FDA Named The Manufacturer, Not The Vendor

So, What Actually Happened?

Wednesday, and the thing I keep circling back to is a warning letter. Not a funding round. A warning letter. We scanned 190,000 articles this week so you don't have to, and the money was the loud part: a Dutch chip startup raised more than €200 million to attack what AI costs to run, not what it can do. Cisco put Nvidia compute behind Splunk and started selling the ability to watch your agents. A startup nobody has heard of shipped agent observability the same morning. Three separate desks, three separate stories. Then underneath all of it, the FDA named an AI tool in a GMP warning letter and addressed the letter to the manufacturer, not to whoever sold the software.

The Bottom Line: Nobody paid for a smarter model yesterday. They paid for proof of what the model did, and for a name to put on it when it goes wrong.

The Tracks That Matter

1. Eindhoven Raised €200M Against What AI Costs To Run

EUCLYD, out of Eindhoven, raised more than €200 million to go at what its founders call the AI efficiency wall, with Samsung and EQT in the round. Read the pitch and it is not about making models smarter. It is about the memory bandwidth that decides what a single served request costs you. That is an unusual thing for European capital to write a cheque that size for, and it tells you where the pain moved. The training bill is a decision maybe six labs in the world make. The inference bill is a decision you make, every day, on every request, and it lands in the same line as your cloud spend. The team comes out of the Dutch chip cluster, ASML's neighbourhood, which is the part most of the coverage buried below the fold.

Here's what works: Ask your platform team what a thousand requests cost to serve today against January. If nobody owns that number, that is the finding.

2. Cisco Is Selling You The Ability To Watch Your Agents

Cisco put Nvidia compute behind Splunk, and the pitch attached to it is not speed. It is governing AI at scale, which in practice means keeping a record of what your agents did. Then look at who shipped the same morning. Airrived, a fraction of Cisco's size, launched observability for enterprise AI agents with very nearly the same sentence in the release. When the incumbent and the startup describe the product identically on one Tuesday, the category has already formed and you missed the forming. What is being sold here is not monitoring. It is the ability to answer a question after the fact, which is exactly what an auditor asks and exactly what your model vendor cannot answer on your behalf.

Here's what works: Name the person who has to answer ”what did it do at 3am” for your next agent pilot. Check they have a tool, not a log file.

3. The FDA Named The Manufacturer, Not The Software Vendor

A warning letter is a dull document until you read who it is addressed to. The FDA put an AI tool in a GMP warning letter and the accountability landed on the manufacturer running it, not on the company that sold it. Every vendor contract you signed this year carries shared-responsibility language. A regulator just demonstrated what shared means when it is tested: it means yours. The same week, UK insurance lawyers started publishing on liability for AI harms, which is what it looks like when a question leaves the ethics panel and reaches the underwriters. The pharma trades worked this through carefully. The AI press mostly skipped it, because it arrived as paperwork instead of as an argument.

Here's what works: Take one AI-assisted step in a regulated workflow and ask who signs. If the answer is the vendor's contract, you have not answered.

Quick hits:

• Buildots raised $130 million to chase construction delays. The Israeli firm closed $130 million for AI that tracks build progress on site, which is AI money going where the failure mode is a schedule, not a chatbot.
• Keewano raised $12 million for a database built for agents. The agent-focused database launched with $12 million behind it, and you only build storage for agents once agents generate enough state to need it.
• Nuance Labs raised $50 million to fix awkward AI avatars. Lightspeed backed $50 million for AI avatar quality, which is a lot of money for the last mile of an interface most enterprises have not deployed yet.

Signal vs. Noise

🟢 Signal: security teams taking the AI governance file. The Register ran a trust framework for enterprise agents and models on its security desk rather than its AI desk, and Seyfarth pulled in-house leaders together on converging AI, privacy and data risk. Buying authority for agent controls is moving to the people who already own access and audit. Most coverage still files this under strategy.

🔴 Noise: the AI governance framework. Frameworks were published faster this week than anything was deployed, and almost none of them name who is accountable when the thing misfires. One warning letter did more to define that in a day than a quarter of principles documents managed.

From the 190K

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

cynapse.ai showed agent-driven video intelligence at a security trade show, CyberAttack.ai launched a platform that attacks your own systems continuously, and Empromptu tied up with Skyrelis to put agents inside regulated healthcare software. Three names appearing for the first time on the same day, in three trades that never read each other.

