huawei hardware, even when developer tooling and throughput still favor Nvidia. The friction is precisely where strategy lives.
Beijing’s Two-Track Position on Nvidia: Optics and Incentives
The past two months crystallized Beijing’s seemingly contradictory stance. After Washington allowed a resumption of exports of Nvidia’s downgraded H20 chip to China in July, Chinese regulators swiftly signaled caution. The Cyberspace Administration of China summoned Nvidia on July 31 to probe alleged “backdoor” risks in the H20, and state-affiliated outlets urged firms to avoid the chip. Nvidia, for its part, flatly denied that any backdoor exists and reiterated that the H20 is not designed for military or government infrastructure. Meanwhile, China’s industry ministry separately pressed leading internet platforms on their procurement choices, reinforcing political pressure to buy domestically. The surface narrative is security, but the underlying objective is to nurture a locally controlled stack even when it costs performance or time-to-market.
Washington’s own posture shifted in parallel. In May, the Commerce Department briefly told global firms that using Huawei’s AI chips anywhere in the world could implicate US export law, a sweeping interpretation that drew immediate pushback and was revised within weeks. Then, in July and August, the administration allowed Nvidia and AMD to resume certain sales to China under an unusual arrangement: export licenses paired with a 15 percent revenue share of China sales to the US government. However one judges the merits, these quick pivots underscore why blunt on-or-off controls are brittle in a fast-moving market.
What the Stack Fight Is Really About
Beneath the headlines is an adoption race. Nvidia’s chips come bound to Compute Unified Device Architecture (CUDA), a proprietary programming model, and to deeply optimized PyTorch builds, container images, drivers, and libraries. That stack lock-in is the culmination of nearly two decades of investment since CUDA’s 2006 debut, and it is a central pillar of Nvidia’s durable advantage. For a Chinese AI developer, choosing Nvidia typically means choosing CUDA; it also means benefiting from the richest third-party ecosystem in model tooling and inference frameworks. Huawei’s Ascend ecosystem—Compute Architecture for Neural Networks (CANN), MindSpore, and growing PyTorch compatibility layers—has improved quickly but still demands migration work, retraining of teams, and sometimes lower throughput per dollar for cutting-edge training workloads. That is why, absent political pressure, many Chinese labs would still rather train on Nvidia.
Beijing knows this, and it is why the government couples nationalist exhortation with selective sticks. The objective is not only to buy Chinese but also to ensure that the Chinese ecosystem becomes good enough, soon enough, to avoid long-term dependence on US standards. That strategy fits neatly with Xi Jinping’s call for “self-reliance and self-strengthening” in science and technology and with infrastructure-style programs like Eastern Data, Western Computing, which build the grid, data center, and networking foundations for a sovereign AI stack. A diffusion-first approach—open weights, rapid deployment in public services, and bundling AI with digital-infrastructure exports—helps spread the Chinese stack across Belt-and-Road markets even before its top-end training performance matches the United States.
Why Static Bans Backfire—And What Dynamic Precision Looks Like
Total bans or permissive green lights are the wrong instruments for a domain where capability is a moving target. Since October 2022, US controls have increasingly tied restrictions to technical metrics—compute density, interconnect bandwidth, memory, and system scale—because those parameters correlate with training frontier models and large-scale inference. The January 2025 revisions went further, calibrating thresholds to the compute needed to train the most advanced weights and addressing model-diffusion risks beyond individual chips. That scaffolding is the right foundation for a dynamic precision strategy.
The best strategy is dynamic precision: allow and encourage sales of chips that should be good enough to outperform China’s domestic hardware options to ensure Chinese developers remain on US platforms. At the same time, such exports should be calibrated low enough to maintain a significant performance gap with hardware systems available to American developers, keeping the United States one or two generations ahead of China.
In practice, dynamic precision means three things.
First, continuously updated thresholds keyed to measurable capabilities: effective floating-point operations per second (FLOPs) available to a cluster, interconnect characteristics that enable high-efficiency parallelism, and high bandwidth memory (HBM) capacity that supports large-context inference. Thresholds should move as the frontier moves, with notice periods that minimize whiplash for allies but leave little oxygen for arbitrage. The aim is not to freeze China in amber—that is impossible—but to slow its climb where it matters most.
Second, ecosystem awareness. A control that nudges Chinese developers off Nvidia and onto Huawei ultimately accelerates Beijing’s sovereign stack. By contrast, allowing a degraded, tightly audited Nvidia part into commercial (non-state) Chinese clouds—paired with licensing conditions that foreclose military or critical-infrastructure use—can keep Chinese developers within the CUDA/PyTorch orbit. That preserves US leverage over standards, documentation, and tooling, even as the raw chip remains one to two generations behind. Nvidia’s H20 episode, and Beijing’s own internal split over whether to shun it, is the case study in miniature.
