Tekever flew drones over Ukraine while Washington drafted rules for machines that don't exist

So, What Actually Happened?

Thursday, and I keep circling two things that landed on the same Wednesday and do not look like they belong in the same newsletter. In Washington, Sanders moved to ban superintelligence and build a federal Department of AI. In Lisbon, a drone company raised more money than most European software rounds see in a year. We scanned 190,000 articles this week so you don't have to, and the money kept pointing the same direction. Tekever hit a $6.4 billion valuation on aircraft already flying over Ukraine. Basecamp Research raised $140 million to design medicines. A Chinese robotics seed round priced physical AI at $500 million before shipping anything. Nobody in the hearing room is arguing about any of that.

The Bottom Line: The rules being drafted this week describe a machine nobody has built. The capital went to machines that already fly and molecules already in a lab.

Advertisement

The Tracks That Matter

1. Tekever Raised $580M For Drones Already Flying

Tekever closed $580 million and came out at a $6.4 billion valuation, which makes the Portuguese maker one of Europe's larger defence-tech names on paper. The multiple is not the interesting part. The reference is: these aircraft already fly reconnaissance over Ukraine, so the round is priced on delivered capability rather than a roadmap, which almost no AI round can claim. CEO Ricardo Mendes put sovereign technological capability at the centre of the pitch, and the announcement tied the raise to European industrial scale rather than to model benchmarks. Every valuation comparison your team runs for an AI company now sits next to one judged on whether the thing comes back.

Here's what works: Before your next AI vendor review, ask what the product does when it fails. A wrong sentence and a wrong flight path are not the same purchase.

2. Washington Produced Two Opposite AI Laws In One Week

Sanders unveiled a bill banning artificial superintelligence outright, with licensing, liability and a federal Department of AI behind it. Days earlier the President stood at the UN and rejected AI regulation, comparing it to climate rules, then announced he would rename the whole field superintelligence. California is not waiting either: Newsom's order requiring independent human monitors and a kill switch sets up a direct fight with the federal position. So that is three incompatible regimes drafted inside one week. Not one of them mentions the agent that already holds your production credentials.

Here's what works: Write your AI control document against your own risk, not against a bill. Every draft on the table regulates capability, none of them regulates access.

3. Your Coding Agent Already Holds The Repo Keys

Z.ai made the uncomfortable part concrete: an AI coding assistant holds the keys to the repo, the same credentials a senior engineer holds, and uses them faster and with no memory of why they were granted. The same day, Arista urged immediate patching of an exploited zero-day in its orchestrator, which is the familiar shape of this problem. The new shape is worse, because the thing with the keys now acts on its own schedule. Meanwhile SOCi put 400,000 agents into production for local marketing. That is 400,000 identities nobody is offboarding.

Here's what works: Pull the list of non-human identities with write access to production this week. If the agents are not on that list, the list is wrong.

Quick hits:

• A seed round priced physical AI at half a billion. Xirang Kaiwu raised at a $500 million valuation before shipping product, which tells you what robotics capacity is worth to Chinese capital right now.
• Sweden priced a breach of 2.2 million people at $183,000. The regulator fined Miljödata that amount, roughly eight cents per exposed person, which answers the question of whether fines change supplier behaviour.
• Your reporting layer may have changed owners. Progress Software completed its acquisition of Domo's platform business, so anyone with Domo in the stack now has a new roadmap and a new renewal conversation coming.

Advertisement

Signal vs. Noise

🟢 Signal: agents that hold real credentials. Agents gained more genuine hold on everything else this week than any other idea we track, and the evidence is operational rather than rhetorical: 400,000 of them running in production at one marketing vendor, coding assistants carrying repo-level access, consultancies signing deals to convert existing work into agent workflows. Most coverage is still scoring agent demos on capability.

🔴 Noise: superintelligence. The word pulled enormous volume this week, a Senate bill, a UN speech, a run of doomsday explainers, and none of it attaches to a system anyone can point at today. The incidents operators actually handled this quarter came from ordinary agents holding ordinary passwords.

From the 190K

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

Basecamp Research raised $140 million to finish a trillion-entry gene atlas, Ambitious Bio launched a healthy human protein atlas, and Flagship Pioneering took a Department of War contract for biological countermeasures, all inside one 48-hour window.

