Google + Intel: Why the "boring" CPU just became strategic again for AI

Google + Intel: Why the "boring" CPU just became strategic again for AI

Silicon Valley has been obsessed with GPUs for two years, but Google’s latest move with Intel is a reminder: the real power in AI may lie in the less glamorous parts of the stack. By doubling down on Intel’s CPUs and co-developing custom infrastructure chips, Google is quietly shaping how all AI workloads will run at scale — not just the headline-grabbing training runs.

In this piece, we’ll unpack what the new deal actually changes, why CPUs and IPUs suddenly matter again, how this fits into the cloud arms race, and what it means for European organisations trying to keep up without losing digital sovereignty.

The news in brief

According to TechCrunch, Google and Intel have agreed to extend and deepen a multiyear partnership focused on AI infrastructure in Google Cloud.

Key points:

• Google Cloud will continue to rely on Intel Xeon processors, including the newest Xeon 6 line, for AI, cloud and inference workloads.
• The companies will expand co‑development of custom infrastructure processing units (IPUs) – specialised chips that take over networking, security and data-movement tasks from CPUs, especially in data centres.
• This IPU collaboration began in 2021 and now explicitly targets ASIC-based custom designs, meaning fully tailored silicon rather than generic, programmable hardware.
• The agreement lands at a moment when CPUs are in tight supply globally, while demand for AI inference and general cloud workloads is surging.
• TechCrunch notes that Intel declined to disclose any pricing or financial terms.
• The deal follows Arm’s announcement of its own AGI-branded CPU, signalling a broader scramble to secure AI-grade general-purpose compute.

On the surface, this is another cloud–chip vendor tie-up. Underneath, it is a bet on how future AI infrastructure will be architected.

Why this matters

Most of the AI hype has been about GPUs for training giant models. But in commercial reality, the expensive training phase is only a fraction of the story. The real money — and the real infrastructure load — sits in inference: serving those models billions of times per day inside search, ads, productivity tools, and enterprise applications.

Inference is where CPUs and IPUs quietly dominate. Google’s renewed commitment to Intel essentially says:

• Google doesn’t want to run its entire AI future on Nvidia’s terms.
• Custom accelerators like TPUs are vital, but they need a balanced system around them: CPUs for control logic and lighter models, IPUs for data movement and isolation.

Winners:

• Intel gets validation that it is still central to AI infrastructure, even as Nvidia steals headlines. It locks in a hyperscaler design win across multiple Xeon generations and deepens its custom-silicon credentials.
• Google Cloud gains more control over its cost structure and performance profile, by jointly steering the design of the IPUs that sit between its CPUs, GPUs and TPUs.

Losers or at least pressured players:

• Nvidia and AMD’s data-centre ambitions face a subtler form of competition. If Google and Intel can offload more tasks to IPUs and tuned CPUs, they can reserve GPUs for only what truly needs them, reducing overall GPU spend.
• Smaller cloud providers without the scale to co-design chips will compete on commodity hardware while hyperscalers build vertically integrated stacks.

The immediate implication: AI is moving from a “buy more GPUs” mindset to full-stack optimisation. Whoever controls the interaction between CPUs, accelerators and networking will control the economics of AI.

The bigger picture: from GPU land‑grab to system design war

This deal slots into several major industry trends.

1. Hyperscalers want custom silicon everywhere.
Amazon has Graviton, Trainium and Inferentia. Microsoft is rolling out Maia and Cobalt. Google has TPUs and custom networking silicon. Extending IPU co-design with Intel is the next logical step: optimise not just the compute chip, but the data path.

In that sense, Intel isn’t just selling Xeons; it is renting out its foundry and design expertise to a cloud giant whose main business is not hardware. The line between merchant silicon vendor and semi-custom design house is blurring.

2. Data-centre architecture is being re-plumbed.
Traditional servers assumed the CPU does almost everything. Modern AI data centres look different: the CPU orchestrates, GPUs/TPUs accelerate, and IPUs/DPUs handle networking, storage virtualisation, security and isolation. Nvidia has BlueField, AMD has Pensando; Intel needs credible IPUs in the field, and Google’s traffic is a perfect proving ground.

