Google dominates AI compute

Training an artificial intelligence model takes warehouses of chips and electricity, and one fresh estimate says more than 60% of that capacity now sits inside a handful of hyperscalers, with Google holding the biggest single share. (networkworld.com) A hyperscaler is a company that runs cloud data centers at such a large scale that other firms rent computing from it instead of building their own. Google, Microsoft, Meta, and Amazon are the names at the center of this market. (networkworld.com) (epoch.ai) Google got to the front by building its own chips instead of buying only Nvidia chips. Epoch AI says Google’s lead came from custom Tensor Processing Units, which are chips built specifically for machine learning workloads. (epoch.ai) That changes the usual power balance in artificial intelligence. Nvidia still dominates the market for selling chips, but Google controls a huge share of the places where those chips, and its own chips, are actually running. (networkworld.com) (datacenterfrontier.com) Google has kept extending that in-house route in 2026. This week it expanded infrastructure deals with Broadcom, Anthropic, and Intel around future Tensor Processing Units, networking gear, and central processors for artificial intelligence data centers. (crn.com) (newsroom.intel.com) (cnbc.com) When one company owns a large chunk of the machines, access starts to look like airport gates controlled by a few airlines. Startups can still fly, but prices, waiting times, and technical choices are shaped by whoever owns the gates. (networkworld.com) At the same time, Google researchers are showing what they want all that compute to do. A new system called PaperOrchestra breaks one writing job into five smaller agents that handle outlining, plotting, literature review, section drafting, and refinement. (arxiv.org) (nerdleveltech.com) A multi-agent system works like a newsroom with specialized desks instead of one reporter doing everything. PaperOrchestra takes a rough idea summary and experimental logs, then produces a submission-ready LaTeX manuscript with figures and checked citations. (arxiv.org) (marktechpost.com) Google says the system is flexible because it can start from messy research materials rather than a fixed template. The paper reports that the design targets the slowest part of many labs’ workflow: turning notes, tables, and plots into a finished paper. (arxiv.org) Put those two pieces together and the pattern is hard to miss. The same company that owns an outsized share of the computing is also building software that splits work across many specialized agents, which pushes users toward bigger platforms with deeper infrastructure. (networkworld.com) (arxiv.org) That does not mean smaller firms disappear. It means the next fight in artificial intelligence may be less about who has the smartest chatbot and more about who owns the factories, the power lines, and the workflow tools everyone else has to rent. (networkworld.com) (epoch.ai)