Power Is the New Bottleneck: What 2,000+ Earnings Calls Say About the Grid

August 17, 2026

For two years the AI buildout was a chip story. You measured it in GPU allocations and HBM supply, and the companies that mattered were the ones etching silicon. Sometime over the past few quarters, that stopped being the binding constraint — and you can watch the shift happen in earnings call language.

I spend my working hours inside a database of 253,129 earnings call transcripts covering 12,799 companies, roughly 10,000 calls per quarter. When a constraint moves, the words move with it, across every sector at once. This summer the words moved to electricity.

The macro backdrop is familiar from public reporting: sell-side estimates have the top hyperscalers nearly doubling capex in 2026, with total spend heading toward $1 trillion in 2027. I traced the spending side of that story in an earlier piece on the AI capex supercycle. What that surge does, mechanically, is bid away finite resources — grid capacity, engineering talent, construction labor, key materials. The transcripts show exactly where it bites first.

A chip company now talks like a utility

Start with the strangest data point of the quarter. On its August 11 call, AMD — a company whose earnings calls used to be about process nodes and roadmaps — described the gating items for hyperscale AI deployments like this:

"...making sure we have land and power and shell and capital commitments for all the folks to actually deploy this stuff. We'd love to be able to do the first full gigawatt with Anthropic in 2027."

Read that word order again: land and power and shell — then capital. A semiconductor company is walking through a real estate and utilities checklist before it gets to money, and it's sizing a customer deployment in gigawatts, a unit that belongs to power plants. The vocabulary of chip calls has become the vocabulary of grid interconnects, and vocabulary on earnings calls is never accidental. Management teams say what they believe explains the numbers.

If one call were doing this, it would be a curiosity. The corpus says it's everywhere.

What ten quarters of transcripts show

I ran three full-text searches over every earnings call in the archive, quarter by quarter. The counts below are companies whose calls matched each query in that quarter. One methodology note, once: this is keyword matching, so the counts are an upper bound that includes incidental co-occurrences — a call that mentions "data center" in one paragraph and "power" in another still matches. That noise is roughly constant over time, which is why the trend is the signal. I ran the two narrower phrases at checkpoints rather than every quarter, hence the gaps.

Three things in that table.

First, the level shift. Data-center-plus-power language sat in a band around 1,500 companies per quarter through 2024. It crossed 2,000 in Q4 2025 and has stayed there — 2,290 in Q1 2026, 2,174 in Q2. That's roughly a 50% expansion in ten quarters, and it's where the 2,000+ in the headline comes from.

Second, "power constraints" is at an all-time high — twice in a row. 940 companies in Q1 2026, 944 in Q2. Both are the highest readings in the archive. Against 687 in early 2024, that's over a third more companies telling investors that electricity availability is a named constraint on their business.

Third, the dips are mostly mechanical. The 2025 Q2 trough partly tracks corpus volume — that quarter had 8,270 calls against a typical ~10,000. Q2 2026, by contrast, sits on 10,053 calls, so the 944 reading needs no excuse.

The theme doesn't live in tech

Here's the part that would be invisible if you only read hyperscaler calls. The Q2 2026 sector split for "grid capacity" mentions:

• Industrials: 264 companies
• Utilities: 140
• Materials: 131
• Information Technology: 106
• Energy: 67
• Financials: 43

Information Technology is fourth. The companies talking about grid capacity are overwhelmingly the ones who build, wire, and feed the grid — not the ones renting compute on top of it. That's what a supply-side constraint looks like in transcript data: the conversation migrates from the buyers to the suppliers.

The top individual mentioners since April 2026 make the same point: Siemens Energy, DTE Energy, BorgWarner, Bentley Systems. A grid equipment maker, a utility, an auto supplier, an infrastructure software company. Not one of them would appear on a hand-picked list of AI names, and that's precisely why the full-corpus query earns its keep. Siemens Energy's August 5 call is the single most frequent "grid capacity" mentioner in the archive since April — which is exactly what you'd expect when transformers and switchgear become the scarce input.

