OpenAI Stargate: A Full Guide to the Abilene Campus, Sites, Partners, Power and Limits

OpenAI Stargate started as one giant number: $500 billion over four years for AI infrastructure in the United States, with $100 billion deployed immediately. Fifteen months later, OpenAI said it had already passed its 10 GW U.S. capacity goal. The largest dated, delivered figure anyone has published is far smaller: 618 MW across six buildings in Abilene, Texas.

That gap between contracted and delivered capacity is the real story of Stargate. This full guide covers what Stargate is today, the Abilene flagship, every announced U.S. site, how the money works, the NVIDIA, AMD and Broadcom chip deals, the power plans, the international projects and the honest limits, all as of October 18, 2026.

If you care most about the electricity side, keep Inside Deep Tech's behind-the-meter generation guide open alongside this one.

Last updated: October 18, 2026.

Key Takeaways

• OpenAI Stargate is now an umbrella for OpenAI's compute buildout, which OpenAI calls its overarching AI infrastructure platform, not a single joint venture.
• Abilene is the proof point: Oracle says six of eight buildings, 618 MW and 75% of the campus, were delivered by September 2026.
• OpenAI says it passed its 10 GW U.S. target in April 2026, but most of that capacity is contracted, not built.
• Chip deals span NVIDIA, AMD at 6 GW and Broadcom at 10 GW, and NVIDIA's $100 billion letter of intent became a $30 billion equity stake.
• Honest limit: OpenAI has dropped or paused Stargate UK, its Norway offtake and the Abilene expansion.

What OpenAI Stargate actually is in 2026

The January 2025 launch described Stargate as a new company, with SoftBank holding financial responsibility, OpenAI holding operational responsibility and Masayoshi Son as chairman. The initial equity funders were SoftBank, OpenAI, Oracle and MGX. The technology partners were Arm, Microsoft, NVIDIA, Oracle and OpenAI.

That structure didn't survive contact with the build. By September 2025, OpenAI was describing Stargate as its overarching AI infrastructure platform, and sites were being signed deal by deal: Oracle-developed campuses on one track, SoftBank and SB Energy campuses on another.

Here's why that matters. When a headline says "Stargate," it can mean an Oracle lease, a SoftBank-affiliated developer, a chip supply agreement or an OpenAI for Countries project abroad. Each one has a different counterparty, financing model and risk.

Who does what

• OpenAI: the anchor customer and the designer of the compute program. It signs leases and capacity contracts and pays as capacity comes online.
• Oracle: the cloud operator for Abilene, Shackelford County, New Mexico, Wisconsin and Michigan, under an agreement for up to 4.5 GW that exceeds $300 billion over five years.
• SoftBank and SB Energy: developers for Lordstown, Milam County and PORTS-Pike. OpenAI and SoftBank each invested $500 million in SB Energy in January 2026.
• Developers: Crusoe and Lancium in Abilene, Vantage in Shackelford County and Port Washington, and Related Digital in Michigan.
• Chip suppliers: NVIDIA, AMD and Broadcom, each with its own gigawatt-denominated agreement.

OpenAI Stargate by the numbers

Source: OpenAI's Stargate site announcement, OpenAI's April 2026 compute update, Oracle's fiscal Q1 2027 earnings call and DatacenterDynamics.

Read those rows carefully. The first four measure plans. Only the Abilene rows measure hardware a customer has actually accepted.

The Abilene flagship: Stargate's only campus at scale

Abilene is where Stargate became physical. Lancium owns the land and electrical infrastructure, Crusoe built and operates the data center buildings, and Crusoe's customer is Oracle, which runs the capacity on Oracle Cloud Infrastructure for OpenAI.

Crusoe's first phase was two 100 MW buildings energized in under one year, and the second phase adds six more buildings to reach 1.2 GW. Crusoe told Abilene officials the eight buildings cover about 4 million square feet, with roughly 9,000 craft workers on site every day and construction running through the first quarter of 2027.

The delivery pace is the headline. On its September 10, 2026 call, Oracle said it delivered 131,000 GPUs at Abilene in its fiscal first quarter, 1.9 times the prior quarter, and that six of eight buildings, 618 MW and 75% of total capacity, are now with the customer. Oracle added that customer acceptance has compressed to 24 hours.

That's fast by any standard.

What Abilene has trained

OpenAI says GPT-5.5 was trained at Abilene on NVIDIA GB200 systems, and Oracle says GPT-6 Astra, released in September 2026, was trained there too. That's the strongest evidence that Stargate capacity is doing frontier work, not just filling press releases. For how those racks are built, see Inside Deep Tech's GB200 NVL72 full guide.

