Çelik here.

Cognition raised a billion dollars at $26B this week, and the line everyone repeated was that Devin, its own product, writes roughly 90% of Cognition's code.

Two days later Marc Benioff said Salesforce hired zero net new engineers this fiscal year because coding agents covered the growth. In between, a startup founded by three Croatians got large enough to buy a tax product off PwC.

And this morning Jensen Huang stood on a stage in Taipei and put a number on why none of this slows down: every token is now profitable. Automation is landing first on the people who build it, and the economics underneath it just stopped being a question. Here is the week, and what to do with it before Monday.

Cognition closed more than $1B at a $26B valuation on May 27, led by Lux, General Catalyst, and 8VC, with Founders Fund along for the ride. The valuation more than doubled from $10.2B eight months ago, on a revenue run-rate that went from $37M to $492M in a year. The detail that traveled was not the money. It was that Devin, the autonomous engineer Cognition sells, now writes around 90% of Cognition's own code.

The positional shift is that the proof point is now the maker itself. Cognition is not pitching a future where agents write most of the code. It is running on one, and showing the cap table the gross margins that come with it. When the company building the tool reorganizes its own engineering around it first, that is the strongest possible signal about where every other software team is headed.

If you ship software, the question stops being whether an agent can write the function. It becomes what share of your stack should be agent-written by Q4, and which engineer reviews the output of a hundred subagents instead of typing it. The builder who only writes code is the one being repriced. The builder who reviews, architects, and decides is the one Cognition just made more valuable.

Most AI newsletters summarize headlines. MavSource is different.

We aggregate updates from all major AI newsletters, podcasts, company news, AI labs, public and private company activity, GitHub projects, funding rounds, earnings calls, and investor letters — hundreds of sources every day. Then we summarize what matters, analyze emerging trends, and add our own founder commentary so you understand why a development may matter — not just what happened.

One 5-minute email, every morning. Built for investors, founders, and operators who want to understand AI as a business, technology, and market trend — not just another news cycle.

The daily email is free. It's also the entry point to a deeper intelligence product covering watchlists, public-company read-throughs, startup trackers, and investor-letter tracking.

Subscribe Now For Free

Fonoa, founded by three Croatian Uber alumni, raised $110M and bought PwC's Indirect Tax Edge platform outright, with the round led by Headline and Eurazeo and Forestay joining. The plan is to fold a Big Four compliance product into Fonoa's own AI tax layer and run the full indirect-tax lifecycle on one data model. PwC stays on to deliver services on top. The startup owns the software now.

Sit with the direction of that sale. For two decades the regional story was Balkan and CEE engineers building the back office for somebody else's brand. Here a company with Croatian founders and customers like Netflix, Booking.com, and Canva is the one acquiring, and a global consultancy is the one selling its product and staying on as the service arm. The audit and tax class is exactly the knowledge work agents are commoditizing fastest, and the incumbents are starting to offload the product before the margin erodes.

For founders in the region, the lesson is that the acquisition arrow flipped. The next few quarters will surface more incumbents quietly selling the software layer they can no longer defend. Watch who becomes the seller in your own category, because that is where a small CEE team can take the wedge without fighting a hyperscaler for it.

Marc Benioff said the part out loud on May 28: Salesforce is not hiring more engineers this fiscal year "because I'm using coding agents." Engineering headcount has sat flat at roughly 15,000 for two years while agents absorbed a near-30% productivity lift. Support shrank from 9,000 toward 5,000. The one function still hiring hard is sales, up close to 20%.

The freeze is the real signal, not a layoff headline. Salesforce is not cutting engineers in a dramatic round; it is simply declining to add them, year after year, and letting attrition and agents do the math. That is quieter and more durable than a reorg. It tells you which work the agents have absorbed (writing code, answering tickets) and which work they have not (closing deals, owning the relationship that signs the contract).

Run the same audit on your own org chart. The roles that produce output an agent can now generate are the ones that plateau first, and the people in them deserve to hear it from you before they read it in a Benioff quote. The roles that carry judgment, trust, and the customer relationship are the ones to keep adding. A hiring mix that still looks like 2024 is the tell that the rethink has not happened yet.

This week at GTC Taipei, Jensen Huang put Vera Rubin into full production, the platform that supersedes Blackwell, and ran a keynote on a single line repeated until it stuck: agentic AI is here, it works, and every token is now a revenue unit.

Behind the slogan sits the Vera Rubin NVL72, which NVIDIA says delivers roughly 10x lower cost per token, and Spectrum-X photonics switches at 5x the power efficiency. The phrase Huang kept returning to was "compute is revenue, now compute is profit."

The positional shift is that the company selling the picks and shovels just declared the gold real. For three years the bear case held that agents were a subsidized demo and the token math never closed.

Huang's argument was the inverse: cost per token has fallen far enough that running agents at scale is margin-positive, not a loss leader. That is the same claim the Cognition raise and the Salesforce freeze make from the demand side, and now the supplier is making it from the other end, with the hardware to back it.

Watch the cost-per-token line, because it now governs your AI budget more than model choice does. If inference keeps getting an order of magnitude cheaper each generation, the build-versus-wait math flips: the workflow too expensive to automate last quarter lands inside budget the next one.

Re-price the agent projects you shelved on cost, and assume the floor drops again before you finish the spreadsheet.

Three things to install or test this week. Design and marketing sat out a thin week.

Productivity: Perplexity now multitasks next to your real work. Perplexity's May update adds iPad Stage Manager support, so you can research alongside Notion and Slack instead of bouncing tabs, plus quiz and flashcard generation and ETF holdings breakdowns in Finance. If you run research-heavy mornings, pin it beside your docs for a week and see whether the context-switching tax drops.

Sales: Visa wired payments straight into Replit. Visa invested in Replit on May 28 and is embedding its Trusted Agent Protocol so agents built there can take payments, with self-serve enterprise contracts now open to $200k. If you are building anything agent-driven that needs to transact, this is the first commerce rail from a payments incumbent worth a prototype.

Dev: Grok Build is the cheap coding agent in the price war. xAI opened Grok Build 0.1 on its API at $1 per million input and $2 per million output tokens, with MCP support and parallel subagents, no SuperGrok subscription required. Run the same migration task through it and your current agent and compare the bill; the floor on coding-agent pricing is still dropping.