Cooley's $30B Dealmaker: AI Prices Are Unmoored

So, What Actually Happened?

Tuesday, and I cannot stop thinking about a support thread. Somebody's web analytics stopped recording on August 12 and every check they ran came back green: tags firing, requests sent, HTTP 200 the whole way. The data simply never arrived at the other end. We scanned 190,000 articles this week so you don't have to. Same day, a benchmark showed frontier agents failing across seven business functions when someone hides an instruction inside the documents they read. And the lawyer who papers the biggest AI deals in the market called them unmoored to numbers. Three unrelated desks. Every one of them describing a check that passed while the thing underneath it did not.

The Bottom Line: Everything that cleared its check this week failed somewhere the check was not looking. Approval is not evidence.

The Tracks That Matter

1. Hidden Instructions Beat Frontier Agents In Every Department

A new enterprise benchmark, ENT-IPI, tests what happens when an instruction is buried inside the material an agent reads: an invoice, a support ticket, a CV, a supplier email. Failure rates are high, and here is the part that should bother you, they are not concentrated anywhere. The same weakness turns up in finance, legal, HR, sales, IT, marketing and customer support. So this is not a model that is weak at legal work. This is every function that lets an agent read text somebody outside your company wrote, which is every function. The enterprise press is separately writing up accountability for agent actions as the next imperative. Same finding, politer language.

Here's what works: List every place an agent reads text you did not write. That list is your attack surface. Nothing else is.

2. The Lawyer Closing The Biggest AI Deals Says Prices Have No Anchor

Cooley's dealmaker, who papered an $11 billion purchase of Confluent and a $30 billion financing round inside the same year, describes this market as unmoored to numbers. Not unmoored from revenue multiples. Unmoored from numbers, plural, meaning there is no shared arithmetic left underneath the price. The structure is drifting with it: licensing agreements are quietly doing work that acquisitions used to do, because a licence attracts less review than a purchase. You can watch the same detachment in public markets, where Micron trades 76.8% above its estimated intrinsic value while insiders sold $167.8 million of it in three months and bought none.

Here's what works: At your next AI vendor renewal, ask what the price is a multiple of. If nobody can answer, you are the anchor.

3. Singapore Regulates AI While It Runs, Not Before

Singapore's monetary authority published a runtime governance framework for AI in financial services. Read the word runtime slowly. Every governance regime built so far grades a system before it ships: model cards, risk tiers, an approval committee, a signature. This one starts from the assumption that the system drifts once it is live, and asks what you are watching while it drifts. The timing is not accidental. Singapore also took the largest share of Southeast Asia's AI funding as regional investment surged, so one city is now writing both the cheques and the supervision rules. Whoever ships a workable runtime standard first sets the shape everybody else builds against.

Here's what works: Take your AI approval checklist and mark every line you could still evidence at 3am on a Sunday. The unmarked lines are paperwork.

Quick hits:

• A police data breach broke the anonymity of abuse survivors. The UK regulator reprimanded the Metropolitan Police over data protection failures that exposed people who had come forward, and the training was already in place before it happened.
• Novo Nordisk and AWS paired up on drug discovery. The partnership targets discovery speed, which is the version of this technology that gets graded on whether a molecule works rather than on how the demo felt.
• Higgsfield raised $400 million for AI video. The round landed on the generation layer, which still prices like scarcity while everything downstream of it prices like a commodity.

Signal vs. Noise

🟢 Signal: Governance and risk ownership. Both climbed hard in real weight this week, and risk management climbed further than anything else we track. That is the audit and risk side of the house taking the pen on questions that used to sit with engineering. Most coverage is still reading vendor announcements and missing that the buyer changed.

🔴 Noise: ”Generative AI” as a category. It pulled heavy volume again while losing its hold on everything else that moved. Data quality is doing the same thing, plenty of mentions and slipping influence, because the argument moved on from whether the records are clean to who answers when a system acts on them.

From the 190K

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

A web analytics stream stopped delivering data on August 12 while every check reported green, a benchmark found agents failing across seven business functions that had all passed their capability tests, and a police force broke survivor anonymity after its data protection training was already documented.

Three desks, three rooms that never talk. The analytics thread is a support ticket. The benchmark is a security paper. The police story is a legal note about a regulator's reprimand. Read them on one morning and they collapse into a single finding: the green light is the failure mode. In each case the check was real, it ran, it passed, and it was measuring something adjacent to the thing that actually mattered. Tags fired and returned 200, which proves a request left the browser and proves nothing about whether data landed. The agents scored well on capability, which proves nothing about what they do when a hostile sentence is sitting inside a supplier invoice. The training was delivered and recorded, which proves attendance, and the regulator's reprimand says the rest out loud.

