Apple Intelligence on Google Cloud: What PCC's 2026 Expansion Means for AI Meeting Notes

Published October 25, 2026

Key takeaways

Quick answer: In June 2026 Apple extended Private Cloud Compute (PCC) beyond its own data centers to run on Google Cloud using NVIDIA Blackwell GPUs. Apple says its stateless, no-logging privacy guarantees still apply, but for meeting notes containing attorney-client, MNPI, or PHI material, the only architecture with zero vendor-side exposure is on-device transcription that never leaves your Mac or iPhone.

October 25, 2026 · 11 min read

If your firm uses Apple Intelligence features to generate meeting summaries, the question “where is this request actually being processed?” got materially harder to answer in 2026. On June 8, 2026, Apple published a post on its Apple Security Research blog announcing that it was expanding Private Cloud Compute (PCC) beyond its own data centers to run on Google Cloud, using NVIDIA Blackwell GPUs with confidential-computing capabilities. For most consumer use cases that is a technical footnote. For meeting notes containing attorney-client, material non-public information (MNPI), SEC/FINRA-sensitive discussions, or protected health information, it reshapes the vendor-exposure diagram.

What Actually Changed in June 2026

Apple’s expansion was reported by MacRumors on June 8, 2026: Apple is “partnering with Google and NVIDIA to run Apple Intelligence workloads on Google Cloud.” Before this, PCC had been limited to Apple-silicon servers in Apple-owned data centers. Help Net Security framed it as the first time PCC has reached third-party data centers.

The scope is narrow but meaningful. According to coverage of Apple’s WWDC 2026 announcement, the arrangement uses NVIDIA Blackwell GPUs with confidential-computing capabilities to handle “the most demanding tasks, including agentic tool-use and complex reasoning.” Everyday things like rewriting a sentence or summarizing a notification continue to run on-device. But the architectural line Apple spent two years defending — “our chips, our racks, our buildings” — has moved.

What Apple Says Did Not Change

Apple’s position, documented on its security research blog, is that the PCC privacy properties travel with the workload. Requests are processed statelessly, meaning data is used only to generate a response and then deleted immediately afterward, with no logging and no persistent storage. The stateless design, confidential-computing hardware, and research-reviewable architecture are the guarantees — not the physical landlord.

Why This Matters for Meeting Notes Specifically

Transcripts are among the richest personal data an organization generates. A one-hour deposition prep, M&A diligence call, or patient consult contains more sensitive content than a quarter’s worth of Slack messages. For regulated users, the salient question has always been: who touches the audio or its derivatives, where, and under what legal regime?

Before June 2026, the honest answer for Apple Intelligence meeting summarization was “on-device, or in an Apple-controlled PCC node.” After June 2026, for the most demanding requests, the answer includes “a Google Cloud data center.” Apple’s stateless guarantees still apply, but the vendor diagram you have to document in a DPA, a FINRA WORM review, or a HIPAA risk assessment now has one more company’s name on it. For a lot of buyer-side risk reviewers, that is not a shrug.

Apple PCC on Google Cloud vs Google Workspace Gemini: Not the Same Thing

It’s easy to conflate “Apple PCC running on Google Cloud” with “Google’s own Gemini meeting notes.” They are fundamentally different lifecycles.

Google Workspace’s Gemini note-taker in Meet writes outputs into the organizer’s Drive. Per Google’s August 2026 Workspace feature drop, Meet now automatically organizes every meeting artifact — notes, transcripts, and recordings — into a parent folder called “Google Meet” in the host’s My Drive. That is designed for discoverability, not for ephemerality.

Retention is admin-controlled, not zero. A review by Fellow notes that Gemini’s “Take Notes for Me” generates transcripts that are stored indefinitely in Google Drive until manually deleted, with no built-in mechanism to automatically remove transcripts after a summary is created and no option to delete them immediately upon meeting completion. For regulated industries, that creates the opposite of what you want.

Comparison Table: Three Architectures, Three Different Risk Surfaces

Dimension Apple PCC on Google Cloud Google Workspace Gemini (Meet) On-Device (Basil AI model)
Processing location Device + PCC node (Apple or Google-owned DC) Google Cloud, within Workspace tenant Your Mac/iPhone only
Persistent storage of content Stateless (Apple states no logging/retention) Yes — notes/transcripts in Drive Local only; you control deletion
Third-party cloud operator in path Yes (Google Cloud, NVIDIA, Intel hardware) Yes (Google is the operator) No
Model-training use Apple states PCC data is not used for training Not used for training outside your domain without permission (Workspace) Nothing to train on — audio never leaves device
eDiscovery / subpoena surface Minimal (stateless), but PCC operator footprint exists High — Drive files, Vault holds Device only (standard legal hold on endpoint)
Zero-day retention option N/A (nothing retained by design) Not available You delete locally whenever you want

What Google’s Own Gemini Policies Say

If a buyer-side risk reviewer is looking at Google Workspace Gemini for meeting notes, the primary source is Google’s Generative AI in Google Workspace Privacy Hub. It specifies that Gemini in Workspace prompts and responses are retained from 90 days to indefinite as determined by admins, that inserted or generated content follows the Cloud Data Processing Addendum, and that admins can manage retention using Vault.

