It's no longer a corridor whisper: Mark Zuckerberg said it openly. Meta is seriously considering turning its computing power into a cloud business, renting out capacity to third parties. A spokesperson confirmed to Bloomberg that selling access to this infrastructure “makes sense,” blessing with the CEO's name a plan that had been rumored since early this month.

The project, known internally as Meta Compute, would mark the company's entry into a market dominated by AWS, Google Cloud and Microsoft Azure. But this isn't just a service announcement. It's a symptom of structural change: AI computing power is becoming a commodity like electricity. And when a giant like Meta decides to sell its own “electricity,” industry balances break.

A new geometry for the cloud market

Meta owns an immense computing infrastructure, built over the years to train and serve its own artificial intelligence models. We're talking about tens of thousands of high-performance GPUs, an asset that today is used intermittently: peak workloads during training, idle capacity during lulls. Monetizing that capacity is a logical step financially. But the implications go far beyond the balance sheet: if one of the world's largest hardware consumers also becomes a supplier, the AI cloud value chain breaks at an unexpected point.

For companies that need to train or run inference on LLMs, Meta's arrival could lower costs and reduce dependence on the current oligopolies. In a genuinely competitive scenario, per-GPU hour prices could fall, and capacity availability would become less concentrated. But there's a flip side. Entrusting your workloads to Meta means sharing data, architecture and operational logic with a company whose core business model is collecting and monetizing personal information. Hardly the ideal partner for those operating in regulated industries or for anyone who enshrines data sovereignty as a guiding principle.

On-premise as antithesis (and confirmation)

Here's where the discussion dovetails with deployment choices. The on-premise and self-hosted movement isn't just about total cost of ownership or latency: it's a governance decision. Those who install servers with GPUs in-house do so to keep data under lock and key, to avoid lock-in contracts, and not to depend on an external provider for daily operations. Meta's move, far from eroding this trend, could strengthen it. How? By making it even clearer that the cloud — even when offered by a non-traditional cloud giant — remains potentially hostile territory for sensitive data.

Meta Compute is not a service designed for air-gap or GDPR compliance at enterprise level: it's a commercial lever to exploit underutilized assets. This means that privacy policies, audit and control will not be at the core of the offer — or rather, they will be adapted to Meta's needs, not the customer's. For anyone evaluating a local deployment framework, this move provides a new benchmark: does it really make sense to pay a rent to Meta, or is it better to invest in your own hardware and manage it with tools like vLLM or TensorRT-LLM?

A thermometer for the AI labor market

There's a third, more subtle but crucial layer. If Meta enters the cloud compute market, it's because the gap between build cost and market value of capacity has become too wide to ignore. In other words, hardware has become so powerful — and at the same time so costly — that renting becomes a standalone business even for those not born as service providers. This says something about the future: the enormous demand for AI compute is no longer met solely by traditional infrastructure builders. Hyperscalers, social networks and now even chip makers like Nvidia with its DGX cloud are crowding into the same space.

In this scenario, the losers are second-tier cloud service providers and perhaps even pure on-premise operators who can't compete with Meta's scale and cut-rate pricing. The winners, paradoxically, could be end users who, with more options, will be able to choose with greater awareness — provided they have the tools to evaluate real costs, performance and security.

The game has just begun. It remains to be seen whether regulators, mindful of Meta's privacy track record, will let it run or impose restrictions. In the meantime, for those who must decide today where to run their LLMs, one certainty stands: the confusion between AI consumer and provider has just deepened.