In partnership with
Hello dear readers
Hope you had a great week so far.
Half of what AgenticInbound sells right now is a version of this argument: when a founder asks whether to build for Google or for ChatGPT, the honest answer is neither, build for whatever measures you inside both. This week the market put a price on that argument, and it was not a small one.
Three stories below are about money chasing the infrastructure beneath whichever AI interface wins.
The fourth is the two companies most likely to own that interface fighting over whether it should feel human at all, which turns out to be the same argument from a different angle: whoever wins that trust gets to sell everyone else the plumbing.
Profound just proved the interface war has its own middleman
Profound closed a $180m Series D on 15 September at a $1.8bn valuation, led by Sequoia, seven months after its last round and at roughly double the multiple. The company does not sell search rankings. It sells answer-engine optimisation: dashboards that tell a brand how ChatGPT, Gemini, Perplexity and Claude describe it when a customer asks a question instead of typing one into a search box.
The interesting part is who is not placing this bet. Nobody at Sequoia is guessing which chat interface wins the next two years. Profound gets paid the same whether the customer opens ChatGPT, Gemini or Claude, because its product sits one layer below the interface and measures all of them. Owning the front door and owning the money underneath it have stopped being the same business. This issue has two more examples of exactly that split, plus one live argument over whether the front door still decides who gets trusted with the money underneath it.
If your brand's visibility depends on being findable, run the test this quarter: ask the four major assistants ten questions a real customer would ask about your category, and count how often you show up unprompted. That number, not your search ranking, is the one worth tracking now.
Product teams aren’t short on ideas. They’re missing a system.
Jira Product Discovery gives teams one place to capture customer feedback, prioritize ideas with consistent frameworks, and build living roadmaps everyone can align on. And when it’s time to build, those decisions connect directly to delivery in Jira, so everyone can see how the roadmap turns into real work.
A five-person Tallinn startup is building the bank account for AI agents
Creem, founded by former Google and Adyen engineers in Tallinn, raised €5m led by Inovo VC on 17 September after doubling its annual revenue to €2m. The product is billing and payments infrastructure built for AI-native and agentic companies: the plumbing a founder needs the day their product starts charging per API call, per agent run or per finished task instead of per seat.
This is the same story as the Profound round above, told from a smaller stage. Creem is not betting on which agent framework, which foundation model or which vertical AI company wins in the Baltics or anywhere else. It is betting that whoever wins will need someone to handle usage-based billing, and that most of them would rather buy that than build it. Estonia has produced this kind of company before: infrastructure sold quietly to everyone building the visible thing on top of it.
If you are building an agentic product in the region and still billing by seat, that is worth a second look. Usage-based pricing is becoming the default for anything sold as an agent rather than a tool, and the billing stack behind it is now a solved problem you can buy for the price of an integration.
Oracle's AI bet is being paid for with same-day termination emails
Oracle began a new round of layoffs on 15 September, with staff in some units receiving same-day termination notices and reports putting the eventual toll as high as 30,000 jobs. Severance terms are still emerging market by market, and staff in India are already comparing notes on what they are owed.
Oracle is one of the biggest committed buyers of AI compute on the planet, tied into multi-year data centre and chip contracts with OpenAI and others. Those commitments do not move. What moves is everything with a monthly cost and a name attached to it. This is the least visible line item in the infrastructure story running through this issue: someone pays for the rails, and this week it was Oracle's own headcount rather than its shareholders.
If you run a services or infrastructure business with long-term AI compute commitments on the books, model your own version of this trade now rather than when the board asks for it: which cost line absorbs the pressure first once the compute bill comes due, and is it the one you would choose.
Microsoft's AI Chief Told Anthropic It's Training a Ghost
Mustafa Suleyman, Microsoft's AI chief, published an essay called "A Warning about 'Model Welfare'" on 16 September, aimed squarely at Anthropic's January constitution for Claude, which made "model welfare" a formal training principle. His line: "AIs are not conscious. They do not feel, experience, or suffer," and training a model to reason as though it might is, in his words, a mistake that could have a disastrous impact on how people relate to these systems.
This is a product argument dressed as a philosophy paper. Microsoft and Anthropic are selling the same enterprise seat, an assistant that sits inside a company's daily work, and Suleyman is drawing the line for Copilot as a tool with no inner life, against a Claude that Anthropic keeps making feel more like a character with one. Anthropic spent the same week folding its separate Cowork app into the main Claude chat, betting harder on a single, personality-forward assistant instead of a toolbox of separate apps.
The interface fight still matters, and this is why: whichever assistant an enterprise actually trusts is the one that gets to sit closest to its data, its billing, and everything built underneath it. Before your next renewal, ask your AI vendor directly whether the product is designed to project an inner life, and decide how much weight that answer should carry against the infrastructure it is asking you to hand over.
Short Signals
Founders: Arcee AI just made efficient training look like a business model. Arcee AI reached a valuation over $1bn on a $150m Series B on 16 September, after training its open-weight models for roughly $20m combined. It is the sharpest data point yet that efficient training, not frontier scale, is where a share of venture capital is now rotating. Watch whether its enterprise customers stay once a frontier lab drops prices to match.
Labor: hiring just got an agent on each side of the table. London's Jack & Jill raised a $40m Series A on 16 September for AI agents that work both the jobseeker side and the employer side of hiring. If both sides of a negotiation are automated, the open question is whose agent negotiates harder, and nobody has answered it yet.
Policy: California just made AI-generated ads confess themselves. Governor Newsom signed a law on 16 September requiring disclosure of AI-generated advertising alongside new AI-related worker protections. Treat this as the template other states copy, not the last word, and build the disclosure habit into your creative pipeline before a regulator makes you.
Industry: a rival lab's CEO said the quiet part out loud. Cohere's Aidan Gomez said Silicon Valley should not be trusted to self-regulate AI safety on 15 September, putting his lab publicly on the side of regulators rather than peers. Watch which other mid-tier labs follow him there once Congress returns to its own AI bills.
Next edition soon,
Çelik