Cooling and water

Abilene uses closed-loop liquid cooling. OpenAI says the one-time fill for each building is roughly two Olympic-sized swimming pools, and annual water use for the whole cooling system at full buildout should be comparable to about four average households. The data center liquid cooling full guide explains why sealed direct-to-chip loops make numbers like that possible.

The expansion that went to Microsoft

Abilene was supposed to get bigger. OpenAI's September 2025 update listed a potential 600 MW expansion near the flagship site. That plan was dropped in 2026, and Tom's Hardware reported failed financing talks and a late power date as the reasons.

Crusoe then announced a separate 900 MW campus next door for Microsoft, with two buildings of 336 MW of critical IT load each, an on-site power plant and first energization in mid-2027. The whole Abilene site is now projected at 2.1 GW, but only 1.2 GW of it is Stargate.

Every announced U.S. Stargate site and its status

Beyond Abilene, OpenAI has announced six more Stargate campuses with Oracle and SoftBank, plus two 2026 projects that extend the same compute program: PORTS-Pike in Ohio and Project Camellia in Georgia. Here's where each one stands as of October 18, 2026, based on company statements and the latest reporting.

Source: OpenAI site announcements, Vantage, OpenAI on PORTS-Pike, OpenAI on Project Camellia and TechCrunch on Project Jupiter. Capacity figures mix campus power, IT load and utility contracts, so don't add them together.

Add the rows up anyway and you get well past 10 GW. That's the number OpenAI cites. The dates tell a different story: outside Abilene, most first power lands in 2027 or 2028, and the two biggest blocks run to 2032.

Oracle confirmed as much on its September call: none of Shackelford, New Mexico, Wisconsin or Michigan delivered capacity in its fiscal first quarter, even though Oracle delivered 850 MW of AI capacity across its fleet in that quarter.

How OpenAI Stargate is financed

The $500 billion headline implied a single SoftBank-funded vehicle. In practice, the money flows through three channels.

1. Partner balance sheets and leases

Oracle's agreement with OpenAI covers up to 4.5 GW and exceeds $300 billion over five years. Oracle funds much of the build itself: in its fiscal first quarter it spent $28 billion on capex, posted negative free cash flow of $5 billion and completed a $20 billion at-the-market equity issuance.

SoftBank's track runs through SB Energy. At PORTS-Pike, SB Energy will build, own and operate the campus under a 20-year lease, OpenAI pays only as completed capacity becomes available, and NVIDIA will invest $1.5 billion in SB Energy plus credit support for the first 4.25 IT-GW.

2. OpenAI's own capital

OpenAI closed a $122 billion round at an $852 billion post-money valuation on March 31, 2026, anchored by Amazon at up to $50 billion, NVIDIA at $30 billion and SoftBank at $30 billion. It says it's generating $2 billion in revenue per month, but CNBC notes the company is still burning cash and isn't profitable.

3. Supplier financing and circular deals

Several deals send money in a loop. NVIDIA invests in OpenAI, and OpenAI buys NVIDIA systems. AMD granted OpenAI a warrant for up to 160 million AMD shares that vests as deployments scale. NVIDIA backs SB Energy, which leases PORTS-Pike to OpenAI, and that campus will exclusively host NVIDIA AI compute.

The chips: NVIDIA, AMD and Broadcom deals behind Stargate

OpenAI denominates its chip deals in gigawatts, not units. That's a useful tell: power, not silicon, is the binding constraint.

Source: OpenAI and NVIDIA, OpenAI and AMD, OpenAI and Broadcom, AMD's July 2026 update and OpenAI's Jalapeño announcement.

The NVIDIA letter of intent was never a contract. Jensen Huang said in March 2026 that the full $100 billion investment was probably not in the cards, and NVIDIA's $30 billion stake in the March round took its place. NVIDIA still supplies most of the fleet: OpenAI says its training fleet and the majority of its inference stack run on NVIDIA GPUs.

Oracle expects to deliver its first Vera Rubin systems to customers in its fiscal second quarter, which ends in November 2026. For the hardware itself, read Inside Deep Tech's Vera Rubin NVL72 full guide and the AMD Instinct MI350 guide, which covers the generation before MI450.

Inside Deep Tech's take: Vera Rubin, MI450 and Jalapeño all target the same late-2026 window. Expect at least one to slip, and expect that slip to show up as finished Stargate halls waiting for racks.

Inside a campus this size, the network is a bill-of-materials line of its own. Inside Deep Tech's optical circuit switching guide explains how Google rewires spines and TPU pods with light instead of another tier of packet switches.

Power: where Stargate's electricity comes from

Every Stargate site solves power differently, and power now sets the calendar more than chips do.

