Apple brings In Situ and Sovereign AI to the masses
Apple's release of an entirely on-device AI model and the software development toolkit in Swift and Python programming languages is not a simple consumer update. It is a major milestone in technology progress for statecraft, defense, and the economy.
Up until now, using AI mostly meant renting intelligence from centralized clouds through API giants such as OpenAI and others with no full-data privacy, or through cloud providers like Microsoft, Amazon, and Google. Open-source models sound liberating until you realize running them still requires infrastructure and a workforce to design, implement, and maintain them. You rent space on your chosen cloud, use an intermediary open-source provider, and suddenly you are back to square one. You remain dependent on someone else's machines and someone else's uptime.
In industries such as healthcare, pharma, finance, government, defense, military, and intelligence, running in-silo AI is a non-negotiable requirement. No data must leave the environment no matter what.
One option remains: running AI models on your own hardware and premises.
The problem up until the past 18 months has been that running large AI models required massive setup, engineering, and computing power on a different scale than traditional computers. This was the case for my friend Chris Gibson, with whom I co-founded the early-stage Recursion Pharmaceuticals back in the summer of 2013 at Stanford GSB. To run massive foundation models for drug discovery like Phenom-1, Recursion built the BioHive-2 supercomputer in May 2024. It is an NVIDIA DGX SuperPOD powered by 504 H100 GPUs, making it the fastest supercomputer wholly owned by any pharmaceutical company worldwide, and ranked #35 in the TOP500 list of the most powerful supercomputers in the world across all industries as of May 2024, and 76th as of November 2025. This obviously takes millions of dollars in investment, engineering, and tooling, as well as massive electrical bills.
The evolution of running your own models locally was a massive unmet need. My old friend Marco Mascorro, a roboticist and Partner at Andreessen Horowitz whose expertise I saw firsthand when we worked and lived together in Mountain View, solved this. In August 2025, he designed the a16z Personal AI Workstation. He packed four NVIDIA RTX 6000 Pro Blackwell GPUs with 384 gigabytes of VRAM into a wheeled, $39,500 under-desk powerhouse that ran on a standard wall outlet. Just months later, the industry followed his lead. In late 2025 and early 2026, we saw the $3,999 NVIDIA DGX Spark, the Dell GB10, the MSI Edge Expert, and the Asus Essent GX10 miniaturize this concept into desktop devices.
This is where Apple's release matters and is revolutionary, for real.
Apple released the Foundation Models framework, a built-in software toolkit that lets any Apple programmer easily plug into Apple's in-device AI features directly into their own apps. This AI runs entirely on Apple's Neural Engine, a dedicated chip inside your Mac or iPhone designed strictly for AI math. This means no internet or remote back-and-forth. No vendor relationship is required after the initial setup.
Of course, to be an Apple programmer, you must join the Apple developer program, and be vetted as a person or institution. In my experience, Apple has built the most coherent software development ecosystem in the world. Not the most open-source, not the most permissive, but the most coherent and privacy-centric. It is privacy-by-architecture. Data stays on the device by design at the hardware and operating system level. Apple vets every software release to control the complete chain of delivery and reduce vulnerabilities.
Developers use C, C++, Objective-C, and Swift to build natively. This means the code talks directly to the hardware without translation layers, delivering ultra-fast, sub-nanosecond execution. Alternatively, developers can use frameworks like React Native to build applications using familiar web languages, achieving near-native speeds without learning Apple's specific coding dialects.
Apple released a Swift and Python SDK, both memory-safe languages vetted by the NSA. Python is heavily used by researchers and massively adopted - but I still highlight notoriously slow and energy-hungry -. Apple in-device AI is now available for the masses. The features available on Apple's AI framework support text generation, structured output via guided generation, tool calling, and safety guardrails. It is available on iOS, iPadOS, macOS, and visionOS from version 26.0 onwards. It certainly won't be a benchmark crusher to flex on social media hype. And this is my next point.
What Apple did here is genuinely genius because it is what I call "Boring Technological Innovation". It is a compact, optimized model running locally on hardware you already own. There is one test that matters in serious environments: does it work when the network is gone? Yes. Does your data leave the device? No. That is the whole conversation. It is a boring tool that works and does the job, just like a BIC pen or a paper clip. If global network infrastructure collapses tomorrow, the intelligence system on your desk still functions. This is sustainable, lasting innovation. Apple has brought back AI and computers as sovereign tools.
Technological sovereignty, though, requires cognitive sovereignty. AI without human intentionality is augmented stupidity. The real advantage for individuals, institutions, and nations will not go to those renting fragile cloud architectures. It will go to those deploying resilient, off-the-grid systems while fiercely keeping the human habit of thinking.