In October 2022, the US imposed sweeping export controls on advanced semiconductors and chip-making equipment to China. The controls targeted chips above certain performance thresholds, equipment capable of manufacturing such chips, and even US persons working at Chinese chip facilities. It was the most significant technology restriction between major economies since the Cold War.
Two years later, the effects are mixed. China has accelerated domestic chip development but remains generations behind on cutting-edge manufacturing. The controls have constrained China's AI training capacity while creating new incentives for workarounds and domestic alternatives.
Why Chips Matter More Than Algorithms
The transformer architecture that powers modern AI is public knowledge. The training datasets — while large — can be assembled by any well-resourced organization. The algorithms are commoditized. What isn't commoditized is compute: the raw processing power needed to train models at scale.
Training GPT-4 class models requires tens of thousands of advanced GPUs running for months. Each generation of models demands more compute. The organizations that can access cutting-edge chips at scale — NVIDIA H100s and their successors — have a structural advantage that algorithmic innovation cannot easily overcome.
Compute Requirements by Model Generation
The Export Control Regime
The US export controls target three layers: chips, equipment, and expertise. Chips above certain performance thresholds (measured in TOPS and interconnect bandwidth) cannot be sold to Chinese entities. Equipment capable of manufacturing advanced chips (especially ASML's EUV lithography machines) is restricted. US citizens and permanent residents face restrictions on working at Chinese semiconductor facilities.
NVIDIA's response illustrates the constraints. The company developed China-specific chips (A800, H800) that operate just below control thresholds. When those were added to the restrictions, NVIDIA developed new variants. The cat-and-mouse continues, but each iteration leaves Chinese customers with less capable hardware.
China's Domestic Alternatives
China has poured resources into domestic chip development. Huawei's Ascend AI accelerators, produced by SMIC using older process nodes, represent the most advanced domestic alternative. The Mate 60 Pro smartphone demonstrated that Chinese manufacturers can produce 7nm-class chips — several generations behind the 3nm chips from TSMC, but better than many expected given the restrictions.
The gap remains significant. TSMC's most advanced manufacturing uses EUV lithography that China cannot domestically produce or import. Chinese foundries work with older DUV technology, requiring multiple patterning steps that reduce yield and increase cost. The performance gap is real, even if narrower than before.
Compute Supply and AI Content Generation at Scale
The semiconductor bottleneck has implications beyond national competition. The volume of AI-generated content in circulation depends partly on compute availability. When compute is expensive, generation is constrained. When compute is cheap and abundant, generation scales.
What Cheap Compute Means for AI-Generated Content
China's domestic compute capacity — while constrained at the cutting edge — remains substantial for inference workloads. Running trained models requires less advanced chips than training them. As model efficiency improves and older chips become sufficient for many applications, the compute constraints on content generation loosen.
This creates an asymmetry: the US may maintain advantages in training frontier models while China builds massive infrastructure for deploying content generation at scale. The authenticity challenges discussed in our AI Governance analysis may intensify as compute for inference becomes more accessible.
The Infrastructure Behind the Content Flood
The physical infrastructure of AI — data centers, power supplies, cooling systems, chip manufacturing facilities — determines what's possible at scale. Understanding the semiconductor competition is essential to understanding the future of AI-generated content.
The current dynamic: the US controls the cutting edge of chip manufacturing and restricts China's access. China builds alternative capacity with older technology. Both sides invest heavily. The outcome will shape not just AI capability but the volume and nature of synthetic content in global information ecosystems.