Meta Turns 3,000 Employees Into an RL Data Factory While Racing to Out-Compute OpenAI and Anthropic
SemiAnalysis one-year progress report on Meta Superintelligence Labs, published July 9, 2026
A year after Llama 4's disastrous reception forced Mark Zuckerberg to blow up and rebuild Meta's entire AI organization, SemiAnalysis has published a detailed progress report on Meta Superintelligence Labs (MSL) that reframes how investors should think about Meta's position in the frontier AI race. The report, authored by Max Kan, Julien Martin-Prin, Jeremie Eliahou Ontiveros, and Dylan Patel, argues that Meta is quietly assembling the only stack among hyperscalers that is world-class across all three inputs needed to build frontier AI: data, talent, and compute. The most striking finding is not a benchmark score, but a corporate policy: Meta has begun tracking employees' screens, keyboards, and mouse movements, effectively turning its own workforce into a proprietary reinforcement learning data operation.
Meta's employee surveillance program becomes a hidden RL gold mine
SemiAnalysis frames the internal tracking initiative, which drew significant employee backlash and negative press when it surfaced, as one of the most valuable and underappreciated moves Meta has made all year. The logic rests on how modern AI labs actually build reinforcement learning environments. Frontier labs no longer improve models simply by predicting the next token; they train models to complete entire tasks, which requires realistic environments, tools, and verifiers. The report argues that screen recordings of real white-collar work are exceptionally valuable inputs for this process because they are, by definition, representative of real economic tasks rather than the "unnaturally over-specified" scenarios found in benchmarks like OpenAI's GDPval, which the authors criticize directly, noting that some tasks come with prompts a human "would never write" in practice.
What makes this a genuine structural advantage for Meta, according to the report, is scale. External data vendors such as Mercor, Surge, and Handshake have each crossed $1 billion in annual recurring revenue by hiring expert contractors to build these environments, with Mercor logging 2.517 million expert hours in the second quarter of 2026 alone, roughly equivalent to 4,800 full-time workers. Meta, the report argues, is "already in the same ballpark" using its own staff, with likely higher average quality, and can draw from a pool of roughly 70,000 additional employees if the approach proves out. In late May, Meta formalized this into a new "applied AI engineering org," reassigning approximately 3,000 engineers, including 70% of new graduates, to build RL tasks and environments full time. SemiAnalysis is blunt about the strategic import: "We think this is an extremely underappreciated advantage for MSL."
The compute ramp: five simultaneous gigawatt titans
Beyond data, the report details what it calls the most aggressive datacenter buildout ever observed in the industry. Meta is simultaneously constructing five clusters exceeding 1 gigawatt each: Prometheus in Ohio, Hyperion in Louisiana, and three unnamed campuses in El Paso, Iowa, and Indiana. SemiAnalysis notes that no company has previously built more than one full-gigawatt campus at a time, with the closest precedent being AWS's 800-megawatt Project Rainier in Indiana. Meta currently has two gigawatt-scale sites under simultaneous construction. At Hyperion, Meta is building what the report calls the world's largest single buildings at 400 megawatts each, with 1.5 gigawatts under construction today. In Iowa, satellite imagery cited in the report shows Meta going from an empty site to a full gigawatt under construction within a single year.
SemiAnalysis's Tokenomics Model now projects Meta will have more total AI compute than both OpenAI and Anthropic by the end of 2026. The firm cautions that a meaningful share of this capacity will still serve recommendation systems and generative advertising rather than frontier model training, but even under conservative assumptions that isolate specific high-profile datacenter sites for MSL, Meta's training compute is comparable to OpenAI's and Anthropic's through 2027. The financial contrast with Google is explicit: Meta lacks a cloud rental business competing for the same GPUs, and Zuckerberg's willingness to run free-cash-flow negative gives Meta a flexibility neither Microsoft nor Google currently exercises for internal model training.
