Will Open-Source Threaten Anthropic's Business & Do Margins Matter in a World of AI | Matt Murphy
Summary
Murphy’s Anthropic lesson is that access to an outlier matters more than textbook ownership or a tidy entry price. Menlo invested a little over $10 million into a pre-revenue company valued above $4 billion, despite operating a roughly $600 million venture fund. His conclusion was blunt: “There’s never going to be a perfect entry point,” so “don’t overthink it and get in.”
Menlo turned its starter position into a $500 million-plus SPV only after accumulating operating evidence. Anthropic launched its model, began adding revenue month by month, and secured Amazon and Google as capital, technical, and distribution partners through Bedrock and Vertex. An Anthropic presentation electrified Menlo’s November LP meeting; two weeks later, the firm signed the follow-on term sheet.
AI has reset venture’s return hurdle so sharply that yesterday’s exceptional growth can now look ordinary. Harry argued that “triple, triple, double, double” and a progression from roughly $1.5 million to $5 million to $15 million no longer clears the opportunity-cost bar. Murphy agreed: companies reaching “zero to 100 in a year” have shifted what once looked top 5% toward “top 50%.”
Margins still matter, but the winning architecture will mix frontier, open-source, and proprietary models. Murphy expects today’s 20–30% gross margins to mature toward 60–70%, though not necessarily legacy software’s 80–90%. He rejected Harry’s thesis that open source could handle 96% of enterprise workflows: frontier models can justify their cost by improving retention, revenue, and engagement, while cheaper models absorb less demanding calls.
Application companies survive foundation-model expansion only when they own consequential workflows and context. Murphy distinguished Anthropic’s technical-user orientation from Lovable’s mission to turn the “99%” of non-programmers into creators. In legal, he argued Legora’s lawyers and FDEs coordinate corporate teams, law firms, clients, and case-specific context—a multi-constituent system that “a model just coming in” cannot readily replace.
Series A is the hardest insertion point because evidence has compressed while price has exploded. A company may move from five POCs to $1 million of ARR while its valuation jumps from roughly $50 million to $200 million, without supplying much additional signal. Menlo’s answer is a barbell: seed checks of up to $8 million on the spot, then concentrated investment after a company has been “anointed the winner”—or Murphy believes it will be.
Murphy flagged robotics and neo labs—and possibly defense tech—as overheated, while multi-model infrastructure appears under-invested. He sees no plausible future with 60 independent model companies and doubts acqui-hires can rescue all the heavily funded teams. By contrast, routing, observability, agent frameworks, and abstraction layers become more valuable as enterprises optimize across models, chips, latency, performance, and cost.
Deep dive
1. Anthropic looked structurally exceptional before it looked obvious
Murphy traced the introduction to Anjney Midha, who told him, “You gotta meet Dario and Tom. This is the one company.” After speaking with them the next day, Murphy’s personal reaction was immediate: “All right, I’m in.”
OpenAI was “absolutely ripping” as ChatGPT took off, and Murphy noted that Dario had created ChatGPT within OpenAI. Dario left believing OpenAI was doing too many things and that this was the one big opportunity. Murphy also saw Anthropic as the strongest candidate to become an alternative in a market unlikely to be dominated by one player.
Murphy saw an unusually technical leader who could attract researchers resembling him: “That’s the leader they gravitate to.” The quantitative evidence was equally important. Before revenue or a public model launch, Anthropic’s benchmarks were at or above ChatGPT’s level while using roughly “a 50th of the capital,” suggesting something technically distinctive under the hood. Former Splunk CTO Tim Tully helped Menlo diligence that claim with Tom.
The difficult piece was portfolio construction: a pre-revenue company wanted a valuation above $4 billion, while Menlo’s roughly $600 million venture fund normally invested about $15 million per company. Flexible partners let Murphy override the category mismatch; a more rigid “that doesn’t fit” partnership would never have reached the later opportunity.
2. A starter check beat perfect ownership math
Harry’s pushback exposed the conventional math. A little over $10 million at $4 billion, followed by an $80 billion outcome and 50% dilution, might produce only a 10x—or roughly $100 million, about 12% of the fund. Murphy acknowledged that exact objection was present inside the partnership.
Murphy’s rebuttal was strategic: Menlo had committed to building around AI, Anthropic was its best available foundation-model company, and sitting out meant abandoning the market. “Once you’re in,” a breakout creates repeated opportunities to invest more; ownership constraints should not prevent the initial wedge.
