"Sovereign AI" — nations, government agencies and regulated industries buying AI infrastructure they control end to end — spent most of 2024 as a keynote phrase. Something executives said on stage, next to a slide with a globe on it. I spend my days inside a database of 253,129 earnings call transcripts, and the transcripts say that phase is over.

In Q1 2024, 53 companies matched a full-text search for "sovereign AI" on their earnings calls. In Q2 2026, 252 companies did. That's nearly 5x overall — and almost the entire move happened in the six quarters from the start of 2025, when the count ran 93 → 97 → 149 → 177 → 244 → 252. A buzzword doesn't do that. A procurement category does.

One methodology note, once, so we're honest with each other: every count in this post is the number of companies whose earnings calls in a given quarter match a full-text search for the exact phrase "sovereign AI." Keyword matching is blunt by design — these counts are an upper bound, and some matches are incidental co-occurrences rather than substantive discussion. The trend direction survives that noise easily. Individual data points, treat with one grain of salt.

Ten Quarters of Data: A Plateau, Then a Break

Here is the full series, per quarter, since the phrase first showed up in meaningful numbers:

Quarter Companies matching "sovereign AI" QoQ change
2024 Q1 53
2024 Q2 59 +11%
2024 Q3 57 −3%
2024 Q4 71 +25%
2025 Q1 93 +31%
2025 Q2 97 +4%
2025 Q3 149 +54%
2025 Q4 177 +19%
2026 Q1 244 +38%
2026 Q2 252 +3%

Read the shape, not just the endpoints. All of 2024 is a plateau: 53, 59, 57, 71. The phrase existed, a handful of infrastructure vendors used it, and nothing compounded. The break comes in the second half of 2025 — +54% in Q3 2025 alone, then two more big legs at +19% and +38%. Year over year, Q2 2026's 252 companies are 2.6x the 97 from Q2 2025.

And the flattening at the top isn't the story dying. 252 companies in a single quarter, against the roughly 10,000 calls we process per quarter, means about one in forty reporting companies now discusses sovereign AI every three months. That's not a spike. That's a new baseline.

I've watched this exact curve before. When I tracked "agentic AI" spreading through 2,400 earnings calls, the pattern was identical: flat, flat, step change, new plateau. The step change is the tell. It marks the quarter a phrase stops belonging to keynote speakers and starts belonging to sales pipelines — because the fastest way to get a term onto hundreds of earnings calls is for customers to start putting it in RFPs.

When Vendors Segment Their Pipeline By It, It's a Category

Counts tell you the conversation grew. Who is having the conversation tells you what it became.

Since April 2026, the companies mentioning sovereign AI most often in our database are Fortinet, Super Micro, Rackspace, BCE, S&P Global and HPE. Look at that list for a second: a security vendor, a server maker, a managed-cloud provider, a Canadian telecom, a data-and-ratings house, and a systems integrator. Not one of them is a pure GPU story. When the same phrase shows up across security, hardware, hosting, telecom and data, it isn't a product anymore — it's a buyer type that cuts across the entire stack.

The clearest evidence is how vendors now slice their own demand. The standard taxonomy on 2026 infrastructure calls has four buckets: hyperscalers, neoclouds (the GPU-rental clouds), enterprises — and sovereigns. Sovereigns get named as their own line, with their own requirements, the way "federal" has been its own segment in enterprise software for decades.

HPE said the quiet part in prepared remarks. On its June 16, 2026 earnings call, management described the offering in pure compliance language (quote lightly cleaned from the machine transcript): "Our sovereign factories now include federal-grade security hardening, federal compliance readiness, validated encryption standards and global data protection requirements all built in." And then named the buyers outright — defense, government and financial services — telling them "this is the sovereign AI architecture you have been waiting for." That is not a vision statement. That is a spec sheet read aloud to three named buyer segments.

The sell side has caught up, too. On Fortinet's June 2, 2026 call, an analyst set up a question about where demand lands in this AI cycle by pointing out that "this time, you spoke about sovereign AI" — the analysts are now probing for the category by name, quarter over quarter, and tracking whether management's language shifted. When a security vendor gets asked about sovereign AI as a demand driver, the term has fully escaped the GPU aisle.

