Open-weight AI models are moving from the cloud to company servers and desks

Where things stand

For a few years, the most capable AI models have mostly lived behind someone else's interface. You send a request to a provider's cloud and pay per use. That is still the dominant model. But recent releases show how fast the alternative is growing: models whose weights anyone can download and run on their own hardware.

Mistral has released the weights for Mistral Large 4, a model with around a trillion parameters. The French company says businesses and governments can now run it themselves, and it is pitching the model for detecting and responding to cyber threats. Nvidia has launched three open-weight Nemotron models: a large one for enterprise agents and two smaller Bolt models that run locally on PCs and workstations with Nvidia chips. Meta has released Muse Spark Open, an 816 billion parameter model built to run locally on Apple Macs.

"Locally" still has limits. Muse Spark Open needs at least one terabyte of RAM. That is a high-end workstation, not a laptop. The largest open models still need serious hardware.

At the same time, the economics of closed models are becoming more visible. Anthropic's IPO filing shows $16.3 billion in revenue for the first half of the year and a net loss of $80.7 billion. The company itself warns it may not achieve or sustain profitability. TechCrunch argues that customers in Europe should plan for higher prices once public investors ask for margins. Google has announced Gemini 4 Carbon but is holding back the public release until it has gone through US government testing, with no date given.

In Sweden, the capacity to run large models close to home is growing. Sferical AI, formed together with several large Swedish companies, has opened Sferical One, built on 16 Nvidia DGX GB300 systems. The French cloud company Scaleway plans an AI supercomputer in Sweden with AMD Instinct MI355X chips, built by Bull.

In five years

We expect a split market. Closed models from a few providers will likely keep the lead at the very top, for the hardest tasks. But a large share of everyday business use will run on open-weight models that the company controls. Think of summarising documents, searching internal knowledge, sorting support tickets, reviewing code and watching logs for attacks.

The reasons are practical. A company that runs its own model knows what it will cost next year. It can keep sensitive data inside its own walls or in a European data centre. It does not depend on a release schedule set by another country's testing process. For banks, health care, defence and public agencies in Europe, those points weigh heavily.

For technology companies, the work changes. Fewer teams will simply call an API. More will need people who can choose a model, adapt it, run it, secure it and measure its quality. That is closer to running a database than to buying a subscription. Swedish facilities like Sferical One and Scaleway's planned machine fit this pattern, but only if Swedish companies and researchers actually use them. Computer Sweden rightly points to that as the harder part.

For people, the change will mostly be invisible. The assistant in a work tool or in a public agency's chat may run on a server in Sweden instead of in the US. Some features will run directly on the computer, without a network connection. Nvidia's smaller Bolt models point in that direction.

This path depends on two things. Open models must stay close enough in quality to closed ones. And hardware costs must fall so that "local" means an ordinary workstation, not a terabyte machine.

In ten years

Ten years is a long time in AI, and we hold this view loosely. The most likely picture is that access to capable models becomes a commodity layer, much as open source operating systems and databases did. Value moves to data, integration and trust. Companies compete less on which model they use and more on what they connect it to.

We expect strong models on ordinary laptops and phones, handling most personal tasks privately. The very largest models will probably still live in big data centres, with access shaped by government testing and regulation. Google's handling of Gemini 4 Carbon already hints at that.

For Europe, this could reduce dependence on a few American providers. Mistral shows that a European company can compete at the open frontier. Whether Europe builds a lasting position depends on whether its companies turn open models into products people want, not only on hosting compute.

What could change the picture

The biggest uncertainty is whether open releases continue. Meta, Nvidia and Mistral each have their own business reasons to publish weights. Those reasons can change. If leading labs stop releasing, or if governments restrict the most capable models on security grounds, the open path would narrow.

Security cuts both ways. A downloadable model can help defend networks, as Mistral intends. The same weights can be misused, and there is no central off switch. A serious incident could change regulation quickly.

Prices matter too. If closed providers keep prices low despite heavy losses, the cost argument for running your own model weakens. If they raise them, as TechCrunch expects, the shift to open models speeds up.

Finally, nobody knows how fast hardware will get cheaper. Today's one terabyte requirement is a real barrier. If it falls slowly, open models will stay with large companies and data centres longer than we expect.