The infrastructure economy of AI is consolidating at speed. Anthropic committed to spending $30 billion on Azure compute over multiple years, while Nvidia invested $10 billion in equity and Microsoft added $5 billion—a circular deal where Microsoft transforms investment into a customer, then likely buys Nvidia chips to fulfill the order. This structure reveals the real constraint: not models, but compute. It also exposes a structural problem. When your three dominant private labs—OpenAI, Anthropic, Google—are simultaneously customer and investor to the same semiconductor and cloud vendors, the appearance of collusion matters less than the reality of vertical capture. Each reinforces the other. Anthropic's move signals desperation and opportunity in equal measure. They trail OpenAI in consumer adoption despite superior API revenue, a reminder that distribution compounds faster than capability. Jensen Huang was right to call this a scaling moment, but it comes at the cost of dependency.
OpenAI's group chat feature—collaborative, interactive conversations powered by Claude—looks innocuous until you map the strategy. Shared chats drop model selection, memory access, and user customization to defaults. It's a trojan horse for social lock-in. With 800 million weekly users and Sam Altman's track record of bundling features into ecosystem stickiness, the path toward a WeChat-like super-app becomes legible. Music streaming, commerce, file management already exist inside ChatGPT; group chat is the social vertebra holding it together.
Google's Gemini 3 lands where it matters most: coding, multimodal reasoning, benchmarks. They've launched Antigravity, a no-code IDE competing directly with Cursor and Loveable. Google always ships late but ships at scale—90% search market share means distribution others will never match. But this creates a different problem: responsibility. When you power billions of devices and users, a flawed AI rollout damages trust across your entire product suite. Apple has bet this fear too heavily; Meta has lost the game already. Google must thread the needle between innovation speed and genuine safety, a tension their previous AI initiatives have struggled with.
The more structurally interesting move belongs to Jeff Bezos. Project Prometheus, with $6.2 billion in early funding and Bezos as co-CEO, targets vertical AI: science, chemistry, physics, robotics. The bet is that generalist models like GPT-5 waste capacity trying to excel everywhere. Specialized models trained narrowly—a wine-making AI, a protein-folding AI, a materials-science AI—will outperform ten-times-larger generalist systems on their domain. This inverts the current paradigm. Instead of one foundation model powering everything, you get many small, savage experts. It's the professionalization of AI, and it requires the kind of patient capital and technical depth only Bezos or Musk seem willing to deploy.
What ties these threads: the infrastructure game is consolidating (Nvidia wins again), the consumer game is flattening (OpenAI dominates), and the vertical game is just beginning. Most missed the last one.