Three different desks read these apart. Biotech covers the Basecamp round and its EDEN models. Research press covers a more complete map of the human body. Defence press covers a first-of-its-kind countermeasures contract. Read on one morning they say something else: the scarce input in AI for biology stopped being the model and became the measurement. Somebody has to physically go out and record nature, and whoever owns that recording sets the price for everyone training on it. Basecamp is not selling a better architecture, it is selling data nobody else collected. And the buyer at the far end of that chain is now a defence ministry, which changes who gets export-controlled and how fast.

What changes Monday: if your AI strategy assumes data is the commodity and the model is the moat, check which of your suppliers actually owns its own measurements. That is where the pricing power quietly moved.

Advertisement

By The Numbers

• Tekever raised $580 million at a $6.4 billion valuation — Europe's defence-tech benchmark is now a company priced on aircraft already flying, not on a roadmap.
• Basecamp Research hit an $800 million valuation on $140 million — AI-designed medicines funded on a proprietary map of nature rather than a bigger model.
• SOCi put 400,000 agents into production — the largest named agent deployment of the week, and a number your identity team has not seen yet.
• Sweden fined Miljödata $183,000 over a breach affecting 2.2 million people — about eight cents per person, the current market price of losing a national dataset.
• One AI memory stock is up more than 650% in 2026 — the physical side of the AI trade is outrunning the software side, and memory is where the shortage bites first.
• See what's rising across AI and data this week →

Deep Dive: When AI Stops Being Software

I started bringing a PC along to play MP3s, and one night the machine did not boot. I played the whole night off the vinyl I had with me. The new system was better in every way except the one that mattered that evening.

A wrong sentence and a wrong flight path are different products

Software failure has been cheap for twenty years. Roll back, apologise, ship a fix. Tekever's customer cannot roll back a sortie, Basecamp's cannot roll back a molecule, and an agent with production credentials cannot unmake the call it already made. The risk category changed while the vocabulary stayed put.

The moat moved to measurement

Two years of argument about model quality, and the rounds closing this week paid for recordings of the physical world: gene atlases, protein maps, flight telemetry from a live conflict. Models are converging and getting cheaper. What is not converging is who bothered to go out and measure something nobody else has.

The rules are written for the last product

A superintelligence ban, a UN speech against regulation, a state kill-switch order. All three regulate capability, meaning how clever the thing is allowed to become. None regulates access, meaning what it is allowed to touch. Nearly every incident an operator handles this year will be an access failure.

What Actually Works

1. Inventory non-human identities: every agent, service account and assistant with write access to production, each with a named human owner beside it.
2. Ask vendors what failure costs: a tested rollback path is a product feature now. If the answer is ”we retrain”, the risk sits with you.
3. Check who owns the measurement: for each AI supplier, is the advantage a model anyone can rent, or data they collected themselves?
4. Write controls against access, not capability: scope, credentials, and an off switch someone has actually pulled in a drill. Capability rules change with elections. Access rules do not.

The turntables still worked when the computer did not. Keep something in the room that does not need the new system to be up.

What's Coming

Sovereign AI Becomes A Budget Line

Sovereign AI now has its own buyer survey, which is what happens just before something becomes a procurement category rather than a talking point. Expect national-preference language in tenders you are already bidding on, and expect it to arrive as a technical requirement rather than a political one.

Your Auditor Arrives Before The Regulator

Deloitte's 2027 hot topics for technology and digital risk are already organised around AI and autonomous systems. Audit committees move considerably faster than legislatures. The first hard questions about your agent estate will come from inside the building, not from Brussels or Washington.

The Middle Of The Market Gets Starved

Series A megarounds keep climbing and AI startups are absorbing most of the $100 million-plus rounds. Fewer, larger bets means less oxygen for everyone in between. The mid-stage suppliers you already depend on are the ones worth diligencing before renewal season.

For Your Team

Friday's meeting prompt: ”List every AI agent, assistant and service account with write access to something we cannot undo. Who owns each one by name, and what happens if it acts wrong at three in the morning?”

Share-worthy stat: Sweden fined a company $183,000 over a breach that exposed 2.2 million people. That is roughly eight cents a person. Every vendor risk conversation should open with what the downside actually costs the supplier, because right now it costs them very little.

Go deeper: Track where AI money and AI risk are actually moving →

The Track of the Day

”Artificial intelligence, autonomy and sovereign technological capability will be fundamental to the security and resilience of democratic societies in the decades ahead.”
Ricardo Mendes, on the round behind Tekever's drones

A DJ can drop a bad record and the floor forgives it by the next track. Nothing funded this week gets a next track.

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

Published: September 24, 2026 | Curated by Yves Mulkers @ Ins7ghts