xAI sells compute. Does that make it an AI company?

There is a simple test I run on any company that calls itself an AI company. Ask: where does the money come from?

For xAI, the answer is getting complicated. The company behind Grok has been quietly moving into territory that looks a lot like a neocloud: renting out access to its GPU clusters to external customers. That is not a product strategy. That is an infrastructure business.

This matters more than it sounds. When the primary revenue stream of an "AI company" is selling raw compute, the label starts to do a lot of lifting. And it raises a question that applies well beyond xAI: what is an AI company, actually?

The neocloud model is not new. The framing is.

Neoclouds are not a mystery. Companies like CoreWeave, Lambda Labs, and Voltage Park exist specifically to rent GPU capacity to teams that cannot get it from AWS or Google at the scale or speed they need. It is a legitimate business. The margins are thin, the capital costs are enormous, and the differentiation is almost entirely operational.

What is new is an ostensibly frontier AI lab sliding into this model as a meaningful part of its business. xAI built Colossus, reportedly one of the largest GPU clusters in the world, to train its own models. Colossus is real infrastructure, and infrastructure that sits idle is capital rotting on the floor. Renting it out is rational.

But rational and strategic are different things. If your cluster is busy running other people's workloads, it is not busy training your next model. That tension does not disappear just because the spreadsheet balances.

Three types of company wearing the same badge

The term "AI company" currently covers at least three very different business models, and the market treats them as interchangeable. They are not.

Model builders train foundation models and monetize through API access, licensing, or consumer products. OpenAI, Anthropic, Mistral. The differentiation is in the model itself: its capability, safety profile, and cost to run.

Application builders take existing models and build products on top of them. Most of the interesting agency and SaaS work lives here. The model is a commodity input; the differentiation is in the product, the workflow, the data layer.

Infrastructure providers supply the compute or tooling that everyone else runs on. NVIDIA is the obvious anchor. Neoclouds sit one layer above. The differentiation is operational: uptime, pricing, network speed, availability.

xAI started as a model builder. Grok is a real product with real users. But once you are also renting out cluster capacity, you have one foot in the infrastructure camp. That split attention is not impossible to manage, but it changes the incentive structure in ways that compound over time.

Why this matters for anyone building on AI

If you are a founder or a product team deciding where to place your bets, the business model of your model provider is not an academic question.

A pure model builder has one job. make the model better and cheaper. Their roadmap is aligned with yours. A provider that also sells compute has a second revenue stream that competes for internal attention and capital allocation. That does not make them a bad partner. It makes them a different kind of partner.

There is also a subtler risk. Infrastructure businesses attract infrastructure customers: large, price-sensitive, and demanding on SLA terms. Serving those customers well pulls engineering and ops resources toward reliability and provisioning work, away from model research. The culture shifts. Slowly, then faster.

None of this means xAI is abandoning its model ambitions. Elon Musk has never been quiet about wanting Grok to compete at the frontier. But ambition and business model can point in different directions for longer than anyone expects, and the gap eventually forces a choice.

The broader pattern

xAI is not unique here. Several well-funded AI labs have built more compute than their own training runs require, and the economic pressure to monetize that capacity is real. NVIDIA reported data center revenue of $47.5 billion in fiscal year 2025. That number tells you something about where the money in the AI stack is concentrating: at the hardware layer, not the intelligence layer.

Meanwhile, the application layer is where most teams are actually spending their AI budgets. The intelligence, for many practical purposes, is already good enough. The constraint is workflow, integration, and trust, not raw model capability. That gap between where capital pools and where value is created is one of the more interesting structural facts about this industry right now.

For anyone building AI products or advising organizations on AI strategy, the takeaway is straightforward. Do not treat "AI company" as a meaningful category without asking which business model sits underneath it. The label is doing too much work.

What an AI company actually is

I do not think there is one right answer. But there is a useful way to frame it. An AI company is one where the core product gets meaningfully better as AI research advances, and where that improvement is the primary driver of competitive position.

By that definition, a neocloud is not an AI company. It is a capital-intensive infrastructure business that happens to serve AI customers. Serving that market is fine. Calling it AI-native is a stretch.

For xAI, the question is whether Grok remains the center of gravity or becomes one line item among several. Right now it is still the flagship. But business models have momentum. Once compute rental is a material revenue source, the organizational logic that follows it is hard to reverse.

The companies worth watching over the next two years are not the ones with the biggest clusters. They are the ones with the clearest answer to that first question: where does the money come from, and does that source make the product better?