His broader conclusion was categorical: ownership is “by far” less decisive than it used to be. When good exits were $300 million, $500 million, or $1 billion, owning 20% mattered enormously; now a small stake in a true outlier can outperform a large stake in a $300–500 million exit that “just isn’t gonna move the needle.”
Murphy added that rising capital needs and signaling dynamics are making large fundraises more common even outside frontier companies. Fast-growing application companies want capital to play offense, while recurring rounds can support employee retention and create more need for secondary liquidity.
3. Menlo converted access into a $500 million-plus follow-on
After the starter investment, Menlo deployed recruiting, business-development, and relationship support to observe Anthropic closely. The round closed around March, the model launched in April, and revenue began a visible drumbeat—“adding 10 this month, 8 the next”—rather than remaining a benchmark-only thesis.
Amazon and Google then supplied capital, technical partnerships, and distribution through Bedrock and Vertex. Murphy framed Anthropic as the multi-cloud alternative to an OpenAI closely tied to Azure, materially improving both reach and strategic position.
The trigger was a November LP meeting where Anthropic executive Nirav explained the models’ effects on applications and human behavior. LPs and partners emerged saying, “This company is amazing”; Menlo resolved to lead the next round and signed a term sheet two weeks later, aggregating more than $500 million through its first-ever SPV.
Murphy called fundraising that SPV his most nerve-racking period: he personally faced occasional turndowns and “second- and third-order questions” while putting Menlo’s reputation behind an unfamiliar structure. Later shocks such as DeepSeek reinforced his view that AI presents a fresh “crisis and opportunity seemingly every six months.”
4. Margins matter, but model choice becomes a portfolio
Lovable was an “outlier even amongst outliers”: Menlo tried to enter near $30 million of ARR, then participated in a round when the company was around $150 million, while the business was described as going from zero to roughly $300 million in a year. Even decelerating to 3x implied a path from $300 million to $1 billion, supporting Harry’s cited $6.2 billion round valuation.
Founder-market fit completed the underwriting. Murphy described Anton as the category’s voice, with a vision to make the “99%” who never programmed into creators. Combined with unprecedented growth, that made Lovable look capable—conditional on continued compounding—of becoming “one of the most valuable companies of all time.”
Margins nevertheless “matter a lot.” Many leading AI applications currently sit at 20–30% gross margins; credible investments need a path toward 60–70% through inference optimization, reduced dependence on external APIs, and complementary models built with proprietary data. Murphy no longer assumes the 80–90% margins associated with classic software.
Harry proposed that open source might handle 96% of enterprise workflows and shrink frontier-model TAM. Murphy disagreed: cheaper models can serve undemanding calls, but Anthropic’s performance can increase retention, engagement, and revenue enough to outweigh cost savings. The end state is a 50/30/20-style tapestry spanning Sonnet, Opus, “Fable,” open source, and privately trained models.
5. Workflow depth decides which AI applications survive
Murphy rejected the idea that Anthropic and Lovable simply hedge one another. Anthropic approaches creation through the technical user; Lovable starts with the lay user. Some middle overlap is inevitable, but he sees enough market for both—and noted that Cursor prospered despite sitting directly in Anthropic’s path.
Harry’s sharper challenge concerned “Anthropic Legal” displacing Legora. Murphy answered that if a model performs the whole function and an application is not distinctive enough in its workflow or value, it may not be defensible anyway. Legal work, however, crosses organizational boundaries among corporate lawyers, outside firms, clients, and multiple firms on a case.
Legora’s defensibility therefore rests on lawyers and FDEs getting into and understanding workflows and context rather than simply having a model perform a task. Murphy described the coordination as “not quite an N-squared problem, but it’s complicated,” and said the same platform should expand naturally into tax, accounting, and other sophisticated service teams.
6. Series A is the market’s worst risk-adjusted insertion point
Harry characterized Series A as paying 200x ARR—roughly $200–400 million—for companies with $1–3 million of revenue and limited product-market fit. Murphy agreed: competitors may still be unknown, yet investors must price the company as though it has already won.
Menlo’s old “early growth” window was $3–10 million of ARR and once lasted 12–18 months. Today it can last a week; Murphy said that in Max and Legora’s case, it could happen in a day. The firm now waits for outliers above roughly $10 million that appear “anointed the winner,” or that Murphy believes will be, with the exact threshold varying by market.
