Arm confirmed that Google Axion, a custom Neoverse-based CPU, hosts the new TPU 8i Max, one day after AMD confirmed 5th Gen EPYC as the host for TPU 8tx. Futurum’s AI Supply Chain Market Model projects 3.2 million 8i units in 2027 against 3.1 million 8t, which puts the Arm-hosted inference line at volume parity with training silicon.
What Is Covered in This Article:
- Axion as host CPU for TPU 8i Max, 8i, and 8t
- 5th Gen AMD EPYC hosts for confidential TPU 8tx
- Futurum 8i and 8t unit and power projections through 2028
- Axion N4A, GKE Agent Sandbox, and Agent Substrate for agent runtimes
- Custom Arm share of server CPU revenue in Futurum’s model
The News: Arm said on November 10 that Google Cloud pairs its TPU 8i Max system with Google Axion processors, custom CPUs built on Arm Neoverse. Axion has served as host compute across the eighth TPU generation since the April launch, and Google’s technical documentation lists Arm Axion as the CPU host for both 8t and 8i. TPU 8i Max packages two 8i dies for 384 GB of HBM, 17.2 TB/s of HBM bandwidth, and 768 MB of on-chip SRAM. “Delivering AI performance at scale requires co-designing CPUs and accelerators to work seamlessly together,” said Mark Lohmeyer, VP and GM of AI and Computing Infrastructure at Google Cloud.
Axion also runs the agent layer on top of the accelerators. Google says Axion N4A instances deliver up to 2x better price-performance than comparable current-generation x86 virtual machines, that GKE Agent Sandbox on N4A gains 30% in price-performance, and that the open-source Agent Substrate runtime packs sandboxes at 10x the density of standard container runtimes. All three are Google figures. One day earlier, AMD said TPU 8tx, the confidential variant of 8t, runs on 5th Gen EPYC hosts using AMD Secure Encrypted Virtualization.
Google Puts Arm-based Axion Under TPU 8i Max as Inference TPU Volume Exceeds Training
Analyst Take: Google has assigned its host CPUs by workload, with the faster-growing workload favoring Arm. The inference flagship runs on Axion. So do the sandboxes, the orchestration runtime, and the general-purpose instances agents call between model turns. x86 entered the eighth generation through one system, a training-class part sold to regulated buyers on hardware-enforced confidentiality. Futurum views the pairing as the clearest hyperscaler signal yet that agentic inference will run on custom Arm hosts, with x86 kept where its security stack decides the purchase. The line is softer than a clean training and inference split, since standard 8t also ships with Axion. The only x86 socket in the family arrived with a compliance requirement attached.
Inference Silicon Reaches Volume Parity With Training in Futurum’s Model
Futurum’s AI Supply Chain Market Model puts 8i ahead of 8t from the first year of the ramp. The model projects more 8i units in 2027 than 8t, narrowly, with model training remaining a major focus for internal Google workloads and leading customers. Power tells a sharper story. Each 8i draws 2.0 kW of critical IT power against 1.36 kW for 8t, so 2027 8i deployments will surpass training silicon by 2.2 GW. Anthropic accounts for 61% of the 2027 8i units, with Google’s own fleet taking the remaining.
Every one of those 8i and 8i Max sockets sits beside an Axion host. Arm’s exposure to the Google TPU program now scales with inference demand, and inference demand in this generation scales with one external lab. The model shows 8t retaking the unit lead in 2028, as 8i shipments to both Google and Anthropic fall from their 2027 peak. Axion’s host footprint inside the TPU fleet depends on whether a ninth generation keeps the same pairing and whether Anthropic’s commitment extends past the current agreement.
Agentic Inference Pulls the Host CPU Into the Serving Path
The host CPU in an agent deployment does more than feed the accelerator. It prepares data, executes tool calls, runs sandboxed code, and now stores spilled KV state, since TPU Sync moves an idle agent’s working memory from TPU HBM into host DRAM. Futurum’s model raises AI host CPUs per accelerator from 0.36 in 2026 to 0.40 in 2028, while standalone AI CPUs, sockets running agent work with no accelerator attached, climb from 0.42 per accelerator to 0.78 over the same span. Server CPU revenue roughly doubles from $60.4 billion in 2026 to $116.9 billion in 2027 in the model.
