A look at Meta Superintelligence Labs' growth in the past year, including a top-tier RL environment and compute ramp that could catch up to Anthropic and OpenAI
SemiAnalysis
Context & Ripple Effects
Meta’s superintelligence effort was formed as a dedicated, heavily staffed push after Meta had separately examined DeepSeek’s training approach. Its early strategy was still unsettled: leaders discussed relying on Google or OpenAI models for product features, while the lab also considered moving from the open Behemoth project toward a closed model.
By January, the lab had delivered its first notable internal models. The reported advances in reinforcement-learning infrastructure and compute therefore mark a shift from organizational build-out and model-strategy debate toward a more self-sufficient frontier-model development program.
First-order effects
Meta Superintelligence Labs gains a stronger internal training stack—especially for reinforcement learning—and more compute capacity, improving its ability to develop and evaluate models without depending as heavily on external model providers.
Anthropic and OpenAI face a better-resourced Meta contender whose infrastructure could narrow a capability and compute gap, rather than merely competing through Meta’s distribution channels.
Second-order effects
Meta’s product teams have a more credible path to use internally developed models in Meta AI and social-app features, reducing the strategic appeal of licensing or integrating rival models if internal quality continues to improve.
The lab’s apparent emphasis on proprietary frontier development reinforces the trade-off already visible in coverage of Behemoth: Meta’s AI strategy may put less weight on releasing its strongest work openly and more on retaining it as a competitive asset.
Third-order effects
If Meta converts compute and reinforcement-learning investment into sustained model gains, frontier AI competition becomes more concentrated among a small group able to fund both large-scale infrastructure and specialized post-training systems.
The pattern points to reinforcement-learning environments becoming a core competitive layer alongside base-model training: advantage may increasingly depend on the quality of evaluation, feedback, and training loops, not compute alone.
The trend: This is one data point in the shift from broadly distributed open-model competition toward vertically integrated frontier-AI programs built around proprietary compute, post-training infrastructure, and internal deployment.
I'm missing something here How can “there is no data wall” be congruous with “screen recording top engineers will be a moat” Just yesterday current AIs destroyed the world's best programmers in coding competitions The era of human generated coding data is ending man
A possible history of the future: “nobody really knew how AI would play out. But they all knew it was the next big thing. This meant that Meta simply had to stay in the game, no matter the cost. If they'd stayed on the sidelines, they would not have been able to mobilize to take
Meta can have compute, data, and talent to compete among the top AI players. But it has a serious public trust problem that has grown during the past 2 decades of neglect and dark patterns. I'm not sure it's fixable.
I do not think anyone saw this massive comeback from Meta coming. And now they are even receiving high praise from SemiAnalysis. Just to be clear: I have the utmost respect for SemiAnalysis and Dylan Patel. They do truly outstanding research. And if they believe that Meta
Gotta admit, Meta being the force that starts scuppering Anthropic and OpenAI's growth is a really funny way to wrap this up. Zuckerberg cannot build products to make his users happy, but he LOVES building stuff to kill off other companies
I heard a pretty interesting rumor on the ground at ICML. Meta has supposedly already developed an internal model at roughly the Mythos 5 level, and all that remains is deployment within the next few months. Honestly, I was skeptical at first. From my perspective, Meta had not…
The Future of Meta Superintelligence: A 1 Year Progress Update A top tier RL environment startup spawns out of thin air, the most aggressive compute ramp we've ever seen, 2000km+ scale-across, and some advice for Google DeepMind https://semianalysis.substack.com/ ...
SemiAnalysis is out with a deep dive on $META Superintelligence, and they're clearly impressed with Meta's pace of improvement. They think Meta may be the only major AI player on track to be world-class across data, talent, and compute. Meta has reportedly turned its internal [im…
The vibes are shifting. SemiAnalysis confirms the view $META will have the most compute of frontier labs this year. And has the best chance to catch up to OpenAI and Anthropic. [image]