AI comes in-house

Apple and Meta are trying to open up another way to run AI, outside data centres, just as new ones face growing resistance. Apple supplies the hardware and Meta the model, an open one that runs on Mac Studios a company could now own.

It’s called Muse Spark Open.

Anyone can download it, and a company can use it internally for free.¹

It needs about 493 GB of memory, more than one Mac Studio holds, so it runs on two or four linked machines.² The prompts stay on that hardware, out of Meta’s view.³

So a large AI model no longer needs a data centre.

It can run on machines the company owns. That is the path Apple and Meta are pushing. It differs from centralised computing, where a few hyperscalers host the models and sell access to them.

$760 billion is what Meta, Alphabet, Amazon and Microsoft are expected to invest in 2026, most of it in AI data centres.⁴ The two Mac Studios that run Muse Spark Open cost about $44,600,⁵ a purchase one company can make on its own.

Centralised computing may dominate simply because the AI race started inside big tech.

For companies that already ran data centres, putting AI at the centre was the natural way to go.

That choice may look less obvious in the long run.

Every new data centre needs a yes

Centralised computing grows only through new sites, and each one now meets resistance. In one quarter, local opponents blocked or delayed 45 US projects worth about $68 billion over land, noise, electricity and water.⁶

Seven in ten Americans oppose an AI facility near their home.⁷

Where a community says yes, power can still run short, because the grid takes years to expand.⁸

The same pushback undercuts a security argument too. It says AI is safest when it runs inside the country, under the state’s control. That means more data centres at home, and every one needs local consent.

An AI server on the office network

Running AI on the company’s own machines avoids that fight, because they sit on desks companies already have. Apple sells the Mac Studio as a desktop for large models that run on the machine itself.⁹

Several of them, linked together, become an AI server within a company’s existing network and IT systems.

Prompts, files and answers then stay on the company’s own machines, so it knows where its data is. Information still goes out when an AI agent on that server uses outside services.¹⁰

This setup also suits small and mid-sized firms, because basic IT is enough to get started.

Paying for control

The cost of owning the server depends on more than tokens. Apple sells the Mac Studio as a way to use AI without paying per token or watching cloud bills grow.⁹ On token cost alone, owned machines pay off for frequent, predictable work, while work that comes and goes is better paid for per use.¹¹

Yet a company may pay more for the same work on purpose, because the price also buys security for its prompts. For work with sensitive data, that security can outweigh the token bill.

For now, local AI complements the cloud. That could change. The technology is young, and the hyperscalers’ head start is only a few years. The more pushback new data centres meet, the more of a real choice local AI becomes. How much AI work will move onto machines companies own is an open question.