Super Micro completes the picture. On its latest call, analysts pressed management on customer diversification by asking how many of its $1-billion-plus customers are "Neocloud versus enterprises versus sovereign AI customers." Think about what that question presumes: that sovereign buyers are numerous enough, and large enough, to be a standing category among billion-dollar customers. Nobody asks that about a slogan.

This is also exactly what you'd expect if the AI capex supercycle is broadening rather than peaking. After the hyperscalers, the next marginal buyer of AI infrastructure is, almost by definition, someone who couldn't use hyperscaler capacity in the first place — because a law, a regulator or a national security policy says the data and the weights cannot leave a jurisdiction. That buyer was always going to arrive late, and arrive with a compliance checklist instead of a benchmark chart.

93 Financial Firms Aren't Selling Sovereign AI — They're Complying With It

The Q2 2026 sector split is where this post earns its title. Companies matching "sovereign AI" by sector:

IT leading is no surprise — that's the sell side of the trade, the vendors we just discussed. The block that made me stop scrolling is Financials: 93 companies — banks, insurers, asset managers — discussing sovereign AI in a single quarter.

Financial firms are not selling sovereign AI. They have no GPU revenue line. When a bank or an asset manager says "sovereign AI" on an earnings call, it is talking about a requirement: where the models run, where the data sits, who holds the encryption keys, and what the regulator gets to audit. Data residency. Validated encryption standards. Auditability of the full stack. It's the same posture these firms have always had toward core banking systems, extended to AI — and it means a meaningful slice of the sovereign-AI conversation is demand-side, not vendor marketing.

That matters for durability. Compliance demand is the stickiest demand there is. It doesn't churn when the hype cycle cools, because the requirement is written into regulation and risk frameworks, not into a product roadmap. HPE naming financial services in the same breath as defense and government — in the quote above — is a vendor telling you it has noticed the same thing: the regulated buyer signs the longer contract.

Even the tail of the sector split is informative. Real Estate at 11 companies in a quarter is small, but the phrase reaching that sector at all hints at where the physical layer — the buildings, the power, the local footprint — enters the conversation. Two years ago the number would have rounded to zero.

The Playbook: Run the Search on Your Own Watchlist

Aggregate counts tell you a category exists. To act on it, you need to know which companies in your universe are naming actual deals and which are name-dropping. That's a two-step process, and the first step takes about ten seconds:

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

The q parameter supports exact phrases in quotes, AND/OR operators, and - negation — so you can tighten the net. "sovereign AI" AND backlog finds companies putting the phrase near a number. "sovereign AI" AND "data residency" isolates the compliance conversation. Scope by sector or by your own ticker list, and the 252-company haystack collapses to a readable stack of calls. Full parameter reference is in the API docs, and the same search runs through our MCP server if you'd rather have an LLM do the reading with you.

Step two is the part no API does for you: read the hits and sort revenue from rhetoric. My working heuristics, from doing this weekly:

None of this is exotic — it's the same discipline behind the five text signals quants pull from earnings calls. Language on calls changes before segment disclosure does, and the gap between the two is where the information lives.

The Caveat That Keeps Me Honest

I sell transcript search, so believe me when I say it: mention-counting finds the conversation, not the revenue. "Sovereign AI" is also a marketing phrase, and by 2026 every marketing department knows it. Some of the 252 companies are booking sovereign deals with governments and regulated buyers. Some are hoping proximity to the phrase does something for their multiple. A keyword match cannot tell those two apart — a human (or an agent) reading the surrounding paragraph can, in about thirty seconds per call.

Six quarters ago, this was 93 companies and a fair question about whether the phrase meant anything. Today it's 252 companies, a standing pipeline segment at the largest infrastructure vendors, an analyst checklist item, and a compliance requirement discussed by 93 financial firms in a single quarter. The counts tell you where to look. The calls tell you what's real. Read them.

Search all 253,129 earnings call transcripts — including every "sovereign AI" mention back to 2020 — via web, REST API and MCP at earningscalls.dev.