Google wants more Intel inside ... its datacenters, taps Chipzilla for more SmartNICs
Custom ASIC biz now running at a $1B annual pace for Intel
Google will continue to work with Intel, buying SmartNICs for its public cloud rather than blazing its own trail as AWS has done with its Nitro NICs.
Like most hyperscalers today, Google employs SmartNICs, or as Intel prefers to call them, infrastructure processing units (IPUs). These devices are essentially a computer on a network card and are designed to offload networking, security, and storage operations, freeing up CPU resources for tenant workloads.
While Amazon employs custom ASICs from its Annapurna Labs team and Microsoft uses custom logic running on FPGAs, Google tapped Intel to develop an ASIC-based IPU called Mount Evans, which launched alongside its C3 instances in 2022.
On Thursday, Intel announced Google had expanded this collaboration to develop new IPUs in a press release that reads like a desperate attempt to convince the public that its Datacenter and Networking divisions are still relevant.
Intel CFO David Zinsner had alluded to increased demand for these services during the company's Q4 earnings call in January, touting that the company's custom ASIC biz grew more than 50 percent in 2025 and exited Q4 at an annualized revenue run rate above $1 billion.
Intel didn't elaborate on what Google's next-gen IPUs might look like, but given the demand for high speed networking for AI compute clusters, there's a good chance it'll be significantly faster than its 200 Gbps Mount Evans IPUs.
Alongside the expanded IPU collab, Intel was also keen to note that the Chocolate Factory wasn't giving up on its Xeon processors, which will power a variety of general purpose and AI workloads. In other words, it is business as usual.
Like many hyperscalers, Google now has its own Arm-based CPU, codenamed Axion, which runs both internal and customer facing workloads.
However, just as Graviton and Cobalt have replaced Xeon or Epyc processors in Amazon's or Microsoft's clouds, neither Intel nor AMD is at risk of being ejected by Axion any time soon.
In fact, since many customers prefer to run their workloads on x86 cores, either for performance or compatibility reasons, Intel remains useful for putting pricing pressure on AMD and vice versa.
In any case, Intel felt it was necessary to reassure everyone that Xeon remained a key part of Google Cloud, particularly when it comes to AI. But again, this is nothing new.
Xeons have been the CPU of choice for Nvidia's 8-GPU DGX reference designs going back to the H100 in 2022. In other words, many of Google's GPU instances already had Intel inside.
And, while AI workloads also need CPUs to orchestrate agents and execute the code generated by Google's GPUs and TPUs, there's nothing that makes Xeon inherently better for this job than Epyc or Axion. In fact, at cloud scale, the best CPU for agentic AI is probably whatever happens to be sitting idle at any given moment. ®