Who controls AI? The gatekeepers behind the model race
The question of who wins the AI race is inseparable from who controls the infrastructure layers beneath it. And the answer may already be decided.
When Nvidia's market capitalization crossed $3 trillion in mid-2024, briefly making it the most valuable company on earth, most commentary read this as a bet on AI's potential. In reality? It is, more precisely, a bet on who will tax it. NVIDIA, alongside TSMC (the Taiwanese semiconductor foundry on which virtually the entire advanced chip supply chain depends), sits at the deepest layer of what I map in the digital framework as the Rails level: the infrastructure that everyone else runs on. At this layer, it does not matter which model wins, which startup survives the bottleneck, or which software category absorbs the most AI functionality. Every player in the race pays the toll.
Infrastructure incumbents are building down the stack deliberately
Amazon Web Services, Microsoft Azure, and Google Cloud Platform are each pursuing a structurally similar strategy: combining the infrastructure layer — cloud compute, increasingly their own chips (Amazon's Trainium, Google's TPUs, Microsoft's Maia) — with the platform layer above it, through which enterprise customers access and embed AI into their operations. This is a rational response to the bottleneck dynamic, because even if the LLM layer commoditizes and margins compress there, the infrastructure and platform layers remain defensible. Google and Microsoft are not primarily competing with OpenAI; they are competing with each other for the position of collecting a share of every AI transaction running on their infrastructure, regardless of which model performs it.
Market access is itself a bottleneck, and it favors incumbents heavily
Google and Apple represent a different and considerably underappreciated form of gatekeeper advantage: the installed base. Google processes approximately 8.5 billion searches per day. Its user relationship, built over two decades of habit and dependency, gives it a distribution asset that OpenAI cannot replicate through model performance alone. Google's AI integration into Search, Gmail, and Android is less technically spectacular than OpenAI's frontier work, but it reaches a user base that is already there, already habituated, already embedded. Apple's position is even more deliberately constructed: the App Store remains the effective distribution tollgate for AI applications reaching iOS users, which is to say the higher-spending segment of the global smartphone market. Apple has invested relatively little in frontier model research, yet it controls the channel that every AI consumer application company needs to access. This is, in the digital framework's vocabulary, a Closed Garden operating at its most strategically elegant: extracting value from a stack position without having built the technology that runs through it.
OpenAI and Anthropic are in the most exposed position structurally
Pure Players, in our digital framework's vocabulary, have integrated the value chain and removed intermediaries. Even though they haven't yet established the network effects that would make them Platform-level businesses with genuine switching costs. OpenAI's GPT Store and enterprise API ecosystem is an explicit attempt to make that transition. Anthropic, the AI safety company founded in 2021 by former OpenAI researchers, is taking a more defensive stance through its distribution partnerships with AWS and Google Cloud. Maybe it's a smart channel strategy. Maybe it's an acknowledgment that pure platform ambitions at scale are out of reach. Naturally, both companies understand that their current position is structurally vulnerable to the infrastructure players who distribute their models while simultaneously building competing ones.
Same stories, same patterns...
The pattern is one we have documented across multiple technology cycles with our Big Technology Framework: the players who survive the bottleneck in the strongest position are rarely those who build the most impressive technology in the expanding phase of the cone. They are those who, by the time the cone narrows, control a layer of the stack that cannot be bypassed.
Eventually? The hypothesis that OpenAI and Anthropic end up building rather commodity businesses on top of "ancient" (by digital acceleration standards), very profitable businesses is rather plausible.
