The Future of Meta Superintelligence: A 1 Year Progress Update

Max Kan · SemiAnalysis · July 09, 2026 at 19:16 · ⏱ 21 min read  | Read on Substack ↗
Summary
Meta's AI lab (MSL) has rebuilt from scratch after Llama 4's failure and is now positioned to catch OpenAI/Anthropic within 12–18 months, driven by three structural advantages: proprietary human RL data from internal screen-recording and a 3,000-engineer task-creation org, a compute ramp exceeding 5 GW across Titan clusters, and a star-studded talent roster. If Meta maintains resolve, it could surpass Google and challenge the frontier leaders, making its AI buildout a key watch for hyperscale infrastructure investors.
  • Meta spent $14.3B to acquire Scale AI's team and Alexandr Wang, and offered $1B+ packages for top researchers.
  • MSL's first public model, Muse Spark, lagged open-source rivals DeepSeek v4 Pro and Kimi K2.6, but the author argues the slope matters more than the intercept.
  • Meta is turning 3,000 internal engineers (70% of new grads plus seniors) into full-time RL task creators, matching the output of leading data companies like Mercor (~4,800 full-time equivalent).
  • Meta is simultaneously building five 1GW+ Titan clusters (Prometheus, Hyperion, Iowa, El Paso, Indiana) — the largest such build in history, with Hyperion housing 400MW single buildings.
  • Meta's AI-Backbone (AIBB) architecture uses L3/L4 Superspines to interconnect campuses up to 2,000 km apart, enabling asynchronous RL training across massive distances.
  • Key hires include Andrew Tulloch, Jason Wei, Hyung Won Chung, and the OpenAI compute team's three musketeers (though one already quit due to culture).
  • The author projects Meta will have more AI compute than OpenAI and Anthropic combined by end of 2026, even counting only dedicated MSL clusters.
  • Google's Gemini 3.5 Flash is called a 'benchmaxxed prop' that performs far worse than GPT-5.6 and Opus 4.8 in real-world use, and Microsoft is said to have 'completely blown their early lead with GitHub Copilot.'
Read time 21 min
Length 21,810 chars
Category finance
Ideas
Max Kan Substack author, SemiAnalysis
The article builds a bullish case that Meta is the only hyperscaler on track to be world-class at data, talent, and compute simultaneously, with a specific timeline to catch OpenAI/Anthropic by end of
The article builds a bullish case that Meta is the only hyperscaler on track to be world-class at data, talent, and compute simultaneously, with a specific timeline to catch OpenAI/Anthropic by end of 2026. The 3,000-engineer RL task factory, screen-recording data moat, and Titans cluster ramp are cited as unique moats. Risk: Execution risk remains high; the author notes 'success is far from guaranteed' and that any weakening of resolve (selling compute, disbanding RL org) would be a death sentence for MSL.
Max Kan Substack author, SemiAnalysis
The article explicitly criticizes Google's AI product strategy: Gemini 3.5 Flash is called a 'benchmaxxed prop' that underperforms in real-world scenarios, 3.5 Pro is 'not even Opus level on coding',
The article explicitly criticizes Google's AI product strategy: Gemini 3.5 Flash is called a 'benchmaxxed prop' that underperforms in real-world scenarios, 3.5 Pro is 'not even Opus level on coding', and Google is 'far from a compelling agentic coding product' despite the Windsurf acquisition. This suggests Google is losing ground in the frontier AI race. Risk: Google still has massive resources and could rebound; the article's critique is based on current product performance, not structural inability.
Max Kan Substack author, SemiAnalysis
Meta's unprecedented compute ramp — five 1GW+ Titan clusters, with Prometheus alone expanding to 3GW — implies enormous GPU procurement. Since NVIDIA is the dominant supplier for frontier AI training
Meta's unprecedented compute ramp — five 1GW+ Titan clusters, with Prometheus alone expanding to 3GW — implies enormous GPU procurement. Since NVIDIA is the dominant supplier for frontier AI training clusters (the article mentions no AMD or custom ASIC alternatives), Meta's buildout directly supports NVIDIA's data center revenue outlook. Risk: Meta could diversify to in-house ASICs or AMD MI series; the article does not specify GPU vendor, but current hyperscaler buildouts overwhelmingly use NVIDIA.
More from SemiAnalysis

This newsletter, published July 09, 2026, features Max Kan discussing META, GOOGL, NVDA. 3 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Max Kan  · Tickers: META, GOOGL, NVDA