Power & Paperwork Beat AI Models: $1.3B in Deals

So, What Actually Happened?

Saturday, and I keep circling the same odd thing: almost none of this week's money went to a model. A company whose entire job is deciding when a data center throttles its own power draw crossed a billion-dollar valuation. Twenty minutes down the same road, a second power startup sold for $232.28 million. We scanned 190,000 articles this week so you don't have to. Then Brussels moved the AI Act out of drafting and into enforcement, and identity verification, of all the unglamorous categories, repriced to $5.2 billion. I opened the tab expecting chip news. Most of what I actually read was about electricity and paperwork.

The Bottom Line: Buyers have stopped paying for the clever part. They are paying for the electricity, the ID check and the lawyer, because those are the three things that can stop a deployment dead on a Tuesday.

The Tracks That Matter

1. A Data Center's Power Bill Just Minted A Unicorn

Emerald AI sits between a data center and the grid and decides when the building can back off its own draw. That job just reached a $1.05 billion valuation. There is no model in the pitch. Down the same Virginia corridor, Claros, another company built entirely around data center power, sold for $232.28 million. What both of them really sell is permission to plug in, and permission is the scarce good now, because the utility interconnect queue does not care how good your inference stack is. A third round landed on the same floor of the same building: Quintessent took $40 million for datacenter lasers, the unglamorous work of moving light between racks fast enough to keep expensive silicon fed.

Here's what works: Ask your data center provider for their interconnect queue position and the megawatt delivery date. Rack space you cannot power is a line item, not a capacity plan.

2. Socure Hits $5.2 Billion As Agents Break Identity

Socure raised $156 million at $5.2 billion and acquired an AI startup called Fravity on the same day, which tells you this round was not about growth, it was about buying capability before the window closed. Identity verification was a compliance checkbox for a decade: run the passport, tick the box, move on. Agents broke that. When software can hold a conversation, clear a document check and then act on an account, ”is there a human here” stops being a yes-or-no question at signup and becomes a thing you have to keep asking. Inside companies the same gap already has a name: shadow identities, the service accounts and tokens nobody put on an org chart.

Here's what works: Count the non-human identities in your environment this week. Service accounts, API keys, agent credentials. If nobody owns that number, nobody owns the blast radius.

3. Europe Stops Writing AI Rules And Starts Enforcing Them

Brussels moved the AI Act from legislation into enforcement, and the shift is duller and more expensive than the two years of commentary that preceded it. Legislation is a document your counsel can summarise in a memo. Enforcement is a named authority with staff, a docket and an appetite for a first case, and first cases get picked for how legible they are, not how bad they are. Companies are pricing that in already: KPMG rolled out Microsoft's Agent 365 across its global workforce, which stripped of the marketing is an audit trail for what the agents did and who let them. Nobody buys a control plane because it is exciting.

Here's what works: Pick your single highest-exposure automated decision and write down who approved it, what data it used and who reviews it. That page is what an inquiry asks for first.

Quick hits:

• Einride's founders raised €450 million for European deep tech. The team behind the autonomous freight company launched Navisalma with a €450 million fund, pointed at hardware and industrial systems rather than another chat layer, which is a European capital pool aimed squarely at the physical bottleneck everyone just discovered.
• A Chinese voice-AI firm's token business grew 760%. Unisound's interim results show revenue up 38.7% year over year with token consumption running far ahead of it, which is what commercialisation looks like before margin shows up.
• SAP's own customers are building AI without SAP. The DSAG investment report finds most large SAP shops now run production AI on third-party stacks rather than SAP's tools, and there is no technical requirement forcing them back.

Signal vs. Noise

🟢 Signal: Governance moving into the budget. Governance was written about less this week and mattered more, the pattern you get when a subject stops being an opinion column and becomes a standing agenda item with money attached. An enforcement desk in Brussels and a control-plane rollout at KPMG are the same move approached from two directions. Most coverage is still counting policy announcements instead of watching who now has to sign off.

🔴 Noise: ”Data governance” as a phrase. It pulled the heaviest volume of the day and kept losing its grip on the things it used to hold together, including this week's actual governance news. McKinsey published a fresh case for governance that delivers value straight into that gap. When the label is loud and the substance keeps moving elsewhere, the label is where the disappointment turns up later.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

Emerald AI crossed a billion dollars for managing when a data center pulls power, Socure repriced to $5.2 billion for checking whether a human is on the other end, and Brussels moved the AI Act into enforcement, all inside 48 hours.

