The Moat Has Really Moved: Even Anthropic Is Desperately Hiring Salespeople
Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle
Open Anthropic’s careers page and you’ll notice something unusual: the department with the most openings isn’t AI research or product engineering—it’s sales.
You could explain this as an isolated case: a model company needs to drive enterprise revenue, so naturally it has many sales positions. But zoom out a bit, and this detail becomes a microcosm of an industry-wide moat migration: even AI companies with the deepest product barriers are placing the most hiring bets on “how to sell” rather than “how to build.” Behind this is a judgment worth taking seriously. The moat has moved away from the product and into distribution: how you make people aware, gain their trust, and get them to choose you among countless similar products.
Four Relocations of the Moat
The history of enterprise software is a history of the moat constantly moving outward. Each era has a constraining cost variable; whoever controls it has a barrier. When new technology collapses that cost, the moat shifts one step outward.
The first era was capital plus technology. Software meant CDs and server racks, with deployment cycles of 12 to 24 months. The barrier was capital itself—only companies that could afford years of cash burn could get through the customer’s door.
The second era was SaaS cloud. When same-day activation replaced multi-year deployment, the capital threshold collapsed, and the barrier moved to product building and sales systems: multi-tenant architecture, SDR/AE machines, partner ecosystems.
The third era was PLG. The product itself became the sales funnel; users tried before they bought. Slack and Figma both reached $1 billion valuations without a single salesperson. The barrier moved to viral growth and community.
The fourth era is now. The first three relocations had something in common: they all compressed costs “around the product”—deployment, sales, procurement friction. None of them compressed the cost of “building the product itself.” AI is the first. A team of two or three people with the right toolchain can build in a few days what used to take dozens of engineers several months. The window for feature replication has shrunk from months to days. You launch a feature that excites users, and similar versions appear in competing products a week later.
The barrier has been forced to the last place that can’t be moved: the audience, and the ability to reach them.
Distribution Beats Product—This Isn’t New
This pattern wasn’t invented by AI; it’s been playing out repeatedly over the past thirty years.
Salesforce versus Siebel. Siebel’s CRM was deeper, its enterprise relationships stronger, with minimum contracts starting at $5 million. Salesforce’s product was thinner, at $50 per seat per month, but they hired a group of actors in red T-shirts to stand outside Siebel user conferences holding signs and chanting “End Software,” attracting police, passersby, and free coverage from Fortune magazine. The outcome: Siebel sold to Oracle for $5.85 billion; today Salesforce has a market cap exceeding $280 billion—roughly 50 times Siebel’s sale price.
HubSpot versus Marketo. Founded the same year, Marketo’s marketing automation was more mature. HubSpot did something more aggressive: wrote a book, coined a category called “inbound marketing,” then spent half its marketing budget evangelizing that category—annual conferences, certification courses, free CRM, dominating virtually every important industry search term. By the mid-2010s, “inbound” essentially equaled HubSpot in the industry.
Notion versus Evernote. Evernote had a ten-year head start and 200 million registered users. Notion’s playbook was community: a landing page inviting power users to participate deeply, selecting 20 ambassadors from 400 applications. These people created templates, recorded tutorials, answered questions in forums at 3 a.M. Today the vast majority of Notion’s new users come from referrals.
Three stories, identical pattern: whoever names the category first, establishes trust first, and occupies user mindshare first outlives competitors with more similar product features. The difference is that this used to be the exception; now it’s the only durable path.
What the Numbers Say
There’s a batch of public benchmark data to cross-reference. The median AI-native startup reaches about $2.1 million in annual revenue by month twelve and closes Series A nine months after generating revenue. Top quartile companies hit $5.3 million in year one. Ten years ago, the “best-in-class” for SaaS startups was $1 million in a year. What used to be the highest standard is now below the median for AI companies.
Another founder survey shows that the number one issue keeping founders awake at night has become GTM execution: how to sell, how to reach people, how to build brand—ranking ahead of product execution, fundraising, and AI strategy. Founders have voted with their anxiety: being able to build is no longer the hardest problem; being able to make people aware, gain trust, and get chosen is.
This also explains why “brand first, product follows” has become the mainstream playbook. Cursor first had a developer community that couldn’t stop discussing it, then became the fastest company in B2B to reach $1 billion in annual revenue. Legal AI player Harvey locked in more than half of the top law firms before its product could meet all buyer needs; by the time competing products caught up, it was already the default option in the category.
But the Product Must Still Be Excellent
The biggest pitfall here is pushing the conclusion too far.
Counterevidence is right in the open. Model-layer companies’ moats still lie in research and compute. No matter how many salespeople Anthropic hires, its moat isn’t built on sales headcount. The “distribution era” judgment means products no longer automatically form barriers—but the premise is that the product itself must be excellent. The requirement hasn’t changed; only the sequence has. Companies that treat distribution as a silver bullet while their product is terrible will only expose themselves faster.
Another hidden risk is distribution inflation. When all companies are competing for mindshare, creating content, and naming categories, the cost of attention competition will rise again, and the “first-come, first-served” window keeps shrinking. At that stage, what truly survives is often “genuinely trusted in small circles”: names repeatedly validated within doctor circles, lawyer circles, developer circles. Deep trust is harder to replicate than broad awareness.
My judgment: the moat has indeed moved, but where it’s moved to may not be somewhere money can buy. It’s moved to your first batch of users—why they chose you, whether they’ll speak for you, whether you’re worth their continued trust. Products can be replicated; audience and trust cannot. So “who are your first 100 users, and why are you going after them” should be figured out before writing the first line of code—it’s part of building the product.
Key points: Four moat migrations: capital → product & sales → PLG → distribution; AI is the first to compress the cost of building the product itself, shrinking the replication window to days; history repeatedly proves category naming and mindshare preemption beat superior products; but excellent products remain the prerequisite, and distribution will inflate—deep trust within small circles is the ultimate barrier.