Bending Spoons Bought Miro for Eight Cents on the Dollar

So, What Actually Happened?

Friday, and I spent the morning staring at two numbers that should not sit on the same page. Miro sold for $1.4 billion, against the $17.5 billion it carried a few years ago. That same morning the tally of companies that crossed a billion for the first time in August came in at 29, most of them AI. We scanned 190,000 articles this week so you don't have to. So one market is minting billion-dollar names faster than anyone can read the list, and the same market just settled an old one at eight cents on the dollar. Then I hit a research note reporting that of 36 apparent gaps between AI models, only three held up once the sample size was done properly. Three unrelated desks, one uncomfortable question underneath all of them.

The Bottom Line: The repricing running through this market is not AI against no-AI. It is claimed against checked.

The Tracks That Matter

1. Miro Sold For Eight Cents On Its Own Dollar

Read the buyer, not the seller. Miro went to Bending Spoons for $1.4 billion against a $17.5 billion peak, and the same Italian acquirer closed its purchase of Airtable in the same stretch. Bending Spoons is not a turnaround story with a view on whiteboards. It is a machine for buying software that once priced itself on growth and now has to price itself on cash. Both targets sold seats, and seats are exactly the meter that agents break. State Street's strategists filed the same observation from the other end of the telescope this week, calling AI a credit-cycle story rather than a technology one. The down round is not the news. A specialist buyer for the category now exists, has capital, and is working through a list.

Here's what works: Check whether your key vendors are priced on seats. If they are, model what your bill looks like when agents do the clicking.

2. Only Three Of Thirty-Six Model Gaps Survived The Statistics

This is the one that should change a meeting. An analyst ran confidence intervals across an agent benchmark suite and found that only 3 of 36 apparent model gaps were real. The other 33 sat inside the noise band created by the eval set's own size. Not wrong results, just too few samples to tell two models apart. Separately, a study of live agent runs traced real money leaking through unmanaged context, with governed metadata cutting token bleed on identical workloads. So the industry publishes decimal-place differences it cannot defend, and leaves the number that actually lands on the invoice unmeasured. That is not a model problem. That is a measurement problem, and you own it.

Here's what works: Ask any vendor for the confidence interval behind a benchmark they quote. No interval, no claim.

3. A Bank Collapsed Nineteen Data Stores Into One

The least glamorous story of the day is the one carrying a receipt. Nationwide consolidated 19 separate data stores into one cloud platform, work it could not avoid after a £2.9 billion takeover of Virgin Money handed it two of everything. No agent framework, no model launch, just a building society deciding it could not answer a regulator in nineteen places at once. Tooling vendors are reading the same room, with ER/Studio wiring data models into semantic layers the same week. Everyone selling AI into banking assumes this plumbing is finished. Mostly it is not, and that is precisely where pilots die between the demo and the audit.

Here's what works: Count how many systems hold the same customer record. That number, not the model you pick, sets your ceiling.

Quick hits:

• Anthropic says state-linked accounts were using its models for the wrong research. It disrupted Russian and Chinese campaigns aimed at its own systems, which means misuse detection is now a published operating metric rather than a policy page.
• Shadow agents reached a bill. The bipartisan Stop Rogue AI Act targets autonomous software running inside companies without anyone approving it, and unapproved agents are already the most common thing security teams find.
• Salesforce is shopping again. Talks to buy Listen Labs for about $2 billion land days after it closed the Fin deal, a pattern of buying capability rather than waiting to build it.

Signal vs. Noise

🟢 Signal: data integration. The dullest phrase in the industry gained more real ground this week than any model launch, and you can see why in the receipts: a bank merging nineteen data stores to survive an acquisition, catalogue vendors racing to connect models to semantic layers. Most coverage is still watching the model leaderboard and missing where the budget actually moved.

🔴 Noise: ”risk management” as a phrase. It pulled heavy volume across the wires again and lost ground in actual developments, living mostly in frameworks and conference decks. The practical question replaced it weeks ago: which system, which agent, which named human signs.

From the 190K

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

ITSecOps, ConsentPlace and Sigma Technology all surfaced in our reading on the same day, none of them in a headline, and all three sell the paperwork around AI rather than AI itself.

Read separately they are nothing: an audit-automation vendor, a consent-management tool tracking the EU AI Act, a sovereign-infrastructure shop. Read together on one morning they are a supply chain forming under a line item most companies have not created yet. Somebody has to produce the evidence that an AI system did what it was supposed to do, and a market is quietly assembling to sell that evidence rather than the system.