Read alone, each one belongs to a different desk. The energy trade press files the first as a grid story. The fintech desk files the second as fraud tooling. The legal press files the third as a compliance milestone and moves on. Put them on the same morning and they stop being three stories. They are the three things that can actually stop an AI deployment: you cannot power it, you cannot prove who is operating it, or you cannot defend it when asked. All three got priced in the same window, and none of that money went anywhere near a model.

What changes Monday is narrow. Take your biggest planned AI workload for next year and write one sentence for each of those three: where the power comes from and when, who or what is authorised to act, and who signs the decision. If any of the three is a shrug, that is your real delivery date, not the one in the roadmap.

By The Numbers

• Emerald AI reached a $1.05 billion valuation for managing data center energy demand — a unicorn built on when you draw power, not on what you compute with it.
• Claros sold for $232.28 million — the second data center power exit in one week, in the same corridor, which is what the start of consolidation looks like.
• Socure raised $156 million at a $5.2 billion valuation and bought Fravity — identity checking is now priced as AI infrastructure rather than as a compliance cost line.
• Navisalma launched with a €450 million deep tech fund — European capital moving toward hardware and industrial systems while US rounds chase the application layer.
• Unisound reported revenue up 38.7% with its token business up 760% — usage is outrunning revenue by a factor of twenty, which is either land grab or margin problem, and you find out later.
• See what's rising across AI and data this week →

Deep Dive: The Generator Behind The Main Stage

I played an outdoor set years ago where the entire main stage ran off a diesel generator parked behind a fence, out of sight of every person on that field. Halfway through the night it browned out. Twenty thousand people, a rig worth more than the farmland it was standing on, and all of it stopped because of a machine nobody had looked at since load-in.

The money moved behind the fence

Emerald AI is a billion-dollar company for scheduling a power draw. Claros sold for $232 million on the same problem. Quintessent raised $40 million to move light between racks. None of them touch a model, and all three were funded by people who have worked out that the model is not what runs out first.

The second generator is a lawyer

The other thing that browns out is permission. Brussels now has enforcement staff instead of drafting committees. Socure just repriced because proving a human is present became a live question rather than a signup form. Both are the same constraint wearing different clothes: you can have the capability and still not be allowed to run it.

Nobody puts the generator on the poster

Which is why ”data governance” keeps pulling volume while the actual governance work quietly relocated into power contracts, identity systems and audit trails. The phrase stayed on the poster. The load moved behind the fence, and almost nobody rewrote the budget to follow it.

What Actually Works

1. Split your AI plan by constraint, not by use case: power, identity, liability. Every stalled deployment I have seen this year failed on one of those three, never on model quality.
2. Put a date on your megawatts: contract signature and physical delivery are two different numbers, and only one of them turns on.
3. Inventory your non-human identities: agents, service accounts, tokens. Name an owner for each before somebody else names them for you.
4. Write the one-page decision record now: what the system decided, on what data, who approved it. Build it while it is cheap, not while it is being requested.

Nobody has ever bought a ticket to see a generator. Everybody goes home early when one fails.

What's Coming

The Power Layer Starts Consolidating

Two data center power exits in one corridor in one week is not a coincidence, it is a bid table forming. After Claros sold for $232.28 million, expect the hyperscalers and the large developers to start buying grid-interface capability outright rather than contracting for it. The independents in this space have about two quarters of leverage left.

Enforcement Gets A First Name

The AI Act's move into enforcement means somebody has to open the first file. Watch which sector it lands in, because the first case sets the reading of the text for everyone else, and regulators pick openers they can win. Employment screening and credit decisioning are the two most legible targets on the board.

Agent Identity Becomes A Budget Line

KPMG putting Agent 365 across its global workforce is the professional services firms doing what they always do: buying the control layer first, then selling the readiness assessment. Expect agent identity and audit tooling to show up as its own procurement line in Q4 budgets, split out from the AI spend it used to hide inside.

For Your Team

Monday's meeting prompt: ”Take our biggest AI commitment for next year. Can we say, in one sentence each, where the power comes from and when, what is authorised to act on our behalf, and who signs the decision? If any of those three is a shrug, is our timeline real?”

Share-worthy stat: A company that manages when a data center draws electricity is now worth $1.05 billion, and a company that checks whether a human is on the other end of a session is worth $5.2 billion. Neither one trains a model. Both were repriced this week.

Go deeper: Track where AI infrastructure money is actually landing →

The Track of the Day

”Token panic”: the spend on generative model tokens is starting to outweigh the value coming back out.
E3 Magazine, on why SAP's own customers are building their AI without SAP

That is the sentence sitting underneath everything above. When the meter runs faster than the return, the money stops shopping for cleverness and starts shopping for whatever keeps the meter running at all. This week that was electricity, identity and a paper trail.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 29, 2026 | Curated by Yves Mulkers @ Ins7ghts