• Abilene: grid power through Lancium's land and electrical development. The Microsoft campus next door adds a 900 MW on-site plant paired with battery storage.
• New Mexico: Bloom Energy fuel cells running on gas, with the supply pipeline delayed to February 1, 2027 and an air permit decision due by November 23.
• Michigan: DTE Energy serves the campus using existing excess transmission capacity, with any upgrades funded by the project.
• Wisconsin: new solar, wind and battery capacity, with WEC Energy Group adding nearly two gigawatts of new energy, 30% of it for other customers.
• PORTS-Pike: the first 800 MW largely uses existing AEP infrastructure, and the rest needs new natural gas generation and transmission.
• Georgia: 3.2 GW from Georgia Power delivered in phases from 2028 to 2032, with a design that cuts load during grid peaks.

The pattern is clear. The fastest sites reuse existing grid capacity. The biggest sites need new generation, which means air permits, pipelines and utility commission approvals. Inside Deep Tech's behind-the-meter generation guide compares gas turbines, fuel cells, batteries and nuclear for exactly this problem.

Further out, fusion is the longest-dated firm-power option for AI campuses. The Helion Energy full guide covers the Microsoft power purchase agreement and the Orion plant targeting initial operations in 2028, and the Oklo Aurora guide covers advanced fission on a similar horizon.

Stargate international: UAE, Norway, UK and what's left

OpenAI exported the brand through its OpenAI for Countries program. The results so far are mixed.

• UAE: a 1 GW cluster in Abu Dhabi with G42, Oracle, NVIDIA, Cisco and SoftBank, with 200 MW expected to go live in 2026. Inside Deep Tech found no public confirmation that the first phase was live as of October 18, 2026.
• Norway: planned at 230 MW with 100,000 NVIDIA GPUs targeted by the end of 2026. In April 2026, OpenAI dropped its direct offtake, and Microsoft took the capacity with more than 30,000 NVIDIA Rubin GPUs.
• UK: OpenAI paused Stargate UK in April 2026, citing energy costs and regulation; its commitment had been to explore the offtake of 8,000 NVIDIA chips.

International Stargate depends on cheap power and a committed anchor customer. Where either was missing, OpenAI chose to rent from Microsoft or wait.

Announced vs built: the honest limits of OpenAI Stargate

First, capacity claims mix units. OpenAI's 10 GW figure counts planned capacity across campus power, IT load and utility contracts. Oracle's 618 MW at Abilene is accepted capacity. They aren't the same thing.

Second, schedules are wobbling at the edges. Oracle sent a force majeure notice on Project Jupiter in September 2026, which would let it defer payments if the site misses 2028, while saying the project remains on schedule.

Third, OpenAI is retreating where the economics don't work. It walked away from the Abilene expansion, its Norway offtake and, for now, Stargate UK.

Fourth, leadership keeps turning over. In October 2026, OpenAI put former xAI executive Uday Ruddarraju in charge of Stargate, replacing Sachin Katti, who took over the effort in March. DCD reports that in-house data center projects remain slow to come online.

Fifth, the bill is enormous relative to revenue. OpenAI projects about $750 billion of compute spending through 2030, against about $2 billion of monthly revenue as of March 2026. That's a bet on growth, not a funded plan.

Sixth, local opposition is real. Project Jupiter has drawn opposition from residents and environmental groups, and new sites now arrive with water, rate and tax pledges, like OpenAI's $80 million community commitment in Georgia.

Inside Deep Tech's take

Who should care most? If you sell power equipment, cooling, construction services or GPUs, Stargate's contracted pipeline is real demand, but it pays on delivery dates that can move. If you're an investor, watch accepted megawatts and lease start dates, not announced gigawatts.

What to watch next

Five checkpoints will tell you whether Stargate stays on track:

• Oracle's fiscal Q2 2027 results, scheduled for December 14, 2026, and any update on the last two Abilene buildings and Shackelford.
• New Mexico's air permit decision for Project Jupiter's fuel cells, due by November 23.
• First production racks of Vera Rubin, AMD Helios and Jalapeño at OpenAI sites.
• Whether the first 200 MW of Stargate UAE goes live before year end.
• Progress toward the first 800 MW at PORTS-Pike in 2028.

FAQ

The bottom line

The next 12 months decide whether Stargate becomes a fleet or stays a flagship. If you're planning power, cooling or capacity around AI campuses, watch Shackelford and Milam in 2027 more closely than any new gigawatt announcement.

For the power side of that question, start with Inside Deep Tech's behind-the-meter generation guide.

Behind-the-Meter Generation for AI Data Centers: A Full Guide

Gas turbines, fuel cells, batteries and SMRs for AI campuses, the interconnection queue reality, and what ships in 2026 to 2027 versus later.

Inside Deep Tech