3. The CPU is being re-imagined for AI.
Arm’s AGI CPU announcement — highlighted by TechCrunch as a response to CPU shortages — is one sign. Intel’s Xeon 6 push, promising better performance-per-watt and AI-friendly instructions, is another. The direction of travel is clear:

• Fewer “fat” general-purpose cores.
• More parallelism, vector units and attached accelerators.
• Tighter coupling with the network and storage fabric.

4. The cloud race is now a capex and efficiency race.
As another TechCrunch piece pointed out earlier this year, Google and Amazon are in an AI capex sprint. But the question is no longer who spends the most, it’s who converts capex into usable AI capacity most efficiently. Custom IPUs and tuned CPUs are levers to squeeze more inference per rack, per watt, per euro.

This is not just a chip announcement. It is a signal that the system design war has begun in earnest.

The European angle: sovereignty on someone else’s silicon

For European organisations, this partnership lands in the middle of three converging debates: AI adoption, cloud dependence and digital sovereignty.

On one hand, this is good news:

• Google Cloud regions in the EU, coupled with energy-efficient Xeon 6 and custom IPUs, mean denser AI capacity in European data centres.
• Enterprises that don’t have the scale of a hyperscaler can effectively “rent” cutting-edge infrastructure design without touching hardware.

On the other hand, Brussels has been clear: strategic sectors shouldn’t rest entirely on non-European stacks. The EU Chips Act is pouring billions into attracting fabs from Intel, TSMC and others to European soil, but the control plane — who designs the systems, who sets APIs, who controls telemetry — still largely lives in the US.

This deal underscores that tension:

• Intel is heavily investing in fabs in Germany and Ireland, supported by EU subsidies. Yet the architecture decisions for these AI systems are being made in California and Oregon.
• The EU AI Act, GDPR and the Digital Services Act all push for stricter control over how data and AI models are handled. But when you consume AI “as a service” from Google, you inherit not just compliance guarantees but also vendor lock-in at the silicon level.

European cloud players like OVHcloud, Scaleway, Hetzner or Deutsche Telekom’s Open Telekom Cloud can’t yet match the custom IPU game. Their differentiation has to come from open standards, data residency guarantees and transparent governance rather than proprietary chips.

For policymakers, the question is becoming sharper: is it enough that AI runs on European soil, or must it also run on architectures that Europe can meaningfully influence?

Looking ahead: what to watch

Several things are worth tracking over the next 12–24 months.

1. Whether other clouds align with Intel or look elsewhere.
   If Microsoft or Oracle deepen their own CPU/IPU partnerships — with Intel, AMD, Nvidia or even start-ups — it will confirm that the balance-of-power has shifted from single vendors (Nvidia) to multi-chip ecosystems tuned for each cloud.

2. How far Google pushes IPUs into its product stack.
   Today, IPUs are mostly invisible infrastructure. Expect Google to start quietly marketing lower latency, better isolation and cheaper inference in services like Vertex AI or Gemini-based APIs – benefits that are enabled by these custom chips.

3. Standardisation versus lock-in.
   If Google and Intel keep interfaces proprietary, we move further into vertically integrated silos. If they choose to support open standards for things like confidential computing, networking offload or AI inference orchestration, European enterprises will have an easier time multi-clouding without massive re-architecture.

4. Regulatory scrutiny.
   As AI infrastructure becomes more central to economies, regulators may start looking at dependency chains: not just on U.S. software, but on specific chip vendors. The EU already thinks in terms of “systemic platforms” under the DMA; a similar lens applied to AI infrastructure would be a logical next step.

From a technical perspective, expect the line between CPU, IPU, GPU and NIC to blur. The long-term trajectory is towards composable data centres, where workloads are dynamically bound to pools of compute, memory and networking, all mediated by programmable infrastructure chips.

The bottom line

Google’s extended partnership with Intel is not about nostalgia for Xeons; it is a calculated move to own more of the AI cost and performance curve. In a world intoxicated with GPUs, the real strategic game is shifting to how you orchestrate everything around those accelerators. For Europe, the question is whether it is comfortable letting that orchestration logic — and the silicon beneath it — be designed elsewhere. As AI becomes basic infrastructure, how much control are we willing to outsource?