Power is the AI bottleneck everyone agrees on - so why are Vertiv and Eaton so quiet?
13 August 2026 · 11 min read
The short version: electricity is the most widely agreed constraint on AI infrastructure. Transformer lead times run to three or four years, the median US project reaching commercial operation in 2025 spent 61 months in the interconnection queue, and heavy-duty gas turbines are sold out into the 2030s. The two companies on our board that sell directly into that constraint - Vertiv and Eaton - are compounding backlog at extraordinary rates. They are also, on our own measurements, among the eight quietest names of the 34 we track and the only two in their segment with negative volume components.
That is not a mispricing, and this article is not going to claim it is. It is a duration mismatch, and it is the most under-discussed thing about the AI power trade. All board figures are as at the close on 12 August 2026.
The split inside our own power segment
We classify four names as power and systems. On the market sentiment score the segment looks strong - a median of 73 against 66 for the whole board, second of the six segments. Break it in half and it stops looking like one thing:
The two names that build the box score 84 and 80. The two names that deliver the electricity into it score 66 and 60. On the latest session that is a 19-point gap between the two halves of a single segment.
One honest caveat before this becomes a bigger claim than it is. Across the thirteen sessions from 27 July to 12 August, the electrical pair actually led the server pair on six of them and trailed on seven. They alternate. What does not alternate is the second measurement, and it is the one that matters.
Nobody is trading them
The volume component of our score compares a stock's turnover with its own normal level. On 12 August:
- Super Micro +7.18 of a possible 7.20 - effectively at the ceiling
- Dell +0.36
- Eaton -3.69
- Vertiv -4.50
Vertiv and Eaton are the only two names in the segment with negative volume readings, on a session when the board's median score was 66 and the whole complex was rising. Turnover in the power names was below their own baseline while AI infrastructure was having a good day.
The attention data says the same thing far more starkly. We count cashtag posts on X ourselves, one row per complete New York day. On 12 August the board carried 32,161 posts. Vertiv drew 88 and Eaton 55 - both inside the eight quietest tickers of the 34. Between them the two power names took 0.44% of the day's conversation. Dell and Super Micro took 6.8%.
That is not a one-day artefact. Across every complete day from 3 to 12 August, the two power names never once exceeded 0.63% of the board's daily posts. The server pair ranged between 2.1% and 9.2%.
One qualification, in our favour rather than against it: $ETN is also the cashtag of a cryptocurrency, so we apply a narrow exclusion filter to that query. Residual contamination would inflate Eaton's count, not suppress it. The true figure is if anything lower than 55.
Meanwhile, the businesses are compounding
This is what makes the quiet strange. These are not struggling companies waiting for a thesis to arrive. Their most recent quarters are among the strongest in the entire AI supply chain.
Eaton, second quarter 2026: revenue $8.5 billion, up 21%, with 14% organic growth. Data centre revenue in Electrical Americas grew 65% organically against an underlying market growing 23%. Electrical Global backlog was up 103% on June 2025. Rolling twelve-month orders rose 41% in Electrical Americas and 33% in Electrical Global, on a book-to-bill of 1.2. The negotiations pipeline was up 60% year to date. The company has committed $1 billion to 24 capacity expansion projects.
Vertiv, second quarter 2026: net sales $3,274 million, up 24% with 18% organic growth. Adjusted operating margin 22.6%, an improvement of 410 basis points. Operating cash flow of $1,100 million, up 241%. Full-year guidance raised to 30-32% organic growth and adjusted EPS of $6.65-6.75, roughly 60% above 2025.
A company guiding to 31% organic growth at the midpoint, with a margin expanding four points, is not a stock the market should have to be reminded about. And yet it drew 88 posts on a day when a $24 billion server maker drew 1,572.
The constraint is real, and it is the most documented fact in AI infrastructure
Nothing above depends on the power story being overstated, because it is not. The physical evidence is unusually well documented:
- Transformers. Lead times for large power transformers have stretched to roughly three to four years. Transformers and switchgear - not turbines, not chips - are now the equipment most likely to delay a US interconnection.
