Apple just did something they've never done before by running Google Cloud's Private Cloud Compute (PCC) workloads outside of their own data centers. Not AWS, not Azure, but Google. And what's even crazier is that they're using NVIDIA Blackwell GPUs, Intel TDX (Trust Domain Extensions), and Google's Titan chip to pull it off. If you're building anything that touches sensitive data at scale, this architecture decision is worth reading and understanding, not just bookmarking for later.
What Private Cloud Compute (PCC) Is
PCC is Apple's framework for running AI inference on cloud infrastructure while keeping user data private, even from Apple itself. The core promise is that your data gets processed in a cryptographically verifiable environment where no one, including Apple, can access it in plaintext. It's a hard problem, and most companies just slap a "zero trust" label on things and call it a day.
But, Apple's approach is more interesting. They maintain an independent, append-only hardware ledger and use dual-vendor attestation roots, meaning two separate parties have to sign off on the integrity of the hardware environment before any workload runs. Neither vendor can unilaterally vouch for the system (This is the part that stood out to me most).
I've worked on systems that handle large volumes of sensitive insurance data, and attestation is one of those things that sounds simple until you're actually implementing it. Getting one party to sign off on environment integrity is already nontrivial, but getting two independent vendors to agree on attestation roots, in a way that's auditable and consistent, is a hard coordination problem.
Why the Hardware Stack Matters
The choice of Intel TDX alongside NVIDIA Blackwell GPUs is deliberate. TDX creates hardware-isolated execution environments (Trust Domains) that protect workloads from the hypervisor itself. That means even if Google's underlying cloud infrastructure were somehow compromised, the workloads inside TDX enclaves stay protected. Blackwell, meanwhile, is NVIDIA's current flagship data center GPU, the same architecture I run locally on my DGX Spark for inference workloads.
The interesting tension here is that GPU-based confidential computing is still maturing. Most confidential computing frameworks were designed for CPU workloads, and extending those guarantees to GPUs, where data has to move between CPU memory and GPU VRAM, introduces complexity. NVIDIA has been working on Confidential Computing support for H100 and Blackwell GPUs, but getting that to work inside Intel TDX enclaves, at Apple's scale, is not something you just set up in an afternoon.
Let's not forget that Google's Titan chip also plays a role here. Titan handles secure boot and hardware attestation at the physical server level. Combined with TDX, you get a layered attestation chain where Titan vouches for the hardware, TDX vouches for the execution environment, and Apple's ledger keeps an auditable record of it all. Three independent verification layers before a workload even runs.
What This Tells Us About Enterprise AI Infrastructure
The fact that AWS and Azure are explicitly not part of this collaboration is something to pay attention to. It doesn't mean they're less capable, but it likely means Google made specific commitments or had specific technical capabilities that matched Apple's requirements better. My guess? Google's Confidential Computing portfolio, specifically their work on confidential Google Kubernetes Engine (GKE) Nodes and confidential virtual machines (VMs), gave them a head start in this conversation.
For those of us building enterprise applications, especially in regulated industries like insurance, healthcare, or finance, there's a lesson to learn here. Privacy-preserving compute isn't just a compliance item to check off before deploying to production anymore. It's becoming a design requirement that affects which cloud vendors you can work with at all. I've seen this shift firsthand where early conversations about cloud architecture were mostly about cost and latency. Now the topics of data residency, attestation, and encryption-in-use come up in the very first meeting.
The append-only ledger Apple is maintaining is also worth borrowing conceptually to make sure your systems are auditable. In a few of the data pipelines I've built, we've used append-only audit logs for similar reasons. Being able to prove, after the fact, exactly what happened and in what order is VERY important, and isn't something that is suggested by a paranoid cybersecurity consultant. When you're processing any amount of sensitive data at scale, that kind of auditability isn't optional.
This Looks Like a Template For The Future
I don't think this is a one-time experiment for Apple. I think it's a proof-of-concept for how they'll scale PCC as AI features slowly expand across their product line. They needed burst capacity beyond their own data centers, they needed it to meet their privacy guarantees, and they figured out how to do it without compromising the model in a reusable pattern.
If you're architecting AI workloads that touch sensitive data, here's what to take from this: don't treat confidential computing as an afterthought. Look at Intel TDX and AMD Secure Encrypted Virtualization-Secure Nested Paging (SEV-SNP) now, even if you're not ready to use them. Understand what attestation means in your cloud provider's stack, because most providers have documentation on it and only a few engineers read it. And if you're in a regulated industry, start asking your cloud vendor what their roadmap looks like for GPU-based confidential compute, because that's where this is heading.
Apple picked Google because Google had the right answers to those questions up front. When you're evaluating infrastructure for your next AI workload, make sure you're asking them too.