Meta’s 5GW AI Supercluster Is a Utility-Scale Reckoning — And It’s Only Getting Started

It’s an easy question to ask, and an almost impossible one to answer: What does a 5-gigawatt AI cluster even look like?

Not the spreadsheet numbers, not the slide decks. The physical thing. The humming, heat-spewing, mile-long concrete beast that will suck more power than some small nations and turn Richland Parish, Louisiana — population 20,000 — into the unlikely ground zero of the AI arms race.

On July 13, 2026, Meta dropped the news: its Hyperion facility is expanding to 5GW, with total infrastructure investment exceeding $50 billion. The chip bill? Maybe another $200 billion, according to Bloomberg, if you count all the processors and networking gear needed to fill those halls. Suddenly, the largest corporate AI bet in history got a whole lot bigger.

A 5-gigawatt campus consumes electricity on a utility-wide scale. Entergy Louisiana, Meta’s partner, is building 10 new power plants specifically for Hyperion — seven natural gas, three grid-scale batteries — plus 1.5GW of solar and storage, 240 miles of transmission lines, and nuclear capacity upgrades. The combined generation will exceed 7GW, more than 30% of the state’s entire grid.

To put a finer point on it: just one of Hyperion’s buildings will occupy a chunk of land comparable to a large slice of Manhattan, and the full campus spans nearly 10 million square feet. When you hear “supercluster,” you should picture a self-contained industrial city optimized for one thing: turning electrons into tensor cores, nonstop.

All of this is paid for by Meta. Under a “Fair Share Plus” agreement, the company covers every dollar of the new generation, transmission, and water infrastructure. Entergy projects $2 billion in customer savings over 20 years from the arrangement. But try telling that to the residents who’ve seen their rent triple since construction began.

Erika James, a 34-year-old mother of two living in a mobile home park about 30 minutes from the site, told Fortune, “We are now having to entertain the idea of leaving the area completely. There is nowhere to go if you can’t pay triple prices.” Her story is echoed across the parish, where median incomes hover around $42,000 and monthly rents have jumped from $600 to $2,500 in some cases. Housing prices surged 52% year-over-year, according to Redfin, against a national average of 1.2%.

Two Americas, One Supercluster

The economic split is almost too tidy. Teachers in Richland Parish received bonuses of up to $50,000 — about four times the prior year’s level — fueled by a temporary 1% sales tax windfall from the construction boom. Scott Holmes, who runs a local charter bus company, expanded from 40 coaches to 102; his drivers are earning $80,000-plus annually. “Nothing compares to what Meta’s project has meant for us,” he said.

But for every Holmes, there’s a family staring at eviction notices. Traffic is punishing. Water usage for Hyperion is projected at 23 million gallons per day, nearly doubling the parish’s existing draw from local aquifers. Earthjustice filed to investigate the financing; community monitoring projects are already tracking air and water quality. In a place where farming is still the economic backbone, the idea of a data center competing for water with crops doesn’t sit well.

A local observer captured the contradiction: “Meanwhile, there is literally a sign outside welcoming Meta workers while local families are left wondering where they’re supposed to go.”

The Silicon Math: $250 Billion and a Custom Chip Gamble

The $50 billion construction number doesn’t include the chips. When you factor in GPUs, networking gear, and custom silicon, the total could reach $250 billion — a number that makes even hyperscaler CFOs blink. Sherwood News reported that Meta will spend roughly $35 billion on GPUs for Hyperion alone.

But Meta isn’t betting solely on Nvidia. In March 2026, it unveiled the MTIA 500, an inference-optimized accelerator built in-house. Specs: 30 petaflops of MX4 performance, 27.6 TB/s HBM bandwidth, up to 512GB of HBM capacity, and a chiplet architecture with four compute dies plus two networking dies. It’s twice as fast as its predecessor, though it also draws 40% more power. By 2027, Meta plans to deploy four MTIA generations, and in September 2026 it will begin mass-producing its “Iris” custom AI chip.

This dual-sourcing approach — Nvidia and AMD for general workloads, MTIA for tailored inference — is designed to keep costs manageable at 5GW scale. Still, an analyst at SemiAnalysis summed up the skepticism: “You can build the biggest cluster in the world, but unless you have workloads that actually utilize that capacity economically, it’s just a very expensive heater.”

Paying for It: Off-Balance-Sheet Structures and the Blue Owl SPV

How does a company already guiding $125–145 billion in 2026 capex absorb a $50 billion project without scaring bondholders? Syndication. In October 2025, Meta sold 80% of Hyperion to a Blue Owl Capital vehicle called “Beignet” for about $7 billion in cash, keeping 20% equity and a long-term lease with a four-year exit clause. BlackRock bought over $3 billion in bonds issued through Morgan Stanley. It’s an elegant bit of financial engineering that keeps the debt off Meta’s balance sheet while Blue Owl earns steady yields.

