OpenAI, SoftBank, and Oracle are ahead of schedule on a $500B compute land grab, and the grid math is already compressing Bitcoin mining margins.
Key takeaways
OpenAI, SoftBank, and Oracle announced five new U.S. Stargate data center sites in September 2025, bringing the joint venture's total planned capacity to nearly 7 gigawatts and committed investment to over $400 billion, per the OpenAI blog. Oracle broke ground on a sixth site in Saline Township, Michigan on June 2, 2026, per CNBC. The project, formally announced January 21, 2025 at the White House, was originally structured around a $500 billion, 10-gigawatt target by 2029. As of the September 2025 announcement, OpenAI said it was "on a clear path" to hitting the full commitment by end of 2025, ahead of schedule, a milestone that has since passed without a confirmed public update as of June 2026.
The five September sites span Lordstown, Ohio (SoftBank); Milam County, Texas (SoftBank and SB Energy); Shackelford County, Texas (Oracle); Doña Ana County, New Mexico (Oracle); and a Wisconsin facility developed by Oracle in partnership with Vantage. The SoftBank two-site Ohio-Texas cluster alone is designed to scale to 1.5 gigawatts over 18 months. The OpenAI-Oracle partnership separately covers up to 4.5 gigawatts of additional capacity, with commitments exceeding $300 billion between the two companies over five years, per the OpenAI-Oracle partnership announcement. The flagship Abilene, Texas campus, 875 acres, eight AI factory halls, more than 400,000 NVIDIA GB200 GPUs planned, had its first two buildings operational as of September 2025. Buildings 3 and 4 remain behind schedule as of June 2026.
Stargate's ambitions are not abstract. They are landing on an already-strained grid with measurable force.
ERCOT's large-load interconnection process had more than 225 gigawatts of large loads going through it as of December 2025, per an ERCOT announcement, nearly quadruple the 63 gigawatts recorded at end of 2024. Data centers and AI companies now account for more than 70% of new large-load applications in Texas, per ERCOT's December 2025 board filing. Bitcoin miners, who once set the marginal price for power in that market, have been displaced as the dominant demand signal. Energy prices in Texas mining regions rose 15-20% in 2025 as AI demand absorbed available capacity.
The PJM capacity market tells a sharper version of the same story. Auction prices jumped from $28.92 per megawatt-day in the 2024/2025 cycle to $269.92 in the 2025/2026 cycle, an 833% increase in a single year, per PJM's 2026/2027 Base Residual Auction Report. Data center load growth was a key driver of those price increases, according to PJM.
These are settled auction results.
The structural driver is a revenue gap that miners cannot close by operating more efficiently. AI data centers generate substantially more revenue per kilowatt-hour than Bitcoin mining operations, giving AI operators the ability to bid for new grid connections at prices that make the economics of a marginal mining operation untenable.
Bitcoin mining economics were already stressed before Stargate accelerated the timeline. The energy shock now has a second engine running alongside geopolitical supply risk.
U.S. data center grid-power demand is expected to grow substantially through 2030 across forecasts. Large power transformers currently carry an average 128-week lead time. Gas turbines run up to a seven-year wait. The infrastructure to serve this demand does not exist yet, and the queue to build it is measured in years.
Three companies, backstopped by government sanction and sovereign wealth capital (Abu Dhabi's MGX is a founding partner), are building what amounts to a national compute monopoly. SoftBank holds financial responsibility; OpenAI holds operational responsibility. The sites were selected from over 300 proposals across more than 30 states. This is coordination at a scale that does not happen without state alignment.
Sam Altman framed it plainly in the five-sites announcement: "AI can only fulfill its promise if we build the compute to power it. That compute is the key to ensuring everyone can benefit from AI."
The contrast with Bitcoin's architecture is not rhetorical. Bitcoin's hashrate is distributed across thousands of operators globally with no single point of control over the protocol.
The falsifiable version of this thesis: if Bitcoin miners successfully operate as flexible load partners at all-in power costs below $0.04 per kilowatt-hour at scale, the displacement thesis softens. If Stargate's unfilled funding gap, the full $500 billion is not yet committed, and behind-the-meter builds are running at $19.2 billion per gigawatt versus $9-11 billion for grid-connected alternatives, leads to material project cancellations, power price pressure eases. Watch ERCOT queue approvals and Bitcoin hashrate distribution against all-in power costs each quarter.
The Michigan groundbreaking is the most recent proof that Stargate is moving concrete, not just capital. The next indicators that matter: whether Abilene buildings 3 and 4 come online in 2026, whether ERCOT approves the queued gigawatts sitting in interconnection requests, and whether any of the $500 billion commitment beyond the initial $100 billion gets formalized with named lenders. If AI capex cools materially, it is the most direct relief valve for mining power costs. If it doesn't, the structural repricing of U.S. grid power has years left to run.
Yes, and the numbers are already visible. ERCOT large-load requests jumped to more than 225 gigawatts in 2025, nearly quadrupling from end of 2024; AI and data center companies now account for more than 70% of new large-load applications in Texas; energy prices in key mining regions rose 15-20% in 2025 alone. AI data centers generate substantially more revenue per kilowatt-hour than Bitcoin mining, giving them bidding power miners cannot match for new grid connections.
The initial $100 billion is being deployed, and the remainder requires ongoing fundraising from banks and asset managers. SoftBank completed a $41 billion investment in OpenAI in December 2025, per the SoftBank Group press release, which addressed early skepticism about the project's viability, but the full $500 billion over four years remains contingent on capital markets and execution. Some builds are running behind schedule, and behind-the-meter construction costs are running nearly double grid-connected alternatives per gigawatt.
The Ohio site is in Lordstown, where SoftBank broke ground as part of the September 2025 five-site announcement. Per OpenAI's blog at the time, it was "on track to be operational" in 2026. No updated confirmed online date has been publicly announced as of June 2026.
