<p>In the past two months, the &#34;increment&#34; of this round of AI investment has begun to spill over from chips and servers to harder-to-supply segments: electricity. Major cloud providers have raised their capital expenditure and R&amp;D budgets, with more aggressive deployments of computing power for training and inference, directly pushing up the forward slope of electricity consumption in data centers. Meanwhile, market concerns are shifting: it's no longer about &#34;whether electricity is needed,&#34; but rather &#34;whether the supply chain can deliver electricity to the data center on time.&#34;</p>
<p>According to news from the Chasing Wind Trading Desk, Goldman Sachs Global Investment Research analyst Brian Singer wrote in a report on the 23rd: <strong>&#34;We have raised our forecast for global data center electricity demand in 2030 relative to 2023 from an increase of 175% to 220%.&#34;</strong> This adjustment is particularly focused on the United States: about 60% of the new electricity demand is expected to come from the U.S., and the forecast for data center capacity has also been significantly raised.</p>
<p>The more troubling aspect is that an increase in demand does not equate to a clear path forward. The timelines for grid interconnection, transmission and distribution, and equipment delivery are all lengthening, leading to the emergence of &#34;behind-the-meter&#34; solutions—primarily using natural gas to get data centers running first, with plans to connect back to the grid later. Goldman Sachs has also raised its forecast for the growth rate of electricity demand in the U.S. to an annualized 3.2% before 2030, with data centers contributing 2 percentage points.</p>
<p>In terms of investment, Goldman Sachs' stance is not conservative: although stocks related to the electricity supply chain for data centers have significantly outperformed, the report maintains a bullish outlook, driven by a larger narrative—investments in infrastructure are entering a longer cycle to avoid &#34;reliability incidents&#34; related to electricity, water, networks, and supply chains. However, this cycle is not without boundaries: if AI shifts from &#34;hopes and dreams&#34; to &#34;execution,&#34; the constraints of budgets and returns will become stricter, and stock price drivers will shift from themes to more brutal individual stock selection.</p>
<h2>The electricity increment for 2030 is being repriced: 905 TWh, with the U.S. taking 60%</h2>
<p><strong>Goldman Sachs estimates the global data center (AI + non-AI) electricity increment for 2030 at 905 TWh (relative to 2023), corresponding to a 220% increase compared to 2023.</strong> The previous assumption was a growth of 175%. The reasoning for the upward revision is straightforward: the TMT team has raised its expectations for AI server shipments, the deployment ratio of higher-power servers on the inference side has increased, and the capacity expansion of data centers is accelerating.</p>
<p>Structurally, the weight of the U.S. continues to rise. Goldman Sachs expects that about 60% of this 905 TWh increment will occur in the U.S. (up from about 50% previously). Corresponding forecasts for data center capacity have also been raised: U.S. data center capacity is expected 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). The expansion of AI and data centers is still defined as a global phenomenon, but the U.S. is still the first to access electricity.</p>
<h2>The reinvestment rate of major cloud providers approaches 90%, shifting the discussion focus from &#34;investment&#34; to &#34;returns&#34;</h2>
<p>A key signal from the report is: <strong>the speed of budget revisions is too fast.</strong> In the past two months, Goldman Sachs analysts have raised their forecasts for hyperscaler capex + R&amp;D for 2026-27 by over $300 billion; they also expect that the capex + R&amp;D of major global hyperscalers will double by 2029 compared to 2025
What is more worth paying attention to is the reinvestment rate (capex + R&amp;D / operating cash flow). Goldman Sachs expects it to reach 87%/83% in 2026/27 (previously 79%/76%). Money is certainly still being invested, but the space for free cash flow available to shareholders has been squeezed — this is also the reason why the report repeatedly emphasizes &#34;AI revenue growth&#34; and &#34;quantifiable value&#34;: when the intensity of investment increases, the market will more frequently question what AI has actually achieved.</p>
<p>Goldman Sachs cited a direction that is somewhat &#34;quantifiable&#34;: <strong>AI accelerating drug development.</strong> The medical team referenced recent data pointing to two changes — a 370 basis point increase in success rates (from 6.4% to 10.3%), and a reduction in the R&amp;D cycle from about 13 years to about 10 years, estimating that the present value of a 10-year drug pipeline could increase by $8.3 billion (21% discount rate) to $41.2 billion (8% discount rate). Such cases are more like answers to the question of &#34;where does pervasiveness actually land.&#34;</p>
<h2>The growth rate of electricity demand in the U.S. has been raised to 3.2%, with data centers contributing 2 percentage points</h2>
<p>On the electricity side, Goldman Sachs has raised its forecast for the growth rate of electricity demand from the U.S. &#34;grid + post-meter electricity&#34; to an annualized 3.2% before 2030 (previously 2.6%). Breaking it down, the grid side is annualized at 2.6%, with post-meter electricity contributing 0.6%; within the 2.6% on the grid side, data centers alone contributed 2 percentage points — this explains why the market's sense of tension regarding electricity, transmission and distribution, and grid resources is rapidly escalating.</p>
<p>Goldman Sachs also pointed out a reality: <strong>A significant portion of the new load is being borne by post-meter electricity, primarily natural gas, even though hyperscalers still prefer grid power in the long term.</strong> It is not difficult to raise electricity demand, but the challenge lies in &#34;delivering&#34; the electricity, which is precisely constrained by transmission, distribution, and construction capabilities.</p>
<h2>Efficiency is improving, but &#34;each server consumes more electricity&#34; is also happening: the inference side becomes a variable in 2026</h2>
<p>The report provided a more detailed breakdown of whether &#34;efficiency can suppress electricity consumption&#34;: <strong>The new generation of servers is indeed more efficient, but the industry's demand for computing power is growing faster.</strong> Taking NVIDIA servers as an example, Goldman Sachs noted that the latest Vera Rubin generation has a 16% improvement in computing speed corresponding to maximum power in training scenarios compared to Blackwell, with a cumulative improvement of over 650% across four generations; however, at the same time, the maximum power of a single Vera Rubin server is 68% higher than that of Blackwell, with a cumulative increase of over 250% across four generations.</p>