What changes on Monday is who you invite to the vendor review. The compliance evidence layer is being bought by someone in your organisation right now, probably on a departmental card, probably without procurement seeing it.

By The Numbers

• Miro sold to Bending Spoons for $1.4 billion, down from a $17.5 billion valuation — the largest collaboration-software exit of the year is also a 92% haircut.
• Only 3 of 36 model performance gaps held up under confidence intervals — most published model comparisons are measuring their own sample size.
• Positron reached a $5 billion valuation in its latest AI chip round — inference silicon is being funded like infrastructure while application software gets repriced downward.
• 29 companies joined the unicorn board in August, led by AI — nearly one new billion-dollar name per day, in the same month an old one sold for scrap.
• Alibaba led a $300 million round for UniPat AI at a $2.5 billion valuation — patent work, one of the most document-heavy professions there is, now carries a mid-cap price tag.
• See what's rising across AI and data this week →

Deep Dive: What The Sleeve Says And What It Sells For

There is a record in my collection with a price written on the sleeve in biro by a dealer in Antwerp, a number that was ambitious in 1994 and is comedy now. Next to it sits a battered 12-inch I paid almost nothing for, and it is the one I actually reach for at 2am when the room needs saving. The sleeve price and the working value have almost nothing to do with each other. This week the AI market found that out in three different rooms.

The sticker was never the price

Miro's $17.5 billion was a real number in the sense that someone once wrote it down. It was never a price, because nobody had to pay it in cash. $1.4 billion is a price, because someone did. The gap between those two things is where a lot of 2022 balance sheets currently live.

Most of the gaps were noise

Thirty-three of thirty-six model differences dissolved once someone asked how many samples were behind them. Nobody lied. The number was simply repeated more often than it was checked, and repetition is not verification, however confident the chart looks.

The boring column is the one that moved

While all that was being argued about, a building society quietly merged nineteen data stores because a regulator was going to ask. That work has no valuation, no benchmark and no press release, and it is the only thing on this page that definitely worked.

What Actually Works

1. Separate announced numbers from settled ones: keep two columns in any vendor review, what was claimed and what someone actually paid or measured.
2. Demand the sample size: any benchmark without an interval is marketing with a decimal point.
3. Meter what you are billed for: token spend, integration hours, review capacity. Those are the numbers your CFO will read back to you.
4. Fund the plumbing before the pilot: if the same customer exists in nine systems, no model fixes that, and every demo will keep lying to you.

Anybody can write a number on a sleeve. The price is whatever somebody hands over at the counter.

What's Coming

Capital Keeps Moving To Cheaper Ground

Vietnam's AI funding jumped 13 times in two years, a curve that used to belong only to established hubs. Watch where engineering salaries and data-residency rules line up next, because that is where the next tranche of infrastructure money lands. Southeast Asia is about to stop being a delivery location and start being a decision one.

ROI Evidence Becomes The Procurement Gate

Gartner is now publishing what multimillion-dollar agentic returns actually required instead of what they might be worth. When the analyst houses shift from projection to post-mortem, buyers follow within a quarter, and the phrase ”we are still in pilot” stops being an acceptable answer at board level.

Open Models Have To Start Billing

Chinese open-weight models are winning adoption, and the open question is whether they can turn that adoption into revenue. If the answer is no, expect a wave of consolidation in open-model startups; if it is yes, expect Western pricing to come under pressure fast.

For Your Team

Monday's meeting prompt: ”Take the three biggest numbers in our AI business case. For each one, say who produced it, what sample or transaction it came from, and what happens to the case if it is off by half. Anything we cannot trace back to a source gets marked as an assumption, not a forecast, and we present it that way.”

Share-worthy stat: A collaboration company once valued at $17.5 billion sold this week for $1.4 billion, eight cents on its own dollar, in the same month 29 new companies crossed a billion for the first time. Both numbers describe the same market.

Go deeper: Track where AI valuations and AI evidence are actually moving →

The Track of the Day

”When business rules are retyped by hand, drift is not a risk. It is a certainty.”
On what happens between a system that works and its replacement

Most of the AI failure I see up close is not the model. It is a rule somebody knew by heart in 2011, typed into a new platform slightly differently in 2026, and never checked again. The demo runs beautifully. The number it produces is quietly wrong, and it stays wrong until a regulator or a customer finds it for you.

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

Published: September 11, 2026 | Curated by Yves Mulkers @ Ins7ghts