The Twenty Minute VCWill Open-Source Threaten Anthropic's Business & Do Margins Matter in a World of AI | Matt Murphy
CHAPTERS
- 0:00 – 1:22
Why open source won’t displace frontier models (and what it will replace)
Matt opens by arguing that open-source models will be useful for many tasks, but won’t match the performance ceiling of frontier models like Anthropic’s for the most demanding workflows. He frames the likely future as multi-model usage, where companies mix premium frontier calls with cheaper open-source or in-house models based on the task.
- •Open source can be “functional” but may not reach frontier performance
- •Displacement risk is overstated for top-tier reasoning and quality
- •Enterprises will likely run a blended model stack, not one-model-only
- •Performance-driven outcomes (retention, engagement, revenue) can justify higher-cost models
- 1:22 – 7:14
How Menlo got into Anthropic: conviction vs. fund constraints
Matt explains the introduction to Dario and Tom via Anj Mittal and why he quickly developed conviction. The challenge was structural: a pre-revenue company at a multi‑billion valuation didn’t fit traditional venture check sizing or ownership targets, requiring flexibility from the partnership.
- •Warm intro through Anj Mittal; immediate call with Dario/Tom
- •Early signals: elite founder pedigree + strong technical credibility
- •Valuation/fund-fit tension: venture fund norms vs. $4B+ entry price
- •Menlo’s partnership flexibility enabled the initial investment
- 7:14 – 8:21
Why ownership matters less now: outliers, dilution, and modern venture math
The conversation shifts to whether ownership and price matter as much as they used to. Matt argues returns are now driven by massive outliers, making small stakes in huge outcomes more attractive than large stakes in modest exits, even with higher dilution across the market.
- •Outcome sizes have expanded dramatically vs. prior venture eras
- •Small % of mega-winners can beat large % of $300–$500M exits
- •Dilution is increasing broadly, not only in frontier labs
- •Rigid ownership targets can be a liability in today’s market
- 8:21 – 10:06
When price actually matters and why capital intensity is becoming normal
Matt clarifies that Menlo’s constraints are less about valuation purity and more about fund-size guardrails and the ability to keep investing where it matters. He notes a structural shift: fast-growing AI companies raise more often for offense, signaling, and retention dynamics, which normalizes heavier capitalization.
- •Stage/vehicle limits matter more than ‘cheap vs expensive’ pricing
- •Companies raise to play offense and keep pace with market signaling
- •Secondaries and frequent rounds are becoming part of the norm
- •AI compute/inference costs shift gross margin expectations
- 10:06 – 12:21
Sizing up: relationship-building, proof points, and leading the next round
Matt lays out the playbook Menlo followed after the first check: embed with the team, add value, and watch execution closely. As Anthropic launched and built revenue momentum—plus strategic partnerships with Amazon and Google—Menlo gained the confidence (and LP support) to lead the next round quickly.
- •Post-investment value-add: recruiting, BD, closer founder ties
- •Observed the revenue ‘drumbeat’ after model launch
- •Amazon + Google partnerships strengthened distribution and credibility
- •LP meeting became the catalyst; term sheet followed within weeks
- 12:21 – 13:43
SPVs in 2025: when to use them and the “bad actor” problem
Matt explains SPVs primarily as a tool to stay within fund mandates while still playing offense in exceptional rounds. He also discusses how the secondary/SPV ecosystem has attracted misleading ‘access’ claims, prompting founders to push back to control their cap tables.
- •SPVs help exceed per-company fund limits without resizing the fund
- •They can be used to win rounds with more capital + stronger support
- •Founder pushback is driven by uncontrolled marketing of private stock
- •Market has ‘bad actors’ claiming access and raising before securing allocation
- 13:43 – 14:55
Taking money off the table: liquidity vs. letting winners compound
Matt describes a bias toward holding onto true outliers rather than selling early, especially in an environment where winners compound faster and longer. Limited selling may happen to satisfy older-fund liquidity needs, but the default is to keep exposure and even add.
- •In a winner-takes-more world, selling winners early hurts returns
- •Partial liquidity can make sense for LP or fund-life reasons
- •Menlo spends more time on partnering and follow-on than on trimming
- •Secondary decisions depend on maturity and public-market timing
- 14:55 – 18:46
The nerve-wracking moments: first-ever $500M+ SPV and fast-moving AI shocks
Matt recounts the stress of running Menlo’s first SPV—especially before Anthropic was widely recognized—while fielding investor diligence and rejections. He also references periodic market ‘shock’ moments (e.g., DeepSeek, internal drama cycles) that create recurring spikes of uncertainty in frontier AI.
