Meta AI Infrastructure, October 2026: Agent Compute, Hyperion, and an 87% Capex Surge

Meta's capital expenditure reached $69.69 billion in FY2025 — up 87.1% year-on-year and equal to 35% of revenue — as the company builds dedicated compute for AI agents, advances a $50 billion Louisiana campus, and targets proprietary silicon by mid-2027.

When OpenAI, Meta, and xAI began carving out dedicated computing infrastructure for AI agents — rather than routing agent workloads through shared inference pools — the move signalled something more than an engineering preference. For Meta specifically, it marks the latest phase of a capital programme whose scale has outpaced even optimistic forecasts. Meta's Muse agent reportedly surpassed 2.8 million downloads within two weeks of launch, a rate reported to have surpassed ChatGPT's 2023 launch pace. The architecture reveals the commercial logic plainly: each Muse user receives an isolated sandbox on AMD EPYC Turin processors, provisioned with two virtual cores and 8 GB of memory. According to Meta's own engineers, recommender-feed workloads deliver roughly a 100-times cost advantage over generative-chat inference at comparable scale. The implication is that the next major wave of AI compute is being sized not around model training but around the lighter, more numerous, always-on demands of agents — and Meta has already built a product that illustrates the trade-off in production.

The infrastructure assembled to support that thesis is extraordinary in scope. Meta's capital expenditure reached $69.69 billion in FY2025, an 87.1% increase from $37.26 billion in FY2024, equivalent to 35% of the company's $200.97 billion in revenue for the year; among the seven hyperscalers Slicast tracks, that intensity ratio ranks third, behind only Oracle. The largest single quarter on record is the three months ended 31 March 2026, at $19.00 billion. Goldman Sachs has projected that Amazon, Alphabet, Microsoft, Oracle, and Meta will together invest $1.2 trillion in AI infrastructure by 2027 — a 50% rise from 2026 — calling the cycle the largest since 19th-century railway expansion. At the centre of Meta's physical build is Hyperion, a Louisiana campus that reportedly expanded to a 5 GW capacity target at an estimated $50 billion, with full completion scheduled for 2032. Alongside it sit a reported $6.6 billion, 20-year lease agreement with CleanSpark for a Georgia facility, a CoreWeave contract reportedly grown to $21 billion, and a $1.2 billion campus in Kuna, Idaho, which went live in September 2026.

Energy procurement has become both a strategic constraint and a source of friction. Meta has signed a 20-year power-purchase agreement with Vistra covering more than 2,600 MW from three nuclear plants, and a separate agreement reportedly backing an Illinois nuclear facility. A deal with TerraPower covering eight Natrium advanced nuclear reactors has also been reported. On the renewable side, a 144 MW solar PPA with Apex Clean Energy in Texas and a broader Texas solar purchase have been disclosed; Inside Climate News noted, however, that Meta is simultaneously powering some AI data centres with natural gas — a divergence the company has not fully addressed publicly. The geographic footprint extends further: Meta has contracted AI-computing capacity at Firmus data centres in Southeast Asia, extending an existing Australian arrangement, while reportedly joining Google and Apple in a combined $25 billion infrastructure programme in Australia. The global network layer is anchored by Petal, a newly announced transatlantic undersea cable spanning some 4,300 miles between the United States and France with petabit-scale raw bandwidth capacity.

The financial picture is drawing scrutiny from multiple angles. Analysts at CFO Dive have identified Meta as one of the Hyper 5 hyperscalers subject to rigorous cash-flow examination given the pace of capital deployment. A reported tax approach — classifying AI data-centre assets as experimental equipment to reduce federal obligations — has attracted attention, and multiple US states have moved to reclaim incentives previously granted to Meta, Amazon, and Google. On the silicon front, Meta reportedly plans to deploy its proprietary Arke and Astrid chips in 2027 for inference workloads — with Arke targeted for the first half of the year — with the stated goal of reducing dependence on Nvidia GPUs. Benzinga analysis noted that should hyperscaler capex slow, Nvidia would absorb the demand shock before Meta. JPMorgan upgraded Meta's equity rating in September. Separately, Anthropic was reportedly recruiting sales staff at packages reaching $3.2 million specifically to serve the Meta relationship — a signal that Meta has become a material customer of external model providers even while constructing its own silicon stack.

The risk profile is not straightforward. One analyst argued that Muse compute may be structured as a developer subsidy rather than a revenue-generating product — a concern that, if validated at scale, would complicate the return-on-investment case for commitments of Hyperion's magnitude. Talent dynamics add a further variable: network chief Ola Kasali departed to Groq, and a former OpenAI Stargate executive spent only a brief period in Meta's Compute organisation before joining Anthropic. Meta's share closed 11.3% higher in the 21 September 2026 session; no coverage item explains the move. Three concrete signals will test the thesis over the coming quarters: whether Arke and Astrid achieve production deployment on their stated 2027 schedules; whether capex intensity holds, falls, or continues to climb in the next reported fiscal period; and whether Meta's energy portfolio — nuclear, solar, and gas — can be consolidated into a coherent and publicly defensible power strategy before regulatory and investor scrutiny intensifies.