SemiAnalysis founder: By 2028, most of the world's AI computing power will belong to only two companies

Author | Xu Chao

      Editor | Hard AI

In the latest episode of the well-known tech podcast Dwarkesh Podcast, SemiAnalysis founder and chief analyst Dylan Patel  predicted: By 2028, laboratories OpenAI and Anthropic may directly control 70% to 80% of the world's incremental computing power and most effective FLOPs.

The deeper impact lies in financial markets—due to the remarkable investment returns shown by AI data centers, tech giants’ extreme competition for capital will raise discount rates and debt issuance rates throughout society, even causing a systemic rise in global borrowing costs.

Key points overview

The main content of this interview includes: extreme centralization of computing power, changing laboratory business models, AI data centers driving up borrowing costs, and the impact on traditional asset valuations and sovereign debt risk.

Over the past year, AI infrastructure has not only become the core engine of US GDP growth, but is also brewing an unprecedented concentration of computing power at its foundation.

According to calculations by Dylan Patel, the computing power scale of OpenAI and Anthropic continues to expand, and their share in the world's incremental computing power keeps rising. It is expected that by 2028, these two will account for 70% to 80% of the global incremental computing power.

This concentration is not slowing down, but rather accelerating. Traditional cloud giants and other corporations with balance sheet capabilities are also building computing infrastructure and then subleasing it to top laboratories.

Because new generations of chips have dramatically improved performance per watt, the leading labs’ control over incremental computing power essentially means they dominate most of the world’s effective computing capacity.

In the past, the operating costs of cutting-edge models were high, and labs were once operating at negative gross margins. But as model capabilities and business efficiency improve, computing output rises significantly, and labs can bear higher computing prices.

From 2024 to 2029, the cumulative capital expenditure of the global AI industry chain is expected to reach about 11 trillion USD. Part of this funding needs to be raised through credit markets. Given the high expected returns of AI infrastructure, tech giants may be willing to accept higher debt issuance rates, which will drive up the funding costs for society as a whole.

The raising of borrowing costs and the central level of risk-free rates will directly reshape the global logic of asset pricing. Traditional assets reliant on stable future cashflows may face valuation pressure, and some highly indebted countries may encounter higher refinancing costs and sovereign default risks.

Frontier labs may reduce their computing power allocated to external commercial inference in the future and invest more into internal model research and recursive self-improvement. The reason is that the long-term capability and potential value gained from internal R&D may exceed the income generated from directly selling tokens to external users.

The interview also discussed the computing power supply chain, data center construction, capital expenditure, regulatory limitations, and possible changes in economic structure brought by AI development. The interviewee believes that the trend of centralizing computing power, capital, and model capabilities may further strengthen the market influence of a handful of top labs.