The moat that had to be bought
Salesforce, Adobe, and SpaceX spent more than $65 billion over eight months to re-secure switching costs that agentic AI had arbitraged away.
For ten years, I ran support teams whose accumulated knowledge was the only reason nobody ripped out the costly CRM underneath them, and not once did anyone in those rooms call that a moat. Between November 2025 and June 2026, four enterprise software companies sold at a premium because that particular moat stopped holding, and every one of them was still growing on the day the deal was signed.
Two routes back to a barrier
Hamilton Helmer’s 7 Powers names seven conditions that produce persistent differential returns: Scale Economies, Network Economies, Counter-Positioning, Switching Costs, Branding, Cornered Resource, and Process Power.
Acquisition is not among them, and the omission is deliberate. Power in Helmer’s model requires two things at once: a Benefit that materially improves cash flow and a Barrier that stops competitors from arbitraging the Benefit away, and his own non-arbitraged test disqualifies any company that gains preferential access to a desired resource but pays a price for capturing the value it was buying. In an efficient market for companies, the premium goes to the seller. Which is what makes a $3.6 billion cheque interesting: nobody writes one unless the Barrier needs attention.
When a Barrier starts to erode, there’re two routes back to it.
The Rebuild Path means constructing the new layer yourself, on your own roadmap, with your own engineers, and accepting that the market will move underneath you while you ship. Salesforce took this path with Agentforce, which reached $1.2 billion in ARR in Q1 FY27, up 205% year over year, and which asks enterprises to build their own agents on a customisable platform. Adobe took it with LLM Optimizer, a product aimed squarely at brand visibility inside language models. Both worked. Neither was fast enough.
The Repurchase Path means buying the layer customers are already anchoring their habits to and absorbing the Barrier wholesale. What is striking about the last eight months is that the same companies walked both paths in sequence, rebuilding first and repurchasing second, which tells you the rebuild was the thing they tried before the clock ran out.
Switching costs were never built by the software. They were built by the effort users spent learning it. Agents absorb the effort, and the barrier leaves with it.
The test any operator can run on their own product this week: name what a customer would have to relearn if they left you this quarter, then ask honestly whether an agent could do that relearning for them over a weekend. If the answer is yes, what you have is not a switching cost. It is a habit, and habits are arbitraged.
Fin, acquired by Salesforce for $3.6 billion
Fin is optimised for what I’d call packaged resolution and does its job extremely well.
Primary intent: Resolution without deployment.
Job it does:
“I want tickets to stop reaching my team at all. And I don’t want to spend two quarters building the agent that makes that happen.”
It translates to:
Ingest the knowledge base
Resolve across every channel
Escalate only the residue
Report cost-to-serve
Best users:
Mid-market support leaders
Post-Series B ops teams
Companies with high ticket volume and thin headcount
Anyone whose CFO has asked about cost per contact
Fin’s agent, running on a proprietary model called Apex, resolves an average of 76% of support volume end-to-end across live chat, email, WhatsApp, SMS, phone and Slack, and handles more than two million conversations weekly across 30,000 organisations. Salesforce did not buy the capability because Agentforce had the capability. It bought the packaging, and packaging is where habit forms, because the mid-market never builds what it can deploy on Tuesday.
Intercom was founded in 2011 by Eoghan McCabe, Des Traynor, Ciaran Lee and David Barrett, did roughly $200 million in revenue by 2021, and then did something almost nobody with a decade of brand equity does: it renamed the entire company after one product, the agent, and pushed past $400 million in ARR on the strength of it. In March 2026 it raised $250 million in venture debt from Hercules Capital. In June it sold.
Churn watch: the risk arrives at the resolution ceiling. Fin looks extraordinary against a world of all-human support, but once 76% automated becomes the floor, the buyer takes it for granted and starts judging Fin on the hard quarter it cannot close, the ambiguous 24% where every vendor is weakest. And because everyone now clears roughly the same 76%, the headline number stops being a reason to pick Fin over a rival; the lead rests on a model that is only briefly ahead. Category leadership, measured in resolution rate, has a shelf life exactly as long as the gap between frontier models.
Semrush, acquired by Adobe for $1.9 billion
Semrush is optimised for what I’d call visibility measurement and does its job extremely well.
Primary intent: Discoverability assurance.
Job it does:
“I want to know whether anyone can still find us now that our customers ask a model instead of a search box. And I don’t want to infer it from analytics that only count the traffic which already arrived.”
It translates to:
Crawl the index
Benchmark against competitors
Diagnose the visibility gap
Hand the fix to the publishing stack
Best users:
In-house SEO teams
Independent agencies
Enterprise brand marketers
Content operations leads
Adobe’s portfolio was built to act rather than to see, with Experience Manager to publish, Commerce to sell, Experience Platform to unify and Brand Concierge to converse, and none of it able to tell a brand whether that work was being surfaced at all. Semrush closed the loop, which is why Adobe’s own framing points at customers running LLM Optimizer alongside Semrush AIO to move from visibility intelligence to deployed optimisation to measured outcome, against a backdrop of AI traffic to US retail sites rising 269% year over year as of March 2026. The price was $12.00 per share in cash, a premium near 77%, against $471.4 million of ARR. Roughly four times revenue for the instrument panel of the agentic web.
