Elon Musk’s real AI bet may not be Grok

Still, JAX remains a general-purpose framework. It needs to offer flexibility to researchers and engineers working across varied environments. SpaceX seems willing to sacrifice part of that flexibility to gain performance on one very specific hardware configuration.

This is exactly the kind of bet you see in extreme infrastructure: when scale becomes gigantic, even a few percentage points of performance improvement can represent millions of dollars in savings. If the gain really gets close to 10x, the impact would be massive.

The real point: training faster, not just training better

In AI, the quality of a model does not only depend on the original idea. It also depends on the number of experiments the team can run.

A lab that trains faster can test more architectures, correct mistakes faster, launch more variants, adjust datasets and start over more often. Training speed therefore becomes a product advantage.

Grok is not seen as the best model on the market. OpenAI, Anthropic and Google remain very strong, especially for professional use cases, coding and agents. But SpaceX/xAI seems to want to move the competition elsewhere: instead of fighting only on the chatbot users can see, the company is trying to build one of the most integrated training infrastructures on the market, one that these competitors could rely on to train their own models. Whether Grok is better or not, SpaceX wins either way.

Colossus, Claude and the new compute economy

This move fits into a broader context. SpaceX and xAI already have Colossus, a massive AI infrastructure. Anthropic has signed an agreement to use Colossus 1 capacity, notably to improve Claude’s limits.

That point is interesting, because Anthropic is one of the strongest players in agentic AI, especially with Claude Code. If SpaceX rents part of its infrastructure to a direct competitor to Grok, it shows that compute is becoming a business in itself.

AI is no longer just about models. It is also becoming a battle over energy, chips, networking, cooling, data centers and orchestration software.

In that logic, SpaceX could become more than an AI lab. The company could become an infrastructure provider for frontier AI.

A promise to take with caution

For now, the 10x gain against JAX is an Elon Musk claim, not an independent benchmark. No detailed technical documentation has been published. We do not yet know which types of models, which training sizes or which exact metrics this comparison is based on.

It is also worth remembering that a very low-level stack can become difficult to maintain. Frameworks like JAX or PyTorch also exist because they allow teams to work faster, debug more easily and reuse an existing ecosystem.

So the question is not only: “Is it faster?”

The real question is: “Does the performance gain compensate for the additional complexity?”

At SpaceX scale, the answer could be yes. For most companies, probably not.

Why it matters

This announcement illustrates a deeper trend: the biggest AI players increasingly want to depend less on generic layers.

Google controls its TPUs and a large part of its software stack. OpenAI works with Microsoft Azure on giant clusters. Anthropic is multiplying compute agreements. Meta is investing heavily in its own infrastructure. SpaceX/xAI is now pushing an even more vertical approach: hardware, networking, data center, training software and final model.

If Musk’s promise is confirmed, V1.0 of this stack will not just be an internal tool for Grok. It could be a sign that the AI competition is moving to a new layer: the training factories.