20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

What It Covers

Matan Grinberg, cofounder of Factory (valued at $1.5B), discusses how enterprises should allocate tokens versus headcount, why model-agnostic application layers beat vendor lock-in, the three phases of enterprise AI adoption, and why labor displacement fears are overstated given the volume of unsolved problems software can address.

Key Questions Answered

• •Enterprise AI Adoption Phases: Enterprises move through three distinct stages: board pressure forcing an AI strategy, unconstrained "token maxing" where usage is measured as a performance metric, then an ROI hangover when bills arrive with no clear business impact. Companies like Uber now cap per-user AI spend. Leaders should front-load routing and cost governance before phase three hits, not after the shock.
• •Token-to-Salary Ratio: Within three years, enterprise token spend will reach the same order of magnitude as developer salaries. Today Salesforce spends roughly 3.8% of dev salary on Anthropic. The ratio varies wildly by role — some engineers delegating to dozens of parallel agents will spend multiples of their salary in tokens, while others who deliver value through customer contact may spend near zero.
• •Model-Agnostic Routing: Approximately 80–90% of software development tasks can be handled by open-source models; only planning and high-stakes decision steps require frontier models. Enterprises that route tasks to the appropriate model tier — open-source for implementation, frontier for architecture decisions — can cut token costs dramatically while maintaining output quality and avoiding single-vendor lock-in.
• •Core Competency Resource Allocation: The correct framework for AI investment is to identify the business's core competency, then measure every resource — headcount, dollars, tokens — against output metrics that directly move that competency forward. Kirkland spending $500M to build internal AI tools illustrates the failure mode: building software is not a law firm's core competency, making the spend likely to validate outsourcing to specialists like Harvey.
• •Full-Stack Engineer Redefined: The highest-leverage engineers in an agent-native environment own end-to-end business outcomes, not feature counts. They write marketing copy for releases, enable salespeople on new capabilities, and monitor product metrics — not just ship code. Competitive programming credentials and syntax memorization become weak signals; agency, ownership, and cross-functional range become the primary hiring filters.
• •Sales and Marketing as Product: Companies that treat engineering as first-class and sales or marketing as secondary will face compounding disadvantages when AI commoditizes technical differentiation. No legendary company has a poor sales or marketing team. Factory seats engineers and salespeople together, uses shared language ("we closed a deal," "we shipped a feature"), and treats the full customer journey from first brand contact through tenth renewal as the product.