Google and Intel have expanded a multiyear partnership centred on the less glamorous but increasingly strategic layers of artificial intelligence infrastructure, betting that the next phase of AI growth will depend not only on headline-grabbing accelerators but also on the processors and networking hardware that keep large systems running efficiently. Under the agreement announced on April 9, Google Cloud will continue deploying Intel Xeon processors, including Xeon 6, across AI, inference and general-purpose workloads, while the two companies broaden co-development of custom ASIC-based infrastructure processing units, or IPUs. That focus matters because the industry’s centre of gravity is shifting. Training frontier models still consumes huge sums of capital and power, but hyperscalers are now grappling with an equally pressing challenge: how to serve those models at scale once they are deployed. Reuters reported that rising demand for inference and for agentic AI systems, which carry out multi-step tasks, is renewing demand for general-purpose CPUs even as specialist accelerators remain central to the stack. Deloitte said inference is expected to account for roughly two-thirds of AI compute in 2026, underlining why cloud operators are paying closer attention to orchestration, data movement and system efficiency.
Intel is trying to use that shift to reclaim relevance in a market where Nvidia has dominated the AI boom and where custom silicon from cloud groups has eroded the role of standard server processors. Lip-Bu Tan, Intel’s chief executive, framed the partnership as a statement that “AI doesn’t run on accelerators alone”, arguing that balanced systems built around CPUs and IPUs are critical to performance, efficiency and flexibility. Google’s Amin Vahdat, senior vice-president and chief technologist for AI infrastructure, said Intel had been a trusted partner for nearly two decades and pointed to confidence in the Xeon roadmap as Google works to meet growing workload demands.
The IPU side of the partnership is especially significant because it speaks to how hyperscale data centres are being redesigned. Intel describes IPUs as hardware that offloads networking, storage and security functions from host CPUs, improving utilisation and performance while isolating tenant applications from provider services. In practical terms, that means freeing more of the central processor for customer workloads while shifting infrastructure housekeeping to dedicated silicon. Intel and Google have been working on this concept for years: Google Cloud’s C3 machine series, introduced in 2022, was built around 4th Gen Intel Xeon processors and Google’s custom Intel IPU, with Google saying the architecture delivered more predictable compute, accelerated networking and stronger isolation.
The latest agreement extends that model into a broader AI infrastructure strategy. Intel said Google Cloud will align across multiple generations of Xeon processors to improve performance, energy efficiency and total cost of ownership across Google’s global estate. The company also said Xeon 6 is already powering Google Cloud’s C4 and N4 instances. Google’s own cloud material shows C4 virtual machines based on Xeon 6 offer up to 30% gains for general compute, up to 60% for machine-learning recommendation workloads, and lower latency on storage-heavy configurations, all of which support Intel’s case that CPUs still have a strong commercial role in AI-heavy environments.
For Google, the deal appears less about dependency on a single supplier than about keeping a broad infrastructure toolkit. The company already has a long history of building specialised silicon, most notably its Tensor Processing Units, while also using other merchant and custom components across its fleet. That makes the Intel tie-up part of a wider pattern in which large cloud operators mix proprietary accelerators, merchant CPUs and purpose-built networking or offload chips to reduce bottlenecks and manage costs. Google said as far back as the C3 rollout that it had spent more than two decades designing workload-optimised infrastructure, combining purpose-built hardware with prescriptive architectures and an open ecosystem.
For Intel, the announcement lands at a moment when the company is trying to show investors that it still has routes into the AI build-out beyond the race for top-end training chips. Reuters noted that the Google pact came as Intel sought to strengthen its position after losing ground in the early years of the AI surge. Market coverage after the announcement pointed to a favourable investor response, but the harder test will be whether such partnerships translate into durable share in hyperscale infrastructure as cloud operators continue to diversify suppliers and design more of their own silicon.
Topics
Technology