20VC: Is Anthropic's Coding Business Worth $2 Trillion? | Should American Enterprises Work With Open-Source Chinese Models? | Why 80–90% of Neo-Labs Die in the Next 18 Months? with Eno Reyes, Co-Founder @ Factory

What It Covers

Eno Reyes, co-founder of Factory, breaks down the AI value stack with Harry Stebbings, arguing that frontier model valuations are overstated, open-source models will handle 99% of workflows within three years, 80-90% of Neo-Labs will collapse within 18 months, and sovereign intelligence ownership becomes the defining enterprise challenge of the next five years.

Key Questions Answered

• •Outcome-based model pricing: Stop measuring AI cost by token price and start measuring by outcome cost. A premium model completing a code review in 1,000 tokens beats a cheap model consuming 50 million tokens on the same task. For cognitively demanding work, the highest-quality model frequently becomes the lowest-cost option when total spend per completed outcome is calculated rather than per-token input rates.
• •Frontier model TAM overestimation: Current $2-4 trillion valuations for Anthropic and OpenAI embed assumptions that token prices can double while customers stay loyal. Both companies face a structural trap: staying model-locked limits their ability to deliver best-in-class outcomes, while opening to competing models erodes their core margin story. Investors should pressure-test whether any path to margin expansion exists beyond simply raising prices.
• •Sovereign intelligence as the core enterprise risk: Any enterprise outsourcing its workflows entirely to a model provider risks that provider replicating and eventually competing against its core business. Within five years, the critical question becomes who owns the learning generated from your workflows. On-premise deployments matter less for technical reasons than for the contractual and strategic assurance that intelligence assets remain under company control.
• •Open-source model framing is a competitive psyop: Labeling DeepSeek and similar models as "Chinese models" functions as a marketing strategy by frontier labs to discourage adoption through fear. Enterprises should evaluate all models, including Anthropic and OpenAI, on three identical criteria: what content does the creator censor, does it solve the specific task, and can you switch providers within six to twelve months if the model is discontinued or altered.
• •Neo-Lab survival filter — three durability questions: Before investing in or building an AI application company, apply three tests: Is the workflow durable and proprietary? Will frontier model improvements make the workflow obsolete without access to the company's specific data? Would a new entrant immediately find a superior approach? Legal-focused Neo-Labs pass all three tests. General knowledge-work tools operating in Excel or Jira-style environments fail and represent the majority of the 80-90% projected to collapse within 18 months.
• •Hiring via micro-acquisition over traditional recruiting: Factory plans to source close to 100% of future hires by acquiring small teams and solo founders already building directly in the autonomous software development space. A founder who quit their job and spent eight months building an open-source project demonstrates more verifiable conviction than any interview process can surface. Pedigree and competition wins signal rule-following, not the capacity to operate outside existing system boundaries.