Is xAI a neocloud now?

On Wednesday, xAI and Anthropic introduced a surprise partnership that has the Claude-maker shopping for out “the entire compute capability at [xAI’s] Colossus 1 knowledge middle,” roughly 300MW that allowed Anthropic to right away elevate its utilization limits. It’s an enormous deal for xAI, possible price billions of {dollars}. Extra importantly, it instantly monetized one of the company’s most impressive accomplishments, turning xAI from a client to a supplier of compute. 

It’s tempting to see the association as a shot at OpenAI amid the continued lawsuit. However Musk’s rationalization on X was that xAI had already moved training to a newer data center, Colossus 2, and xAI merely didn’t want them each. 

Within the quick time period, there’s an apparent logic at work. xAI’s current merchandise are principally targeted on Grok, which has seen plummeting usage for the reason that picture technology debacles earlier this yr. If xAI’s knowledge middle buildout is that rather more than what Grok must function, partnering with Anthropic provides quite a lot of inexperienced to the steadiness sheet. That is particularly helpful as the corporate, now mixed with SpaceX, speeds towards an IPO. Extra broadly, having Anthropic lined up as a buyer makes it simpler to imagine that SpaceX’s orbital knowledge middle play might actually work. 

However past the short-term profit, the Anthropic partnership sends an uncommon message about the place Elon Musk’s priorities actually lie. It suggests the corporate’s actual enterprise could also be extra about constructing knowledge facilities than coaching AI fashions. 

It’s uncommon to see a serious tech firm deal with compute assets this fashion when corporations like Google and Meta, who’re additionally coaching fashions, are constructing extra knowledge facilities. It’s a simple level to overlook, as a result of so many of those corporations are working as enterprise AI distributors, on-line providers, and cloud suppliers . However when pressured to select between promoting extra obtainable compute to prospects and preserving some to construct their very own instruments, they reliably select door No. 2. 

Simply final month, Sundar Pichai admitted on a name that Google Cloud income was decrease than it may have been as a result of the corporate was “capability constrained” — and when given the selection of renting out their GPUs or utilizing them to develop AI merchandise, Google selected the AI merchandise. 

Fb has confronted a extra excessive model of the identical constraint, spinning up a wholly new cloud equipment simply to make sure they might have sufficient GPU energy to chase Mark Zuckerberg’s AI ambition. As he put it when saying Meta Compute in January, “How we engineer, make investments, and associate to construct this infrastructure will grow to be a strategic benefit.” 

The important thing phrase there may be “strategic.” Each Zuckerberg and Pichai are wanting towards a future the place AI is powering the preferred and profitable methods on the earth. Computing energy isn’t only a approach to fulfill immediately’s inference demand, however to construct tomorrow’s merchandise — and operating quick on compute means lacking out on that probability.  

By specializing in knowledge facilities (earthbound and in any other case), xAI is positioning itself extra like a neocloud enterprise: shopping for GPUs from Nvidia and renting them out to mannequin builders like Anthropic. It’s a much more troublesome enterprise, squeezed by each chip suppliers and the shifting cycles of demand. The valuations for many lively neoclouds replicate that actuality: xAI was valued at $230 billion in its January funding spherical; CoreWeave, which oversees a comparable amount of computing energy, is price less than a third of that. 

Musk’s model of a neocloud is extra bold, as you may anticipate. Among the knowledge facilities could be in house — at the very least by 2035, if issues go in keeping with plan. xAI can be making its personal chips at the Terafab, which can take away some however not all of Nvidia’s pricing energy. However none of it adjustments the fundamental economics of the neocloud enterprise. 

As not too long ago as the February all-hands, xAI had actual ambitions in software program. That was the presentation that unveiled the orbital knowledge middle challenge, but it surely additionally teased vital ambitions in coding (since bolstered by the Cursor partnership) and attention-grabbing concepts like leveraging pc use into full-scale digital twins (within the sadly named Macrohard challenge). These are the type of long-horizon tasks that want dedicated computing assets to succeed. So long as xAI is promoting giant portions of compute to its opponents, it’s exhausting to assume such new ambitions have a lot of a future.