The Intelligence Stays Here
Why the New Wave of AI Hardware Changes Everything for Businesses That Want to Keep Their Data Private
By Mark Presnell, Convergence Ltd
Something significant happened this week in the technology world — and it is directly relevant to every growing business that stores sensitive data and is wondering whether AI can genuinely work for them.
Jensen Huang, founder and CEO of NVIDIA, and Satya Nadella, chairman and CEO of Microsoft, announced a landmark partnership that is being described as a reinvention of the personal computer. The announcement, made at NVIDIA’s GTC Taipei event on 31 May 2026, centres on RTX Spark — a new superchip that delivers one petaflop of AI compute, up to 128GB of unified memory, and a full NVIDIA CUDA and RTX software stack, all within a slim laptop or compact desktop PC.
This is not a marketing exercise. It is a hardware and software shift that has direct consequences for the way businesses can deploy private AI — and it arrives at exactly the moment when the case for keeping your AI on your own dedicated infrastructure has never been stronger.
What Jensen and Satya Actually Announced
“The PC is being reinvented,” Jensen Huang said at the announcement. “For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask — and the PC does the work.”
Satya Nadella’s framing was equally clear: “Our goal is to deliver unmetered intelligence to every home and every desk with Windows.”
The partnership goes well beyond a new chip. NVIDIA and Microsoft are co-building a native Windows platform for AI agents — including new security primitives and the NVIDIA OpenShell runtime, which gives users the ability to:
- Define what agents can and cannot access
- Intelligently route queries to local models based on privacy policies
- Disguise personal and commercial information in any queries that do go to cloud models
That last capability is telling. Microsoft and NVIDIA are not building this platform for hobbyists. They are building it for businesses and individuals who have sensitive data they cannot afford to have processed by third-party cloud AI systems — and who now, with RTX Spark, have the hardware to do something about it.
Compact RTX Spark laptops and desktops will be available from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI this coming northern hemisphere autumn.
The Case for Private AI — and Why It Has Become Urgent
Let me be direct about what private AI actually means in a business context, because the term gets used loosely.
Private AI means deploying a large language model — the kind of technology that powers ChatGPT, Claude, or Gemini — on your own dedicated infrastructure, querying only your own data, with no data leaving your building or your private network unless you have specifically and deliberately approved it. No public AI models in play. No third-party processing of your customer data, your financial records, your supplier terms, your pricing strategy, or your staff information.
The alternative — sending business queries to a public cloud AI — creates risks that many business owners have not yet fully reckoned with.
When you ask a cloud AI tool about your business, you are potentially feeding commercially sensitive information into a system operated by a third party. You have limited visibility over how that data is stored, processed, or used. You cannot guarantee it will not be used to train future models. And under New Zealand’s Privacy Act 2020 — reinforced by Information Privacy Principle 3A, which came into force in May 2026 — your obligations around how data is shared with third-party systems have never been more defined or more enforceable.
Private AI resolves all of this. Your AI. Your data. Your dedicated infrastructure. Full stop.
The ROI Is No Longer Theoretical
For a long time, the private AI argument was a privacy argument first and an economics argument second. Public cloud AI was cheap and easy. Private AI required hardware investment. The business case needed work.
That calculation has now fundamentally shifted.
The numbers are striking:
- On-premise AI inference is up to 18 times cheaper per million tokens than premium cloud AI APIs. (Lenovo/NVIDIA TCO analysis, 2026 edition)
- Break-even on on-premise AI hardware is now achievable in under four months for businesses with sustained, predictable workloads — down from 12–18 months just two years ago. (Lenovo, Luminix AI)
- Organisations deploying Dell AI Factory with NVIDIA infrastructure report 1,225% four-year ROI — $25.9 million in savings against a $1.96 million investment. (DreamFactory, citing Dell data)
- 40% of organisations report cost savings from enterprise AI adoption, and 66% report productivity and efficiency gains. (Deloitte State of AI in the Enterprise, 2026)
- Enterprise AI inference on-premises has grown from 12% to 55% in three years — a 4.6x increase driven by data sovereignty, regulatory compliance, and performance requirements.