- The queue. Berkeley Lab counted roughly 8,200 projects actively seeking US grid interconnection at the end of 2025 - about 1,312 GW of generation plus 749 GW of storage. The median project that actually reached commercial operation in 2025 had spent 61 months in the queue, against 22 months in 2008. Natural gas capacity in the queue rose 86% to 253 GW, which is the gas-to-power pivot showing up in the data rather than in press releases.
- Turbines. GE Vernova's heavy-duty gas turbines are sold out through roughly 2030-2031 against a backlog near 100 GW. Siemens Energy's gas turbine backlog sits near 70 GW, and its 50% capacity expansion does not arrive until 2030. Siemens Energy's grid order backlog reached a record €51 billion.
- The demand side. The IEA projects global data centre electricity consumption more than doubling from 415 TWh in 2024 to around 945 TWh by 2030, with data centres driving close to half of US electricity demand growth over that period.
- The rack. Nvidia's VR200 NVL72 draws roughly 190-230 kW against 120-130 kW for Blackwell, and the 2027 Kyber generation is specified near 600 kW, with megawatt-class racks behind it. That is the reason 800VDC architecture exists at all.
So the constraint is not in doubt, the suppliers are not underperforming, and the market is not paying attention. Those three things are only compatible under one explanation.
The view: this is a duration trade wearing a momentum trade's clothes
What follows is judgement, not measurement. Our score reads price and volume and has no forward-looking component; nothing in this section comes out of it.
Equity prices discount roughly the next twelve to twenty-four months. The power constraint does not resolve on that clock - it resolves on a three-to-seven-year clock, because that is how long transformers, interconnects and turbines take. A constraint that binds hardest in 2029 does not move a stock in 2026. It moves the stock of whoever converts an order into revenue this quarter, which is exactly what Dell and Super Micro do and exactly what Eaton and Vertiv cannot.
Eaton's own management said this on the July call, and it is the most important sentence in this article. Describing a US data centre backlog of 307 GW - about fifteen years of work at 2025 build rates, up from twelve years at the previous update - they noted that only roughly 20% of this backlog converts near term. The majority will translate to 2028 and beyond deliveries.
Read that as an investor rather than an engineer. A fifteen-year backlog is a magnificent fact about the business and a difficult one about the stock. It says revenue is rate-limited by factory output rather than by demand, which caps the rate at which good news can arrive. There is no quarter in which Eaton surprises by shipping three years of backlog early.
Look at what the score picked up without knowing any of that. Eaton's momentum component is 0.72 out of a possible 12.00 - essentially flat - while the stock sits at 96.3% of its 52-week high. That is the signature of a share price that has already absorbed its story and has nothing left to do but wait for delivery. Super Micro, by contrast, has momentum at the 12.00 ceiling and relative strength at 9.60 of 9.60 while trading at just 64.0% of its high. One of those is a stock discounting the future; the other is a stock that has finished discounting it.
The asymmetry that actually matters
Here is the part we think is genuinely under-appreciated, and it is a structural point rather than a trading one.
Power is the only input in AI infrastructure that money cannot accelerate. Pay up for HBM and you get allocation next quarter. Pay up for advanced packaging and a foundry finds you slots. Pay up for optics and Coherent adds a laser fab. Every other bottleneck in this build-out responds to capital on a one-to-two-year lag.
A grid interconnect does not. Nor does a transformer, a transmission right-of-way, or a turbine slot in 2031. You can pay any price you like and the queue does not move, because the constraint is regulatory approval, physical construction and a supply chain that has to build its own factories first. Hyperscalers have committed over 9.8 GW of nuclear capacity across thirteen announced deals, and only about 1.92 GW of that is operational; the first new capacity, the 835 MW Three Mile Island Unit 1 restart, is targeted for the second half of 2027, and Western SMRs largely arrive from 2029.
That asymmetry is why we think power ultimately caps the pace of the entire build-out - and simultaneously why the listed power names trade like industrials rather than like semiconductors. The same fact produces both outcomes. Most commentary picks one and ignores the other.
