AI Capital Expenditure Surges, Power Grids Under More Strain! Goldman Sachs Significantly Raises Global AI Electricity Demand Forecast: 220% Surge Expected by 2030

In the past two months, the “incremental” investment from this round of AI has started to spill over from chips and servers to a harder-to-supplement segment: electricity. Hyperscale cloud providers have revised up their capital expenditure and R&D budgets, and have become more aggressive in deploying compute power for both training and inference, directly pushing up the long-term slope of data center electricity consumption. At the same time, market concerns have shifted: it’s no longer “is electricity needed,” but rather “can the supply chain deliver electricity to the data centers on time.”

According to Wind Trading Desk, Brian Singer, analyst at Goldman Sachs Global Investment Research, wrote in a report on the 23rd: “We are revising up the increase in global data center electricity demand by 2030 relative to 2023 from 175% to 220%.” The focus of this upgrade is on the United States: about 60% of the new electricity demand comes from the U.S., and data center capacity forecasts have also been significantly raised.

More troublesome, an upward revision in demand does not mean a clear path. The grid’s connection, power transmission and distribution, and equipment delivery timelines are all extending, which has pushed “behind-the-meter power” — a transitional solution mainly using natural gas to get data centers running before later reconnecting to the grid — to the foreground. Goldman Sachs has also raised its forecast for U.S. electricity demand growth to an annualized 3.2% before 2030, with data centers contributing 2 percentage points.

In terms of investment, Goldman Sachs is not conservative: even though stocks related to the data center electricity supply chain have already significantly outperformed, the report still maintains a bullish stance, based on a broader narrative — to avoid “reliability accidents” in electricity, water, networks, and supply chains, infrastructure is entering a longer investment cycle. But this cycle is not without bounds: if AI shifts from “hopes and dreams” to the “execution phase,” budget and return constraints will become tougher, and stock price drivers will shift from thematic to more brutal individual stock selection.

2030 Power Consumption Increment Repriced: Of 905TWh, the U.S. Takes 60%

Goldman Sachs estimates the global incremental electricity demand for data centers (AI + non-AI) by 2030 at 905TWh (relative to 2023), representing a 220% increase from 2023 to 2030. The previous assumption was a 175% increase. The reason for the upgrade is straightforward: the TMT team raised expectations for AI server shipments, the proportion of higher-power servers on the inference side is rising, and data center capacity is expanding faster.

Structurally, the U.S. weight keeps rising. Goldman Sachs estimates that about 60% of the 905TWh increment will occur in the U.S. (previously about 50%). The forecast for data center capacity has also been raised: U.S. data center capacity is expected to reach 95GW by 2030 (32GW in 2025), while overseas capacity is expected to reach 72GW (42GW in 2025). AI and data center expansion are still defined as global phenomena, but the U.S. will be the first to “consume the power.”

Hyperscale Cloud Providers’ Reinvestment Rates Near 90%, Focus Shifts from “Investment” to “Return”

The report signals a key point: budget upgrades are happening too quickly. In the past two months, Goldman Sachs analysts raised capex + R&D for hyperscalers by more than $300 billion for 2026-27; they also expect total global hyperscaler capex + R&D to double by 2029 compared to 2025.

More noteworthy is the reinvestment rate (capex + R&D / operating cash flow). Goldman Sachs expects it to reach 87%/83% in 2026/27, respectively (previously 79%/76%). Money is still being invested, but the space for free cash flow available to shareholders is being squeezed — this is why the report repeatedly emphasizes “AI revenue growth” and “quantifiable value”: as investment intensity increases, the market will more frequently question what AI is actually delivering.

Goldman Sachs gives a relatively “quantifiable” example in the report: AI accelerating drug discovery. Its healthcare team cites recent data pointing to two changes — a 370 basis points increase in success rate (from 6.4% to 10.3%), and a reduction in R&D cycles from about 13 years to about 10 years. Based on this, they estimate the present value of a 10-year drug pipeline could increase by $830 million (21% discount rate) to $4.12 billion (8% discount rate). These cases are intended to answer the question of “pervasiveness” — where is it actually landing?

U.S. Electricity Demand Growth Raised to 3.2%, Data Centers Contribute 2 Percentage Points

On the power side, Goldman Sachs has raised the expected growth rate of U.S. “grid + behind-the-meter” electricity demand to an annualized 3.2% before 2030 (previously 2.6%). Breaking it down, the grid side will grow at 2.6% annually, with behind-the-meter contributing 0.6%; within the grid’s 2.6%, data centers alone contribute 2 percentage points — explaining why market concerns over electricity, transmission, distribution, and grid connection resources are escalating rapidly.

Goldman Sachs also points out a reality: a significant portion of new load is being absorbed by behind-the-meter power, mainly natural gas, even though hyperscalers still prefer grid power in the long run. Raising electricity demand is not hard, the hard part is “delivering the electricity,” which is precisely stuck at transmission, distribution, grid connection, and construction capacity.

Efficiency Is Improving, But “Each Server Uses More Power” Is Also Happening: Inference Becomes a 2026 Variable

The report breaks down whether “efficiency can suppress electricity use” in more detail: the new generation of servers is indeed more efficient, but the industry’s compute power demand is growing even faster. Taking Nvidia servers as an example, Goldman Sachs writes that the latest Vera Rubin generation achieves a 16% higher compute speed per unit maximum power in training scenarios compared to Blackwell, with a cumulative four-generation improvement of over 650%; however, the maximum power per Vera Rubin server is 68% higher than Blackwell, with a four-generation cumulative increase of over 250%.

