AI agents · OpenClaw · self-hosting · automation

Quick Answer

Meta Hyperion $50B Louisiana Expansion: What It Means (2026)

Published:

Meta Hyperion 5GW Louisiana Expansion: What It Means (July 2026)

On July 13, 2026, Meta announced that its Hyperion AI training campus in Richland Parish, Louisiana will expand to 5 gigawatts of compute capacity at a cost of $50 billion+ — up from a $27 billion / 2GW plan earlier in 2026 and the original $10 billion / 2GW announcement in December 2024. That is a 5× cost jump in 18 months and puts Hyperion just behind Stargate as the largest AI infrastructure buildout on Earth.

Here is what the expansion means, how it stacks up against Stargate and Colossus 2, and what it signals about the frontier-lab compute arms race.

Last verified: July 15, 2026

The Numbers

Per the July 13, 2026 announcements from Meta, Louisiana Economic Development, CNBC, and Fortune:

MetricOriginal (Dec 2024)Q1 2026 updateJuly 13, 2026
Power capacity2 GW2 GW5 GW
Total investment$10 billion$27 billion$50+ billion
Campus footprint~4M sq ft~4M sq ft~10M sq ft
First phase online202620272027
Full ramp20282028-20292029-2030

That is a 2.5× capacity expansion at 5× the cost — reflecting Blackwell-generation GPU pricing, Meta’s own MTIA silicon integration, and the cost of self-funding the 5GW of power infrastructure (new gas turbines, transmission upgrades, and grid interconnect).

How Hyperion Compares

The four biggest single-project AI campuses being built as of July 2026:

ProjectOperatorLocationCapacityEst. costFirst online
StargateOpenAI + SoftBank + OracleAbilene TX + more10+ GW total~$500B over 4 yrsPhase 1 late 2026
HyperionMetaRichland Parish LA5 GW$50B+2027
Colossus 2xAIMemphis TN~2 GW target~$18BAlready scaling
TeraWulf/AnthropicAnthropic + Google Cloud + TeraWulfMultiple sites~700 MW~$5B2027

Meta is second only to Stargate on single-project scale — but Meta has a real completion advantage. Stargate is a multi-site portfolio spread across Abilene, Wisconsin, and future sites; Hyperion is one campus in one place, which is easier to power, cool, and network than a federated buildout.

What Meta Trains on Hyperion

The July 13 announcement confirmed Hyperion will host:

  • Muse Spark model family — Meta’s post-Llama frontier model line, currently a mid-tier competitor to Claude Sonnet 5, GPT-5.6 Terra, and Gemini 3.5 Flash
  • Muse Video 2 — successor to the July 2026 Muse Video release, competing with Sora 2 Pro, Runway Gen-4, and Google’s Gemini Omni video models
  • Meta AI recommendation and ranking — the always-hungry consumer AI workload behind Instagram, Facebook, WhatsApp, and Threads
  • Future frontier training runs — Meta has publicly stated Hyperion will train “the largest single-cluster AI training run ever” by mid-2028

The AI Andreessen Horowitz quipped after the July 13 announcement: “Meta is building a chip factory for models, not a data center.” That is exactly right — Hyperion is designed as a training-first campus, not a hyperscaler general-purpose facility.

Why the Cost Jumped 5×

  1. Blackwell (B200) pricing — B200 systems cost 2-3× per watt vs. H200-era hardware. Meta signed for hundreds of thousands of B200-class accelerators plus its own MTIA v3.
  2. Power self-funding — 5GW is more than the entire electricity load of ~4 million US homes. Meta is co-funding new gas turbines (with utility Entergy), transmission upgrades, and dedicated substations. In 2024 those items were expected to be utility-funded; in 2026 they are on Meta’s balance sheet.
  3. Land, water, cooling — 10M sq ft of campus footprint, new water infrastructure for direct-liquid cooling at 5GW, and community mitigation payments (schools, roads) are a nine-figure addition each.
  4. Scope expansion — the original 2GW plan was for training only. The 5GW plan integrates on-campus inference for consumer Meta AI traffic, which adds networking, edge cache, and product-side infra.

Fortune reported on July 13, 2026 that “the town of Holly Ridge is split in two” over the project — a preview of the political fight AI data-center buildouts face across US markets in 2026-2027.

What the Expansion Signals

Meta is fully committed to being a frontier training operator. After the Llama-to-Muse-Spark pivot and the muddled Muse Image launch (Instagram AI image tool killed after 3 days in July 2026), Hyperion is Meta’s way of saying “we will out-spend our way to relevance.” The MSI (Meta Superintelligence Initiative), announced June 2026, gets its physical home here.

The frontier is capital-intensive, not just algorithmic. OpenAI has Stargate. Meta has Hyperion. xAI has Colossus 2. Anthropic has TeraWulf. Google has already-built TPU capacity plus new campuses. Every US frontier lab now needs a multi-gigawatt training campus or a similar-scale leased footprint. This is the moment frontier AI became a heavy-industry business.

Louisiana becomes the AI capital of the South. Richland Parish + Meta’s earlier Rebecca campus + Google’s planned Louisiana site put ~7-8GW of AI capacity in one state — more than most countries. Expect tax-incentive politics to become a national debate: Louisiana granted Meta ~$1.5B in tax breaks for Hyperion, per Fortune.

Nvidia gets ~$25-30B of the $50B. At B200-class pricing, roughly half the Hyperion budget is chips. That means one Meta project alone is a mid-single-digit percentage of Nvidia’s annual data-center revenue. Nvidia’s grip on frontier training is not loosening — Meta’s MTIA integration is real but complementary, not replacement.

What to Watch Next

  1. 2027 first-phase power-on — first 1GW online is the credibility milestone
  2. Meta MTIA v3 share — how much of Hyperion is Nvidia B200 vs. Meta silicon
  3. The next lab to announce a 5GW+ single campus — Google, Microsoft/OpenAI-Stargate expansion, or Anthropic
  4. Local politics — Fortune’s “split town” report is a warning sign; expect environmental and community pushback to scale with the campus footprint

Sources