From Data to Compute to Power: Rethinking the Geopolitics of Artificial Intelligence

Introduction

In Chip War, Chris Miller argues that advanced computing power now underpins economic competitiveness, military capability and scientific progress. The race for AI leadership is therefore also a race for semiconductor supremacy.

That claim sits awkwardly with an older assumption. If AI runs on data, countries with the largest populations, India and China, should lead it. They don't. The binding constraint has moved from data to compute, and it is now moving again, toward the electricity that powers compute. "Compute is king" is probably a transitional phase.

The Early Assumption: More People, More Data, More AI Power

For a decade the logic seemed airtight. More people generate more digital activity, which yields more data and better AI. That logic was built for an earlier generation of systems.

Recommendation engines, ad targeting, facial recognition and fraud detection learn from behavioural data such as clicks, payments, faces and location trails. Population is a real advantage there, and China's e-commerce and payments ecosystem exploited it.

Frontier language models learn differently. They are trained mostly on public text, code, scientific literature and media, which is a shared global pool. A billion additional smartphone users add little to it. That is why population stopped being decisive, and why the United States leads frontier AI with a fraction of the demographic base.

Why Compute Became the Bottleneck

Once data is largely commoditised, the scarce input is the ability to process it. Data is the raw material and GPU clusters are the factory. The factory rests on a very narrow industrial base.

In the second quarter of 2026, TSMC captured 72.5% of foundry revenue among the top ten foundries, on sales near $40.2 billion, with its 5/4nm and 3nm capacity fully booked by AI server GPUs and other accelerators. Upstream, ASML is the only company in the world able to produce EUV lithography machines, the tools used to make the most advanced chips. Design, fabrication and tooling are concentrated in three ecosystems, American, Taiwanese and European, and no one else can currently replicate all three. China is trying. Reuters reported in December 2025 that a Chinese EUV prototype had been developed by a team that included former ASML engineers, though a prototype is a long way from volume production. Export controls on chips and tools have become the central instrument of U.S.-China competition. (Press Center +2)

China's Constraint, and Its Counter-Example

China shows both the strengths and the limits of demographic scale. Its datasets were enormous, but in the foundation-model era it hit the same hardware dependency as everyone else, compounded by sanctions.

It also complicates the "compute is king" story. After U.S. rules barred sales of its top chips to China, Nvidia modified the H100 into a China-legal version, the H800, which mainly halved chip-to-chip data transfer speed. DeepSeek's V3 model was trained on 2,048 of these H800 GPUs, consuming about 2.8 million GPU hours in total. At an assumed rental rate of $2 per hour, that works out to roughly $5.576 million. The figure needs a caveat. DeepSeek itself says it excludes earlier research and ablation experiments, so it is the price of the final run, not of building a frontier lab. Even so, the episode suggests that engineering efficiency can partly substitute for raw hardware, and that constraint pushed Chinese labs toward it. (www.malaymail.com +3)

India: Data and Talent Are Not Enough

India has a population comparable to China's, a deep engineering workforce and a vast digital footprint. It has not produced a frontier model. The Stanford AI Index 2025 put India second globally in AI skill penetration, behind only the United States, yet found it lags in retaining talent, attracting private investment and building intellectual property. (tele) (communicationstoday)

The IndiaAI Mission is an attempt to close the compute gap. It has an outlay of ₹10,372 crore over five years and has onboarded more than 38,000 GPUs, well past its original target of 10,000. But as industry experts quoted by Business Standard note, closing the capability gap will require much more than hardware access. Chips alone do not produce frontier teams, models or customers. (Centre To Set Up 600 Data Labs, Deploy 38K GPUs: IndiaAI Mission CEO +2)

Compute's Limits: Scaling Is Changing Shape, Not Ending

The case for compute rests on scaling laws, the observation that larger models trained with more compute perform better. Diminishing returns on any single axis are inevitable, and each tenfold increase in spending may buy less than the last.

But so far the pattern is a shift in where compute is spent, not a plateau. Reinforcement-learning post-training and inference-time reasoning have opened new scaling axes rather than exhausting the old one. The defensible claim is therefore a narrower one. Returns to pre-training may flatten while compute demand migrates toward inference and post-training. Compute stays central, but a given quantity of it buys a smaller edge, and efficiency, algorithms and data quality gain weight.