Inference is another turning point. Goldman Sachs maintains the assumption that “inference servers overall consume less power than training,” but also acknowledges that inference power intensity is being revised upward, as higher-power servers take a larger share on the inference side. The report views 2026 as a key observation window: will inference roll out on a large scale with low power consumption, or will inference, inference models, and automation drive higher energy consumption? The debate is not yet settled.

“Willingness to Pay a Premium for Reliability” Is Becoming a Contract Clause: $40-48/MWh Green Composite Premium

Power is not just a supply issue, but is also becoming a “price + policy” issue. Goldman Sachs summarizes this as the “Green Reliability Premium”: in the U.S., the average supply cost of a clean energy mix that meets data center baseload reliability is about $40/MWh higher than the benchmark, and about $48/MWh if IRA incentives are phased out.

More important is the comparison: if this premium is roughly applied to the global data center incremental electricity consumption between 2023 and 2030 (905TWh), Goldman Sachs estimates the industry spend would be about $3.7-4.3 billion. This is not dramatic in the context of hyperscalers’ profit and loss statements: equivalent to 3.4%-4.0% of Goldman’s estimated total 2027 hyperscaler EBITDA ($1.079 trillion), impacting average 2027 CROCI by about -0.8% to -0.9%. This is why the report concludes that hyperscalers can still afford to pay for “time-to-market” and “reliability.”

On the policy side, the keyword mentioned in the report is “ring-fence”: the costs and reliability risks brought by data center expansion should be contained as much as possible, without passing them on to other electricity customers. Goldman Sachs expects all parties to push for more contractual designs to isolate these impacts, and data center operators will also be required to provide clearer commitments on flexibility, bearing infrastructure costs, and even supplying power back to the grid.

Generation Equipment Is Not the Only Bottleneck — “People” Are

If you had to pick a tougher constraint between “equipment” and “people,” Goldman Sachs votes for the latter. The report estimates that to meet U.S./European electricity demand growth from 2023-2030, about 510,000 new U.S. electricity and grid-related positions and about 250,000 European electricity positions need to be created.

The highest risk is concentrated in the transmission & distribution (T&D) segment: Goldman Sachs estimates that the U.S. alone needs about 207,000 new T&D and grid connection jobs, implying a 22% labor force growth, and these positions typically require 3-4 years of training. For comparison, currently there are about 45,000 active apprentices in U.S. energy-related industries; to close the gap and cover retirements, Goldman Sachs believes the “running speed” of active apprentices may need to increase by 20,000-30,000.

Labor constraints in turn explain two things: why behind-the-meter power is more attractive in the short term (less transmission and grid connection required), and why contractors, utilities, automation, and grid optimization providers with labor acquisition advantages are being repriced.

“Reliability Supercycle” Gives the Supply Chain a Second Leg: Not Just Upgrading the Grid for AI

At the stock level, Goldman Sachs casts a wide net : “reliability investment” for electricity, water, networks, and supply chains as demand rises and infrastructure ages. The report’s quantifiable lever: based on their estimates for listed companies with Green Capex tailwinds, the reliability theme corresponds to annualized capex growth of over $80 billion.

This also explains a market phenomenon: data center electricity supply chain stocks and hyperscaler stocks have diverged. Since 2025, Goldman Sachs calculates, the data center electricity ecosystem has outperformed the MSCI ACWI by 41 percentage points and hyperscalers by 36 points; power generation equipment companies performed best, leading other supply chain peers by 196 points, with solar products, electrical components, and cooling solutions also significantly outperforming.

Goldman Sachs is straightforward about when the cycle will end: when AI’s competitive threat fades, corporate returns and free cash flow deteriorate significantly enough to reduce investment capacity, or surplus investment is deemed sufficient. As long as none of these three are triggered, reliability investment is unlikely to suddenly stop.

AI Still in the “Hopes and Dreams” Phase, but Three Indicators Will Decide When It Enters Execution

Goldman Sachs places AI within the “innovation cycle” framework: at present, it’s still in the “Appraisal / Hopes & Dreams” phase, which is most favorable for infrastructure investment and valuation expansion, but the capex upgrades are heating up debate about whether we are nearing the Execution phase. The report gives three triggers: financial flexibility constraints, declining corporate returns, and product oversupply.

As of now, Goldman Sachs believes the first two are starting to show “marginal change,” but not enough to constitute an inflection point: reinvestment rates are rising and squeezing free cash flow, but hyperscaler balance sheets remain strong, with net debt/EBITDA at about 0.3x (2026); on the return side, Goldman expects CROCI to weaken by 2028, from “slight” to “more evident,” but not yet falling to the low end of its historical range (24%-31%). As for oversupply, Goldman Sachs clearly writes: there is no evidence yet of excess compute or token demand.

This leads to a more realistic conclusion: in the short term, electricity and infrastructure chains remain in the “sweet spot”; but the market will increasingly demand to know where AI’s revenue and cash flow are, who can retain value, and who is merely paying for their competitors.