The productivity numbers are equally compelling:
- Knowledge workers using AI report saving 40 to 60 minutes per working day on average, with data-intensive roles saving 60 to 80 minutes.
- Integrating AI into core workflows delivers 27–74% time savings without reducing output quality.
For a business with 10 staff, 40 minutes saved per person per day across a year equals roughly 2,400 hours of recovered time. At an average fully loaded employment cost of NZD $40–$50 per hour, that is $96,000–$120,000 in productive capacity recovered annually — from a single AI productivity layer applied to their own data.
Now consider that the same private AI layer that generates those productivity gains is also eliminating per-query cloud AI costs entirely. The economics are becoming undeniable.
Why the NVIDIA–Microsoft Announcement Makes This More Accessible
Here is the significance of the RTX Spark announcement for businesses of the size I typically work with.
Until now, running a capable private AI model on local infrastructure meant procuring a serious GPU server — hardware that is powerful, but not always practical or affordable for an SME. The RTX Spark changes that boundary substantially.
One petaflop of AI performance and 128GB of unified memory in a compact desktop or laptop is enough to run highly capable open-source models — Llama, Mistral, DeepSeek — with genuine business-grade performance. Ollama, the runtime we use in our CODI for AI deployment stack, runs natively on this class of hardware. The open-source models available today match GPT-4 on most business tasks to within 10% after fine-tuning on your own data.
What NVIDIA and Microsoft have done is compress the hardware access point significantly. The same AI capability that previously required a data centre rack can now run on a device on your desk or in your office.
More importantly, the NVIDIA OpenShell platform — built natively into the new Windows architecture — means that the privacy controls are baked into the operating system, not bolted on afterwards. You define what the AI agent can access. You define what stays local. You define what, if anything, is permitted to route to the cloud. This is private AI as a first-class Windows citizen.
The Missing Piece That Most Businesses Overlook
Here is where I need to add something important — because the hardware and software story, compelling as it is, does not work on its own.
Private AI is only as good as the data it can access. A large language model deployed on your own dedicated infrastructure, pointed at disconnected, siloed, poorly structured business data, will not deliver on its promise. You will end up with an expensive box that cannot answer the questions your business actually needs answered.
The questions that make private AI genuinely valuable are questions like:
- *”Which of our customers have reduced their order frequency in the last 90 days, and what is their combined annual value?”*
- *”What is our current stock position on our top ten product lines, and do we have enough to support a promotional campaign this month?”*
- *”Compare our actual margins this quarter against our targets — which categories are underperforming and why?”*
- *”Which suppliers are we most dependent on, and what is our current outstanding exposure to each of them?”*
These questions require your ERP, your CRM, your eCommerce website, your inventory management system, and your accounting software to be connected — bidirectionally, in real time, with clean and consistent data flowing between them. Without that integration layer, your private AI is answering questions about a fragment of your business, not the whole picture.
This is exactly what CODI — our Convergence Optimised Data Integration platform — provides. CODI has been connecting eCommerce websites with backend ERP, CRM, accounting, and inventory management systems for growing businesses since 2009. The hub-and-spoke architecture means your data flows through a single validated, translated integration layer. When CODI for AI is added on top of that, your private AI model has clean, connected access to your full business data — and the questions above become natural language queries that anyone in your business can run, in seconds, on your own dedicated infrastructure.
No public AI models in play. No data leaving your network without your explicit approval. Your private AI, querying your connected data, on your dedicated infrastructure.
What “Approved to Go Outside” Actually Means
I want to address one nuance directly, because it matters.
Private AI does not have to mean completely air-gapped from the world. There are legitimate use cases where you might want your private AI to query approved external sources — current freight rates, currency exchange data, public market pricing, regulatory updates.