Solving scale-across: AI-Backbone and the 2,000-kilometer network
The report offers a rare level of technical detail on how Meta is networking these campuses together, a problem the industry calls "scale-across." Prometheus, rather than being a single site, is described as a constellation of 27 datacenters spread across six campuses, five clustered within 6 kilometers of each other and a sixth roughly 75 to 80 kilometers away. Meta's answer is a new architecture called AI-Backbone, an evolution of its existing 10X Backbone network, which uses layered superspine and aggregation hubs to deliver approximately 22 petabits per second of bidirectional bandwidth across the full Prometheus cluster. The connections between campuses rely on a mix of long-range optics and dense wave division multiplexing systems, depending on fiber distance.
This architecture is not without tradeoffs. SemiAnalysis notes that latency within a single scale-out region runs 1 to 10 microseconds, but reaching a site 100 kilometers away cannot fall below roughly 500 microseconds due to the physical limits of light propagation in fiber, forcing Meta to run pretraining synchronously within a single region while distributing reinforcement learning workloads asynchronously across the globe. The report says future titan sites will push this design further, linking campuses up to 2,000 kilometers apart.
Assembling, and occasionally losing, the superteam
On talent, the report tracks continued high-profile recruiting beyond last year's headline-grabbing $14.3 billion Scale AI deal that brought in Alexandr Wang, including the hiring of Thinking Machines cofounder Andrew Tulloch along with several of that startup's founding team, and former OpenAI researchers Jason Wei, Hyung Won Chung, and Zhiqing Sun. Meta also brought on Dina Powell McCormick as President and Vice Chairman to help build out its compute fleet and poached OpenAI's three-person compute leadership team in April. However, the report notes one of those three hires has already departed, citing Meta's internal culture issues within the infrastructure organization, a reminder that assembling a superteam on paper does not guarantee cohesion in practice.
SemiAnalysis is careful to temper its optimism. "We commend them for marshaling the resources and balls necessary to take a true shot at building RSI, but now they have to do the actual work," the authors write, adding that any sign of weakening resolve, such as signing a long-term compute sale with no clawback provisions or letting top researchers walk, would be "tantamount to a death sentence for MSL."
Muse Spark 1.1: catching up, not caught up
On the product itself, the analysts had early access to Muse Spark 1.1 ahead of its official release and assess it as roughly on par with Anthropic's Opus 4.6 or Zhipu's GLM 5.2 for general agentic use cases, a notable improvement from the original Muse Spark launch in April, which lagged open-source competitors DeepSeek v4 Pro and Kimi K2.6 on most benchmarks. The firm believes Meta's pricing of the model just under GLM 5.2 was likely a deliberate positioning choice. Even so, the report flags functional shortcomings, including a tendency to ignore code warnings rather than fix them and improper use of editing tools, and states plainly that none of SemiAnalysis's own internal token volume will shift to Muse Spark 1.1. The firm does not expect Meta to reach parity with Anthropic or OpenAI before the end of 2026, even in its bull case.
Google's "loser mentality" and the fight for third place
The report's sharpest language is reserved for Google. Despite Gemini 3 Pro and Nano Banana briefly putting Google in the frontier conversation, SemiAnalysis argues DeepMind is being structurally starved of the compute needed to compete, with the majority of Google's incremental datacenter capacity over the next two years projected to serve its infrastructure-as-service and third-party API businesses rather than internal model training. The firm notes Google recently issued $85 billion of equity to fund additional AI infrastructure, but believes most of that new capacity will ultimately be rented out to customers including Anthropic. "This is loser mentality from Google," the authors write, adding that DeepMind will have less training compute than OpenAI, Anthropic, and MSL going forward, and that Google has continued losing key reinforcement learning researchers to Anthropic because its RL efforts remain too decentralized.
Citing Alexandr Wang's recent podcast comments that true frontier labs are built on the conviction that superintelligence is imminent and that all business decisions must flow from that belief, the report argues Google's leadership does not genuinely hold that conviction, unlike the founder-driven urgency at OpenAI and Anthropic. SemiAnalysis's advice to Google, and by extension to Microsoft AI and Amazon AGI, is to immediately redirect far more compute to internal model development and put engineers to work generating RL tasks rather than continuing to subsidize competitors' training runs. The report closes with a pointed reordering of the competitive landscape: in its view, the race for third place in frontier AI is now between Meta and SpaceX's xAI, not Google.