U.S.–China competition in chips, AI, and advanced manufacturing is crystallizing into a long-term structural race: the United States holds a widening hardware and fabrication lead, while China is choosing to sacrifice performance for sovereignty, betting that domestic capacity and energy advantages will eventually erode Washington’s leverage. The Trump administration’s decision to approve exports of Nvidia’s H200 AI chips to China turns this race into a high‑risk experiment in “addiction strategy”: using access to U.S. hardware to lock Beijing into an American tech stack even as China’s policy response is explicitly designed to break that dependence.
The result heading into 2026 is not a clean U.S. “win,” but a deeply interdependent, fragile equilibrium—one that hinges on Taiwan’s fabs, Western control over manufacturing equipment, and China’s willingness to eat short‑term performance losses to avoid a long‑term strategic chokehold.
—
The performance gap in AI chips is now the fulcrum of this strategic contest. According to detailed modeling of commercial roadmaps, the best U.S. AI accelerators are already about five times more powerful than Huawei’s leading chips, with that gap projected to widen to seventeen times by the second half of 2027. This is not a marginal edge; it is a compounding advantage anchored in access to cutting‑edge fabs, advanced manufacturing equipment, and a mature ecosystem of design tools and software.
Huawei’s own public roadmap underscores how severe these limitations are. Its next‑generation chip slated for 2026 is expected to be less powerful than its current flagship, a rare backward step that reflects the architecture and process constraints imposed by export controls and lagging fabrication technology. Even under optimistic assumptions about output, Huawei’s total production will still deliver only a sliver—on the order of a few percent—of the aggregate AI computing power Nvidia is expected to ship over the same period.
From Washington’s perspective, this is proof that the semiconductor choke point is working. U.S. and allied restrictions on advanced lithography, etching, and process tools have frozen China’s top‑end fabs several nodes behind Taiwan Semiconductor Manufacturing Company (TSMC), where Nvidia’s most advanced accelerators are produced. The gap is not only qualitative—smaller, denser, more power‑efficient chips—but also quantitative: TSMC’s volume and yield are far beyond what China’s domestic fabs can match under current constraints.
Yet that very success creates a dilemma. If China is compute‑constrained while its demand for AI training surges, the U.S. must decide whether to keep the pressure on—or to convert that leverage into a different kind of dependency.
—
The Trump administration’s December 8 decision to approve exports of Nvidia’s H200 chips to China is where strategy takes a counterintuitive turn. The H200 is not Nvidia’s absolute bleeding‑edge part, but it is the most powerful AI chip ever cleared for export to China, and substantially more capable than the earlier H‑series products that Chinese firms had broadly matched.
According to modeling cited by the Council on Foreign Relations, if the U.S. authorizes exports of roughly three million H200 chips to China in 2026, Chinese firms would gain more AI computing power from those imports than their domestic industry can produce until 2028 or 2029 at the earliest. That volume would be enough to support some of the largest AI data centers in the world, giving Chinese labs and infrastructure providers a temporary leap in training and deployment capacity.
The official logic behind this move is not purely commercial. Trump administration officials and aligned analysts describe a deliberate “addiction” strategy: by making Chinese AI development structurally dependent on U.S.‑origin chips, software stacks, and design ecosystems, Washington aims to ensure that Beijing can never fully break away from American technology. Access can be tightened, conditioned, or withdrawn as circumstances demand; the more embedded U.S. hardware becomes in Chinese infrastructure, the more leverage Washington retains.
Supporters of this approach frame H200 exports as *addictive but non‑decisive*: powerful enough to keep Chinese AI firms on the U.S. stack, not powerful enough to eliminate the underlying U.S. performance lead or to let Huawei catch Nvidia at the frontier. Critics, however, argue that the exports undermine the very advantage export controls created, by giving Chinese AI labs the compute they need to narrow the gap with leading U.S. models and to compete globally in AI cloud and infrastructure services.
For Chinese firms such as DeepSeek, which publicly identify access to compute as their single largest bottleneck, a sudden influx of H200s could be transformative. It would allow them to train larger, more sophisticated models on timelines closer to U.S. counterparts, closing not only performance gaps but also learning‑curve gaps in scaling laws, optimization techniques, and deployment architectures.
