Meta just wrapped up the first year of its Superintelligence division with a jaw-dropping revelation: not only has Meta scaled its computing infrastructure to unprecedented levels, but it has also spawned a startup in the Reinforcement Learning (RL) environment sector that's now running independently. While OpenAI and Anthropic focus on expanding language models, Meta is dedicated to building infrastructure that could revolutionize AI training as we know it today.
A progress report shared by Mark Zuckerberg this week reveals nearly unbelievable figures: a 40x compute scale-up in just twelve months, distributed training architectures processing 2.8 sustained petaflops, and a team that grew from 120 to 890 specialized engineers. However, the most stunning aspect isn't the numbers. The intriguing part is that Meta has transformed its entire RL simulation stack into a standalone product that already has three business clients, each paying $240K annually.
The Most Aggressive Compute Ramp of the Decade Wasn't Planned
When Meta announced its Fair AI Research Superintelligence Group (FAIR-SI) in January 2025, their internal projection was to reach 15,000 H100 GPUs by year-end. However, reality exceeded those expectations as early as February. By December 2025, FAIR-SI was running 68,000 H100s across four data centers, with usage peaks at 94%. Honestly, that would bring tears of joy to any MLOps team at Google or Microsoft.
The key was an architectural decision made in July. Instead of following OpenAI's model of vertically scaling a single massive model, Meta opted for a distributed training system where multiple specialized models train simultaneously and share gradients through a low-latency network specifically designed for RL. The result? Kubernetes' scheduler was rewritten, PyTorch was modified at the backend level, and a checkpointing system was designed to store the state of 40 concurrent models without sacrificing throughput.
This approach not only increases speed but also offers operational flexibility. While OpenAI's GPT-5 requires weeks of continuous training on a monolithic cluster, Meta can iterate on architectural experiments in days, change reward shaping strategies in hours, and launch specialized models into production without halting the rest of the pipeline. To me, this explains how they managed to develop three specialized superintelligence models in just twelve months.
The Startup That Had No Name But Already Grossed $720K Annually
In September 2025, a team within FAIR-SI posed an operational problem: their RL simulation environments consumed 30% of the infrastructure budget but were only utilized at 60% capacity. The solution wasn't to optimize but to externalize.
By November, this simulation stack, based on modified MuJoCo with extensions for fluid dynamics and contact dynamics, became a standalone product called Tensor Gym. The proposition is simple yet powerful: RL environments as a service, with guaranteed latency under 8 ms for complex simulations, native integration with Ray, and the ability to scale from 100 to 100,000 agents in under two minutes.
The three initial clients aren't simple startups. They include two enterprise robotics labs (one in automotive, another in logistics) and a hedge fund training algorithmic trading agents. Each pays $240K annually for priority compute access, dedicated support, and the ability to design custom environments without modifying a line of MuJoCo. Notably, the setup is so specific that one client migrated entirely from OpenAI Gym in just four weeks.
Not only is the product interesting, but so is the spin-off strategy. Meta retains 100% ownership but operates Tensor Gym as an independent entity with separate P&L, an autonomous roadmap, and the potential to raise external capital if it scales faster than expected. This is the same strategy they used with PyTorch in 2018, and it worked so well that PyTorch is now the de facto framework for AI research.
Why Meta Chose RL While Everyone Else Bets on LLMs
The decision to focus Superintelligence on Reinforcement Learning rather than scaling Large Language Models like GPT or Claude isn't by chance. It's an architectural choice. LLMs have a predictable ceiling: eventually, more data and more parameters yield diminishing returns. RL, on the other hand, doesn't have that ceiling. In RL, the limit is the complexity of the environment and creativity in reward design.
Meta has identified three domains where RL crucially outperforms LLMs: control of physical systems (robotics, manufacturing, autonomous driving), dynamic multi-agent optimization (logistics, networks, financial markets), and adversarial environments where models must adapt to intelligent opponents (cybersecurity, applied game theory). In these cases, an LLM can suggest strategies but cannot execute them, adjust them in real-time, or learn from its errors without human intervention.
Meta's first deployed superintelligence model, internally called Athena-v1, doesn't generate text. It optimizes delivery routes for Instagram Shopping in emerging markets where road infrastructure changes weekly. Athena-v1 reduced average delivery times by 18% in Brazil and 22% in Indonesia during the last quarter of 2025, simply by learning to predict disruptions before they occur and proactively reassigning routes.
The second model, Hermes-v1, manages real-time content moderation. It doesn't label posts as good or bad. It learns behavior patterns that precede policy violations and adjusts algorithmic visibility before problematic content escalates. In A/B tests during October, Hermes reduced user reports by 31% without increasing false positives. That's RL working where GPT-4 would simply tag content after the fact.
The Architecture Enabling Training of 40 Models Simultaneously
The technical stack supporting FAIR-SI is where the report really dives into the technical details and also where it's revealed why no other player can easily replicate this. Meta has built a custom orchestration system on Kubernetes called Hydra Scheduler, specifically designed for distributed RL workloads.
