At NVIDIA GTC Taipei, during Computex week, NVIDIA unveiled RTX Spark: a new Arm-based superchip that combines a Blackwell RTX GPU and Grace CPU in a single package.
Laptops and compact desktops powered by RTX Spark are expected later in 2026 from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and GIGABYTE models to follow.
The hardware numbers are significant. NVIDIA says RTX Spark delivers up to 1 petaflop of AI performance and up to 128 GB of unified memory. It is designed to run local AI agents, large language models, creative workloads, and developer workflows directly on Windows PCs.
NVIDIA also says RTX Spark can run 120B-parameter large language models with up to 1 million tokens of context locally.
That raises an important question for businesses: if powerful AI can now run locally, does that make it safe for sensitive data?
Important warning: local AI does not automatically mean safe AI
Local AI is a major step forward for privacy concerned organisations. But can local AI leak private information? Is local AI compliant with privacy laws? Is on-device AI safe for highly confidential documents? The short answer is: it can reduce cloud exposure, but local execution is not the same thing as governance.
The RTX Spark announcement describes new Windows security primitives for identity, containment, policy, and end-to-end security. NVIDIA also describes OpenShell as a runtime layer that defines what agents can and cannot do, routes queries to local models based on privacy policies, and disguises personal information in queries sent to cloud models.
In short: Microsoft handles the OS-level security sandbox, NVIDIA handles the AI routing and masking layer. Together they form a perimeter around what the agent can see, do, and send. The gap is that neither sits inside your document workflow. No interception, no redaction, no governance before sensitive content reaches the agent. That is a different problem entirely — and that is exactly what InBay is built for.
But most organisations we are working with have moved past the simple question of: Did someone upload the file to ChatGPT? With local AI agents, the attack surface now spans agent permissions, data access scope, cloud fallback paths, telemetry retention, prompt injection, and redaction timing — and most of it is invisible until something goes wrong.
A model running locally can still create risk if it can access the wrong folder, summarise the wrong document, store sensitive prompts in logs, retrieve restricted content for the wrong user, or generate output that gets shared externally without review.
Local AI should therefore be treated as a controlled workflow, not just a hardware upgrade. A powerful local model reduces cloud exposure, but it still needs boundaries: access control, audit logs, redaction rules, export controls, model routing, retention settings, user permissions, and review steps.
A stronger position goes further than: Run AI locally. It is: Run AI locally with clear data boundaries, practical governance, and workflow-level controls.
This is what Berrysbay Labs already does great: helping organisations understand data governance, sensitive data flows, where AI exposure happens, and where local-first tools, redaction, policy, or training can reduce that risk.
The practical upside: local-first products become more capable
The exciting part is that local-first products like InBay and KnowledgeBay become significantly more capable as hardware catches up.
The core trade-off until now has been:
Cloud models are powerful, but they create real risk for sensitive data.
While local models are private, but they have often been smaller, slower, or harder to run at scale. RTX Spark changes that equation.
With up to 128 GB of unified memory and 1 petaflop of on-device AI performance, the gap between local and cloud AI is closing fast. For Berrysbay Labs, this matters because our products from the day one have been designed to benefit from this shift.
InBay is our local-first sensitive data discovery and redaction tool. InBay combines redaction pipelines, compliance-grade audit records, on-device media editing, and governance frameworks — enhanced with local AI review and document analysis that just got even more capable.
KnowledgeBay is Berrysbay Labs air-gapped internal knowledge assistant, becomes a more realistic option for organisations that want fast document search, policy Q&A, and internal knowledge retrieval without any external data transfer.
The Buzz in the Tech Community
Across the broader tech community and developer circles, the upcoming RTX Spark architecture is already being hailed as a definitive "MacBook Pro Killer" for local AI agents. Apple’s unified memory monopoly for on-device AI is officially being challenged on Windows turf. Industry analysts are highlighting the massive structural shift of running data-center-grade models entirely offline.
The Data Sovereignty Reality Check
This hardware evolution arrives at a critical moment for global compliance. Sending data to foreign cloud servers isn't just a security risk—it’s an international regulatory minefield. Cross-border data transfers increasingly clash with strict data sovereignty laws. Here in Australia, recent reforms to the Privacy Act have introduced massive financial penalties for data mishandling, while the Australian Signals Directorate (ASD) warns that keeping data onshore with foreign-owned cloud providers still exposes enterprises to jurisdictional risks and foreign surveillance laws like the US CLOUD Act. By shifting heavy AI workloads directly to the local device, you effectively eliminate the foreign cloud trap, ensuring your data never crosses international borders or leaves your physical custody.
The Bottom Line: Hardware is Only Half the Battle
As local AI moves from specialist infrastructure into everyday business hardware, the opportunity for organisations handling sensitive data is real. But the hardware is only one part of the answer. The workflow around it still needs to be designed, governed, and controlled. PII and enterprise secrets still needs to be redacted.
That is the work Berrysbay Labs does through practical reviews, assessments, workshops,custom local tools and platforms like InBay—ensuring that as your hardware gets smarter, your data boundaries remain entirely within your control.
Ready to move your AI workloads on-device safely? If your team is exploring AI but cannot afford accidental data exposure, get in touch with Berrysbay Labs today.
