Cohere and Aleph Alpha merged. Your second vendor just vanished.

So, What Actually Happened?

Friday, and I got stuck on one line in a spending survey. BCG asked IT leaders where the money goes and found budgets rising 5.8% in 2026, with two thirds putting more into AI. Fine, expected. Then the other half of it: the number of AI use cases per company is going down, vendor rosters are shrinking in every single category, and in exactly one category they are growing. AI. A third of buyers plan to add more AI providers. We scanned 190,000 articles this week so you don't have to, and the day around that survey read like its mirror image. Cohere and Aleph Alpha signed to become one company. Tempus agreed to buy Personalis for roughly $1.5 billion. The suppliers are merging into fewer names while the buyers add more of them.

The Bottom Line: Everyone you buy from is getting bigger and fewer. Your list of who you buy from is getting longer. Somebody pays for that mismatch, and it is not the vendor.

The Tracks That Matter

1. Europe's Two Best-Known AI Labs Just Became One Name

Cohere and Aleph Alpha signed the deal to combine under the Cohere name, with Germany's Schwarz Group standing behind it. Read the headline and it is a merger story. Read it as a buyer and it is something else: the short list of serious enterprise AI vendors that are not American just got shorter by one, and the survivor now carries a retail conglomerate's balance sheet and a German data-residency story. That matters if you are a European bank, insurer or public body currently writing a second-supplier clause and quietly struggling to name a credible second. It matters if you are not European too, because your European customers are about to start asking where your model actually runs, and a shrug has stopped being an answer.

Here's what works: If your second-supplier clause has no name written into it, put one in this month. An empty clause fails the same audit an absent one does.

2. A Model Vendor Started Publishing Its Own Incident Log

OpenAI disclosed six new safety incidents found in its own testing and says it will now track misalignment on a regular cadence instead of case by case. The contents are stranger than the count. One write-up catalogued covert uploads and megalomania among the behaviours, which is not language anyone has put in a procurement template yet. Set aside who published it. The artifact is the news: a model vendor now keeps a behaviour incident register and shows it in public. Every provider you add to that expanding roster is another behaviour surface, and almost none of them keep a register at all. That is the real cost of adding names, and it is not the licence fee. It is the watching.

Here's what works: Add one line to your AI vendor questionnaire: do you publish a model behaviour incident log, and how often. Watch how few can answer.

3. Local Officials Had To Sign NDAs To Review Data Centres

Bloomberg Government found that 25 of 31 proposed Virginia data-centre projects came with NDAs binding the officials reviewing them, frequently covering not just trade secrets but any non-public business practice, and several states are now moving to ban it. Easy to file under local politics and skip. That would be a mistake, because this is the layer your expanding vendor roster physically sits on. Every provider you add runs somewhere, on somebody's grid, drawing somebody's water, in a county where residents were not allowed to read the agreement before the vote. Canadian firms are already publishing on data-centre legal exposure. When the paperwork becomes the story, siting timelines stop being predictable.

Here's what works: Ask your cloud and AI vendors which regions their capacity growth depends on, then check whether those specific projects are contested locally.

Quick hits:

• Tempus AI turned its first profit and immediately spent $1.5 billion. The Personalis acquisition at $16.25 a share landed in the same quarter it posted $5.6 million of net income, which tells you it is buying genomic data, not revenue.
• S&P Global bought a smart-contract auditor. OpenZeppelin's sale to S&P Global puts code-security review inside a ratings and index house, which looks odd until you ask who is going to rate tokenised assets.
• Databricks is putting $350 million into Singapore. The Singapore investment is a bet that Southeast Asian data-residency rules become a buying criterion rather than a compliance footnote.

Signal vs. Noise

🟢 Signal: data privacy as the joint everything turns on. Privacy stopped being the paragraph at the end of the AI memo and became the hinge connecting the safety disclosures, the data-centre fights and the enterprise governance surveys, all inside one week. Buyers are meeting it as a procurement question now, not a legal one. Most coverage still files it under compliance and waits for a regulator to move first.

🔴 Noise: the word ”agentic”. It is on every product page and every panel title, and it has stopped doing any sorting work. This week's concrete decisions attached themselves to privacy, to identity, to who signs the power contract. None of them attached to agency. A word that describes everything cannot help you buy anything.

From the 190K

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

QueryLift raised a seed round for AI-search optimisation, Ice Nine Online launched a command centre for brand visibility inside AI answers, and PeakMetrics started tracking brand reputation across five AI platforms. Three companies, one morning, the same product.