What would actually re-rate these stocks
Not another AI demand headline. Demand is the part that is fully priced. These are the specific things that would change the picture, and each is checkable:
- Lead times shortening. Counter-intuitively bullish for the equity. Shorter lead times mean capacity has landed and backlog can convert faster, which is the only mechanism by which a fifteen-year book becomes near-term revenue. Eaton's $1 billion across 24 expansion projects is the thing to watch. There is already a first hint of it: Berkeley Lab found active queue volume fell 10% year on year, as high withdrawal rates cleared out speculative projects.
- The near-term conversion share rising above 20%. Eaton has quantified it. If that figure moves, the duration mismatch is closing.
- Behind-the-meter generation reaching scale. On-site gas is deployable in 12-18 months against years for grid upgrades. Every megawatt built behind the meter routes around the queue entirely.
- The 800VDC transition landing on schedule. Full-scale 800VDC production is meant to coincide with Kyber in 2027. That is a content-per-rack increase for the power vendors, and it arrives inside the window equities actually discount.
- Turnover normalising. Our own volume component answers this every session. Two names with negative volume readings in the middle of an infrastructure rally is the anomaly. Its disappearance would be the first sign the market has re-engaged.
The limits of this
- The score is price-only and backward-looking. It contains no forecast, no news and no fundamentals. It cannot tell you Eaton is cheap or Super Micro is expensive, and it is not being used here to say either.
- Four names is a small segment, and two of them - Dell and Super Micro - are server makers rather than power companies. The comparison is deliberate, but it is a comparison across a segment boundary that we drew ourselves.
- Low attention is not a buy signal. We measured this across the whole board and found post volume has a rank correlation of -0.00 with the direction of the next day's move - the working is here. A quiet stock is not thereby a cheap one. The quiet is evidence about positioning, not about value.
- Two weeks is not a trend. Our sentiment history began on 7 August 2026, with earlier sessions reconstructed from daily bars.
- Third-party figures are cited as published, not as facts about the future, and lead-time and queue estimates in particular vary by source and by region.
- None of this is a recommendation, and it takes no account of anyone's circumstances.
The power and systems segment page carries live scores for all four names, and the daily board has all 34, rebuilt every session.
Frequently asked questions
Is power really the bottleneck for AI data centres?
Yes, and it is unusually well documented. Large power transformer lead times run to roughly three or four years, the median US project reaching commercial operation in 2025 had spent 61 months in the grid interconnection queue according to Berkeley Lab, and GE Vernova's heavy-duty gas turbines are sold out into 2030-2031. The IEA projects global data centre electricity consumption more than doubling to around 945 TWh by 2030. Transformers and switchgear, not turbines or chips, are the equipment most likely to delay a US project today.
What are the best AI data centre power stocks?
The two US-listed names on our board that sell directly into datacentre electrification are Eaton (ETN), which supplies electrical gear and switchgear, and Vertiv (VRT), which supplies power distribution and cooling. The wider universe includes GE Vernova and Siemens Energy in generation and grid equipment, and Schneider Electric in electrical infrastructure. We do not price the last three, so we do not score them.
Why are Vertiv and Eaton stock quiet if power is the AI bottleneck?
Our view is that it is a duration mismatch rather than a mispricing. Equities discount roughly the next twelve to twenty-four months; the power constraint resolves over three to five years or more - the median US project reaching commercial operation in 2025 spent 61 months in the interconnection queue alone. Eaton's management said that of its 307 GW US data centre backlog - about fifteen years of work at 2025 build rates - only roughly 20% converts near term, with the majority translating to 2028 and beyond deliveries. Revenue is rate-limited by factory output, which caps how fast good news can arrive.
How big is Eaton's data centre backlog?
Eaton described a total US data centre backlog of 307 GW on its second-quarter 2026 call, equivalent to about fifteen years of work at 2025 build rates and up from twelve years at the previous update. Its Electrical Global backlog was up 103% on June 2025, data centre revenue in Electrical Americas grew 65% organically, and rolling twelve-month orders rose 41% in Electrical Americas on a book-to-bill of 1.2.