The Next Bottlenecks

1. Power, the binding input to compute. Energy is not a rival to compute. It is the constraint compute runs into. The IEA's central projection is that data-centre electricity consumption will more than double to around 945 TWh by 2030, from 415 TWh, about 1.5% of world electricity use, in 2024. The scale at the facility level is striking. A typical AI-focused data centre uses as much electricity as 100,000 households, and the largest now under construction will use 20 times as much. The IEA also stresses that countries able to deliver the energy needed at speed and scale will be best placed to benefit from AI. (Cite report +3)

The demand is concentrated. The United States accounted for 45% of data-centre consumption in 2024 and China for 25%. Globally, data centres account for around one-tenth of electricity demand growth to 2030, but in the United States they account for nearly half. The projections are also uncertain. By 2035, the IEA's scenarios range from 700 to 1,700 TWh. The contest after chips is over who can build generation and grid capacity fast enough to run the chips they have. (Cite report +3)

2. High-quality knowledge. As routine web text is exhausted, scarce, curated and domain-specific material (scientific literature, engineering practice, clinical and industrial data) becomes more valuable than volume. Countries with deep research institutions hold an advantage that population alone does not provide.

3. Algorithms. Transformers, reinforcement learning and mixture-of-experts architectures each delivered gains that hardware alone could not. If scaling returns soften, research quality becomes the differentiator again, and DeepSeek's efficiency work is an early example.

4. Human capital, and the ability to keep it. Populous countries could gain influence in a more knowledge-driven phase, but only if they retain their talent. On Stanford's measure of net AI talent migration, India scored -1.55, a net outflow of roughly two AI professionals per 10,000 LinkedIn users. The picture at the very top is changing slowly. In a study by the Paulson Institute's MacroPolo, nearly all top Indian AI researchers worked abroad in 2019, but by 2022 one-fifth stayed in India. Scale in scientists and engineers becomes an advantage only when domestic institutions, capital and compute give them reason to stay and build. (tele) (paulsoninstitute)

Conclusion

Miller is right that semiconductors are the strategic resource of the AI age. Today, access to advanced compute largely determines who can build frontier models, and the supply chain behind it is concentrated in a few firms and geographies.

But the sequence matters. Data made the AI revolution possible. Compute is now the constraint, and the next binding constraint is likely to be power, followed by the quality of knowledge and the pace of algorithmic innovation. Population matters only to the extent that a country converts it into institutions, capital, energy and retained talent.

The winners of the AI age may not be those with the largest GPU clusters, but those who turn computation into knowledge and knowledge into durable economic and strategic power.

Sources

TrendForce, "Global Foundry Revenue Approaches US$53.49 Billion in 2Q26," 9 September 2026 (trendforce.com/presscenter/news/20260909-13225.html)

IEA, Energy and AI, April 2025 (iea.org/reports/energy-and-ai)

Paulson Institute / MacroPolo, Global AI Talent Tracker update, March 2024

Stanford HAI, AI Index 2025

DeepSeek-V3 technical report, with independent analyses (Epoch/Interconnects, Bernstein)

Reuters reporting on the Nvidia H800 (March 2023) and on the Chinese EUV prototype (December 2025)

AP/Euronews on ASML's EUV position

Government of India / IndiaAI Mission announcements; Business Standard, 2026

Verification notes

Opened at the primary source: the TrendForce and IEA figures.

Cross-checked across several outlets but not opened at the primary source: the Stanford AI Index figures, the DeepSeek V3 numbers and the IndiaAI figures. I'd open the originals before publishing. The Stanford 2025 edition is also the latest I verified, so check whether a 2026 edition updates it.

Left out because only one source supported them: TSMC's "~90% of advanced nodes," the IndiaAI GPU utilisation rate, India's cumulative private AI investment, and the "80% of India-educated researchers work in the U.S." claim.

Analysis, not fact: the scaling and algorithms sections are argument and carry no figures.

Generated by Sonnet 5.5 and Copilot