The other end of Menlo’s barbell moved earlier. A three-partner seed strategy can write up to an $8 million check “on the spot,” versus $3 million previously. Murphy sees little signal when five POCs become $1 million of ARR but valuation rises from around $50 million to $200 million.
Harry argued sub-$100 million boutique seed funds risk being “too big to be friendly and too small to lead” against full-stack firms. Murphy agreed that traditional swim lanes and fears of institutional negative signaling have faded; larger rounds, preemptions, and renewed syndication increasingly make collaboration across stages more viable.
7. Full-stack venture requires focus, not institutional sprawl
Murphy framed Menlo’s roughly $3 billion scale as a deliberate choice rather than a reason to become a much larger organization. He prefers roughly 12 partners in a “small and mighty machine,” with fluidity across two funds and two investment committees, rather than five teams whose uneven results weaken collective alignment and agency.
Fluidity does not eliminate specialization. Murphy wants investors spending roughly 80% of their time in one stage or domain: a partner cannot simultaneously scout Stanford laboratories and chase the world’s 20 best growth rounds. In semiconductors, two or three years can pass before a chip ships, and an apparent design win can evaporate.
Geography supplies a similar information advantage. Murphy estimates Bay Area context can be “10 or 100X” greater because talent and technical conversations are concentrated there. Yet Lovable and Legora changed Menlo’s view of Europe: founders operating on “hard mode” demonstrate unusual grit, even if Menlo is not putting permanent boots on the ground.
The most common deal loss is arriving weeks before a round when another investor has cultivated the founder for a year. Menlo’s Anthology fund counters that with more than 50 company investments of $100,000–$1 million; graduates included OpenRouter, Whisper, and Axiom Math. Murphy estimates even a small wedge makes meaningful next-round participation “10X more likely.”
8. AI has reset growth hurdles and investor psychology
Harry defended rejecting a company growing from roughly $1.5 million to $5 million to $15 million: “triple, triple, double, double” is no longer exciting enough. Murphy agreed that zero-to-$100-million companies have changed the comparison set; what historically ranked in the top 5% can now resemble the top 50%.
Murphy’s changed mind is about how large companies can become and how boldly Menlo should pursue them. He said the firm needs free thinkers willing to take those risks.
Murphy’s Plaid miss once felt “existential” when Menlo lost at the “one-inch line.” His revised lesson is that no single loss defines a career: investors must keep pursuing major cycles rather than allowing one defeat to create defensive decision-making.
Harry argued richer investors become better investors because they optimize upside rather than mitigate downside. Murphy agreed at both firm and individual levels: doubt makes investors “dramatically worse,” while high trust permits failure, honest recognition, and helping founders “land the plane” instead of pretending reality has not changed.
After Murphy said Menlo’s Anthropic position was “north” of $10 billion, Harry estimated eventual carry at $2–3 billion. Murphy’s answer to complacency was Menlo’s challenger mentality: “We arrived. We’re here. What do we do with that?” The ambition is to compound its AI position, not treat the monetary outcome as completion.
9. Capital is overheating labs while infrastructure and healthcare reopen
Murphy flagged robotics and neo labs as overheated, and possibly defense tech, even though he likes all three. With more than 60 neo labs and Menlo invested in seven, he distinguishes focused efforts such as Chai in drugs and antibodies or Axiom in mathematics from generic teams promising to assemble researchers and discover a purpose later.
His arithmetic is unforgiving: “There’s no way in hell” 60 independent model companies survive alongside frontier and open-source alternatives, and acqui-hires cannot rescue them all. Many have nevertheless raised oversized rounds that create concentrated exposure for their investors.
Infrastructure may have suffered the opposite mistake. Early observability, agent-framework, and developer-stack startups struggled when customers used one model; multi-model optimization now creates demand for spend management, routing, and abstraction. Murphy says OpenRouter’s edge is its organic developer activity, trust, and intelligent inference marketplace—not merely being an underlying cloud provider—and calls it “wildly profitable.” Gimlet abstracts underlying chips and stacks such as CUDA.
At sufficient scale, model providers may also optimize specific training or inference workloads with custom chips, although Murphy called the chip business difficult and requiring a special team.
Over ten years, Murphy is most excited by therapeutics and healthcare delivery, citing Chai, Zaera, Villia, and Assort Health among Menlo’s efforts. Having lived through four or five major technology cycles, he believes this is the largest: what AI transforms over five or ten years could be “more mind-boggling than what we’ve seen in our society.”