The Thirty-Second War
Somewhere in eastern Ukraine, an AI-enabled drone has just compressed the entire history of military strategy into thirty seconds. Target identified through satellite imagery and analyzed by machine learning algorithms. Flight path calculated. Jamming countermeasures anticipated. In the final phase of flight, the drone's AI locks onto its pre-identified target independently—Russian electronic warfare having forced Ukrainian engineers to build backup systems that can function when GPS and communication links fail. From detection to strike: thirty seconds.
This is not a thought experiment. According to CSIS analysis of the Russia-Ukraine conflict, Ukrainian forces have reduced targeting cycles to just over thirty seconds through AI-assisted systems, fundamentally compressing the observe-orient-decide-act loop that has defined warfare for generations. Ukraine has employed AI-powered facial recognition to identify over 250,000 Russian soldiers operating in the country, creating a database used for both military targeting and psychological warfare—notifying families in Russia of casualties.
To be sure, critics from the tradition of critical technology studies would argue that I'm presenting this as inevitable technological progress rather than a political choice to remove human judgment from life-and-death decisions. They're not entirely wrong. The thirty-second targeting cycle represents a choice—by Ukrainian forces under existential threat—about acceptable trade-offs between speed and deliberation. What strikes me is not that this choice was predetermined by technology, but rather that once one side makes it, the pressure on adversaries to match that capability becomes intense. This is the security dilemma playing out at machine speed.
What the Ukraine example reveals is the broader transformation of artificial intelligence from a commercial technology into a fundamental instrument of state power—and the question of whether that transformation leads to permanent fragmentation or merely contested integration within an interdependent system remains genuinely uncertain. We are witnessing extraordinary capital concentration, geopolitical tensions creating real friction in technology flows, and vertical integration at unprecedented scale. Whether these forces prove self-reinforcing (driving toward separate technological ecosystems) or self-limiting (creating counterpressures for integration) will depend on political choices not yet made.
This is not the only future possible from where we stand today.
The Cycles of Integration and Specialization
I've written before about how the personal computer and cloud computing eras were defined by horizontal specialization—Intel made chips, Microsoft made operating systems, Dell assembled hardware, and thousands of software vendors built applications atop these standardized layers. This modular architecture enabled extraordinary innovation because companies could specialize in specific layers without mastering the entire vertical.
That era is ending—or more precisely, being transformed. But this is not the first time we've witnessed this pattern. The history of management theory offers a useful lens.
In 1937, Ronald Coase asked the foundational question: why do firms exist at all? If markets coordinate economic activity efficiently through prices, why organize production inside companies? His answer: firms exist when the transaction costs of market coordination exceed the costs of internal management. When it's cheaper to coordinate internally than to negotiate thousands of contracts with suppliers, you integrate vertically.
Alfred Chandler's "The Visible Hand" (1977) documented how this played out in the late 19th century. Railroads, steel companies, and oil refineries integrated vertically because the transaction costs of coordinating complex, interdependent operations through markets were prohibitive. Carnegie Steel owned iron mines, coke ovens, railroads, and mills because the performance requirements of making high-quality steel at scale demanded end-to-end control.
But Clayton Christensen observed that industries cycle between integration and modularity. When performance is insufficient, companies integrate to optimize the full stack. When performance becomes "good enough," modular specialists emerge, offering flexibility and cost advantages. The PC industry followed this pattern: early computers (IBM mainframes) were vertically integrated, then PCs became modular, then smartphones shifted back toward integration (Apple's vertical approach).
The question is: where does AI infrastructure fall in this cycle?
The capital expenditure numbers suggest we're entering an integration phase of unprecedented scale. In 2025, the major hyperscalers—Amazon, Microsoft, Google, and Meta—spent $405 billion on AI infrastructure, exceeding analyst forecasts by $130 billion. This represents 38-40% of all S&P 500 capital expenditures. These four companies are allocating capital at levels that rival small nations' GDP.
More revealing is how this capital deploys. Approximately 50-60% flows directly to processors and specialized chips—but what's changing is who makes those chips. Amazon's custom Trainium chips have reached multi-billion-dollar run rates, growing 150% quarter-over-quarter. Google's Tensor Processing Units number in the hundreds of thousands. Microsoft's CTO stated the goal of "mainly Microsoft silicon in datacenters"—remarkable for a company that spent decades as Intel's most important software partner.
This shift represents vertical integration at a scale unprecedented in the technology industry. The reason is not primarily cost, though Amazon claims Trainium delivers up to 70% lower cost per inference. The reason is optimization—and here Coase's logic applies directly.
AI workloads appear to have transaction costs that make market coordination inefficient. Google's Ironwood TPU delivers 42.5 exaflops across 9,216-chip superpods—twenty-four times more powerful than the world's largest supercomputer—precisely because hardware and software were co-designed. The coordination required to achieve this performance through negotiated contracts with specialized chip vendors would be prohibitively complex. The transaction costs of specifying requirements, managing interfaces, coordinating updates, and optimizing across layers exceed the costs of bringing chip design in-house.