Over the past two months, the incremental focus of AI investment has shifted from chips and servers to a more challenging bottleneck: power supply. Hyperscale cloud providers have increased their capital expenditures and R&D budgets, with more aggressive deployment of computing power for both training and inference, directly steepening the long-term trajectory of data center electricity consumption. Concurrently, market concerns have evolved: the issue is no longer whether electricity is needed, but whether the supply chain can deliver it to server rooms on schedule.
According to sources, Goldman Sachs analyst Brian Singer wrote in a report dated the 23rd: "We have raised our forecast for the increase in global data center power demand by 2030 relative to 2023 from 175% to 220%." This revision heavily focuses on the United States, which is expected to account for approximately 60% of the new electricity demand, and data center capacity projections for the country have also been significantly raised.
Compounding the issue, increased demand does not equate to a clear path forward. Grid-side challenges, including interconnection, transmission, distribution, and equipment delivery timelines, are lengthening. This has pushed "behind-the-meter" power generation—primarily using natural gas as a transitional solution to get data centers operational before later connecting to the grid—to the forefront. Goldman Sachs has also raised its forecast for annualized U.S. electricity demand growth to 3.2% through 2030, with data centers contributing 2 percentage points of that increase.
Regarding investment strategy, Goldman Sachs maintains a non-conservative stance: despite significant outperformance by stocks related to the data center power supply chain, the report remains bullish. This is underpinned by a larger narrative—a prolonged infrastructure investment cycle aimed at preventing reliability failures related to power, water, networking, and supply chains. However, this cycle is not without limits: if AI transitions from a phase of "hope and dreams" to an "execution" phase, budget and return constraints will tighten, and stock price drivers will shift from thematic trends to more rigorous individual company selection.
The 2030 electricity demand increment has been repriced: the U.S. accounts for 60% of the 905 TWh increase. Goldman Sachs estimates the global data center (AI and non-AI) electricity consumption increment by 2030 (relative to 2023) will reach 905 TWh, corresponding to a 220% increase from 2023 levels, up from a previous assumption of 175% growth. The reasons for the upward revision are straightforward: the TMT team raised its forecasts for AI server shipments, the deployment share of higher-power servers for inference is increasing, and data center capacity expansion is accelerating.
Structurally, the weighting of the United States continues to rise. Goldman Sachs estimates that approximately 60% of this 905 TWh increment will occur in the U.S. (up from a previous estimate of about 50%). Corresponding data center capacity projections have also been increased: U.S. data center capacity is forecast to reach 95 GW by 2030 (up from 32 GW in 2025), while overseas capacity is expected to reach 72 GW by 2030 (up from 42 GW in 2025). AI and data center expansion remain a global phenomenon, but the U.S. is still positioned to "get the power first."
The reinvestment rate for hyperscale cloud providers is approaching 90%, shifting the discussion focus from "investment" to "return." A key signal from the report is the rapid pace of budget increases. Over the past two months, Goldman Sachs analysts have raised their combined capex + R&D estimates for hyperscalers for 2026-2027 by over $300 billion; they also project that the major global hyperscalers' capex + R&D will double by 2029 compared to 2025 levels.
More critical to watch is the reinvestment rate (capex + R&D / operating cash flow). Goldman Sachs forecasts this rate will reach 87% and 83% in 2026 and 2027, respectively (up from previous estimates of 79% and 76%). Investment continues unabated, but the free cash flow available for shareholders is being squeezed—this is why the report repeatedly emphasizes "AI revenue growth" and "quantifiable value": as investment intensity rises, the market will more frequently demand evidence of what tangible results AI is delivering.
Goldman Sachs cited a relatively quantifiable example in the report: AI-accelerated drug discovery. Recent data referenced by their healthcare team points to two changes—success rates improving by 370 basis points (from 6.4% to 10.3%) and R&D cycles shortening from approximately 13 years to about 10 years. Based on this, they estimate the present value of a 10-year drug pipeline could increase by $8.3 billion (using a 21% discount rate) to $41.2 billion (using an 8% discount rate). Such examples serve to illustrate where the "pervasiveness" of AI is actually materializing.
U.S. electricity demand growth forecast raised to 3.2%, with data centers contributing 2 percentage points. On the power side, Goldman Sachs has raised its forecast for annualized U.S. electricity demand growth (grid + behind-the-meter) to 3.2% through 2030 (up from 2.6%). Breaking this down, grid-side demand is expected to grow at an annualized rate of 2.6%, with behind-the-meter contributing 0.6%. Within the grid-side 2.6% growth, data centers alone contribute 2 percentage points—this explains why market concerns about power, transmission, distribution, and interconnection resource constraints are escalating rapidly.
Goldman Sachs also highlights a reality: a significant portion of the new load is being met by behind-the-meter generation, primarily natural gas, even though hyperscalers' long-term preference remains grid power. Increasing electricity demand forecasts is straightforward; the difficulty lies in delivering the power "to the right place," which is precisely where bottlenecks in transmission, distribution, interconnection, and construction capacity exist.
Efficiency is improving, but "more power consumption per server" is also happening: Inference becomes the key variable for 2026. The report provides a detailed analysis of whether efficiency gains can curb electricity consumption: while new-generation servers are indeed more efficient, the industry's demand for computing power is growing even faster. Using Nvidia servers as an example, the report states that the latest Vera Rubin generation offers a 16% improvement in compute speed per unit of maximum power compared to Blackwell in training scenarios, with a cumulative improvement exceeding 650% over four generations. However, simultaneously, the maximum power per Vera Rubin server is 68% higher than Blackwell, with a cumulative increase exceeding 250% over four generations.