Hydra solves three problems that traditional schedulers can't manage:
-
Heterogeneous resource affinity. In RL, some models need GPUs for quick inference, others need high-memory CPUs for simulation, and others require TPUs for massive gradient operations. Hydra dynamically assigns resources based on each model's training phase, moving workloads across clusters in real-time without restarting processes.
-
Continuous checkpointing without degradation. Saving a 70B parameter model's state every hour consumes bandwidth that slows down training. Hydra implements incremental checkpointing where only deltas from the last snapshot are saved, compressed with a custom algorithm Meta developed (based on Zstandard but optimized for sparse tensors). The result: checkpoints every 15 minutes with less than 2% throughput overhead.
-
Fault tolerance without complete rollback. When a node fails mid-batch, traditional schedulers restart from the last checkpoint, which can mean losing hours of computation. Hydra redistributes the failed node's workload to other workers using a replica set system with state sharing. If a node dies, work continues on another node that already has 80% of the necessary context. The average loss for failure is 4 minutes instead of 90.
Hydra's code isn't open source, but Meta published a technical paper in December detailing the architecture. Three startups have already tried to replicate it, but none have managed to make it work at scale without rewriting much of Kubernetes.
Three Startups Will Copy Tensor Gym Before Q2 2026
The problem with creating such a specific and effective product is that it inevitably attracts competition. Tensor Gym solves a real pain: RL simulation environments are slow, hard to scale, and require deep computational physics knowledge for customization. This indicates there's a market, and where a market is confirmed by a top-tier player, competitors appear.
Signs are already visible. Sim Foundry, a YC W26 startup, raised $4.2M in January to develop exactly this: RL environments as a service. Its founder is a former DeepMind member, involved in AlphaGo, and their pitch deck literally mentions Tensor Gym as market validation. Neural Sandbox, another startup backed by a16z, pivoted in December from game engines to RL simulation after one of their advisors showed them Meta's report.
The third competitor is less obvious but more dangerous: Google DeepMind is considering turning its internal simulation infrastructure (used for training manipulator robots) into a cloud product. If they do, it will likely be part of Vertex AI, aggressively priced to quickly capture market share, and will probably offer $10K in free credits for developers migrating from Tensor Gym.
Meta has a temporal advantage of 6-9 months, but it's not an insurmountable technical advantage. The real edge lies in the data: every client using Tensor Gym generates telemetry about the most used environment types, the reward functions that converge fastest, and which agent architectures fail under which conditions. This data directly feeds back into FAIR-SI, which uses it to improve its own superintelligence models. That's a virtuous cycle competitors can't replicate without years of production use.
What Meta Isn't Saying About the Next Year
The progress report is transparent in metrics but mysterious about its future strategy. There are three things it doesn't mention, but internal sources confirm are in active development:
-
A superintelligence model for internal infrastructure. Meta operates one of the world's most complex distributed systems: 3.6 billion users, petabytes of daily generated content, recommendation algorithms that must adapt to radically different cultures in real-time. They are training an RL model that optimizes data center resource allocation: what content to cache where, which queries to process in which regions, how to balance latency vs. operational cost. If successful, it could reduce infrastructure costs by 15% to 20% annually. We're talking billions of dollars.
-
Integration of Tensor Gym with Reality Labs. The Quest 4 headsets Meta will launch in Q3 2026 will include eye and facial tracking sensors much more precise than previous generations. Those data are perfect for training RL models that predict user intention, not just what you're looking at, but what you'll do next. Tensor Gym could become the platform where external developers train AI agents for adaptive VR/AR experiences. It's a move to position Meta as the AI infrastructure for spatial computing before Apple does the same.
-
A grant program for academic research labs. Meta is preparing $50M in Tensor Gym credits for universities researching applied RL. It's the same strategy Google used with TPUs for research: give away computing in exchange for publications that validate your technical stack and generate talent familiar with your tools. When those PhDs graduate, they know how to use Tensor Gym better than any alternative. Some join Meta, others found startups that become clients. All amplify your ecosystem.
In Closing: When Scaling Infrastructure Creates New Products
What's fascinating about Meta's report isn't that they've scaled computing 40x—that's simply spending money on GPUs. The truly amazing part is that by doing so, they've discovered a completely new commercial product that no one else is building, and launched it before the competition even noticed the space.
Tensor Gym is what happens when internal infrastructure is so efficient that it's more valuable as an external product. PyTorch followed that path. So did React. Meta has a history of turning internal tools into industry standards, and now they're applying the same strategy to RL.
The question isn't if other players will try to replicate Tensor Gym. The question is if Meta can scale fast enough to establish network effects before that happens. With $720K ARR and three months on the market, the clock is already ticking.
Want to bet OpenAI announces something similar by June?