Explainer · updated October 2026
Super intelligence (SI), on your own GPUs
AI has a new name: SI, super intelligence. Here is what the term means, why so many products now live on .si domains, and how to run SI models privately on hardware you control.
Why AI is now called SI
On 22 September 2026, speaking at the UN General Assembly, the US president said the word “artificial” made the technology sound fake and that AI would be officially called “super intelligence”. On 29 September an executive order told federal agencies to use “SI” and “super intelligence” in their communications.
The models did not change: SI covers the same large language models everyone called AI a month earlier. What changed is the vocabulary, first in government and then in product names, press coverage and search.
Why .si domains took off
.si is the country-code domain of Slovenia, and its two letters match SI. Slovenia's registry counted fewer than 2,000 new .si names in August 2026 and more than 44,000 in September, with over 20,000 in a single day after the executive order. .si now sits next to .ai as the natural address for an intelligence product, at a fraction of the price.
Many of those names are parked or resold. gpuos.si is a working product: the GPU operating system for super intelligence, built and hosted in the EU.
Private SI: run it where your data is
Most SI usage goes through third-party APIs, which means prompts, documents and customer data leave your company. Open-weight models are now good enough to run in-house. gpuos turns your NVIDIA GPUs into a private SI cloud in three steps:
- 01
Connect a GPU
One install command on your server, homelab or rented machine. No inbound port to open.
- 02
Deploy an SI model
Pick Qwen3, GLM-4, gpt-oss or another tested model. A VRAM gauge checks the fit first.
- 03
Call one API
Use the OpenAI SDKs with your gpuos key. Keys, quotas and usage metering come built in.
SI models you can run today
- Qwen3 32B≈ 22 GB VRAM · Apache 2.0 · Internal assistants, RAG over company documents
- GLM-4 32B 0414≈ 20.5 GB VRAM · MIT · Tool calling and agents, Structured JSON output
- gpt-oss 20B≈ 14 GB VRAM · Apache 2.0 · Coding agents, Tool use
- Qwen3 8B≈ 6.5 GB VRAM · Apache 2.0 · Classification and extraction, High-volume tasks
An SI that acts: gpuOS and cpuOS
A model that only answers questions is half of super intelligence. Agents also act: they write code and run it, open web pages, edit files. gpuOS runs the thinking on GPUs you control. Its sibling cpuOS runs the acting: every task gets its own Firecracker microVM, ready in under a second, isolated by its own kernel and hosted in the EU. Together they form one operating system for super intelligence that stays on infrastructure you choose.
cpuos.siSandboxes where SI agents run code safelyGlossary
- SI (super intelligence)
- The name the US government uses for AI since September 2026. In practice it covers the same systems: large language models and the products built on them.
- ASI
- Artificial superintelligence: the research notion of a system that exceeds human ability across most tasks. Today's SI products are not ASI.
- Open-weight model
- A model whose weights you can download and run yourself, like Qwen3, GLM-4, gpt-oss or Mistral Small.
- Inference
- Running a trained model to answer a request. This is what your GPUs do when gpuos serves an SI model.
- VRAM
- The memory on the GPU. It decides which models fit: about 22 GB for a 32B model at 4-bit, about 6.5 GB for an 8B model.
- AI agent
- A model in a loop: it plans, calls tools, reads the results and decides the next step. It needs inference (gpuOS) and somewhere safe to run what it decides (cpuOS).
- Sandbox (microVM)
- A throwaway virtual machine with its own kernel where an agent can run code, browse or test a repository without touching your servers. cpuOS creates one per task.
- OpenAI-compatible API
- The request format popularized by OpenAI and used by most SDKs and tools. gpuos speaks it, so existing code works unchanged.
Questions
- What is the difference between gpuOS and cpuOS?
- They are the two halves of one operating system for super intelligence. gpuOS (gpuos.si) runs the models: open models on your GPUs behind an OpenAI-compatible API. cpuOS (cpuos.si) runs what the models decide to do: code, browsers and tests in Firecracker microVM sandboxes. Use either alone, or point an agent at both.
- What does SI stand for?
- SI stands for super intelligence. On 22 September 2026 the US president said AI would be officially called super intelligence, and an executive order on 29 September told federal agencies to use SI and super intelligence instead of AI.
- Is SI different from AI?
- Not technically, for now. The models are the same: what changed is the name used in official US communication and, increasingly, in product names. Searches, docs and SDKs still mostly say AI, so gpuos uses both.
- Why are AI companies moving to .si domains?
- .si is the country domain of Slovenia, and its letters match SI. Registrations jumped from under 2,000 in August 2026 to more than 44,000 in September, according to the Slovenian registry. A .si name also costs far less than .ai.
- Can I run super intelligence on my own GPU?
- You can run today's open SI models on your own hardware. A 24 GB card runs 32B models such as Qwen3 32B or GLM-4 32B; an 8 GB card runs Qwen3 8B. gpuos installs the engine, deploys the model and exposes it through an OpenAI-compatible API.
- What is a private SI cloud?
- SI models served from machines you control instead of a third-party API, so prompts and documents never leave your infrastructure. gpuos adds the cloud features on top: API keys, quotas, usage metering and team access.
Sources
Run your own super intelligence
Free for one GPU. Connect a machine, deploy an SI model and call it through one OpenAI-compatible endpoint.