<p>The inference side is another turning point. Goldman Sachs maintains the assumption that &#34;the overall power of inference servers is lower than that of training,&#34; while also acknowledging that the intensity of inference power is being revised upwards, due to the increasing proportion of higher power servers in inference. The report views 2026 as a key observation window: whether inference will be &#34;widely deployed at low power&#34; or will move towards higher energy consumption due to inference, inference models, and automation is still a matter of debate.</p>
<h2>&#34;Willing to pay a premium for reliability&#34; is becoming a contract term: a green comprehensive premium of $40-48/MWh</h2>
<p>Electricity is not just a supply issue; it has also begun to transform into a &#34;price + policy&#34; issue. Goldman Sachs summarizes this change with the term &#34;Green Reliability Premium&#34;: in the United States, the average cost of clean energy portfolios that meet the base load reliability for data centers is approximately $40/MWh higher than the benchmark, and this is expected to rise to about $48/MWh after the IRA incentives taper off.</p>
<p>More importantly, in terms of comparative metrics: if this type of premium is roughly applied to the global data center electricity increment from 2023 to 2030 (905 TWh), Goldman Sachs estimates the corresponding industry expenditure to be around $37 billion to $43 billion. The scale of this in the hyperscalers' profit and loss statements is not exaggerated: it accounts for about 3.4% to 4.0% of Goldman Sachs' estimated total EBITDA for hyperscalers in 2027 ($1,079 billion), with an impact of approximately -0.8% to -0.9% on the average CROCI for 2027. <strong>This is also why the report concludes that hyperscalers still have the capacity to pay for &#34;time to market&#34; and &#34;reliability.&#34;</strong></p>
<p>On the policy front, the key term mentioned in the report is &#34;ring-fence&#34;: the costs and reliability risks brought about by the expansion of data centers should not spill over to other electricity customers. Goldman Sachs expects all parties to push for more contract designs to isolate these impacts, while data center operators will also be required to provide clearer commitments regarding flexibility, infrastructure cost bearing, and even the ability to supply back to the grid.</p>
<h2>Generation equipment is not the only bottleneck; &#34;people&#34; are</h2>
<p>If one had to choose a harder constraint between &#34;equipment&#34; and &#34;people,&#34; Goldman Sachs casts its vote for the latter. The report estimates that <strong>to meet the electricity demand growth in the U.S. and Europe from 2023 to 2030, approximately 510,000 new jobs related to electricity and the grid will need to be created in the U.S., and about 250,000 in Europe.</strong></p>
<p>The risks are more concentrated in the transmission and distribution (T&amp;D) segment: Goldman Sachs estimates that the U.S. will need about 207,000 new jobs related to T&amp;D and grid connection, which means about 22% of the labor force growth demand, and these positions typically require 3-4 years of training. In comparison, there are currently about 45,000 active apprentices in the U.S. energy-related industry; to fill the gap and cover retirements, Goldman Sachs believes the &#34;operational speed&#34; of active apprentices may need to increase by about 20,000 to 30,000.</p>
<p>Labor constraints will, in turn, explain two things: why behind-the-meter electricity is more attractive in the short term (less transmission line and grid connection processes), and why contractors, utility companies, and automation and grid optimization solutions that have a labor acquisition advantage will be repriced.</p>
<h2>The &#34;Reliability Supercycle&#34; gives the supply chain a second leg: not just fixing the grid for AI</h2>
<p>On the stock level, Goldman Sachs broadens the theme <strong>: the &#34;reliability investment&#34; in electricity, water, networks, and supply chains under rising demand and aging infrastructure.</strong> The quantitative metric provided in the report is that based on its estimates of publicly listed companies with Green Capex tailwinds, the reliability theme corresponds to an annualized capex growth of over $80 billion
This also explains a market phenomenon: the divergence in stock prices between data center power supply chain stocks and hyperscaler stocks. According to Goldman Sachs, since 2025, the overall performance of the data center-related power ecosystem has outperformed MSCI ACWI by about 41 percentage points and hyperscalers by about 36 percentage points; among them, companies related to power generation equipment performed the strongest, leading other supply chain segments by about 196 percentage points, with solar products, electrical components, cooling solutions, etc., also showing significant excess returns.</p>
<p>Goldman Sachs provides a straightforward condition for &#34;when the cycle ends&#34;: the diminishing competitive threat of AI, a significant deterioration in corporate returns and free cash flow leading to a decline in investment capacity, or redundant investments being deemed sufficient. As long as these three conditions are not triggered, reliability investments are unlikely to suddenly stop.</p>
<h2>AI is still in the &#34;Hopes &amp; Dreams&#34; stage, but three indicators will determine when it enters the execution phase</h2>
<p>Goldman Sachs places AI within the framework of the &#34;innovation cycle&#34;: currently still in the &#34;Appraisal / Hopes &amp; Dreams&#34; stage, which is most favorable for infrastructure investment and valuation expansion, but the upward revision of capital expenditures has intensified the debate on &#34;whether we are close to the Execution stage.&#34; The three triggers provided in the report are: <strong>financial flexibility constrained, corporate returns declining, and product supply surplus.</strong></p>
<p>Based on current evidence, Goldman Sachs believes the first two are beginning to show &#34;marginal changes,&#34; but are not enough to constitute a turning point: the rising reinvestment rate is compressing free cash flow, but the balance sheets of hyperscalers remain strong, with net debt/EBITDA at about 0.3 times (2026); on the return side, Goldman Sachs expects CROCI to weaken by 2028, with the magnitude shifting from &#34;slight&#34; to &#34;more pronounced,&#34; but it has not yet fallen to the lower end of its historical range (24%-31%). As for supply surplus, Goldman Sachs explicitly states: there is no evidence yet of a surplus in computing power and token demand.</p>
<p>This leads to a more realistic conclusion: in the short term, the power and infrastructure chain remains in the &#34;sweet spot&#34;; however, the market will increasingly demand to know where AI's revenue and cash flow are, who can retain value, and who is merely footing the bill for competitors</p>

