Failing to make breakthroughs with models, Zuckerberg is now setting his sights on infrastructure.
The trigger was a series of setbacks for Meta: Gemini model usage was restricted, Zuckerberg admitted internal AI agent tech was progressing slower than expected, employee morale hit a 20-year low...
It's just been a rough year.
But it's okay, Zuckerberg had a flash of inspiration and came up with a Plan B.
If our in-house models can't keep up, then we can just sell GPUs!!
According to Bloomberg, Meta is considering launching Meta Compute, opening its vast AI infrastructure to external customers.
Wow, it seems everyone just wants to be in the shovel-selling business...
Meta Is Going to Sell GPUs
Now that they're selling shovels, just how many shovels does Meta have?
According to SemiAnalysis, Meta's data center and compute procurement isn't slowing down; it's actually accelerating.
In just the first half of this year, Meta has already signed contracts for over 5GW of capacity in cloud and colocation data centers. That's not even including its accelerated push into self-built data centers.
The two largest data center campuses Meta is currently building represent a combined 2.5GW of capacity.
And since the beginning of 2024, Meta has signed deals for data centers and compute capacity nearing 10GW.
Those densely packed dots on the map are Zuckerberg's confidence in selling GPUs.
This massive compute power has a few potential destinations:
First, continue feeding it to its own models, like Alexander Wang's MSL's recently launched Muse Spark, and the next-generation model Watermelon currently in training.
Second, use it on the ad recommendation system. SemiAnalysis believes Meta might want to increase the complexity of its ad recommendation system by another 10x, using more training and inference compute to boost ad revenue.
Third, engage in neocloud deals similar to SpaceX's, leasing out a portion of compute at a premium to external clients.
If calculated based on high-compute leasing contracts like SpaceX's, each GW could generate annual revenue of about $50 billion.
If Meta allocates just 200MW of compute to external clients, it could bring in $10 billion in annual revenue, and at extremely high margins.
Tsk tsk, the potential profit is really not small~
And SpaceX has pioneered a new model: a three-year contract, but either party can cancel with 90 days' notice—essentially making it a three-month, auto-renewing deal.
This means Meta could reclaim the compute power for MSL's use at any time.
Fourth, host third-party models.
SemiAnalysis even judges that Meta is in final negotiations with Anthropic to gain private instance access to Claude.
In the future, Meta would create a model service platform similar to Amazon's Bedrock, Microsoft's Foundry, or Google's Vertex.
In other words, Meta could deploy third-party models like Claude on its own infrastructure and then package and sell them to enterprise customers.
For Meta, this has at least three uses:
First, of course, is internal use.
Google just restricted Meta's use of Gemini, and Meta might turn around and use Claude as a replacement.
After all, Meta's own AI projects require massive amounts of high-quality model tokens.
And Claude happens to be one of the strongest models currently available.
Second is external sales. Meta could sell Claude-as-a-service like Amazon's Bedrock.
Customers wouldn't need to contract with Anthropic, deploy, or maintain the model themselves; they could just call the model through Meta's platform.
Third is vertical applications. Meta could leverage its own ad platform to build sales and marketing SaaS, integrating cutting-edge AI Agents.
SemiAnalysis expects that Meta may soon announce a similar agreement, with Anthropic being the prime candidate, but OpenAI or Google could also join.
If Meta's compute business takes shape, then its competitors won't just be model companies like OpenAI, Anthropic, and Google.
It will also stand opposite AI cloud providers like AWS, Azure, Google Cloud, as well as neocloud companies like CoreWeave and Nebius.
As soon as the news broke, the capital market reacted immediately.
Meta's stock price surged nearly 9%, while neocloud companies like CoreWeave and Nebius saw sell-offs.
Wall Street clearly understood Zuckerberg's new story:
Our models might not be winning yet, but our GPUs can start earning money now!
Why Sell Compute? Building Models is Too Expensive
The most direct reason for Zuckerberg's pivot from models to selling shovels is:
Developing models really burns too much cash!!!
Meta's official capital expenditure guidance for 2026 has already been raised to $125-$145 billion.
For comparison, Meta's Q1 capital expenditure this year alone reached $19.84 billion.
But looking at Meta's model progress, it's hard not to feel nervous:
The Llama series is open-source and has significant ecosystem influence, but it's hard to directly translate that into revenue.
And Meta's latest in-house model, Muse Spark, hasn't truly put Meta back in the top tier yet.
Now Meta is internally training the next-generation model Watermelon, reportedly involving an order of magnitude more compute investment than Avocado.
Alexander Wang states: Don't worry everyone, Watermelon has already caught up to GPT-5.5 levels.
Meanwhile, the current version of Muse Spark is about to be updated, with major improvements expected in coding capabilities and agent intelligence.
When users ask Wang when Meta will release a model comparable to Claude Opus, he says:
Very soon!
(Come on Xiao Wang, stop talking and just release it already)
Ultimately, Meta's AI ambitions have always revolved around a simple goal:
Catch up with OpenAI, Anthropic, and Google.
To achieve this, Zuckerberg has spared no expense. Chips, data centers, talent—almost everything has been invested at the highest level.
But the problem is, the money has been spent, yet Meta hasn't truly convinced developers and customers that its models are at the industry's cutting edge.
When model progress can't be immediately monetized, compute power becomes the asset most easily understood by Wall Street.
Because GPUs and data centers can at least be priced.
These resources can be rented out, can host models, can sell APIs, can serve advertisers, can be used for AI agent SaaS, and can also internally continue to improve the ad recommendation system.
It's like Meta was originally telling the market a distant story:
Trust us, we'll create superintelligence.
But now the story sounds much closer:
Even if superintelligence doesn't arrive that quickly, these GPUs aren't sunk costs.
Of course, selling compute doesn't mean Meta is giving up on its own models. Zuckerberg's Plan A is still superintelligence.
Keep poaching talent, keep stacking GPUs, keep training bigger models, keep chasing the big three.
On the path to ASI, Zuckerberg never gives up!
It's just that the uncertainty in the frontier model competition is too high, and a little compromise along the way is inevitable~
References:
[1]https://newsletter.semianalysis.com/p/meta-compute-everyone-wants-to-be
[2]https://www.bloomberg.com/news/articles/2026-07-01/meta-is-building-a-cloud-business-to-sell-excess-ai-compute
This article is from the WeChat public account "QbitAI", author: Tingyu
