The industry is fragmenting into layers, each controlled by different players with conflicting incentives. At the top, model makers are racing to lock in distribution and compute while their products still command premium pricing. Below them, a new class of compression and optimization companies is quietly making those models cheaper and faster, undercutting the margin story. And underneath it all, the real leverage is shifting to whoever controls the data, the inference infrastructure, and the ability to turn AI from a question-answering tool into an agent that actually executes transactions.
Nvidia's networking business pulling in $11 billion last quarter with minimal attention reveals the real profit center in AI infrastructure. Multiverse Computing's compression work on models from OpenAI, Meta, DeepSeek, and Mistral AI shows that frontier model capability is becoming a commodity faster than anyone expected. Meanwhile, Microsoft is acquiring entire teams like Cove to fold collaboration tools directly into its stack, and Snowflake is building Project SnowWork to move from answering questions about data to executing decisions on data. The message is clear: whoever owns the execution layer owns the customer relationship. OpenAI's $50 billion AWS deal is less a vote of confidence in Amazon than a defensive move to keep Microsoft from controlling all three layers at once. Walmart's pivot away from OpenAI's Instant Checkout feature and back to embedding Sparky directly into ChatGPT and Gemini shows that integrations matter more than specialized agents when the underlying models are interchangeable.
The incentive structure is also exposing cracks in the "fair use" narrative. Patreon's CEO is calling out AI companies' licensing hypocrisy: they claim fair use when training on creator content but license from major publishers when they need quality data. The DOD labeling Anthropic a supply-chain risk because it might refuse to disable its safety measures during warfighting reveals that the real tension is not between safety and capability but between corporate control and state control. And the Arena leaderboard being funded by the companies it ranks is the infrastructure version of the same problem: the judges are players in the game. Entry-level software developers face a 20% hiring decline and wage pressure as AI adoption accelerates, which means the labor market is already pricing in the assumption that junior talent is less valuable. The industry is not democratizing anything. It is consolidating power while distributing the cost.
Sloane Duvall