- •Menlo’s first SPV was also massive, increasing personal pressure
- •Capital raising created empathy for founders’ constant scrutiny
- •Frontier AI has recurring ‘crisis/opportunity’ cycles every ~6 months
- •Reputational and execution risk felt highest during the SPV build
- 18:46 – 21:58
Lovable and the margin question: underwriting hypergrowth in AI apps
Matt explains how he underwrote Lovable’s extraordinary revenue ramp and why the founder vision mattered alongside metrics. He then addresses whether margins still matter, arguing that many AI companies start with low gross margins due to compute, but must have a credible plan to reach healthier levels over time.
- •Lovable’s growth trajectory looked ‘outlier among outliers’
- •Founder-led category vision (making non-coders creators) was decisive
- •Margins still matter; many start at ~20–30% and need a path to 60–70%
- •Optimization strategies include infra improvements and complementary modeling
- 21:58 – 28:47
Open source vs Anthropic: routing, optimization, and multi-model reality
Harry presses the idea that open source could cover most enterprise workflows and shrink frontier TAM. Matt disagrees: frontier models can drive better business outcomes, while open source will serve cheaper calls that don’t need top-tier performance—making routing/optimization layers increasingly important.
- •Open source won’t take 96% of workloads in Matt’s view
- •Frontier models can increase retention and engagement, not just quality
- •Companies will optimize by task: price vs latency vs reasoning/performance
- •Wave 2 is multi-model sophistication; routing becomes strategic
- 28:47 – 30:44
OpenRouter as a “beast”: why the routing layer is defensible
Matt argues OpenRouter benefits from developer trust, organic pull, and a marketplace position that infrastructure providers may struggle to replicate. He suggests routing is especially valuable for teams that want to stay multi-cloud and abstract away underlying provider complexity.
- •Developer distribution and trust are core assets for OpenRouter
- •Nebius offering routing may work for its own users but isn’t a neutral layer
- •Routing helps companies optimize across providers and models
- •Matt claims strong profitability and momentum (without citing private numbers)
- 30:44 – 34:33
Legora vs ‘Anthropic as a threat’: why workflows beat raw model capability
Matt explains why Legora’s defensibility comes from deep legal workflows, multi-stakeholder complexity, and GTM execution, not just model access. He notes the market narrative has shifted from ‘SaaS is dead’ to sorting which apps have real workflow moats, and he believes Legora is in the latter camp.
- •Application defensibility depends on workflow distinctiveness, not just UI
- •Legal work spans firms/clients/contexts—hard for a model-only product to replace
- •Legora’s edge: embedded understanding of lawyer processes + execution speed
- •Broader trend: “SaaSpocalypse” cooled; now it’s about durable application value
- 34:33 – 39:52
Why Series A is brutal: the barbell strategy (go earlier or go bigger)
Matt agrees the Series A insertion point is especially hard due to compressed timelines, unclear competitive landscapes, and steep valuation jumps with limited new signal. Menlo’s response is a barbell: more seed activity (fast, flexible checks) and later concentrated ‘winner’ rounds once leadership is clearer.
- •A-round pricing often assumes category leadership before it’s proven
- •Signals between seed and A have weakened relative to valuation step-ups
- •Menlo seed strategy: faster decisioning, larger early checks, more optionality
- •Later-stage focus: pay up when a company is clearly breaking out
- 39:52 – 55:17
Why Menlo raised ‘only’ $3B: culture, alignment, and avoiding platform sprawl
Harry challenges why Menlo didn’t raise more given recent wins; Matt emphasizes the internal cost of scale. He argues very large platforms can dilute alignment and collaboration, while Menlo wants to stay small, flexible, and high-trust even with multiple vehicles.
- •More capital changes culture: staffing, structure, incentives, and cohesion
- •Menlo prefers “small and mighty” with tight partner alignment
- •Even with two funds/ICs, they aim for cross-vehicle fluidity
- •Goal is to compound AI advantage without becoming a sprawling platform
- 55:17 – 1:06:44
Wealth, risk-taking, and what’s overheated/underinvested in AI
Matt discusses how big wins can improve investor behavior by reducing downside fear and enabling bolder upside optimization. In rapid-fire, he flags neo-labs/robotics/defense as overheated due to volume and capital, while arguing AI infrastructure tooling (observability, frameworks, abstraction layers) is underappreciated as multi-model complexity rises.
- •Richer investors can focus more on upside than downside mitigation
- •Overheated: neo-labs proliferation, plus robotics and defense capital intensity
- •Underinvested: infra/dev tooling for multi-model ops (observability, frameworks, abstraction)
- •10-year excitement: healthcare breakthroughs and the unknown scale of AI transformation
