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
Eno Reyes, CTO and co-founder of Factory, discusses how the smartest AI models will become the cheapest when measured by outcomes rather than token costs, predicts 80-90% of AI neo-labs will fail within 18 months, and argues that 99% of workflows will run on open-source models within three years while frontier models capture only 1% of usage but 30-40% of economic value.
Summary
In this in-depth conversation with Harry Stebbings, Eno Reyes presents a contrarian view of the AI landscape that challenges prevailing assumptions about frontier model providers like OpenAI and Anthropic. The discussion begins with Reyes's central thesis: the cheapest model isn't necessarily the cheapest system. He illustrates this with the example of code reviews—a sophisticated model that completes a review accurately using 1,000 tokens is cheaper than a basic model requiring 50 million tokens to achieve the same outcome. This frames AI economics around outputs and outcomes rather than inputs.
Reyes argues that the TAM for frontier models is significantly overweighted in current valuations. While many investors assume one to three companies will dominate the intelligence era, he contends this is unlikely due to margin compression and competitive dynamics. He specifically questions the $2 trillion valuation being discussed for Anthropic's core code business, noting that code completion is not difficult to switch away from and represents one of the most competitive application markets in dev tools.
The conversation explores the emerging model ecosystem, where specialized, post-trained models will dominate over time. Rather than millions of models, Reyes predicts "quite a lot" of specialized models—where enterprises take commodity models and fine-tune them on proprietary data for internal use only. This requires democratizing access to post-training tools, which Reyes believes will happen through platforms similar to those that democratized software development.
On verifiability, Reyes emphasizes it as the single most important property of AI system success. He explains how frontier AI systems are now focused on building verification where none exists, allowing progression into tasks previously considered too difficult. He uses law firm hiring as a metaphor: gut decisions work at scale only if systematized into written frameworks that define "good" outcomes.
Reyes discusses Factory's approach to building harnesses—the logic layer where state is maintained and AI integration happens. He contrasts this with gateway routing solutions like OpenRouter, arguing that true stateful intelligence allocation requires integration within agentic workflows rather than external routing. The harness, not the model, becomes the application layer.
A critical theme is sovereign intelligence—the ownership of learnings and workflows generated from business tasks. Reyes warns that outsourcing intelligence to model providers creates existential risk, citing OpenAI and Anthropic's explicit statements about entering industries they provide intelligence for. This is why on-premise solutions matter: they provide assurance of ownership and control.
On open-source models, particularly Chinese-developed ones, Reyes reframes the narrative, calling concerns about security a "psyop" by frontier labs. He argues open models are fundamentally similar to frontier models, just with different creator biases. For most tasks like code review, there's no meaningful security difference. The real distinction is that frontier labs have explicit incentive misalignment when model-locked to their own technology.
Reyes predicts that in three years, 99% of workflows will run on open models, but the remaining 1% will capture 30-40% of economic value. This bifurcation reflects how frontier models will concentrate on truly frontier problems—bio research, AI development, security—while commodity tasks migrate to open alternatives at lower cost.
On neo-lab survival, Reyes applies three durability tests: Is the business attached to a durable workflow? Will that workflow change if frontier models improve? Will the workflow survive introduction to new businesses? Legal AI passes all three; general computer use fails, making many AI startups likely to consolidate or fail within 18 months.
Factory's hiring strategy diverges from traditional approaches. Rather than optimizing for pedigree, Reyes emphasizes hiring founders and mission-aligned builders already working on the problem. He explains that people who've built something in their spare time demonstrate more conviction than interview performance alone reveals. The company avoids subsidizing free-tier users, betting that superior product and open-model alignment will eventually drive adoption without needing to artificially inflate usage metrics.
On enterprise sales, Reyes shifts from persuasion to discovery, positioning sales conversations as joint problem-solving rather than convincing customers. He emphasizes that enterprises value partners who help identify real unsolved problems over those pushing predetermined solutions.
Reyes contrasts Factory's human-AI partnership model with competitors positioning themselves as labor replacements. He argues the future of software development isn't about swapping humans for AI, but building new development methodologies where humans and AI collaborate—a fundamentally different narrative from agents replicating engineer behavior.
On valuations and growth, Reyes believes the $2 trillion+ assessments for frontier model companies significantly underestimate the scale of transformation, yet these valuations are still at risk due to margin compression and competitive intensity. He positions Microsoft as best-positioned among hyperscalers due to its independence from any single model provider, having hedged bets across OpenAI, Anthropic, and open models.
