Enter a market an incumbent already leads by adopting the opposite posture on the value they compromise, then bake that stance into your structure so copying it would cost them their identity.
Anthropic
Anthropic is an AI safety company founded in 2021 by former OpenAI leaders Dario and Daniela Amodei, and it builds the Claude family of models. Its through-line is a single bet: that safety could be built not as a brake but as a governance structure, a research method, and a procurement asset, letting a later entrant win the buyers who treat trust as a purchase requirement. The honest read on its enterprise lead is that two engines run at once. Coding capability wins the developer, and a credible-commitment governance structure wins the risk committee. Each story below isolates one of them.
Counter-Positioning: a governance structure the incumbent could not copy, priced in when the incumbent's own structure failed
The problem. Anthropic launched into a market where OpenAI already had the lead, the Microsoft capital (a $1 billion investment in 2019), and the mindshare. Competing on raw capability or first-mover speed was a losing game for a company that did not exist until Dario Amodei left OpenAI in December 2020.
The approach. Rather than match OpenAI's model, Anthropic made its structure the differentiator and bet it could not be copied. It incorporated in April 2021 as a Public Benefit Corporation, then built the Long-Term Benefit Trust: five financially disinterested members holding no equity, who elect a growing share of the board (rising from one seat toward a majority over time). This is the strategic point. A fast-commercializing incumbent with existing equity investors cannot bolt a disinterested trust onto itself later without slowing down and alienating those investors. The commitment is only credible because it was made before there was money to protect.
How it solved it. The counterfactual arrived in November 2023, when OpenAI's board fired Sam Altman, over 700 employees threatened to quit, and the company nearly came apart in five days. Per Time's reporting, watching that meltdown convinced Amodei his structure "was the right approach." The event did what no marketing could: it demonstrated to enterprise buyers what governance risk looks like when it is unpriced. OpenAI later attempted its own contested nonprofit-to-for-profit conversion, a charitable-trust fight Anthropic had structurally avoided by never having a nonprofit to convert. The distinction from a classic mutual-ownership commitment like Vanguard's is the payoff surface. Vanguard's structure lowers price invisibly. Anthropic's structure is an auditable governance narrative a buyer inspects and buys, which is why the next two stories are about who was buying.
Category Creation: making AI values auditable instead of internal
The problem. Model alignment was, across the industry, a black box. The rules a model followed lived in internal documents no outside buyer, regulator, or risk committee could read. For enterprises in healthcare, defense, and law, "trust us" was not a purchasable answer.
The approach. Anthropic defined a category of publicly legible AI governance. It invented Constitutional AI, training Claude to critique and revise its own outputs against an explicit written constitution rather than relying solely on human labeling, then published that constitution in full under a Creative Commons CC0 dedication so anyone could read or reuse it. It paired this with a Responsible Scaling Policy, a voluntary catastrophic-risk framework it has maintained and versioned for years, and later ISO/IEC 42001 certification and a public Trust Center.
How it solved it. By turning alignment rules into a document a law firm's risk committee or a defense contractor can actually read, Anthropic made the alternative look like "internal rules nobody outside the vendor has seen." Constitutional AI also cut training cost by using model-generated feedback instead of exclusively human labeling, producing a method rivals then studied. The category move mattered most exactly where capability is not the deciding vote: regulated procurement, where audit-ready documentation clears a gate that a better benchmark score does not.
Founder-Market Fit: the people who discovered scaling starting the lab built around it
The problem. Building a frontier lab in 2021 required both the technical conviction that scaling laws were real and would keep working, and the credibility to raise billions and recruit the researchers who could execute. Very few people had both.
The approach. Anthropic was founded by the people who produced the underlying science. Dario Amodei had been VP of Research at OpenAI and was an author of "Scaling Laws for Neural Language Models" and on the GPT-3 author list; Daniela Amodei had been OpenAI's VP of Safety and Policy. As Dario put it, "the first conviction was the scaling laws... I was finding that in 2019 with GPT-2." They left specifically because they believed the risks of scaling were not being taken seriously enough.
How it solved it. The fit showed in recruiting and execution. Roughly 14 other researchers followed the Amodeis out of OpenAI, giving the new company a founding team that had already built GPT-2 and GPT-3. Founders who had personally discovered the scaling relationship, and who left over how to handle it, were uniquely positioned to build a lab whose entire thesis was scaling responsibly.
Moats: two engines, honestly separated, and where each does the work
The problem. A values stance is only a durable advantage if it converts into revenue a competitor cannot easily take back. And the causal question is genuinely hard, because Anthropic also happens to make the best coding model, so capability and trust are entangled in the topline number. The moat claim only survives if you can say which engine is pulling.
The approach. Run both and let the data separate them. On capability, Anthropic aimed a developer wedge at coding, where Claude has led for over 18 months and Claude Code embedded into engineering workflows at Netflix, Spotify, KPMG, L'Oréal, and Salesforce. On trust, it aimed the auditable-constitution and governance posture at regulated and first-time enterprise buyers, where there is no incumbency and the governance narrative, not the benchmark, tips the deal.
How it solved it. Per Menlo Ventures' 2025 enterprise report, Anthropic held 40% of enterprise LLM spend, up from 12% in 2023, while OpenAI fell from 50% to 27%. Here is the honest split. The 54% coding share versus OpenAI's 21% is mostly capability. Claude simply codes better, and no trust story is needed to explain a developer picking the better tool. But the collapse of OpenAI's overall share is where trust does causal work capability alone cannot, because OpenAI is the clean same-conditions counterfactual. It had more capital, a first-mover lead, and no capability deficit large enough to shed 23 points of share on merit. What it lacked was a credible-commitment structure, and it demonstrated the lack publicly in November 2023. The tell is the segment with no switching costs: among first-time enterprise buyers choosing head-to-head, Anthropic wins roughly 70% of the time, the exact place capability parity hands the decision to governance. Capability wins the coder; the structure wins the risk committee; the moat is that only one company is running both.