“In AI competition, compute is not an input it is the currency. Whoever controls the hardware controls the frontier.” AI compute refers to the raw processing power required to train machine learning models and run them at inference scale measured in floating-point operations per second (FLOPs) and has emerged as the primary strategic chokepoint in national AI competition.
Executive Summary
Training a frontier AI model requires extraordinary computational resources: GPT-4 reportedly used an estimated 10^24-10^25 floating-point operations during training, running on tens of thousands of specialized GPUs for months. This concentration of compute requirements has made access to advanced AI chips primarily Nvidia H100 and A100 GPUs the central lever of US technology export control policy toward China. Compute is also the organizing concept behind AI capability governance: the EU AI Act uses a 10^25 FLOPs threshold to classify general-purpose AI models, and the US Executive Order on AI required companies training models above this threshold to notify the federal government. Understanding compute is essential for interpreting the entire landscape of AI regulation and geopolitical competition.
The Strategic Mechanism
- Training compute: The FLOPs consumed to train a model from scratch. Frontier models (GPT-4, Gemini Ultra, Claude 3 Opus) require 10^23-10^25 FLOPs, translating to $50-200 million in cloud compute costs.
- Inference compute: The FLOPs consumed each time a model responds to a user query. At scale, inference cost exceeds training cost over a model’s deployed lifetime, making efficiency a core commercial and environmental variable.
- GPU clusters: Modern AI training requires thousands of GPUs operating in parallel with high-speed interconnects. Nvidia’s H100 GPU (2023) delivers 4x the training speed of its A100 predecessor, making cluster architecture a strategic variable.
- Compute efficiency: Research advances like distillation, quantization, and mixture-of-experts (MoE) architectures allow capable models to be trained with less compute. DeepSeek’s reported training of a GPT-4-class model for $6 million challenged the assumption that capability requires ever-escalating compute budgets.
- Compute governance: Using compute thresholds as a proxy for AI capability in regulation, since training runs are observable and quantifiable even when model capabilities are difficult to evaluate directly.
Market & Policy Impact
- Nvidia’s data center revenue grew from $15 billion in fiscal 2023 to $47.5 billion in fiscal 2024, driven almost entirely by AI training GPU demand illustrating compute’s commercial centrality.
- US export controls imposed in October 2022, October 2023, and updated in October 2024 banned sales of Nvidia H100s, A100s, and advanced AI chips to China and a growing list of restricted entities, making compute access the primary US-China AI competition lever.
- Microsoft committed to $80 billion in AI data center investment for fiscal 2025, Amazon Web Services to $75 billion, and Google to $50 billion reflecting how cloud compute infrastructure has become a strategic asset class.
- China’s Huawei responded with the Ascend 910B AI accelerator, estimated at 60-80% of H100 performance, as part of a state-backed effort to circumvent compute export controls through domestic chip development.
- The Compute Index, maintained by Epoch AI, tracks training compute used by frontier models across time, showing a roughly 4x annual increase from 2010-2022 the empirical basis for compute threshold governance frameworks.
Modern Case Study: The H100 Export Control and China’s Compute Gap, 2022-2025
In October 2022, the US Bureau of Industry and Security imposed export controls on advanced AI chips, including Nvidia’s A100 and H800. The controls were tightened in October 2023 and again in October 2024 to close workarounds. The stated goal was to prevent Chinese labs from acquiring the compute needed to train frontier AI models. By 2024, the controls had measurably constrained Chinese access to frontier training clusters: while US labs operated clusters of 16,000-100,000 H100s, leading Chinese labs faced fragmented, smaller-scale infrastructure. However, DeepSeek’s January 2025 release of R1 a model claiming to match OpenAI’s o1 at a reported training cost of $6 million suggested that compute efficiency research could partially offset hardware access gaps. The episode forced a recalibration of compute-based governance: export controls remain a significant constraint but are not a permanent barrier to Chinese AI capability development.