Meta MSL Builds Five Titan Clusters; Muse Spark Ties Opus 4.6

A year after Mark Zuckerberg reorganised Meta around superintelligence, the interesting question isn't whether the company has bought scale, it plainly has, but whether the scale is buying model quality yet. SemiAnalysis published a July 9 progress report on Meta Superintelligence Labs, and the honest read is three impressive charts and one soft one.

On compute, the piece describes five gigawatt-scale 'Titan' campuses in flight. Prometheus in Ohio has grown from roughly 1GW to over 3GW, Hyperion in Louisiana is 1.5GW under construction, an Iowa lease covers another 1GW, and two more 1GW-plus campuses are going up in El Paso and Indiana. Prometheus alone spans 27 datacenters across six campuses, stitched together with an AI-Backbone network SemiAnalysis pegs at 22 petabits per second bi-directional. The outlet's own Tokenomics model projects Meta will have more AI compute than both OpenAI and Anthropic by end of 2026, though a real slice is still going to recommendations and ads.

On data, the reporting describes roughly 3,000 engineers, about 70% new graduates, working full-time on building RL environments after a May 2026 restructuring, alongside a supplier stack that now includes Mercor, Surge and Handshake at over $1B ARR each. On talent, the piece walks through the $14.3B Scale AI deal that brought Alexandr Wang across, a stream of ex-OpenAI and Thinking Machines hires, and Dina Powell McCormick joining as President and Vice Chairman in January 2026.

The soft chart is the model itself. April's Muse Spark launched, on SemiAnalysis's read, 'roughly on par with Opus 4.6 or GLM 5.2 for general agentic use cases,' priced just under GLM 5.2, and lagged DeepSeek v4 Pro and Kimi K2.6 on most benchmarks at launch. The piece also flags a bad habit of ignoring warnings and misusing the edit tool. The honest caveat is that this is one outlet's read of one shipped model, and while SemiAnalysis projects parity with Anthropic and OpenAI by end of 2026, it does not show the pre-release Muse Spark 1.1 numbers behind that call. What the reporting doesn't give you is a clean read on retention risk after the late-May employee screen-tracking rollout, or how Zuckerberg is measuring return on those billion-dollar packages. If Meta hits parity, the Llama distribution flywheel gets very interesting; if it doesn't, this is the best-funded second tier in the industry.