181 GW of Gas for Data Centers – How Much of It Is Just a Place in Line?

What this is about

On November 11, 2026, BloombergNEF (BNEF), Bloomberg's research arm, published a number that made headlines instantly: 181 gigawatts of gas power plants are slated to go up right next to data centers. According to the Guardian, that's almost four times the UK's entire peak demand. A day earlier, SoftBank had reported a loss of ¥597.2 billion (about $3.7 billion). According to the Wall Street Journal, it stems mainly from gas turbines that SoftBank reserved and then let go.

The two stories belong together. Our thesis: a large share of the “gas wave” for AI data centers isn't power plants at all, but options. Companies use them to lock in a place in line for turbines and grid connections. That distorts emissions forecasts and turns many buildout plans into a bet. We lay out the numbers stage by stage.

Stage 1: Announcements – the number grows faster than reality

BNEF itself says the 181 GW are “a count of announcements, not a forecast that these facilities will be built.” The projects are at very different stages, and the figure should only be read as an “indication of the potential scale.” The report is titled Powering the Data Center Boom and was commissioned by Bloomberg Philanthropies.

It gets interesting when you put BNEF's earlier counts next to it:

So in eight months the list grew by roughly 60 percent. Measured against the 2 GW actually running in March, that's the factor of 57 Bloomberg Law reported at the time. Another number matters more: in July, BNEF could assign a start date to only just over half the projects (about 69 GW). The press release doesn't say whether that share is similar for the new 181 GW.

The counter-calculation also comes from BNEF. So far, the US grid has connected at most about 10 GW of new data center load per year. If that doesn't speed up, roughly 48 GW of on-site gas would need to be built by 2035 to cover BNEF's base forecast for data center demand (194 GW in the US). In other words, announcements are already almost four times what BNEF itself thinks is needed.

Stage 2: Slot reservations – paid places in line

If you want a large gas turbine, you don't just buy one. First you sign a slot reservation agreement: in exchange for a deposit, you secure a manufacturing slot in a specific year. The firm order with all the details comes later, or never. Manufacturers report these reservations separately from firm orders, which shows you how much of the total is pure optionality:

Until mid-year, then, more than half of GE Vernova's gas pipeline consisted of reservations. Only in the third quarter were 13 GW converted into orders. CEO Scott Strazik called it an “important inflection point” that the firm backlog now exceeds reservations. Siemens Energy expects its 39 GW to turn into firm orders in fiscal 2027, “at more attractive margins.” Slots can be reassigned if a project misses the agreed timeframe.

A rough cross-check (our math): about 20 percent of GE Vernova's 122 GW works out to roughly 24 GW for data centers, reservations included. Even if you generously add Siemens and Mitsubishi, that's far below 181 GW. The comparison is incomplete, though. Many on-site plants rely on gas engines or smaller turbines from other suppliers. BNEF has published a separate note on this titled “Engines Beat Turbines.”

Then there are lead times. GE Vernova has already sold slots for 2032 and expects slots through 2030 to be nearly sold out by the end of 2026. According to Bloomberg, Mitsubishi Power is talking about deliveries “around 2031,” with 2028 and 2029 almost full. Prices are rising too: at GE Vernova, gas orders in 2026 cost about 50 percent more per kilowatt than in 2025.

The SoftBank case: what happens when you book too many slots

SoftBank shows how this system works in practice. According to the WSJ, the company at its peak had reserved more than 13 GW of turbines, the equivalent of six Hoover Dams, mostly with GE Vernova. The centerpiece is a 9.2 GW gas plant for the PORTS Pike campus in Ohio, an 8 GW data center leased to OpenAI. Even after subtracting Ohio and other sites, SoftBank had reserved far more turbines than it had projects. When further payments came due, the WSJ reports, SoftBank dropped the surplus reservations and also unwound contracts that were further along.

The numbers from the quarterly report, broken down based on reporting by mlq.ai and an analysis on Substack:

• Paid to date to the project company Silver Bands 7 for equipment: ¥1,500.9 billion ($9.48 billion)
• Total loss: ¥597.2 billion, including ¥457 billion in forfeited, non-refundable deposits (about $2.9 billion) and ¥140.2 billion for contract modifications and settlements
• Still on the balance sheet: about $6.6 billion in deposits expected to flow back via project financing
• As a consolation prize: discounts of up to $2.1 billion on future purchases over ten years, subject to conditions

On November 6, Elon Musk claimed on X that SoftBank was “in default” on its turbine orders. He offered no evidence. SoftBank rejected this as “simply not accurate at all.” CFO Yoshimitsu Goto said canceled orders are not a default. Ohio won't be scaled back, he said, but will be “thoroughly reassessed,” including equipment procurement. The company has to “balance the accelerator and the brake.”

GE Vernova said the news has “no negative or positive financial impact.” That's plausible, since deposits on orders and reservations are non-refundable, according to the company. JPMorgan noted that freed-up slots could go to waiting customers. As recently as October 28, when asked by a Morgan Stanley analyst, Strazik said no customer wanted to back out of reservations. SoftBank hasn't disclosed exactly which manufacturers are affected.

Our take: for manufacturers, the system is nearly risk-free. Anyone who walks away loses the deposit, and the slot gets resold. For buyers, a reservation is an expensive option, and that's exactly why they get booked in abundance while money is cheap. In the third quarter alone, $5.3 billion of GE Vernova's cash inflow came mainly from such deposits.

Stage 3: What this means for the emissions forecast

BNEF runs the numbers: if all announced US plants were built and operated under BNEF's assumptions, they would emit about 345 million t CO2e (CO2 equivalents, meaning all greenhouse gases converted to CO2) per year. That's 152 million t more than the same computing power would cause using average US grid electricity. According to BNEF, the 152 million t equal about 10 percent of US power sector emissions in 2025.

What you can work out from that: - On grid power, emissions would be around 193 million t. On-site gas would therefore be about 1.8 times as harmful to the climate. - The 345 million t would equal just under a quarter of US power sector emissions (about 1,485 million t, according to the EIA). - Scaling roughly proportionally (our back-of-the-envelope estimate, not BNEF's): if only the 48 GW BNEF needs for its base case were built, you'd be looking at around 100 million t.

In August, Bloomberg had arrived at 318 million t for 126 GW, assuming 60 to 100 percent utilization with simple-cycle gas turbines. So the projection grows with the length of the announcement list. The real risk lies elsewhere: according to BNEF, plants meant only as a bridge until a grid connection arrives could keep running afterward, because generating your own power pays off. That's called lock-in, committing to fossil infrastructure for decades.

Europe: under ten percent, but concrete

The UK and continental Europe together account for less than 10 percent, so at most around 18 GW. BNEF doesn't publish a country breakdown in the accessible materials. Individual projects are known:

• UK: Wapseys Wood in Buckinghamshire plans its own 270 to 350 MW gas turbine plant for up to 300 MW of IT load. The government called in the application in March. There's also Quest Park in Bedfordshire. For the two sites combined, the NGO Foxglove estimates more than 4.5 million t of CO2; the government calls the assumption of full load from day one “misleading.” According to Future Energy Networks, there were more than 100 requests from data centers for gas connections in 2024/25. The “Data Centre Covenant” of October 28 sets out a narrowly limited path to temporary on-site generation for “exceptionally strategic” projects. What counts as strategic is still open.
• Germany: According to Global Energy Monitor (cited by cleanthinking.de), up to 1,950 MW of planned gas capacity is tied to data centers. EdgeConneX plans 170 MW in Maintal near Frankfurt and isn't due to get a grid connection until 2037. CyrusOne and E.ON plan 61 MW in Frankfurt-Sossenheim, and in Birstein a power plant is set to supply a 200 MW data center. According to JLL, larger facilities in the Rhine-Main region wait up to 13 years for a grid connection.
• Ireland: Here, on-site generation is actually mandatory. Since the CRU decision of December 2025, data centers of 10 MVA and up need dispatchable on-site generation or storage covering their full connection capacity. According to energy researcher Hannah Daly, an approved project in Ennis is set to get about two-thirds of its power from on-site gas.

Our verdict

The 181 GW isn't a made-up number. But it describes the crowd at the door more than what's actually getting built. Between announcement and a running plant sit several filters: a missing start date (44 percent of projects in July), reservations instead of orders (at times more than half at GE Vernova), lead times stretching to 2031/32, and financing. Nobody knows how much will be left in the end, not even BNEF.

Our take: if you argue with 345 million t CO2e, you're describing the worst case, not the likely path. That's still no all-clear. Even the 48 GW BNEF considers necessary would be a fossil burden for decades if the plants keep running after the grid connection arrives. SoftBank also shows that the bet can go wrong financially. For now, the winners are the turbine makers, who keep the deposits and resell the slots. Whether faster grid connections materialize, which BNEF recommends as the more climate-friendly solution, will decide the rest. According to BNEF, there's still little solid evidence on which policies actually work.