Custom Arm silicon takes the largest share of that growth. Futurum projects Arm-based custom CPUs rising from 38.9% of server CPU revenue in 2026 to 49.2% in 2027. Google is one of several hyperscalers on this path. NVIDIA hosts Vera Rubin on Vera, its 88-core custom Arm CPU, while AWS builds on Graviton, Microsoft on Cobalt, and Arm now sells its own AGI CPU as merchant silicon. A custom host lets the cloud set core count, memory capacity, and coherency with its own accelerator, which is the co-design argument Lohmeyer makes. x86 vendors answer with core density. AMD’s EPYC Venice and Intel’s Xeon 6+ both push core counts higher for the same agent fan-out.
x86 Keeps the Workloads Where Its Trust Stack Is Proven
TPU 8tx explains why x86 stays in the family. AMD SEV has years of production deployment in public clouds, and 8tx uses it to keep model weights, prompts, and intermediate data encrypted in CPU memory and in transit. Regulated buyers audit what they already know. Arm’s Confidential Compute Architecture has yet to reach comparable deployment depth in hyperscale fleets, in Futurum’s assessment, and Google chose the proven stack for the one system whose value proposition is attestation. AMD says EPYC hosts accelerators at more than 10 hyperscalers and frontier labs, a count the company supplied.
The Arm case still rests on vendor numbers. The 2x N4A price-performance figure compares against unnamed x86 instances, the 30% and 10x sandbox claims have no third-party test, and Google has published no comparison of Axion against an x86 host inside a TPU system. Axion is also built on Neoverse V2, a generation behind the cores Arm now licenses. The implication is that Google has made an architectural commitment ahead of the independent evidence, and buyers renting 8i Max capacity will see the host choice only through the tokens per dollar the full system delivers.
What to Watch:
- Whether Google publishes Axion versus x86 host benchmarks inside a TPU system
- Whether a next-generation Axion on newer Neoverse cores ships with 8i Max
- Whether Google offers a confidential 8i variant and which host CPU it uses
- Whether AWS and Microsoft pair their next accelerators with Graviton and Cobalt hosts
Read more details on the company website.
Sources
- Accelerating the agentic era: How Arm and Google Cloud co-design infrastructure for TPU 8i Max, ARM, November 2026
- The $246 Billion CPU Opportunity’s Winning Chip Has Not Been Designed Yet
- State of the Market Report: Semiconductors, Supply Chain, and Emerging Technology, Q3 2026
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights From Futurum:
Arm Q2 FY 2027: AGI CPU Orders Expand Data Center Opportunity
Arm AGI CPU Posts 2x the Agentic Work Per Rack of AMD’s EPYC 9755 at OCP 2026
Lattice Brings FPGAs and AMI Firmware to Arm’s $15 Billion AGI CPU Server Push
Author Information
Brendan is Research Director, Semiconductors, Supply Chain, and Emerging Tech. He advises clients on strategic initiatives and leads the Futurum Semiconductors Practice. He is an experienced tech industry analyst who has guided tech leaders in identifying market opportunities spanning edge processors, generative AI applications, and hyperscale data centers.
Before joining Futurum, Brendan consulted with global AI leaders and served as a Senior Analyst in Emerging Technology Research at PitchBook. At PitchBook, he developed market intelligence tools for AI, highlighted by one of the industry’s most comprehensive AI semiconductor market landscapes encompassing both public and private companies. He has advised Fortune 100 tech giants, growth-stage innovators, global investors, and leading market research firms. Before PitchBook, he led research teams in tech investment banking and market research.
Brendan is based in Seattle, Washington. He has a Bachelor of Arts Degree from Amherst College.