How does Vertiv's growth compare with its share of investor attention?
Vertiv reported second-quarter 2026 net sales of $3,274 million, up 24% with 18% organic growth, and raised full-year guidance to 30-32% organic growth with adjusted EPS around 60% above 2025. On 12 August 2026 it drew 88 cashtag posts on X out of 32,161 across our 34-name board, placing it inside the eight quietest tickers we track. Its volume component that session was -4.50 of a possible 7.20, meaning turnover below its own normal level.
Why can't AI companies just pay more for electricity?
Because power is the one input in AI infrastructure that capital cannot accelerate. Paying up for high-bandwidth memory, advanced packaging or optics buys allocation on a one-to-two-year lag. A grid interconnect, a transformer or a 2031 turbine slot does not respond to price, because the constraint is regulatory approval, physical construction and a supply chain that must build its own factories first. That asymmetry is why power ultimately caps the pace of the build-out.
What is 800VDC and why does it matter for power stocks?
800-volt direct current is the rack power architecture designed for the next generation of AI systems. Nvidia's VR200 NVL72 draws roughly 190-230 kW against 120-130 kW for Blackwell, and the 2027 Kyber generation is specified near 600 kW with megawatt-class racks behind it. Eliminating rack-level AC/DC conversion improves end-to-end efficiency by up to about 5%. For the power vendors it means more content per rack, and full-scale 800VDC production is intended to coincide with Kyber in 2027 - inside the window equities actually discount.
Does low social media attention mean a stock is undervalued?
No. We measured this across all 34 names on our board and found post volume has a rank correlation of -0.00 with the direction of the next day's price move. A quiet stock is not thereby a cheap one. What a low count tells you is something about positioning and crowding, not about value, and it carries no direction at all.
Sources
- ChipSentiment - power and systems segment - the four constituents, live prices and cap-weighted sentiment history
- ChipSentiment - live board - prices, moves, scores and score components for all 34 names, quoted as at the close on 12 August 2026
- ChipSentiment - X mentions board - daily cashtag post counts, counted by this site against the X API; figures quoted for 3-12 August 2026
- ChipSentiment - do X mentions predict stock moves? - the -0.00 rank correlation between post volume and next-day direction
- Eaton - record second quarter 2026 results - revenue $8.5bn up 21%, 14% organic, Electrical Global backlog up 103%, book-to-bill 1.2, orders up 41% and 33% rolling twelve months, full-year guidance
- Eaton - Q2 2026 earnings call transcript - 307 GW US data centre backlog, about fifteen years at 2025 build rates, 'only roughly 20% of this backlog converts near term', 65% data centre organic growth, $1bn across 24 capacity projects
- Vertiv - second quarter 2026 results - net sales $3,274m up 24%, 18% organic, adjusted operating margin 22.6%, operating cash flow up 241%, full-year guidance 30-32% organic growth
- Lawrence Berkeley National Laboratory - Queued Up: 2026 edition - roughly 8,200 active US interconnection requests at end-2025, 1,312 GW generation plus 749 GW storage, median 61 months from request to commercial operation for projects built in 2025, active queue volume down 10%, natural gas capacity up 86% to 253 GW
- IEA - AI and data centre electricity demand - 415 TWh in 2024 rising to around 945 TWh by 2030, and data centres as a share of US electricity demand growth
- Utility Dive - Siemens Energy gas turbine backlog nears 70 GW - turbine backlog, three years or more as normal delivery, capacity expansion arriving 2030
- Power Engineering - GE Vernova orders and turbine slots through 2030 - gas turbine backlog and slot availability into the 2030s
- NVIDIA - building the 800VDC ecosystem for AI factories - 800VDC architecture, efficiency gains and the vendor ecosystem
- DataCenterDynamics - Nvidia prepares the industry for 1MW racks and 800VDC - rack power trajectory from Blackwell through Kyber in 2027
Figures are taken from the public filings and the reporting linked above.