GS-C

GS-A

GS-D

<p>Goldman Sachs' latest report raises the forecast for global data center electricity demand growth in 2030 to 220%, with the United States accounting for 60%. The incremental investment in AI is spilling over from computing power to the electricity supply chain, triggering a "super cycle of reliability" in infrastructure. Despite the surge in capital expenditures from cloud providers squeezing free cash flow, they still have the financial capacity to pay a premium for clean electricity. Currently, the industry's bottleneck is shifting from power generation equipment to labor shortages, and market focus is beginning to shift from "thematic speculation" to the hard constraints of "AI revenue conversion."</p>

- In the past two months, AI-related investment has shifted from chips and servers to electricity, significantly increasing data center power demands, with the US expecting a 220% rise in electricity requirements by 2030.  
- Goldman Sachs now estimates that 60% of the global data center electricity demand increase will occur in the US, driven by heightened expectations for AI server deployment and data center expansion.  
- As supply chain challenges arise, a greater emphasis on "behind-the-meter" electricity solutions is observed, while issues with workforce availability and the adoption of green energy are impacting reliability and costs.

Wallstreetcn

GOLDMAN SACHS GROUP INC DEP SHS REPSTG 1/1000TH PRF SER C

GOLDMAN SACHS GROUP INC DEP SHR REP 1/1000TH PFD SER A

GOLDMAN SACHS GROUP INC DEP REP 1/1000TH PRF D

Goldman Sachs

AI capital expenditures surge, the power grid is under more pressure! Goldman Sachs significantly raises global AI electricity demand forecast: demand will soar by 220% by 2030