Read alone, each is a small product note on a niche wire. cynapse.ai took agentic video intelligence to GSX, CyberAttack.ai unveiled continuous threat validation, and Empromptu and Skyrelis went after regulated SaaS. Read together on one morning and none of them is selling intelligence. Each sells a continuous witness: a camera that files a report, an attacker that never stops testing, a compliance layer riding along inside somebody else's software. The buyer is not shopping for a better decision. They are shopping for something that was watching when the decision happened.

What changes on Thursday is where your own AI budget points. Less at the model that makes the call, more at whatever can testify about it afterwards.

By The Numbers

• EUCLYD raised more than €200 million in Eindhoven — a chip round aimed squarely at what inference costs to serve, with Samsung and EQT participating.
• Buildots closed $130 million against construction delays — nine figures for AI that watches a building site, in a sector where a week of slippage is real money.
• Nuance Labs raised $50 million for AI avatar quality — Lightspeed betting that the interface, not the model, is what blocks deployment.
• Keewano launched an agent-focused database with $12 million — a16z Speedrun and Hetz behind a storage layer that assumes agents now leave a lot behind them.
• Gartner found 63% of organizations lack or are unsure of the right data management practices for AI — two thirds of the market is deploying on foundations it cannot describe.
• See what's rising across AI and data this week →

Deep Dive: The Set Nobody Wrote Down

Years ago I played a wedding in Limburg and got a letter a few weeks later asking for the tracklist. Belgian collecting society, royalties, entirely reasonable request. I had played four hours out of crates and memory and I could not produce a list to save my life. The room had danced until two. On paper, nothing had happened.

Nobody asks what you played until money moves

For most of my DJ years the question never came up. Then a promoter needed to file, or a venue got audited, or an artist wanted their cut, and suddenly the only thing that mattered about the night was whether anyone had written it down. Same shape as an AI pilot. Nobody asks what the agent did until a customer complains or a regulator writes.

The desk that remembers beats the desk that sounds better

I have played on beautiful gear that logged nothing and on ugly gear that logged everything. Guess which one I would take now. The Cisco and Airrived launches are both that ugly gear. Neither makes your agents smarter. Both make the night reconstructable, and that is the feature the buyer is actually paying for.

Two people can play the same set, one can prove it

That is the whole gap opening up right now. Two companies run the same model on the same process. One has a record of what it saw and who acted on it. When the letter arrives, those two companies are in entirely different positions, and it has nothing to do with the model.

What Actually Works

1. Write down who signs: for every AI-assisted step in a regulated or customer-facing process, name a human. Not a team, a person.
2. Buy the boring layer first: observability and audit trail before the next model upgrade. The upgrade is easy to add later; the missing history is not.
3. Price the request, not the project: track cost per served request monthly. Inference is now an operating line, and operating lines get cut when nobody defends them.
4. Test the shared-responsibility clause: ask your vendor, in writing, what they would produce if a regulator asked what their tool did last Tuesday. The silence is the answer.

The set nobody wrote down still happened. It just did not count.

What's Coming

Observability Becomes A Line In The Agent RFP

Cisco and a startup described the same product on the same morning, which is what a category looks like the week before procurement notices. The next version of that conversation is a section in your RFP asking what record the platform produces and how long it keeps it.

The Warning Letter Turns Into A Supplier Questionnaire

The pharma trades are already working through what the FDA letter means for accountability. That analysis becomes a customer asking you to prove it in writing, and regulated buyers move faster on this than anyone expects. Expect the clause before you expect the rule.

Inference Cost Reaches The Board Deck

A €200 million round aimed at AI efficiency is investors saying the running cost is structural, not temporary. When hardware money moves on that thesis, the finance conversation follows within two quarters. Your CFO will ask for cost per request before your CTO volunteers it.

For Your Team

Thursday's meeting prompt: ”Pick our most-used AI-assisted process. If a regulator or a large customer asked tomorrow what it did last Tuesday, who answers, what do they open, and how long does it take them? If the honest answer is 'we would go ask the vendor', we have outsourced something we cannot outsource.”

Share-worthy stat: Gartner found that 63% of organizations either lack or are unsure about the right data management practices for AI. Two thirds of the market is deploying on foundations it cannot describe, and the letters are starting to arrive addressed to the deployer.

Go deeper: Track where AI accountability and agent controls are actually moving →

The Track of the Day

”Stop the events nobody logged from becoming the losses nobody can explain.”
Roboflow, on vision agents for the factory floor

A vendor tagline, and still the sharpest sentence in 190,000 articles this week.

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

Published: September 16, 2026 | Curated by Yves Mulkers @ Ins7ghts