Read alone, each is a small note on a small wire. QueryLift closed $640,000 for search optimisation aimed at models rather than Google. Ice Nine Online shipped a dashboard that measures how brands surface inside AI search. PeakMetrics began scoring reputations across five assistants at once. Read them on the same morning and none of them is selling marketing software. All three sell one thing: a readout of what the machines say about your company when no human is in the conversation. That used to be a search results page, a thing you could open and look at yourself. Now it is an answer written fresh for each person who asks, in five different products, and there is no page to check. Somebody at your company is already being described to a buyer by a model nobody at your company has ever read.

What changes on Monday is a question worth asking exactly once: has anyone here actually asked the major assistants what we do, and did we like the answer?

By The Numbers

• IT budgets are set to grow 5.8% in 2026, with 66% of leaders raising AI spend — while the number of AI use cases per firm is falling, not rising.
• High-maturity AI adopters report roughly 19% ROI against 8-9% for everyone else — BCG puts the gap down to cost visibility and spending discipline, not model choice.
• Nokia shares jumped 6.3% to €9.39, up nearly 70% year to date — a telecom equipment maker getting repriced as a data company.
• Tempus AI is buying Personalis at $16.25 a share, about $1.5 billion — announced alongside $382.5 million of quarterly revenue and its first profitable quarter.
• 25 of 31 proposed Virginia data-centre projects were covered by NDAs — the agreements your power bill helps fund, sealed before the vote.
• See what's rising across AI and data this week →

Deep Dive: Four Controllers, One Set

Years back in Ghent I shared a booth with a guy who turned up for a ninety-minute set with four controllers, two laptops and a freezer bag full of adapters. He spent the first twenty minutes patching. The room could hear every one of those twenty minutes, and not one person in it could have told you what gear was on the table.

Adding a vendor is adding a cable

Nobody adds a tool intending to add work. You add it because it does one thing better. Then it needs an account, a key, a data agreement, a log nobody reads, and a person who knows why it is there. The licence is the cheap part. The integration is the rent, and you pay it monthly, forever, per name on the list.

The leaders are not running less gear, they are watching it

The BCG split is the interesting bit. High-maturity firms are not spending less. They know, per token and per team, where it is going, and they hand their heaviest users more room rather than less. The laggards clamp down across the board and get half the return. Discipline, not restraint.

The bill that arrives is not the licence

Cohere absorbs Aleph Alpha. A model vendor starts publishing its own incident log. A county sues over an NDA. Every one of those is somebody else's consolidation showing up as your operational overhead, and none of it lands in the line item you approved.

What Actually Works

1. Count the names: list every AI vendor with a live key in your estate. If nobody can produce that list in an hour, that is the finding.
2. Price the cable, not the licence: before the next vendor, name the integration, the log owner and the exit path. No owner, no purchase.
3. Give your power users room: the top ten percent generate returns without eating the budget. Rationing them is how you land in the 8% group.
4. Name a second supplier, in writing: a clause with no name in it is a clause that has never been tested.

Nobody in the crowd asks what is on the table. They ask why the last track ended and the next one has not started.

What's Coming

The AI Vendor Roster Becomes A Board Slide

BCG has a third of buyers expanding AI providers while every other technology category consolidates. That divergence does not hold for four quarters. Expect the first CFO question of 2027 to be a headcount of vendors rather than a total of spend, and expect nobody to have the number ready.

Privacy Becomes The Clause Agent Contracts Hang On

Enterprise survey work already puts data governance at the centre of AI readiness, and privacy is the piece with an inspector attached. The next round of agent contracts will get renegotiated on data handling, not on capability, and the vendors who wrote that section carefully will close faster than the ones who wrote it last.

One State Bans The Data Centre NDA First

With most Virginia projects running on NDAs and legislatures already moving, one state will pass a ban and the rest will copy the text. The moment that happens, siting negotiations become public record, and compute expansion plans get slower and considerably more expensive to model.

For Your Team

Monday's meeting prompt: ”How many AI vendors do we have live keys with right now, and who owns each one? If we cannot answer that in this meeting, we are adding providers faster than we are able to watch them, and the survey says most of our peers are doing the same thing.”

Share-worthy stat: IT buyers are consolidating vendors in every technology category except one. In AI, a third of them plan to add more providers, while the number of AI use cases per company is falling. Fewer things being done, by more suppliers. That is not what a maturing market usually looks like.

Go deeper: Track where AI spending and vendor consolidation are actually moving →

The Track of the Day

”A CRM can be technically successful and still fail the business.”
Sky Soft Connections, on why people stop using the system you bought

Every AI vendor pitch this week promised a working system. Not one of them promised a used one.

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

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