Google DeepMind Expands Private AI Compute with Secure Server-Side Memory

Proud to have contributed to this milestone! Bridging long-term AI context with verifiable, hardware-grade privacy is a crucial step toward building persistent AI responsibly. It has been an incredible journey working alongside such talented teams across Google DeepMind, Core, Platforms & Devices, and Cloud. Check out the blog post and technical brief to learn more about the architecture and third-party security audit! #PrivateAI #ConfidentialComputing #GoogleDeepMind #AI #Security

For years, private AI stayed stateless: safe, but incapable of handling truly agentic use cases. Today, that changes. Google’s Private AI Compute is launching secured server-side memory—the industry’s first verifiable, stateful, and privacy-preserving computing paradigm. This represents three years of incredible cross-functional work across GDM, P&D, Core, and Cloud. Sincere thanks to all the teams involved for your perseverance, rigor, and collaboration in making this milestone happen!

Continued investment in Private Computing technology is so critical to maximizing the utility of AI for consumers. And it's really difficult to deliver the cryptographic assurance that extends on device privacy to the cloud, while scaling to multiple billion+ user products. Thanks to all our teams driving this fantastic work.

Perplexity Introduces Hybrid AI Compute for Mac, Enhancing Privacy and Local Processing 🛰️ [AI AGENTS] Perplexity launches hybrid AI for Mac, boosting privacy. Why it matters: This development significantly enhances data privacy for AI users by processing sensitive information locally, reducing exposure to cloud environments. It also democratizes advanced AI capabilities by enabling robust local execution on compatible Mac hardware, potentially lowering operational costs and improving latency for specific tasks. 🤔 How will the increasing demand for local AI processing impact future hardware design and cloud service models? #HybridAI #PrivacyTech #OnDeviceAI #PerplexityAI #MacAI 📡 Follow DailyAIWire for high-signal AI news.

Google DeepMind The Cloud Memory Problem Just Got a Serious Answer — Privacy by Design, Not Privacy by Promise Google DeepMind (with Platforms & Devices, Core, and Cloud) published how Private AI Compute will add persistent, cross-device AI memory while keeping on-device-grade privacy: personal context sealed in encrypted cloud storage, unlock keys held on the user’s devices, temporary decrypt inside hardware enclaves, then re-encrypt. Until now the platform was “stateless” — wipe context when the task ends. Assistants that remember across phone, laptop, and glasses need memory. Memory without a verifiable vault is a compliance non-starter. They are also publishing a tamper-proof public record of server software so devices can verify authenticity before sending personal data, plus independent audit results. From a LatAm bank seat — dollarized market, cross-border USD, correspondent scrutiny — this is the kind of *stack* innovation that matters more than another benchmark chart: • Long-running agent context without “trust us, we don’t look” • Architecture you can ask a vendor to map against your data-residency and key-custody policy • A path where customer-facing AI assistants stop being either dumb (no memory) or radioactive (unbounded cloud memory) Innovation without key custody is just a prettier data leak waiting for an exam. If a vendor sells you “AI that remembers the customer,” can they show you where the keys live — and that even they cannot read the vault? #Privacy #ConfidentialAI #EnterpriseAI #Banking #Innovation Source: https://lnkd.in/efGiqWFK

What happens to the privacy boundary when an AI assistant needs to remember you tomorrow? Three recent designs caught my eye: Google outlines secure server-side memory across devices. Meta lays out persistent context for AI glasses, with encrypted storage and queries inside its Private Processing boundary. Amazon's Bee paper describes a background assistant that builds context over time in attested confidential VMs. All three are tackling continuity: retaining useful personal context between requests while limiting who can access the plaintext. Continuity is tied to TEEs, but it is not a TEE property. TEEs protect execution. The system around them has to decide what state survives, how it is encrypted, when a device releases keys, and which attested code may read it later. That is where confidential AI gets especially interesting to me. The unit of privacy can no longer be one prompt; it has to include the assistant's memory lifecycle. What would you want independently verifiable before trusting an assistant to remember you? I expanded this into a longer piece on key release, software changes, and what it means to delete an AI memory: https://lnkd.in/eRs4pT8w Google: https://lnkd.in/evpZ2EHX Meta: https://lnkd.in/eUeGxbCW Amazon Bee: https://lnkd.in/eBBVkqbG Views are my own.

For the last three years, the AI conversation has been dominated by one assumption: Powerful AI requires powerful data centers. ChatGPT. Claude. Gemini. DeepSeek. Behind all of them are massive cloud infrastructures, expensive GPUs, and huge server clusters. Every prompt goes somewhere. Every document is processed somewhere. Every AI-generated output comes back from somewhere. But that assumption is starting to change. AMD recently demonstrated a mini PC running a 235-billion-parameter AI model locally. No cloud GPU rental. No massive server rack. No data center. Just a compact machine with 128GB of unified memory doing the work. And that's where this becomes bigger than a hardware story. I see it as a governance shift. For years, our default approach to AI has been: ☁️ Send the prompt to someone else's infrastructure 📄 Upload the document for processing 🔑 Trust a third party with sensitive information 💰 Pay based on subscriptions, usage, or tokens 🌐 Depend on connectivity and cloud availability Why? Because local hardware simply wasn't capable enough. Memory was the wall. That wall is starting to come down. Local AI changes the conversation. It's not only about speed or performance. It's about control. Imagine running certain AI workloads entirely inside your organization: Your data stays on your infrastructure. Your sensitive documents don't need to leave the building. Your AI workload isn't entirely dependent on an external API. Your organization has greater control over where data is processed. For organizations operating under frameworks such as Saudi Arabia's PDPL, NCA ECC, GDPR, or other data-governance requirements, this can become a very different conversation. I'm not saying cloud AI is going away. It isn't. Frontier-scale models will continue to require enormous computing infrastructure, and cloud platforms will remain essential. But we may be entering a world where AI doesn't automatically mean sending your data somewhere else. And that's a much bigger change than simply getting a faster AI computer. The real question for organizations is no longer: "Can we use AI?" It's: "Which AI workloads actually need to leave our environment?" That is where the conversation around AI, infrastructure, data governance, cybersecurity, and privacy gets really interesting. Detail Video : AMD CEO Lisa Su just killed Nvidia’s $4,699 AI box with a $1,499 lunchbox. https://lnkd.in/d8hGJW7A #ArtificialIntelligence #LocalAI #GenerativeAI #AIInfrastructure #DataGovernance #Cybersecurity #Privacy #PDPL #SaudiArabia #NCA #AITransformation

Bringing sensitive enterprise data into #AI shouldn’t mean sacrificing control, #compliance or sovereignty. See how Thales, Microsoft and Intel® are combining hardware-enforced isolation, independent verification and customer-controlled keys to help close the trust gap in confidential AI. Read Gaurav Chander’s insights in our latest blog:

Bringing sensitive enterprise data into #AI shouldn’t mean sacrificing control, #compliance or sovereignty. See how Thales, Microsoft and Intel® are combining hardware-enforced isolation, independent verification and customer-controlled keys to help close the trust gap in confidential AI. Read Gaurav Chander’s insights in our latest blog:

Bringing sensitive enterprise data into #AI shouldn’t mean sacrificing control, #compliance or sovereignty. See how Thales, Microsoft and Intel® are combining hardware-enforced isolation, independent verification and customer-controlled keys to help close the trust gap in confidential AI. Read Gaurav Chander’s insights in our latest blog: