AI Data Center Power 🚀: SoftBank’s $3B Gas Bet (2026)12 min read2,433 words

AI Data Center Power 🚀: SoftBank’s $3B Gas Bet (2026)

Learn how AI data center power plans meet real demand—SoftBank’s $3B gas turbine bet, plus grid connections and battery energy storage for smarter flexibility.

AI data center powernatural gas turbinesSoftBank energy investmentdata center electricity demandbattery energy storage
AI Data Center Power 🚀: SoftBank’s $3B Gas Bet (2026)

SoftBank's $3 Billion Turbine Loss Is a Warning for AI's Power Race

AI data centers need huge amounts of electricity, and getting connected to the grid can take years. That reality has led tech companies and developers to chase power they can control themselves. Natural gas turbines - used by utilities for decades - have looked like a practical option because they can generate electricity on demand.

SoftBank went all-in. Its SB Energy unit said it plans to spend $33 billion on a 9.2-gigawatt natural gas power plant. At the same time, SoftBank reportedly put deposits down for more than 13 gigawatts of turbines - more than it ultimately needed. The company has since recognized over $3 billion in losses tied to power-generating equipment as part of a $7.1 billion quarterly loss.

This isn't proof that natural gas will stop powering AI infrastructure. It's a reminder that locking in equipment before your demand is fully clear can end up being an expensive kind of insurance. Here's what went wrong with SoftBank's ordering, what the loss says about how the market works, and how data center operators can think more carefully about grid power, gas generation, and batteries.

Why did SoftBank order so many turbines?

SoftBank was trying to buy time in a market where waiting can mean losing years. Electricity demand from AI data centers grew faster than the power buildout could keep up with. Grid connection queues stretched out, and orders for gas turbines reportedly ran into the early 2030s. If a developer waited for everything to be certain - permission, financing, and load - there was a real risk of ending up with a fully approved project but no equipment ready to run it.

So SoftBank reserved capacity early and at a scale larger than its stated needs. Starting about two years before the reported loss, it put deposits on more than 13 gigawatts of turbines - around 50% more than it expected to need for its plans, including SB Energy's announced 9.2-gigawatt gas plant.

There was a clear business logic. Peter Gardett, CEO of Noreva, told TechCrunch that SB Energy had little choice but to overbook future turbine supply if it wanted to secure turbines quickly. In short: the deposits weren't just a prediction of final demand. They were also a way to grab a scarce spot in a supplier queue.

The value - and cost - of reserving capacity

In procurement, reserving equipment early can be valuable because it gives you an option: build when the hardware is available rather than waiting behind other buyers. But that option only helps if you can adjust without eating the full cost.

SoftBank's deposits were nonrefundable. That turns what sounds like a flexible reservation into a hard financial commitment. If a project changes - or if the company no longer needs the full quantity - it can't easily recover the money. And once the expected demand comes in below the reserved supply, part of that reserved capacity stops being useful in practice.

The trade-off is pretty direct: pay more now to reduce the risk of waiting, or keep more capital liquid and accept the risk of delay. Neither choice is automatically right. It depends on what the project would cost if it slips, how certain the underlying demand really is, and the contract terms.

If you're building multiple facilities with steady demand and approved sites, reserving extra capacity can make sense. But if you're managing a portfolio of projects still working through permitting, financing, and customer commitments, large nonrefundable orders can turn uncertain forecasts into fixed liabilities. SoftBank's loss is a good example of how quickly that gap can show up at gigawatt scale.

What does the $3 billion loss tell us about AI power demand?

The loss highlights how hard it is to turn AI demand projections into actual power infrastructure. It doesn't prove that AI data centers will need less electricity overall. It shows that one investor's equipment commitments outpaced what it could justify - or actually use - under its plans, and that unwinding the mismatch was unusually expensive.

That matters because public discussion often treats data center power like one single number. In reality, projects don't all move forward on the same timeline. A campus can hit permitting delays. A customer can change planned capacity. A developer can decide that a different mix of grid power and on-site generation makes more sense. Meanwhile, overall electricity demand can keep rising even if one buyer ends up reserving too much of one type of equipment.

SoftBank's scale makes the write-down stand out. The company reported a $7.1 billion quarterly loss, with more than half tied to losses related to power-generating equipment. More than $3 billion tied to turbine orders is, in Gardett's words, a costly form of price discovery - where the company learned what it would have to pay for capacity it couldn't fully use.

A market signal, not a market verdict

It would be a mistake to read this loss as evidence that gas plants are about to disappear across the industry. The problem here was SoftBank's over-ordering and the nonrefundable nature of its deposits, not a demonstrated drop in AI electricity demand. The same equipment shortages and long lead times that pushed the original orders are still real for other projects.

And it's risky to treat those shortages as proof that every proposed gas project will go forward. When supply is tight, buyers may lock in equipment before they've fully worked out whether the project is viable. That can inflate commitments without guaranteeing that the final generation capacity will be needed - or economical - once everything is finalized.

The main takeaway is more limited: equipment availability, project demand, and power strategy are related, but they're not interchangeable. A turbine reservation isn't the same thing as a finished plant. A data center's forecast load isn't the same as its real operating pattern. And growth in national or regional demand doesn't automatically mean every individual project's assumptions will hold up.

For investors and infrastructure teams, the key question isn't just whether AI needs more power. It's whether a specific project can turn a power plan into reliable, affordable electricity on a timeline that matches the data center's construction and actual utilization.

Is natural gas still a practical choice for data centers?

Natural gas can work for data centers because it's a familiar, dispatchable source of electricity. Operators can plan around generation that's available when it's needed, rather than relying only on output that varies. In places where grid access is delayed, on-site generation can also help a project move forward sooner.

But dispatchable doesn't mean risk-free or always available. Brent Nelson, senior managing director of markets and strategy at Ascend Analytics, told the Equity podcast that gas supplies can be curtailed in winter, when demand peaks and available supply gets tighter. If a facility assumes its gas plant will always deliver full output, it may find out the hard way that its backup plan is weakest right when the wider system is under stress.

There are also project-level realities. A gas plant requires equipment, development work, fuel arrangements, and a realistic view of how much power the data center will actually use. A plant sized for a peak forecast can be underutilized if ramp-up takes longer than expected or if the facility runs below its projected load. In other words: capacity that looks good on paper still has to earn its place in the operating plan.

The case for a smaller grid connection plus storage

Nelson's alternative isn't grid only or everything else instead of gas. He said many data centers may be better served by a grid connection that's smaller than what they might normally request, with large batteries filling in the gap.

That setup can better match the power solution to the facility's actual needs. A battery can discharge during high-demand periods, reduce peak grid draw, and help a site respond when a utility asks it to curtail consumption. It also adds flexibility without forcing a company to secure enough generation capacity to cover every theoretical peak with dedicated hardware.

Of course, batteries aren't a universal substitute for either generation or grid capacity. Their value depends on how long and how often the shortfall happens, local electricity prices, interconnection rules, and storage costs. A battery that smooths short peaks isn't automatically capable of carrying a campus through a prolonged outage.

So the real comparison isn't gas versus batteries, full stop. It's the cost and reliability of the whole system: grid supply, on-site generation, storage, demand management, and contingency plans. Gas may fit a project with reliable fuel supply and a durable need for dispatchable power. A grid connection plus storage may fit when the main challenge is peak capacity or intermittent constraints. Some projects will need a blend.

How should data center operators plan their power supply?

First, stop treating a peak-load estimate like it's already a procurement order. Forecasts are inputs to planning - not a reason by themselves to buy the maximum amount of equipment you can. Operators should separate what they know from what they're assuming, and then tie big commitments to evidence that improves over time.

A practical decision process has four parts:

  1. Model the load in stages. Estimate demand during construction, early operations, and at full utilization. Keep peak load separate from average consumption; the two point to different infrastructure needs.
  2. Map where the constraint really is. Figure out whether the bottleneck is a delayed grid connection, limited capacity at peak times, local reliability issues, or a lack of generation equipment. Different constraints call for different responses.
  3. Compare complete supply options. Evaluate grid power, gas generation, batteries, and demand flexibility as one system. Include fuel and equipment availability, operating conditions, connection timing, and what happens during outages or curtailments.
  4. Commit in stages. Use decision gates tied to permits, customer demand, financing, and construction milestones when you can. Avoid locking in large nonrefundable orders before the assumptions behind them are solid.

Put a price on flexibility

A cheaper contract isn't automatically the better contract if it leaves the project exposed to years of delay. Likewise, an expensive reservation isn't automatically prudent just because turbine slots are scarce. You have to compare the cost of the commitment with the value of moving earlier, and think through what happens if demand is lower, the schedule slips, or the preferred power mix changes.

This is where cancellation and change terms matter. If suppliers require deposits, ask what you can transfer, reschedule, or reduce. If the money is effectively nonrefundable, treat it like capital at risk - not like a temporary hold. That framing helps finance teams and project decision-makers see the trade-off before procurement momentum takes over.

Operators should also test the plan under bad conditions, not only the most likely scenario. What happens during winter gas curtailments? How much load can batteries cover, and for how long? Can the data center reduce or shift demand when the utility is constrained? Which parts of the facility need continuous power, and which loads have more flexibility?

These questions don't make the plan less ambitious. They make it usable. A data center needs electricity that's available when the business needs it - not just a big capacity number in a proposal. The right portfolio is the one that still works when schedules change and real operating conditions show up - without trapping the project in stranded capital.

Key takeaways

  • SoftBank reportedly reserved more than 13 gigawatts of turbines - about 50% more than it needed - and recognized more than $3 billion in losses tied to power equipment.
  • The loss shows what happens when you make large, nonrefundable commitments before demand and timelines are truly firm.
  • Turbine scarcity and AI's growing electricity needs are real, but they don't guarantee that every planned gas project will be built.
  • Natural gas can deliver dispatchable power, but winter fuel curtailments and supply limits create reliability risks.
  • For some sites, a smaller grid connection paired with batteries may be more cost-effective, especially when peak demand is the main constraint.
  • Operators should compare complete power systems and stage procurement against project milestones, instead of buying equipment based only on peak forecasts.

Conclusion

SoftBank's turbine loss is best read as a warning about procurement under uncertainty - not as a verdict on AI's electricity needs or on natural gas. When equipment lead times stretch for years, reserving capacity can protect a project from delay. But when those reservations are nonrefundable and exceed credible demand, the insurance can turn into a multibillion-dollar liability.

For data center operators, the practical response is to plan around the real constraint, model demand in stages, and compare grid power, storage, and generation as one integrated operating system. Secure scarce equipment when the value is clear - but spell out the cost of changing course before signing.

Frequently Asked Questions

SoftBank reported a $7.1 billion quarterly loss, with more than half related to power-generating equipment. More than $3 billion of the loss was tied to turbine-related commitments. The reported issue was not simply the cost of operating gas plants; it involved deposits on equipment the company no longer considered valuable to its plans.
No. The loss does not show that AI power demand has disappeared or that gas projects will be broadly cancelled. It shows that SoftBank over-ordered relative to its needs and faced losses on nonrefundable deposits. Natural gas may remain part of data center power plans, but individual projects still need to justify their costs and risks.
Turbine supply was scarce, with order queues extending into the early 2030s. Reserving equipment ahead of time could help secure some capacity sooner. Peter Gardett of Noreva said overbooking was effectively necessary if SB Energy wanted turbines quickly. The strategy reduced one risk—waiting—but increased the cost of changing plans.
Not in every situation. Batteries can help shave peak demand, cover short-term gaps, and support curtailment plans, but their suitability depends on how long and how often power is needed. Some facilities may benefit from a mix of grid supply, batteries, and generation rather than relying on a single source.

Related articles