More on the topic...
Generating detailed summary...
Failed to generate summary. Please try again.
The AI arms race has moved beyond just building smarter models. Today’s battle spans four layers: hardware (GPUs and accelerators), compute clusters (data centers and clouds), model architectures (GPT, Claude) and applications (ChatGPT, Cursor). Owning a slice of this stack no longer guarantees profits. Leading labs are pushing up and down the chain—software teams are building custom chips and data centers, while cloud and hardware giants are moving into models and apps to defend their turf and capture new revenue.
OpenAI and Anthropic started with models but found margin pressure so intense they turned to infrastructure. OpenAI teamed with Broadcom to build its own inference chip, Jalapeño, set to ship by late 2026. Anthropic struck a $50 billion deal with FluidStack to build multi-gigawatt “neocloud” centers in Texas and New York. xAI went all in, leveraging SpaceX’s resources to build the Colossus supercomputer and rent out compute—Anthropic reportedly pays $1.25 billion a month for access.
On the application side, Cursor shows how deep UX data and workflow integration can become a moat. After routing millions of coding sessions through GPT and Claude, Cursor trained its own Composer 2.5 model on that data. It now handles routine tasks in-house and uses frontier APIs only for complex queries. SpaceX acquired Cursor’s developer Anysphere in June 2026 for $60 billion, illustrating how infrastructure players need apps to drive usage. Meanwhile, Amazon and Microsoft, each tied to major labs, are expanding up into software and down into custom silicon—AWS with Trainium and Inferentia chips, and Microsoft embedding its Phi and MAI models into Windows and 365.
At the base, Nvidia still rules GPU sales but has begun open-sourcing its Nemotron 3 models—weights, data and all—to lock in demand for its hardware. Any team that grabs an open Nemotron model must buy Nvidia GPUs to run it. And then there’s Google. It manufactures TPUs, runs vast data centers, develops DeepMind and Gemini models, and embeds AI across Search, Workspace and Android. No other company occupies all four layers simultaneously.
Questions about this article
No questions yet.