Utilities at 140 deserve a second look, too. Utility earnings calls were, for decades, the sleepiest documents in the corpus. Now they field analyst questions about hyperscaler load letters and interconnection queues quarter after quarter. When the boring sector becomes the interesting one, the constraint has moved.

The second-order winners are already printing numbers

Constraints create queues, and queues create pricing power for whoever sells the workaround. The clearest example this season is Caterpillar's August 4 call:

"In Power and Energy, sales to users grew a robust 33%. Power generation grew 72%, driven by very strong demand for large gen sets and turbines used in data center applications."

Power generation up 72%, explicitly attributed to data centers. When the grid can't deliver an interconnect on the timeline a data center developer needs, the developer buys generation — large gen sets, gas turbines, whatever ships this year. Caterpillar is what "power is the bottleneck" looks like on an income statement.

The pattern repeats down the supply chain. Grid equipment makers like Siemens Energy field questions about order backlogs. Utilities like DTE Energy get asked about AI-driven load growth on calls that used to be about rate cases. None of these companies designed a chip, and all of them are now part of the AI trade — the transcripts found them before any curated watchlist would have.

The nuclear counterpoint: markets talk about what bites this year

Now the control group. If power scarcity is the theme of 2026, you might expect small modular reactors — the most-hyped long-term fix — to show the same hockey stick. They don't.

Companies matching "small modular reactor": 28 in Q1 2024, 44 in Q2 2026. Over the same ten quarters in which "power constraints" mentions grew by more than a third and crossed 900, SMR talk stayed essentially flat.

The mentions that do exist are real business — Curtiss-Wright's August 6 call noted:

"Growth in the Power and Process market was mainly driven by increased revenues in the commercial nuclear market supporting advanced small modular reactors."

But the divergence is the story. Earnings calls are a quarterly document about the next twelve months, and executives spend words on what bites now. The 2026 constraint is solved with gen sets, gas turbines and grid interconnects; nuclear is discussed as a 2030s solution, in future tense, by a handful of specialists. When SMR mentions start climbing the way "power constraints" did in 2025, that will be a genuinely new signal. Right now the corpus says: not yet.

This gap between a theme's press coverage and its transcript footprint is one of the more reliable reality checks I know — it's cousin to the mention-count signals I covered in five text signals quants pull from earnings calls.

Build your own power-constraint watchlist

Everything above is reproducible with one endpoint. Here's the query that pulls every transcript since April 2026 mentioning data centers and power:

curl 'https://earningscalls.dev/api/v1/search?q="data center" AND power&type=transcripts&date_from=2026-04-01' -H "X-API-Key: $KEY"

The q syntax supports "exact phrase" matching, AND/OR operators, and -negation — so you can tighten the net with something like "power constraints" OR "grid capacity" -solar and rerun it per quarter. The full parameter reference is in the docs, and if you'd rather have Claude run the queries conversationally, the same search is exposed through the MCP server.

A practical setup that takes an afternoon: pick three phrases, query them per quarter for a baseline, then schedule the same queries weekly during earnings season and diff the company lists. New names appearing on the "power constraints" list are companies whose management just started telling investors about an electricity problem — or an electricity opportunity. I walked through the general pattern in how to build an earnings call theme tracker in Python; the power version is the same loop with different phrases.

What I'm watching next

Two numbers, both cheap to track.

Does "power constraints" set a third consecutive high in Q3? Two record quarters is a level shift. Three starts to look like the new normal, and it would say the interconnect queues aren't clearing.

Does the SMR count break out of its 28-to-44 band? Flat mention counts mean nuclear remains a slide-deck solution. The quarter that changes is the quarter the market starts believing the timeline.

The broader lesson is the one this database teaches me every season: reading a single earnings call tells you what one management team wants you to believe. Querying all 253,000 of them tells you where the constraint actually sits. In 2024 the constraint vocabulary was chips. In 2026 it's land, power and shell — and the companies saying it loudest are the ones holding the transformers.

Run these queries yourself at earningscalls.dev — 253,000+ transcripts, full-text search with phrase and boolean support, one GET request per quarter of grid history.