Churn watch: Semrush was not losing customers in any dramatic way, and that is exactly why the risk is easy to miss. The clearer tell is that the customers it already had were barely spending more than the year before, so even though total revenue grew, almost all of that growth came from signing brand-new accounts and nudging existing ones onto more expensive plans, not from the core base becoming more valuable on its own. The fast growth was concentrated at the top, among the biggest spenders. At the same time, the ordinary middle of the customer base stayed flat, which is really a bet that a small number of large corporate accounts will keep carrying the whole company, and that bet only holds while those big accounts keep buying. The deeper problem is what Semrush is actually for: its entire business is helping brands get found when people search, and people are increasingly asking an AI assistant instead of searching at all, so when Semrush launched its own app inside ChatGPT, it was planting itself inside the very shift most likely to make its original job disappear.
Semrush was founded in Boston in 2008, went public in 2021, and by the end of 2025 had grown AI product ARR from $4 million to more than $38 million in twelve months while enterprise ARR quadrupled from $9 million to $37 million across 579 customers. Profitable on a non-GAAP basis, growing 18%, category-leading. It sold anyway, to a buyer whose $20 billion Figma bid had collapsed under regulatory pressure in December 2023 and who therefore had to buy something small enough to survive review.
Cursor, acquired by SpaceX for $60 billion
Cursor is optimised for what I’d call delegated authorship and does its job extremely well.
Primary intent: Engineering throughput.
Job it does:
“I want the tedious sixty percent of this codebase written without me. And I don’t want to hand my repository to a vendor my security team has never heard of.”
It translates to:
Index the codebase
Suggest inline
Run multi-step agents in parallel
Review before merge
Best users:
Enterprise platform teams
Fast-shipping startups
Staff engineers running large refactors
Anyone whose backlog grew faster than their headcount
Cursor’s revenue curve is the most violent in modern software, moving from $500 million to roughly $4 billion annualised inside a year, $2.6 billion of it enterprise. Underneath that curve, its competitive position was slipping: Claude Code, launched only in May 2025, overtook Cursor on developer preference through late 2025, and by May 2026 Ramp corporate spending data put Cursor at roughly 26% of AI coding spend with Anthropic at about 13% and closing. Revenue was doubling while the thing that predicts revenue, who developers actually reach for, was moving the other way.
Churn watch: the moment the agent stops being the reason and starts being the interface. When a developer’s loyalty attaches to the model doing the reasoning rather than the editor framing it, the editor becomes a skin, and skins are switched in an afternoon. Cursor could not build the one Barrier that would have held, its own frontier model, so it rented one.
Anysphere was founded in 2022 by four MIT students, Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger, launched Cursor in March 2023, raised an $8 million seed led by the OpenAI Startup Fund, and by November 2025 priced a Series D at $29.3 billion. In April 2026 SpaceX paid roughly $10 billion for exclusive collaboration rights plus an option to buy the company outright at $60 billion, pre-empting a round pricing near $50 billion. SpaceX exercised on 16 June 2026, four days after the largest IPO in US market history, in all stock, at about fifteen times revenue, for the coding data that trains Grok. The Cornered Resource was never the editor. It was the corpus.
Six mechanics behind these deals
Effort is the barrier, not the software. Every switching cost these three companies enjoyed was denominated in user effort, the ticket macros a support team wrote, the audits an SEO lead ran weekly, the shortcuts a developer memorised, and an agent that absorbs that effort dissolves the barrier without ever competing on features.
Packaging beats capability in a commoditising category. Agentforce could already do what Fin did and still could not reach the mid-market, because a platform that asks you to build the agent loses to a product that arrives having already built it, and the $3.6 billion was the price of that distinction.
Distribution is the terminal moat. Fin defined a category, hit $400 million in ARR, renamed itself after the product, and still sold, because when the category commoditises the surviving Barrier belongs to whoever owns the customer relationship rather than whoever owns the better model.
The measurement layer outlives the execution layer. Adobe could rebuild publishing, commerce and personalisation faster than it could rebuild Semrush’s decade of crawl history, which is why the instrument panel cost $1.9 billion while the things it measures are Adobe’s own products.
Revenue growth is a lagging retention signal. Cursor quadrupled ARR while losing a third of its market share, and the only reason we can see that at all is Ramp’s spending data, because ARR is the number a company reports and share is the number that predicts what it will report next.
The exit multiple prices the Barrier, not the Benefit. Semrush went at roughly four times ARR, Fin at about nine, Cursor at fifteen, and the spread has almost nothing to do with product quality and everything to do with how defensible each company’s position looked to the person writing the cheque.
If you found this useful, here’s what else Do Not Churn has covered:
The collapse of earned activation: Why agents that perform the work destroy the attachment the work used to build, and the one motion worth protecting from automation.
Churn is just graduation you didn’t design for: The churn event that never shows up in cancellation data, because the user outgrew the tool without ever leaving it.
AI takes the thinking out, how do we get it back?: What three AI tutors reveal about the difference between delivering a result and building the model that makes the result stick.