In other words, the U.S. is deliberately feeding a rival’s AI capacity in order to deepen that rival’s dependence on U.S. technology—a gambit that assumes lock‑in will matter more than the boost in Chinese capabilities that inevitably comes with it.
—
Beijing is responding to this gambit with a strategy that, on its face, cuts directly against pure market logic. Chinese authorities are instructing public and private buyers to prioritize domestic chips first, even when imported U.S. parts are demonstrably faster and more power‑efficient. Analysts estimate that this policy amounts to accepting roughly a 15% performance penalty—sometimes more—in order to guarantee that local fabs and design houses have a stable base of demand.
This is not about marginal efficiency; it is about survival under sanctions risk. Chinese policymakers have watched successive rounds of U.S. controls target specific chips, tools, and firms, and concluded that any enduring dependence on U.S. semiconductors constitutes an unacceptable vulnerability. In a scenario where Washington decides to “turn off the tap” entirely—cutting China off from advanced chips and manufacturing equipment—Beijing calculates that its entire digital economy would be at risk.
Accepting slower chips is thus the price of strategic insurance. Domestic demand guarantees help Chinese chipmakers scale up, move down the learning curve, and invest in successive process nodes despite lower initial competitiveness. Foreign chips like Nvidia’s GPUs are used only to backfill the remaining gap between domestic production and total demand, rather than to displace local capacity.
From this vantage point, Washington’s “addiction” theory misreads Chinese intent. Rather than becoming hooked on U.S. hardware, Beijing is openly treating U.S. chips as a temporary stopgap—a bridge to a future in which Chinese fabs, design tools, and alternative ecosystems are sufficiently mature to make Western components optional instead of existential.
The effect is an unusual dynamic: the U.S. is trying to entrench dependence precisely as China is institutionalizing policies to break it, even at a significant short‑term cost to performance and efficiency.
—
Behind both strategies is a basic asymmetry in production and demand. China’s AI chip output is rising, but all available estimates suggest it is growing only linearly, while demand for AI compute—driven by larger models, more parameters, and more intensive inference workloads—is growing exponentially. Even in China‑friendly scenarios, Huawei and other domestic producers cannot manufacture enough accelerators to close the gap with U.S. firms that enjoy access to leading‑edge fabs and unconstrained tooling.
Nvidia, by contrast, already struggles to meet U.S. and allied demand alone, despite commanding the lion’s share of the global high‑end GPU market. When the firm’s output is considered alongside competitors that also tap TSMC’s most advanced lines, the dominant share of cutting‑edge AI hardware remains firmly in U.S. or allied hands.
This creates a paradox. On one hand, China’s attempts to work around export controls—by designing alternative chips, opening domestic fabs, and optimizing software to squeeze more out of inferior hardware—help fund the same global ecosystem that keeps it dependent. Every purchase of an Nvidia accelerator, every license of a Western design tool, every contract for TSMC capacity reinforces the scale and R&D advantages of the U.S.‑aligned side of the industry.
On the other hand, the sheer scale of China’s AI ambitions means that starving it of compute entirely would have global consequences: slowed AI research, fragmented technical standards, and greater incentives for illicit acquisition of equipment and expertise. H200 exports, in this light, look less like generosity and more like an attempt to channel inevitability—to ensure that China’s growth happens on U.S.‑controlled rails rather than through a shadow ecosystem beyond Washington’s reach.
—
The contest is not confined to chips. It is embedded in a broader struggle over global tech infrastructure, where Taiwan remains the most critical—and vulnerable—node. TSMC alone fabricates more than 60% of the world’s semiconductors and nearly 90% of its most advanced chips, a concentration that leaves both Washington and Beijing dependent on an island that is also a military flashpoint.
U.S. policy is now oriented around reducing that single‑point dependency without fully severing it. The CHIPS Act and related incentives aim to bring a meaningful share of advanced fabrication onto U.S. soil and into allied countries, diversifying away from Taiwan while preserving TSMC’s role in the near term. At the same time, Washington’s export controls knit together a de facto coalition of equipment and materials suppliers—from lithography in the Netherlands to deposition tools in Japan—that collectively define the upper bound of what Chinese fabs can achieve.
