Review: Silicon Empires by Nick Srnicek
Gabriella Chronis
The question Nick Srnicek tries to answer with Silicon Empires is, "what form will AI superpower take?" What this book offers us as protagonists in struggles against data centers is information as ammunition. It empowers us to counter the primary myths that data center developers rely on and offers an understanding of which corporate actors have the most power, how they are allied, and where they are in conflict. The basic story he tells is how the "Silicon Valley consensus"—an era of unregulated growth of Big Tech, where US financial policy served global financial interests—is giving way to a new tech-industrial complex where AI progress is inseparable from an empire-building political project.
The first chapter is a good primer on what AI (a.k.a. large language models or LLMs) is and how it works. This chapter is skippable if you already have a baseline understanding of LLMs. The rest of the book organizes the overwhelming stream of AI news into straightforward explanations of the economic structures, alliances, and flows forming around this new “General Purpose Technology.”
Chapter 2 covers four strategies for how different firms seek to profit at different layers in the “AI stack”—AI model research and development, software integrations, and physical infrastructure. Google is making integrated systems for specific industries with the aim of becoming the de facto provider for entire sectors like education or materials science. Amazon is concentrating on infrastructure, aiming to control cloud computing by encouraging platform lock-in, where other companies (including AI companies) design their systems to work only on Amazon servers. The few companies that work on LLM research and development, like OpenAI and Anthropic, concentrate on finding the “best” models, by searching through all possible combinations of parameters.
A few powerful actors control the game, especially at the infrastructure level. The Silicon Valley consensus has meant that US policy has protected Big Tech from anti-trust litigation. Now, these companies have the power to heavily influence the market. Even though US policy towards Big Tech monopolies continues to be lax, these companies still use strategies to dodge potential anti-trust scrutiny. One such strategy is to hire all of the major talent from a small startup and then shut the company down without acquiring it.
This chapter answers the question of who is making money from AI. From model to software to servers to microchips, "value concentrates up the chain": infrastructure (computing hardware, and the cloud service providers like AWS that provision it) is far and away more profitable than AI model or application development. In fact, AI developers like OpenAI and Anthropic still operate at a loss; their biggest investors are cloud infrastructure companies like Amazon and Microsoft. These cloud providers rent out their server space to other businesses, who operate platforms without˚ owning and operating their own servers. Cloud companies are investing in AI to ensure that whichever model gets used the most by the public, whether Claude or ChatGPT or Grok, ends up running on their servers, thereby securing server rent. For example, Microsoft’s massive investments in OpenAI ensure that if ChatGPT is widely adopted, Microsoft will reap the benefits in the form of rent extracted from OpenAI. Microsoft's investments also tip the scales in favor of their horse in the race.
Meanwhile, smaller AI companies without Big Tech investment money are dropping out of developing new models because it is so expensive to do. Despite apparent competition, the AI boom is actually strengthening existing tech monopolies that grew out of the Silicon Valley consensus.
Data centers operate as landlords. By focusing on where the money is made in AI, our movement has already put Srnicek’s analysis into action. He explains why this intuition is correct.
This analysis also reveals that the industry is being driven not by user demand at the bottom of the chain but by speculative finance at the top. Rapid AI development is indeed being driven by user demand. However, the “users” of AI are not the public but rather data centers themselves. Data centers are not being built in response to an increased demand for compute. Rather, by subsidizing unprofitable AI companies, the AI cloud providers are manufacturing demand for the massive amounts of compute they are bringing online with new data centers. The real demand for AI is from all the new servers in need of customers. In other words, the actual driver of AI development is supply-push from the computer-landlords rather than demand-pull driven by the code's benefits.
Chapter 3, "Interregnum," maps the transformation of this consensus as the industry allies with new nationalist projects. Srnicek slots specific US government policies into a larger trend away from the interests of global finance towards more nationally focused growth. He illustrates that the reason we "need" gigawatts is not for diffuse human benefit, not to meet the future that arrives no matter what, but to win a rent extraction race where it's the stock-less, and the global south paying the rent.
Chapter 4 offers a long view of the data center boom and geopolitics. It deconstructs the myth that the US and China are locked in an AI arms race: only the US approach to AI prioritizes pushing the absolute technological capabilities via private capital, centralized mega-compute infrastructure, and market-driven models. The purported arms race is almost exclusively between American companies: OpenAI, Google, xAI, Microsoft, Amazon, Oracle, and Meta, with Alibaba as the lone Chinese exception. China's approach, on the other hand, emphasizes broad economic integration, decentralized application, support for open architectures, and public-sector adoption.
China differs significantly from the US in its AI strategy, pursuing diffusion first (widespread adoption) rather than innovation. This strategy is driven in part by necessity: China is more constrained in its access to natural resources.
The US and China are in competition. Not a competition to create the most cutting edge technology, but a competition for global platform dominance, that will be determined by which power can exert the most global influence through AI. Both the US and China are working to encourage other nations to use their infrastructure, with the aim of making entire nations technologically reliant on them.
Where the old status quo empowered global capital to make whatever alliances were most profitable, the US now encourages domestic self sufficiency. Not all tech companies are equally supportive of the US’ new foreign policy. The narrative of the US-China arms race is pushed especially hard by newer defense tech companies and AI model companies that profit handsomely from increased tensions and conflict between the two nations.
Though he never says it outright, the suggestion is that the US race for cutting edge technological innovation will ultimately fail as a strategy for global influence. More concretely, the data center buildout is less a natural consequence of social progress than it is a physical manifestation of the US government’s policy of being the biggest bully on the world stage.
Srnicek is restrained in weighing in on who will win, instead putting the evidence in front of the reader and letting us connect the dots. One of Srnicek's strongest points is that historically, what he calls "general purpose technologies" (GPTs) impact global power hierarchies by widespread adoption of technologies (like electricity), rather than initial invention. (GPTs, like electricity and steam engines, are a useful concept from economics with a now confusingly overloaded acronym. The GPT in ChatGPT is different, standing for “Generative Pretrained Transformer”).
Srnicek reminds us that the data center movement must be a global fight in addition to a local struggle. All our struggles are connected: a fight against a data center anywhere is a fight against data centers everywhere.