CDN NEWS | 
US NEWS 
(Reuters Breakingviews) – A sprawling global network of silicon has sprung up to power artificial intelligence. Housed within data centers, these chips demand vast sums of electricity to crunch the numbers behind the likes of ChatGPT. Yet both technology giants and energy utilities are incentivized to pad projections for how many of these facilities need to be built and how many electrons they will ultimately consume. When trying to reconcile canceled plans from cloud giant Microsoft with, say, multi-year backlogs for gas turbine manufacturer GE Vernova , it’s worth keeping these distortions in mind.
After two decades of flatlining, the U.S. Energy Information Administration now expects American power demand to grow by 2% annually. Above all, that’s thanks to a wave of data centers being set up as Alphabet, Amazon.com, Meta Platforms and Microsoft aim to spend over $300 billion in combined capital expenditure this year. These server farms are expected to triple as a share of total grid usage, rising to 12% by 2030, McKinsey reckons.
This economy-reshaping boom has dramatically lifted AI-adjacent stocks, but left investors on edge about any potential sign of a slowdown. The success of China’s DeepSeek, for instance, raised questions about whether chatbots really do need the monumental supply of chips being churned out by Nvidia, briefly lopping $600 billion off its market value.
Perhaps the most important bellwether is the company led by Satya Nadella. Microsoft ushered in the AI age by partnering with ChatGPT developer OpenAI in 2016, and its Azure cloud division provides computing grunt to a raft of chatbot developers and users. When an executive said in April that the company had slowed or paused some of its data center build-out, it seemed an ominous sign. A few weeks later, Azure nonetheless reported accelerating growth, and a roughly $80 billion infrastructure spending target was reaffirmed.
This should be a familiar dynamic for tech watchers. The semiconductor industry is notoriously plagued by “double ordering,” or the phenomenon of customers ordering more chips than they need on the theory that cancelling some shipments is better than going without a mission-critical part.
Whether Microsoft and Google are consciously doing the same or not, the worst thing that can happen in the lightning-fast AI race is falling behind. There are plenty of stumbling blocks – like Nvidia’s essential chips, constantly in short supply – that are out of their control. But it’s simple enough to earmark land and requests for power hookups by starting new data center projects. It is far less damaging to cancel any surplus developments, especially early-stage projects where minimal capital has been spent, than be left without capacity to serve customers.
This flows up to utilities and grid operators like ERCOT, PJM Interconnection and MISO, which together monitor over two dozen states accounting for roughly half the U.S. population, including data center hotspots in North Virginia and Texas. The three organizations estimate that “large load demand” – essentially, server projects – will reach 140 gigawatts by 2030, according to nuclear reactor operator Constellation Energy. Yet that figure is double estimated data center consumption for the entire country, according to S&P Global and McKinsey.
Of course, companies like OpenAI are planning data centers as large as 5 gigawatts, roughly equivalent to the power draw of four million homes, based on average annual home use according to the EIA. It therefore takes very few projects to move utilities’ projections around substantially. They also face incentives to over-estimate, though.
These can be as benign as wanting to plan for a bigger buffer of potential power generation on tap, rather than too little, thus avoiding blackouts. On the other hand, as strictly regulated businesses, defraying the cost of capital improvements is among the few means they have of gaining permission to raise prices and earnings.
Whatever the cause, the energy industry’s crystal balls often produce overcooked figures. PJM, which manages the grid across 13 states, overestimated peak summer demand for 17 years in a row before 2024, according to Wilson Energy Economics. The Rocky Mountain Institute reckons that utilities and grid operators similarly forecast growth 12 percentage points too high between 2005 and 2015.
Of course, once projects get started, they ramify up the supply chain. GE Vernova estimates that its gas turbines will be effectively sold out through 2028 by this summer. Similarly, Hitachi Energy recently said that the wait time for new power transformers has now reached four years. This wash of new orders is stretching capacity and upping prices. NextEra Energy reckons that the cost of building new gas-fired generation has tripled since 2022. Laments of a skilled worker shortage abound.
If any of the demand driving this proves illusory, it might also over-egg the apparent need for alternative energy sources. OpenAI boss Sam Altman has joined a scramble to explore shrunken-down nuclear plants. Google said it would commit capital to Elementl Power to advance three nuclear projects – though their location or even basic technology was still to be determined. The outlook for such efforts is uncertain at best.
There are less far-fetched or disruptive ways to handle whatever influx really does arise. The current U.S. energy grid could absorb 76 gigawatts of new load, about 10% of aggregate peak demand, so long as new users agree to switch off their facilities just 0.25% of the time, or less than one day a year on average, according to researchers at Duke University. For any new generation that does get built, utilities and regulators could try to ensure that big tech customers understand that they will be on the hook for at least some payments.
For now, though, all of this simply makes the AI tea leaves more difficult to read. Microsoft might step back from deals, utility load growth might underwhelm – and Azure and its rivals might still continue to grow at a rapid clip all the same. Like a malfunctioning chatbot that suggests combining glue and pizza, tripping up over apparent connections between unrelated data is an investing hazard.
Share This:
COMMENTARY: Even Big Oil Thinks Big Oil Is Too Risky These Days – Fickling