Intel and Google Multi-Year AI Infrastructure Partnership

For the past three years, the AI infrastructure story has been written almost entirely in GPU terms — who can secure the most Nvidia H100s, who can build the biggest training clusters, who can scale GPU interconnects the fastest. In that narrative, CPUs were afterthoughts, the quiet janitors of the data center while GPUs did the glamorous work of training foundation models. The Intel-Google partnership announcement on April 9 signals that this narrative is breaking down, and breaking down fast.

The structural driver is the shift from AI training to inference and deployment at scale, compounded by the rise of agentic AI. As analyst Stephen Sopko of HyperFrame Research put it, "CPUs are no longer seen as background infrastructure; they are becoming the active bottlenecks." The reason is architectural: agentic AI systems — where autonomous agents orchestrate API calls, query databases, manage state, and coordinate with other agents — generate workloads that are fundamentally CPU-bound. These are not matrix multiplication tasks suited to GPU parallelism; they are serial, branching, I/O-heavy operations that general-purpose processors handle best. Constellation Research analyst Holger Mueller reinforced this point: "In the agentic world where agents call APIs and business applications, CPUs are the best to do the job." What makes this consequential is scale. Every GPU-optimized AI cluster still relies on host CPUs for orchestration, pre/post-processing, and system management. As inference volumes explode — driven by millions of AI agents running continuously rather than batch training jobs — the CPU-to-GPU ratio in data centers is being re-examined. Google’s commitment to multiple generations of Xeon 6 is a bet that this ratio needs to tilt back toward more capable CPUs, not just more GPUs.

The Xeon commitment grabbed the headlines, but the more strategically significant element of the Intel-Google deal may be the expanded co-development of custom ASIC-based Infrastructure Processing Units. IPUs offload networking, storage, and security functions from host CPUs — functions that become exponentially more demanding as AI clusters scale. Google and Intel have been building this relationship since the 2022 launch of the Mount Evans IPU, which enabled 200 Gbps networking in Google Cloud C3 instances. The new deal deepens that co-development across future generations.

What makes the IPU angle underappreciated is the business it has already built. Intel’s custom ASIC division grew over 50% in 2025, reaching an annualized revenue run rate exceeding $1 billion. This positions Intel not just as a CPU vendor but as a custom silicon partner for hyperscalers — a business model closer to Broadcom’s highly profitable custom chip operation than to Intel’s traditional merchant silicon approach. The IPU co-development with Google is the flagship example: rather than selling commodity chips, Intel is embedding itself into Google’s infrastructure architecture at the design level, creating switching costs that commodity CPU sales alone cannot. This also opens a competitive front against Nvidia’s DPU (Data Processing Unit) business, which targets the same infrastructure offloading workloads. While Nvidia dominates in GPUs, its DPU penetration at hyperscalers is less entrenched, giving Intel an opening — particularly when the IPU is co-designed with the customer’s specific requirements.

Google’s decision to deepen its Intel partnership should not be read as a vote of exclusive confidence in Intel — it is a calculated diversification play. Google simultaneously develops its own Axion Arm-based processors, uses AMD EPYC chips, designs custom TPUs for AI training, and relies heavily on Nvidia GPUs. The Intel commitment adds another leg to a deliberately multi-vendor strategy designed to prevent dependency on any single supplier. As Amin Vahdat, Google’s SVP and Chief Technologist for AI Infrastructure, noted, "Intel has been a trusted partner for nearly two decades, and their Xeon roadmap gives us confidence" — careful language that endorses Intel’s roadmap without declaring exclusivity.

The timing is revealing. Reports indicate Anthropic is eyeing its own custom chips, and every major hyperscaler is investing in proprietary silicon. In this environment, maintaining strong relationships with merchant silicon vendors like Intel gives Google leverage in multiple directions: it keeps pricing competitive against custom alternatives, ensures supply resilience, and provides a mature ecosystem (software stack, security features like TDX, validated instance types like C4 and N4) that custom chips take years to replicate. For Intel, the risk is that Google’s commitment is broad but shallow — a multiyear deal with no disclosed financial terms could mean anything from billions in procurement to a modest extension of existing volumes. The market’s reaction — a 4.7% stock jump on 154.3 million shares traded — suggests investors are giving Intel the benefit of the doubt, but the absence of financial specifics is a notable gap that bears watching.