Google has announced its first confidential-computing-enabled TPU infrastructure, and it built that infrastructure's root of trust not on its own Arm-based Axion processor, but on AMD's Epyc server chips. For a company that designs its own AI hardware, relying on an outside supplier for the most sensitive security layer casts a small shadow over the 'independence through custom silicon' narrative.
What happened?
Google Cloud announced it is rolling out a confidential computing layer for AI workloads running on TPUs that hides data even from the cloud provider itself. This layer's root of trust was built on AMD Epyc rather than Google's own Axion processor. Axion was the Arm-based server chip Google developed for general-purpose cloud workloads, and it had become a symbol of the company's push to reduce its hardware dependence on outside vendors.
Why does it matter?
Confidential computing is an architecture that keeps the data a processor handles in memory encrypted at the hardware level, inaccessible to the operating system, the hypervisor, or even the cloud provider itself. The technology that makes this possible is the trusted execution environment (TEE) — known as SEV-SNP at AMD and TDX at Intel — and both companies have spent years on hardware-software co-design in this space. Arm's own TEE standard, Confidential Compute Architecture (CCA), is relatively new, and its production support on the server side hasn't yet matured to the level AMD and Intel offer.
When Axion was introduced in 2024, it was positioned around general compute performance and energy efficiency; TEE features built on a hardware root of trust weren't central to that announcement. Bringing a processor's TEE support to a production-ready, auditable level that security researchers actually trust requires a far longer maturation process than general performance optimization does. Google turning to AMD for this step rather than waiting on Axion can be read as a sign of that maturity gap.
The real crack here is in the narrative that 'major cloud companies are achieving independence with their own chips.' AWS Graviton, Google Axion, Microsoft Cobalt — each has moved general compute workloads onto its own silicon and achieved real cost advantages in doing so. But in security-critical layers, especially ones like TEEs that demand long-term trust, cloud companies still rely on AMD and Intel's engineering experience built up over many years.
Who does this affect, and how?
For regulated sectors like finance, healthcare, and government, this is a direct benefit. Organizations that want to train or run AI models on sensitive data can move faster to adopt TPUs once they can verify that their data is hidden even from the cloud provider.
| Chip | Architecture | Focus area | TEE maturity |
|---|---|---|---|
| AMD Epyc | x86 | General + security | High (SEV-SNP, long-established) |
| Google Axion | Arm | General-purpose cloud | Low/unclear |
| Intel (Xeon) | x86 | General + security | High (TDX) |
For Google, this is a pragmatic choice in the short term: promising customers security takes priority over being able to say 'every component is ours.'
What's next?
The thing to watch is whether TEE support gets added to future versions of Axion. If Google shifts to using Axion as its own confidential computing root within a few years, that would show this AMD Epyc choice was a temporary stopgap. If it doesn't, it will become even clearer that Arm-based custom silicon isn't going to challenge the x86 duopoly in security-critical cloud infrastructure anytime soon. Whether AWS and Microsoft make a similar choice for their own accelerators will also signal the broader trend in the industry.
Ultimately, this doesn't mean Google is abandoning Axion; but it does show that its custom-chip strategy isn't advancing at the same pace across every layer, and that security engineering follows a different timeline than general performance engineering. Details on the development were reported by TechRadar.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.