AI race is turning infrastructural
The contest in AI is shifting from model glamour to who controls efficient compute and power at scale. Google struck a multiyear deal with Intel to supply Xeon CPUs and co-develop IPUs for Google Cloud, signalling a push to diversify hardware paths rather than rely on a single accelerator type (verdict.co.uk). Analysts note more than 60% of global AI compute now sits with hyperscalers, led by Google, and engineers are seeing real bottlenecks from data-centre power and transformer shortages — a structural constraint on growth (networkworld.com, []).
Google just did something that looks boring next to a flashy new chatbot: it signed a multiyear infrastructure deal with Intel for Xeon central processing unit chips and custom infrastructure processing units inside Google Cloud. The point was not a new model launch; the point was locking in the plumbing that keeps artificial intelligence systems running at scale. (intel.com, cnbc.com) That matters because artificial intelligence runs on more than one kind of chip. Graphics processing units handle the heavy matrix math, but central processing units still coordinate storage, networking, scheduling, and a lot of inference and serving work, which is why Google said its C4 and N4 cloud instances already use Intel Xeon 6 processors. (intel.com, networkworld.com) Google and Intel also said they will expand work on custom infrastructure processing units, which are chips that move data around a data center the way a traffic system moves cars through a city. In artificial intelligence clusters, that traffic job decides whether expensive accelerators stay busy or sit idle waiting for data. (intel.com, tech.yahoo.com) The backdrop is that compute is concentrating fast. Network World, citing Epoch AI analysis, reported this week that more than 60% of global artificial intelligence compute capacity now sits with hyperscalers, and Epoch AI estimated Google alone held about one quarter of global cumulative capacity by the fourth quarter of 2025. (networkworld.com, epoch.ai) Google got there by not betting on one supplier. Epoch AI says most of Google’s artificial intelligence compute comes from its own tensor processing units rather than Nvidia graphics processing units, and the new Intel deal adds another lane by keeping Xeon central processing units and custom networking silicon in the mix. (epoch.ai, intel.com) That is why the competition is starting to look less like “whose model is smartest” and more like “who can assemble a whole working factory.” Google’s own cloud blog describes its Artificial Intelligence Hypercomputer as a stack of hardware and software built for performance and efficiency at scale, which is another way of saying the winner needs the whole system, not just the fastest chip. (cloud.google.com, networkworld.com) Now the physical bottleneck is shifting again, away from chips and toward electricity equipment. Bloomberg reported on April 1 that transformers, switchgear, and batteries are becoming hard to source for new United States data centers, and market trackers cited in follow-up coverage say close to half of planned 2026 capacity faces delay or cancellation. (bloomberg.com, computeforecast.com) A transformer is the metal box that changes voltage so power can move from the grid into a building without frying everything inside. Compute can be ordered with a purchase order, but a five-year wait for a high-power transformer can leave a billion-dollar site stuck as concrete and steel. (computeforecast.com, finance.yahoo.com) So the new artificial intelligence race is being fought on three layers at once: custom accelerators, general-purpose chips, and the electrical gear outside the building. Google’s Intel deal makes sense in that world because a cloud company that can mix tensor processing units, Xeon central processing units, and custom infrastructure chips has more ways to keep growing when any single part of the supply chain tightens. (epoch.ai, intel.com, bloomberg.com)