What changes for you this week is small and mildly annoying. Take one system you trust and find the check that certifies it. Then ask what that check observes, and what sits between that observation and the outcome you care about. You will usually find a gap nobody owns, because everybody assumed the green light covered it.

By The Numbers

• IBM paid $11 billion for Confluent, and a $30 billion AI financing round closed in the same year — the lawyer who papered both says there is no shared arithmetic left under the prices.
• Micron trades at $1,032.63 against an estimated intrinsic value of $584.16 — 76.8% above, with $167.8 million of insider selling in three months and not a single insider buy.
• Gartner expects AI applications to drive 50% of cybersecurity incident response by 2028 — half the response function moves to software in two years, which is a hiring plan, not a headline.
• Slate locked $245 million while Fulcrum pays out roughly $270 million and keeps 5% of the combined company — that is an exit dressed as a merger, and the cash leaves before the science starts.
• Cloudera says single customers now run over 250 petabytes each, one of them past 300 — the unglamorous storage layer is still where the AI workload actually lives.
• See what's rising across AI and data this week →

Deep Dive: The Soundcheck Lies

Every soundcheck I have ever done was a lie, and everyone in the room knew it. You stand in an empty hall at four in the afternoon, the levels look perfect, the low end is tight, the engineer gives you a thumbs up. Then two thousand people walk in, their bodies swallow the high frequencies, the room warms up six degrees, and the mix you approved no longer exists.

The check ran in the wrong room
Tags fired and returned 200. That is a soundcheck. It confirms the signal left the desk. It says nothing about what the room does to it on the way to the back wall. The analytics thread this week is exactly that shape: every client-side test green, and five days with no data at the other end.

Nobody plays to an empty hall
Capability benchmarks grade an agent alone in the room. The enterprise injection results grade it with a crowd in there, and the crowd includes somebody who hid an instruction inside a purchase order. Same agent, different room, different result. Singapore's regulator just wrote the room into the rulebook.

The engineer who stays
The good ones do not leave after soundcheck. They stand at the desk all night with a hand on the faders, because the room keeps changing and the only useful measurement is the one taken while people are dancing. That is what runtime governance means once you strip the compliance vocabulary off it.

What Actually Works

1. Name what each check observes: write the literal observable next to every green light. Half of them measure a step before the one you care about.
2. Instrument the outcome, not the request: alert on data arriving, deals closing, tickets resolving. Not on the call succeeding.
3. Plant a hostile instruction yourself: drop one into a document your agent reads, and see what it does. Do it before somebody less friendly does.
4. Give the runtime a named owner: approval already has one. The eight hours after go-live usually do not.

Anyone can pass a soundcheck. The job is the four hours after it.

What's Coming

The SOC Stops Trusting And Starts Grading

Security teams are being handed AI triage tools faster than they can decide whether to believe them. A read of this year's SANS survey frames it squarely as a trust problem, and vendors are already answering with reasoning traces on detection triage. Expect ”show me the reasoning” to become a procurement line by Q1.

The Intellectual Property File Gets Reopened

A client alert this week lays out how AI use erodes copyright, trade secret and privilege at once. Privilege is the one to watch. Feed a model your legal reasoning and you may have handed it to the other side. General counsel will reach this before any CIO does.

Vehicle Data Becomes A Regulated Commodity

Europe and Australia are both moving on who owns car data, with Australia consulting on UN cyber security and software update rules as design standards. Any manufacturer still treating telemetry as a free by-product is about to discover it has an owner and a price.

For Your Team

Wednesday's meeting prompt: ”Pick the AI system we trust most. What does its green light actually observe, and what sits between that observation and the outcome we care about?”

Share-worthy stat: Gartner expects AI applications to drive half of all cybersecurity incident response by 2028. Half the response function, inside two years, running on software nobody in the room has audited yet.

Go deeper: Track where governance and risk are moving across AI and data →

The Track of the Day

”When we were a public company, if Frank needed to meet his numbers, he would have to take a deal whether it was good for the company or not, because otherwise the stock price will tank.”
Sudhir Menon, Cloudera

He is describing a check that ran perfectly on schedule and produced a worse decision every quarter. That is this whole issue in one sentence, and he said it about an earnings call, not about AI.

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

Published: August 18, 2026 | Curated by Yves Mulkers @ Ins7ghts