Third, export controls aligned with global diffusion. The question is not just “what stays out of China,” but “what the rest of the world adopts.” The United States cannot win an ecosystem race if its stack is absent from the fastest-growing markets. That is why the new US–EU trade arrangement—which pairs procurement commitments, including $40 billion in AI chips, with a pledge to adopt US-compatible security standards to avoid technology leakage—is strategically important. It ties allied demand to compatible rules, creating a larger secured market for the US stack. Similar frameworks should be pursued with Japan, South Korea, Taiwan, and key partners in the Middle East and Southeast Asia.
Anticipating and Managing the Real Risks
Two objections deserve a serious answer. The first is that resumed sales of a compliant Nvidia part to China could expose sensitive code or enable backdoors. Here, the record matters. When China’s cyberspace regulator raised “backdoor” alarms about H20, Nvidia publicly denied any such capability and reiterated that its China-bound chips are designed for commercial markets, not state or military use. Washington can and should require transparent, verifiable assurances as a condition of any license, but blanket insinuations that commercial parts are secretly weaponized have not been substantiated. Precision controls backed by verification do more for security than rumors do.
The second objection is that any sales into China simply free domestic capital for Huawei and other national champions. That is partly true—and another reason to emphasize degraded performance and tight end-use constraints for any licensed exports, combined with active enforcement against smuggling. US prosecutions this summer against US-based actors illegally shipping H100-class parts to China underscore the scale of the black market. Static bans invite gray-market leakage; smart thresholds plus active enforcement limit it.
What Beijing Wants—And Fears
From Beijing’s vantage point, the logic is stark. The state is willing to absorb short-term performance penalties to build a sovereign stack and rewrite the rules around it. That is why “self-reliance” has become the governing concept in science and technology policy and why national compute-infrastructure programs pair with open-weight model diffusion and state-backed deployments in health care, education, and public security. At the same time, Chinese firms and researchers still prize throughput, developer ergonomics, and time-to-market—areas where Nvidia’s ecosystem retains a lead. That gap, even Huawei’s founder has admitted, remains at least one generation at the single-chip level, mitigated partly by cluster-level engineering. A dynamic-precision US strategy that keeps Chinese developers orbiting the US stack while constraining top-end capability targets Beijing’s core fear: dependence on an external standard at the moment of rule-writing.
Building the Operational Muscle to “Shoot a Moving Target”
Treating export controls as a living system, not an annual rulemaking, requires operational changes. The Commerce Department and interagency partners should institutionalize a technical watchfloor—an analytic cell that tracks frontier model training runs, notable scaling experiments, interconnect architectures, and software breakthroughs on a weekly cadence. Its job would be to recommend timely threshold adjustments and license conditions grounded in empirical capability, not rumor or lagging metrics. Because compute efficiency and software layers (compilers, graph optimizers, inference runtimes) can substitute for raw silicon, this watchfloor must include software engineers and model-systems researchers, not only hardware specialists. The 2023 and 2025 rules established the principle that controls can hinge on what systems can do; now the United States must build the reflexes to update those thresholds as the frontier shifts.
A Multinational Operation By Design
Finally, this cannot be a unilateral exercise. The United States should continue converting partners’ procurement into policy: the more allied governments and cloud providers pre-commit to the US-compatible stack (and to screening for leakage to destinations of concern), the more viable dynamic precision becomes. The EU deal shows how to fuse market creation with security standards; similar compacts should be explored with the United Kingdom, G7 partners, and with leading cloud providers whose services define practical access to compute. In parallel, Washington should support open-weight research outside China—funding shared model-safety testbeds, open evaluation suites, and tooling that flourishes on US hardware. The goal is to keep global innovation energy inside a rules-compatible ecosystem while narrowing the channels by which the cutting edge can be replicated under hostile control.
The Path Forward
DeepSeek’s arc—open-weight shock, attempted pivot to domestic chips, and delay—should not be read as a one-off. It is the template for the next stage of competition: China will tolerate near-term friction to reach long-term sovereignty; the United States must tolerate near-term complexity to preserve long-term ecosystem leadership. Static bans or blanket relaxations both miss the point. Dynamic precision—measurable thresholds updated on cadence, licensing that preserves US developer-stack gravity, and allied procurement wired to interoperable security standards—is how Washington slows Beijing’s advance, discourages all-in self-sufficiency, and expands the reach of US rules and norms.
There is no finish line in a domain where models, compilers, and interconnects reinvent the denominator every quarter. But there is an advantage to be kept and a standard to be set. Getting there means treating export control not as a wall, but as a steering wheel—one that is adjusted as the road turns and that keeps the world’s developers, by choice and by design, on a path the United States helps to write.
About the Author: Dr. Jianli Yang
Dr. Jianli Yang is a research fellow at the Kennedy School of Government of Harvard University. He is the founder and President of Citizen Power Initiatives for China and author of For Us, The Living: A Journey to Shine the Light on Truth and It’s Time for a Values-Based “Economic NATO”.
Image: Pla2na/Shutterstock