The inference side represents another inflection point. While maintaining the assumption that inference servers generally have lower power requirements than training servers, Goldman Sachs acknowledges that inference power intensity is being revised upward due to the increasing proportion of higher-power servers used for inference. The report identifies 2026 as a critical observation window: the debate continues over whether inference will primarily involve "low-power, large-scale deployment" or trend towards higher energy consumption due to complex inference, inference-specific models, and automation.
"Willingness to pay a premium for reliability" is becoming a contract term: A green reliability premium of $40-48/MWh. Power is not just a supply issue; it is also becoming a "price + policy" issue. Goldman Sachs uses the term "Green Reliability Premium" to describe this shift: in the U.S., the average cost of a clean energy portfolio meeting data center baseload reliability requirements is approximately $40/MWh higher than the baseline, potentially rising to around $48/MWh if IRA incentives phase out.
More importantly, for perspective: if this premium is roughly applied to the global data center electricity demand increment from 2023 to 2030 (905 TWh), Goldman Sachs estimates the corresponding industry expenditure would be approximately $37-43 billion. This magnitude is not excessive on hyperscalers' income statements: it equates to about 3.4%-4.0% of their estimated total EBITDA for 2027 ($10.79 trillion), with an impact on the average 2027 CROCI of approximately -0.8% to -0.9%. This is why the report concludes that hyperscalers still have the capacity to pay for "time to market" and "reliability."
At the policy level, a key term mentioned is "ring-fence": the goal is to prevent the costs and reliability risks associated with data center expansion from spilling over to other electricity customers. Goldman Sachs expects stakeholders to push for more contract designs to isolate these impacts, while data center operators will likely be required to make clearer commitments regarding flexibility, bearing infrastructure costs, and even providing capabilities to feed power back to the grid.
Generation equipment is not the only bottleneck; "people" are the critical constraint. If forced to choose between "equipment" and "people" as the harder constraint, Goldman Sachs votes for the latter. The report estimates that meeting U.S. and European electricity demand growth from 2023 to 2030 will require adding approximately 510,000 power and grid-related jobs in the U.S. and about 250,000 in Europe.
Risks are more concentrated in the transmission and distribution (T&D) segment: Goldman Sachs estimates the U.S. alone needs approximately 207,000 new T&D and interconnection-related jobs, implying a workforce growth requirement of about 22%. These roles typically require 3-4 years of training. For comparison, there are currently about 45,000 active apprentices in U.S. energy-related industries; to fill the gap and account for retirements, Goldman Sachs believes the "run rate" of active apprentices may need to increase by about 20,000-30,000.
Labor constraints help explain two things: why behind-the-meter power is more attractive in the short term (bypassing lengthy transmission line and interconnection processes), and why contractors, utilities, and companies offering automation and grid optimization solutions with advantages in labor acquisition are being revalued.
The "reliability super-cycle" provides a second leg for the supply chain: It's not just about building the grid for AI. Regarding equities, Goldman Sachs broadens the theme: investments in the "reliability" of power, water, networking, and supply chains amid rising demand and aging infrastructure. The report quantifies this opportunity: based on its estimates for publicly listed companies benefiting from Green Capex trends, the reliability theme corresponds to an annualized capex growth exceeding $80 billion.
This also explains a market phenomenon: the divergence between the performance of data center power supply chain stocks and hyperscaler stocks. Goldman Sachs notes that since the start of 2025, the data center-related power ecosystem has outperformed the MSCI ACWI index by about 41 percentage points and hyperscalers by about 36 percentage points. Among these, power generation equipment-related companies have shown the strongest performance, leading other supply chain segments by about 196 percentage points, with solar products, electrical components, and cooling solutions also showing significant outperformance.
Goldman Sachs outlines straightforward conditions for the end of this cycle: a reduction in the perceived competitive threat from AI, a significant deterioration in corporate returns and free cash flow leading to reduced investment capacity, or a consensus that redundant investment is already sufficient. As long as these three triggers are not activated, reliability-focused investment is unlikely to halt abruptly.
AI remains in the "hope and dreams" stage, but three metrics will determine its transition to the execution phase. Goldman Sachs frames AI within an "innovation cycle": it currently remains in the "Appraisal / Hopes & Dreams" stage, which is most favorable for infrastructure investment and valuation expansion. However, the upward revisions to capital expenditures are heating up the debate about whether the "Execution" stage is approaching. The report identifies three potential triggers: constrained financial flexibility, declining corporate returns, and product oversupply.
Based on current evidence, Goldman Sachs believes the first two are showing "marginal changes" but are not yet sufficient to constitute an inflection point: rising reinvestment rates are compressing free cash flow, but hyperscaler balance sheets remain strong, with net debt/EBITDA around 0.3x (2026). On returns, Goldman Sachs expects CROCI to weaken by 2028, moving from "slight" to "more pronounced," but not yet falling to the low end of its historical range (24%-31%). As for oversupply, the report explicitly states that it does not yet see computing power or token demand entering an oversupply situation.
This leads to a more pragmatic conclusion: in the short term, the power and infrastructure chain remains in a "sweet spot." However, the market will increasingly demand answers regarding AI's revenue and cash flow generation, scrutinizing which companies can retain value and which are merely funding their competitors' infrastructure.
Comments