Separately, security vendor Jazz Security’s October 16, 2026 analysis summarizes it this way: Workspace Gemini does not train on your data; personal Gemini can, and human-reviewed chats stay up to three years. The important nuance for meeting notes: “does not train on your data” is not the same as “does not retain your data.” A compliant Workspace Gemini deployment still writes meeting content into Drive, and that content lives there until a retention rule or a human deletes it.

Consumer vs Workspace Gemini: Totally Different Risk Profile

Per Secure Privacy’s October 7, 2026 breakdown, consumer Gemini defaults to 18 months retention (configurable to 3 or 36 months), temporary chats or Keep Activity off still retains for 72 hours, and human-reviewed conversations are kept up to three years and are not deleted when the user deletes their activity. If any employee signs into personal Gemini on a work device to summarize a meeting, that is where the data is going.

The Thele v. Google Lawsuit: What Happens When Defaults Flip

The question of consent is not academic. According to a Bloomberg Law report dated July 8, 2026, plaintiffs Thomas Thele and Melo Porter alleged in Thele v. Google LLC, No. 5:25-cv-09704 (N.D. Cal.), that Google silently changed the settings for Gmail, Chat, and Meet in October 2025, turning on its AI tool by default and allowing it to collect users’ private data without their knowledge or consent.

A federal judge dismissed the amended complaint in July 2026 on standing grounds — the court found the plaintiffs had not described which specific data Google accessed or what concrete harm they suffered — but the plaintiffs refiled a second amended complaint in August, and Google filed a new motion to dismiss on September 29, 2026, per Open Class Actions’ case tracker. The case is still live. The privacy theory did not fail; standing did.

Whatever one thinks of the merits, the fact pattern — a default flipped, billions of accounts, AI processing of previously-opt-in communications — is exactly the pattern compliance teams are now writing new vendor-review controls to catch. On-device-only tools are not vulnerable to that failure mode because there is no vendor-side default to flip.

Regulated Users: Why This Shifts the Calculus

Under the GDPR Article 5 data-minimization principle, you process the least data for the least time against a defined purpose. PCC’s stateless design is well-aligned with that principle. Google Drive-resident Workspace meeting notes with admin-defined retention are not inherently misaligned either, but the surface you have to document, defend, and audit is larger. On-device-only transcription collapses the vendor-processor argument because there is no vendor processor to describe.

Under HHS’s HIPAA Privacy Rule, any cloud service that processes PHI is a business associate and needs a BAA. Apple does not sign BAAs for Apple Intelligence consumer features; Google Workspace does for covered services under specific plans. For an on-device recorder that never transmits audio, PHI never reaches a vendor, which simplifies the BAA question considerably — though compliance is still the covered entity’s determination.

And under California’s CCPA, the “sale or share” and “service provider” distinctions matter: a cloud-processed transcript creates a service-provider relationship that must be contractually constrained. On-device processing largely sidesteps the question.

The Developer Angle: PCC Is Not Open to Third Parties

One more wrinkle buyers should know: even if an app wanted to lean on Apple’s PCC privacy properties for meeting notes, it mostly can’t. Per Apple Developer documentation, server-side Apple Foundation Model access via PCC is restricted to specific developer programs with enhanced capabilities. The practical implication: most third-party apps that promise “Apple-level privacy for AI meeting notes” are either (a) running entirely on-device, (b) running on their own cloud, or (c) misrepresenting the architecture. There is no “turn on PCC as a backend” checkbox for third-party developers today.

This is important for buyer diligence. If a competing meeting-notes app markets “Private Cloud Compute privacy” without clearly stating that Apple’s own first-party features are the ones using PCC, read the fine print. Their summarization is almost certainly running on OpenAI, Anthropic, or Google infrastructure under a separate commercial agreement.

How Basil AI Solves This: Transcription That Never Leaves the Device

Basil AI’s architecture answers the PCC-on-Google-Cloud question by making it moot. Audio capture, real-time transcription via Apple’s on-device Speech Recognition framework, speaker diarization, and summarization all run locally on your Mac or iPhone, using the Apple Neural Engine. There is no PCC round-trip. There is no Google Cloud hop. There is no vendor-side copy of the audio that could be subpoenaed, breached, or quietly repurposed via a flipped default.