Lately, Meta has hinted at converting this infrastructure into a revenue engine. It’s building “Meta Compute,” a cloud business that will sell raw AI compute and model hosting — think AWS Bedrock but for Llama. The internal target: $10–15 billion in annual revenue by end of 2027. When Bloomberg reported the plan, Meta’s stock jumped 10% in a day. For the first time, there’s a tangible monetization path for all those GPUs.

Cooling a Beast and Keeping the Lights On

5GW of compute means roughly 5GW of heat. Meta is deploying air-assisted liquid cooling racks, custom substations, and likely waste heat recovery systems. But the real headache is grid stability. Researchers have found that hyperscale AI data centers introduce novel instabilities — inverter-interfaced, voltage-sensitive electronic loads that can swing millions of watts in milliseconds. PJM, one of the largest U.S. grid operators, has said its supply situation is “precarious.”

Meta’s answer? Build your own grid. The 10 new plants, battery storage, and dedicated transmission effectively decouple Hyperion from the public grid’s worst fluctuations. But it also means Meta is becoming a quasi-utility — applying for energy trading licenses, signing demand-response agreements with utilities in Indiana and Tennessee, and shifting training workloads to off-peak hours. As one energy analyst put it, “When your data center consumes more power than the entire city of New Orleans, you stop being a customer and start being a system operator.”

Competitive Omens: Stargate, xAI, and the Race to Win

Hyperion doesn’t exist in a vacuum. OpenAI’s Stargate initiative has broken ground on 10 data centers in Texas, with five more planned. xAI’s Memphis cluster, Colossus, hit 200,000 GPUs in 122 days — Elon Musk’s trademark speed — and is scaling past 1.2GW. But neither matches Hyperion’s single-site density.

Yet size isn’t destiny. A commenter on a data center forum noted, “Building a 5GW cluster is one thing; filling it with workloads that generate more value than the cost of capital is another.” The AI model market is already showing signs of commoditization. OpenAI and Anthropic have had to discount compute in some offerings. Meta’s own CFO Susan Li warned that the company expects a significant capex ramp in 2026, and unlike cloud providers, it doesn’t have an existing customer base paying by the hour.

In an analyst briefing, someone asked Mark Zuckerberg about the return on this investment. He replied, “Some of the offers we get to rent out this compute are so high that it might make more sense to do that than to use it ourselves.” It’s a pragmatic admission that even the largest internal workloads — Llama training, real-time translation, recommendation models — might not soak up 5GW.

The Unspoken Alliance: Broadcom, AMD, and the Chip Story Few Saw Coming

April 2026 brought another twist: Meta and Broadcom revealed a multi-year partnership to co-develop custom AI silicon, spanning chip design, advanced packaging, and networking. The deal covers multiple XPU designs. This means Meta’s chip ambitions go beyond MTIA — Broadcom’s expertise will help Meta build chips tailored to its exact internal workflows, potentially saving billions compared to off-the-shelf GPUs.

Will we see a Meta-branded GPU for sale? Probably not. But developers on GitHub are already eyeing the MTIA software stack and wondering if custom kernels could give them a performance edge if they ever get access to Meta Compute. As one contributor on a related repository mused, “If Meta opens up MTIA instances, even with a weird toolchain, the cost per inference could be a fraction of A100s.”

Louisiana’s Tax Gamble

None of this would be happening without an aggressive tax play. Governor Jeff Landry signed a 20-year sales tax exemption for data centers built before 2029. With Louisiana’s 9.56% combined state and local rate, the exemption is worth over $3.3 billion just on the $35 billion GPU spend — more than the state’s entire police budget for seven years. Critics call it a giveaway; supporters, including Landry, argue that the overall fiscal impact is positive once jobs and infrastructure improvements are counted.

Richland Parish expects to net $30 million in property taxes from Meta over three years, against a prior annual total of $22 million. The $5 million donation to Louisiana Delta Community College for data center training and $200 million in local infrastructure improvements are tangible. But the sales tax exemption means the windfall has a timer. When construction ebbs, so do the temporary spikes.

What Comes Next?

Hyperion won’t be a single flip-of-the-switch event. It is expected to hit 2GW by 2030, with full 5GW capacity around 2032. By then, Meta will likely have multiple other gigawatt-scale clusters online — Prometheus in Ohio is set to be the first 1GW+ cluster this year. The AI supercluster era is already here; Hyperion is just its most extreme expression.

The unanswered question isn’t whether Meta can physically build it. The company has 32 data centers operating or under construction, the cash flow to fund it, and the engineering talent to deliver. The real question is whether the economics of AI will justify the sheer scale. If the bubble bursts, Hyperion could become the world’s most expensive stranded asset. If it works, it cements Meta as the central utility of the AI age.

As a Reddit user put it: “We’re watching a company bet its entire future on the idea that more compute equals more intelligence. Either that’s true, or they just built a monument to a hype cycle.” The answer won’t come from a press release. It will come from the hum of those first buildings, the water running through the cooling towers, and the quarterly earnings that reveal whether 5GW of silicon is an engine of value or a cautionary tale.