China, in turn, is mapping its own paths around these chokepoints:
– Funding domestic equipment makers to replace foreign lithography and etch tools.
– Encouraging Chinese foundries to specialize in “good enough” nodes for AI workloads that can be parallelized across many less advanced chips.
– Using its dominance in other parts of the clean‑tech and manufacturing supply chain as leverage in negotiations.
This last point intersects directly with the energy dimension of the AI race. Analysts note that while the U.S. controls access to cutting‑edge AI semiconductors, China has a distinct advantage in powering AI infrastructure—its capacity to install and scale low‑cost solar, wind, and associated grid upgrades at a pace the U.S. has struggled to match. As U.S. data centers run into power bottlenecks—some facilities now require over a gigawatt of electricity, comparable to a small city—China’s relative strength in energy infrastructure could become a counter‑choke point: a way to host, export, and monetize AI compute, even if some of the highest‑end chips remain out of reach.
The emergent picture is of a two‑level race. At the hardware and fabrication level, U.S. and allied control of advanced nodes, equipment, and design ecosystems provides a tightening grip. At the energy and infrastructure level, China is building the capacity to host AI workloads at giant scale, positioning itself as a potential global hub for compute‑intensive industries even with slightly inferior chips.
—
The most profound uncertainty in this race concerns AI model development and services, where compute is necessary but not sufficient. Access to H200 chips will likely allow leading Chinese labs to train models closer in scale to U.S. frontrunners, lowering the barrier to parity in raw capabilities. With enough GPUs and energy, many of the remaining differences become questions of talent, data, and organizational learning—areas where China has already demonstrated the ability to close gaps quickly in other digital domains.
Supporters of maintaining strict export controls argue that this is precisely why loosening them is dangerous: once Chinese models close the performance gap, they can be integrated into global products, digital platforms, and state‑backed initiatives that embed Chinese AI standards and infrastructure around the world. In this view, H200 exports risk seeding a Chinese AI cloud ecosystem that competes directly with U.S. hyperscalers in emerging markets, especially where Chinese financing and infrastructure packages already carry weight.
Yet denial is not costless either. An uncompromising embargo on high‑end AI hardware would sharpen incentives for technological decoupling, accelerating Chinese efforts to build parallel ecosystems in chips, software tools, and standards that operate beyond Western control. It could also provoke broader retaliation in areas where China holds leverage, from critical minerals to renewable energy components.
The Trump administration’s middle path—allowing exports of powerful but not absolutely top‑end chips, in volumes that can be monitored and conditioned—attempts to square this circle. It keeps Chinese AI firms hooked on U.S. supply, giving Washington visibility into their build‑out while preserving the option to tighten controls later. But it also bets heavily on the assumption that lock‑in will endure even as China invests enormous political and financial capital in breaking that very dependence.
—
In this environment, the phrase “playing checkers while China builds chess” captures less a judgment about sophistication than a contrast in time horizons. Washington is using finely calibrated controls and strategic exports to optimize for the next few product cycles, confident that its current lead in advanced manufacturing and design will persist well into the coming decade. Beijing is optimizing for the day those assumptions no longer hold, willing to spend more and perform less now to ensure it can never again be strangled by a foreign choke point.
By 2026, the scoreboard is mixed but clear in one respect: the United States and its allies still define the frontier in chips, fabs, and many of the tools that make modern AI possible. China’s domestic efforts have narrowed some gaps but run into hard ceilings wherever export controls bite deepest. The H200 decision does not erase that lead—but it does transform it from a simple question of who has better chips into a far more complex contest over who controls the dependencies built on top of them.
Whether the addiction strategy cements U.S. dominance or accelerates the rise of a more autonomous Chinese ecosystem will depend on variables that no export‑control regime fully controls: the pace of breakthroughs in alternative architectures, the geopolitics of Taiwan, the economics of massive‑scale energy, and the choices of firms and governments far beyond Washington and Beijing. What is certain is that by tying chips, AI, and advanced manufacturing so tightly to grand strategy, both countries have ensured that the global tech order heading into 2026 will be as contested as it is interconnected.
—