Concretely, for a one-hour meeting:

For teams who need to document the architecture in a vendor risk review, “processes 100% on-device, no third-party cloud operator in path” is a very short diagram. For comparisons with cloud-first bot-free tools, see our writeup on Granola alternatives where ‘bot-free’ is not the same as ‘cloud-free’, and for a regulated-industry framing see our piece on AI meeting notes for compliance officers in financial services.

What To Ask Before You Pick an AI Meeting Notes Tool in 2026

Whatever vendor you choose — Apple’s first-party features, Workspace Gemini, a cloud notetaker, or an on-device app — the questions worth asking in 2026 are:

  1. Where is the audio processed? On-device, in the vendor’s cloud, or in a third-party cloud operating on behalf of the vendor?
  2. Is the transcript persisted? If yes, where, under whose retention policy, and can I force zero-day deletion?
  3. Is any content used to train a model, now or in the future, under any plan tier or default?
  4. What is the eDiscovery / subpoena surface? Who holds a copy that could be compelled?
  5. What’s the default-flip risk? If the vendor flips an opt-in to opt-out, how do I detect it, and does it change my compliance posture retroactively?

For compliance-sensitive conversations, “on-device, no cloud, no copy” is the shortest acceptable answer. For a deeper walkthrough of what to check, see What ‘Compliant AI Meeting Notes’ Actually Means.

The Bottom Line on PCC Going to Google Cloud

Apple’s expansion of Private Cloud Compute to Google Cloud is a genuine engineering achievement — stateless, confidential-computing inference running on a third-party operator while preserving the stated privacy guarantees. For 99% of consumer Apple Intelligence usage, it’s a shrug. For meeting notes in regulated and privilege-sensitive contexts, it’s a reminder that even the most carefully engineered cloud is still a cloud, and the only way to remove a vendor from your diagram is to never hand them the audio in the first place.

That is the architecture Basil was built around, and it’s why on-device transcription keeps winning the compliance conversation even as the cloud’s privacy ceiling rises.

Get Weekly Privacy Insights

On-device AI tips, privacy news, and Basil AI updates. No spam.

Unsubscribe anytime. Privacy Policy

Try Basil AI — on-device transcription that never touches the cloud

Record, transcribe, and summarize meetings 100% on your Mac or iPhone. No PCC. No Google Cloud. No vendor-held copy.

Download on the App Store Download on the Mac App Store

Frequently Asked Questions

Does Apple Intelligence send my meeting audio to Google Cloud?

Only for the most demanding Apple Intelligence requests — things like agentic tool-use and complex reasoning — does Apple now route workloads to Google Cloud PCC nodes running on NVIDIA Blackwell GPUs. Simple on-device summarization and the Apple Speech Recognition transcription that Basil uses do not leave your device. But if you rely on cloud-dependent features for meeting notes, your audio-derived content can hit a Google-owned data center.

Is Private Cloud Compute the same as Google Workspace Gemini for meeting notes?

No. Private Cloud Compute is Apple's stateless, no-logging inference architecture — Apple says data serves the request and is then deleted. Google Workspace Gemini stores Meet notes and transcripts in the organizer's Google Drive under admin-set retention policies, with no built-in zero-day retention. These are fundamentally different data lifecycles.

Does Google train Gemini on Workspace meeting notes?

Google states that Workspace customer data is not human-reviewed or used to train generative AI models outside your domain without permission, per its Generative AI in Google Workspace Privacy Hub updated October 6, 2026. Consumer Gemini is different: with Keep Activity on, chats can train models and human-reviewed conversations are retained up to three years.

Are AI meeting notes safe for compliance-sensitive conversations?

Cloud-processed notes — even on PCC — create a vendor-side processing surface that regulated users must document in DPAs and vendor risk reviews. For attorney-client, MNPI, SEC/FINRA recordkeeping, or PHI contexts, on-device-only transcription eliminates the third-party processor entirely, which is a cleaner architecture to defend. Compliance is still the customer's determination, not a vendor claim.

What is the Thele v. Google lawsuit about?

Thele v. Google LLC (No. 5:25-cv-09704, N.D. Cal.) is a proposed class action alleging Google turned on Gemini Smart Features by default across Gmail, Chat, and Meet around October 10, 2025, letting the AI access private communications without consent. A federal judge dismissed the first amended complaint July 7, 2026 on standing grounds; plaintiffs refiled, and a second motion to dismiss is pending.

How does Basil AI handle meeting transcription differently?

Basil runs transcription on your Mac or iPhone using Apple's on-device Speech Recognition framework and the Neural Engine. There is no PCC call, no Google Cloud hop, and no vendor-held copy of your audio. The architecture is simple: audio in, transcript on your device, notes you own. That eliminates both the Apple-cloud and Google-cloud processing surfaces from the picture entirely.