Meta raises AI capex to $135-145B as NVIDIA's Rubin goes live in the cloud

Meta goes all in

Meta's third quarter told the buildout story in one earnings report. The company posted record quarterly revenue of $64.2 billion, up 26% year over year on advertising pricing power, and raised its fiscal year 2026 capital expenditure guidance to a range of $135 billion to $145 billion. A $9.9 billion legal charge cut earnings per share to $3.57. Meta will also introduce non-voting Class C stock effective January 2027 while Mark Zuckerberg maintains 99.8% of the voting power through his Class B shares, and it expects Reality Labs losses of $18 billion to $19 billion in 2026. m:…4362

The cost of that posture is showing up on the balance sheet. CNBC's Fast Money noted Meta is experiencing its first period of negative free cash flow in a decade, even as it raised the low end of its full-year capex range. The company plans to monetize the 'Muse' AI platform by taking a cut of user transactions, and expects to improve compute cost efficiency more than tenfold over the next year - a forecast the scale of the spend makes necessary rather than optional. m:…8230

Rubin goes GA

While Meta spends, the hardware it buys is moving from announcement to availability. Google Cloud and NVIDIA announced general availability of NVIDIA Vera Rubin infrastructure: A5X instances accelerated by Vera Rubin NVL72 systems, paired with ConnectX-9 SuperNICs and the Virgo network, engineered for large-scale agentic inference and reinforcement learning. Google Cloud says it plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. m:…8235

The early users are named. Ineffable Intelligence, the London-based technology company founded by AlphaGo creator David Silver, was the first customer to access A5X instances. WPP demonstrated a creative-agent reference architecture on them - in the companies' initial tests, workloads that typically required a multi-node cluster ran on a single A5X instance - using NVIDIA Nemotron models and NVIDIA OpenShell for its creative agents. m:…8235

The neoclouds deploy

CoreWeave is putting Rubin into its cloud this quarter with the NVIDIA HGX Rubin NVL8 platform: an eight-GPU node integrating eight Rubin SXM GPUs on a single baseboard, with 2TB of HBM4 and sixth-generation NVLink. NVIDIA indicates the system delivers up to 10x higher token throughput than previous architectures while matching training performance using 4x fewer GPUs, the company says, and the node delivers 265 TFLOPS of FP64 double-precision performance for scientific simulation and computational finance workloads. m:…5547

CoreWeave says it completed the industry's first bring-up, validation and benchmarking of Vera Rubin NVL72, and that customers are already seeing results: Cognition achieved 4.8x higher SWE-2 inference throughput and 3.8x faster output token speed for RL on the platform. Those are vendor-cited benchmarks, not independent tests - but they are the first third-party numbers attached to Rubin in production. m:…5547

The read

Our read for operators: the supply side of the AI buildout is now shipping at a cadence the demand side has to absorb. Meta is burning cash - its first negative-free-cash-flow period in a decade - to stay in the race, while its primary hardware supplier moves a full platform generation into production and the neoclouds get first deployment. If Rubin's claimed throughput holds in customer hands, the cost-per-token math for agentic inference changes this quarter, not next year. m:…8230 m:…8235 m:…5547