The point is that this should always be a deliberate, controlled, auditable decision — not the default.
The NVIDIA OpenShell architecture is built exactly around this principle: local-first, privacy-first, with explicit and user-defined permissions for anything that routes outward. What CODI for AI does at the integration and data layer, NVIDIA and Microsoft are now doing at the hardware and operating system layer. They are converging on the same philosophy: your AI, your data, your rules.
Businesses that build on this architecture will have a private AI that is both powerful and trustworthy — capable of answering complex questions about their own business in seconds, while never accidentally or unknowingly sharing competitive intelligence with external AI systems.
The Integration Imperative
I have been in systems integration for over 25 years, working with growing businesses across eCommerce, manufacturing, distribution, and professional services. In that time, I have watched three major technology waves arrive — and in each one, the businesses that had their systems properly integrated captured the advantage far faster than those that did not.
The Private AI wave is no different.
The hardware is becoming accessible. The open-source models are now genuinely capable. The Microsoft and NVIDIA partnership has just announced that private AI infrastructure is moving to the mainstream Windows PC. The economics have crossed the threshold where the business case is compelling for most organisations, not just the large enterprise.
What determines whether a given business captures this opportunity or watches it from the sidelines is the same thing it has always been: whether their systems and applications are properly connected, with clean data flowing in real time between the tools they use to run their business.
If your ERP, your eCommerce website, your CRM, your inventory, and your accounting software are integrated through a reliable, real-time middleware platform like CODI, you are already most of the way to having a private AI that can genuinely answer questions about your business. The AI layer is CODI for AI — the natural language intelligence layer that sits on top of your integrated data, running on your own dedicated infrastructure.
If they are not yet integrated — or integrated in ways that rely on manual exports, CSV imports, or fragile point-to-point connections — then the first investment is the integration foundation. Build that properly, and the private AI layer follows naturally.
A Practical Starting Point
If you are considering what a private AI layer might look like for your business, here is where I would recommend beginning:
1. Audit your integration architecture. Which of your systems are connected in real time? Where are you still exporting spreadsheets, manually re-entering data, or relying on overnight batch processes? Every gap in your data flow is a question your private AI will not be able to answer accurately.
2. Identify your highest-value questions. What are the three to five questions you ask most often that require pulling data from multiple systems? These are your use cases — and the ones that will generate the most immediate ROI from a private AI layer.
3. Consider the hardware horizon. The RTX Spark announcement means that the hardware for local AI inference will be commercially available and affordable for growing businesses within months. Now is the right time to ensure your data integration foundation is ready for it.
4. Think about your data sovereignty obligations. Under the NZ Privacy Act 2020 and IPP 3A, what data are you currently sending to third-party AI cloud systems? What are your obligations around that data? Private AI eliminates this risk category entirely.
If you would like to understand what this looks like specifically for your systems and applications, that is exactly the conversation we have with growing businesses through our free requirements assessment — no obligation, even if that is not to proceed with us.
The Bottom Line
The announcement this week from NVIDIA and Microsoft is not just a product launch. It is a signal that private AI — AI that runs on your own dedicated infrastructure, queries only your own data, and keeps your business intelligence where it belongs — is moving from the province of large enterprise to the mainstream.
The economics are now compelling. The hardware is becoming accessible. The open-source models are now genuinely capable. The only question is whether your business data is integrated, clean, and connected in a way that lets your private AI answer the questions that matter.
That is the integration problem. It is the same problem Convergence has been solving since 2009. And it is the foundation that every CODI for AI deployment is built on.
The intelligence stays here. Your data. Your AI. Your dedicated infrastructure.
About the Author
Mark Presnell is Managing Director of Convergence Ltd, systems integration specialists since 1997. Convergence builds and manages integrations between eCommerce websites and backend business software through CODI — Convergence Optimised Data Integration — and offers Integration for Private AI for businesses who want to run their own private AI models on their own dedicated infrastructure. Book a free requirements assessment at convergence.co.nz.