About this episode
<p dir="ltr">Eno Reyes is the co-founder and CTO of Factory, the agent-native software development platform building autonomous "Droids" for enterprise engineering teams. Factory has raised $220 million, most recently a $150 million Series C at a $1.5 billion valuation, from investors including Khosla Ventures, Sequoia Capital, 20VC, NEA, Blackstone, Insight Partners and Nvidia. Before founding Factory, Eno worked as a machine-learning engineer at Hugging Face, training, optimizing and deploying large language models for enterprise customers.</p> <p dir="ltr">AGENDA:</p> <p dir="ltr">00:00 Are We Underestimating AI by an Order of Magnitude? </p> <p dir="ltr">06:35 Why Can the Smartest AI Model Be the Cheapest? </p> <p dir="ltr">18:51 Is Anthropic's Coding Business Really Worth $2 Trillion? </p> <p dir="ltr">33:41 Will Continuous-Learning Models Help or Hurt Factory? </p> <p dir="ltr">40:43 Will 80–90% of Neo-Labs Die in the Next 18 Months? </p> <p dir="ltr">44:33 Should American Enterprises Work With Open-Source Chinese Models? </p> <p dir="ltr">55:42 Must AI Founders Radically Rethink What a Great Outcome Looks Like? </p> <p dir="ltr">1:04:30 Do Pedigree and Credentials Still Matter in AI Hiring? </p> <p dir="ltr">1:19:17 Which Is the Biggest Threat: Claude Code, Codex, Cognition or Cursor? </p> <p dir="ltr">1:24:20 What Seems Crazy Today but Will Be Obvious in Five Years?</p> <p> </p>
Key Insights
- Reyes argues that outcome-based pricing, not token-based pricing, will determine the true cost of AI—a sophisticated model using fewer tokens can be significantly cheaper than a basic model burning through 50 million tokens for the same task.
- The frontier model TAM is overweighted in current valuations because the assumption baked into $2-4 trillion valuations is that prices can double without losing customers, which competitive dynamics and open alternatives make unlikely.
- Reyes predicts 80-90% of neo-labs will fail or consolidate in the next 18 months, with survival determined by whether a business is attached to durable workflows that won't change as models improve and can survive introduction to entirely new industries.
- Sovereign intelligence—owning the learnings and workflow outcomes generated by AI use—will become the critical business concern as enterprises recognize that model providers like OpenAI and Anthropic have explicitly stated intentions to enter the industries they provide intelligence for.
- The harness, not the model or gateway router, is becoming the true application layer because it maintains state and enables stateful intelligence allocation that must happen within agentic workflows, not external to them.
- In three years, 99% of workflows will run on open-source models, but this 1% of frontier model usage will capture 30-40% of the economic value, concentrating frontier models on frontier problems like bio research and AI development.
- Open-source models, including Chinese-developed models, pose no greater security risk than American frontier models for most tasks—concerns are primarily about different creator biases, making the security narrative a marketing tactic by frontier labs.
- Building businesses requires reactive decision-making based on real-time feedback from customers and market signals, not predetermined master plans, making clarity on future outcomes difficult and making founders who operate outside system rules more valuable than credentials.
- Verifiability is the single most important property of AI system success, with frontier AI's new frontier being systems that can build verification where none exists, enabling progress into tasks humans consider too difficult.
- Mission alignment in hiring means candidates already deep in the exact problem space, with demonstrated conviction through built products, are more valuable signals than traditional pedigree or performance in interviews.
- Performative work culture and grinding as signaling are correlated with hiring mistakes, whereas outcome-focused expectations where occasional high-intensity weekend work happens contextually to solve real client problems is optimal.
- Enterprise sales success comes from treating customer conversations as joint problem-solving discovery rather than persuasion, with enterprises valuing partners who help identify previously unsolved problems over those pushing predetermined solutions.
Topics
Transcript
I see a world where the smartest model is actually the cheapest. People are thinking about outcomes in AI, and they're looking at 20, 30, 50, and they're saying, that's ludicrous, that's crazy. That is underestimating by an order of magnitude how massive a transformation this is going to be. The TAM of frontier models is, frankly, overweighted right now. Eight and 10 billion is the new one billion. Two of the largest companies that provide models today have explicitly said we are going to go after every single one of these industries and businesses that we provide intelligence for. I think it could be 80 to 90 percent of Neo labs die in the next 18 months. Calling open…
Full transcript available for MurmurCast members
Sign Up to AccessMore from The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
20VC: Is Seed Investing Dead Without a $1BN Fund? | Does Ownership and Price Matter When Companies Can Be $1TRN Exits | Are AI Revenue Numbers Real and What to Watch Out For with Venky Ganesan, Menlo Ventures
Venky Ganesan from Menlo Ventures discusses how venture capital has fundamentally changed in the AI era, arguing that seed investing still exists but founders must understand that venture is now focused on outlier companies with explosive growth rather than steady revenue growth. He emphasizes the importance of IRR over ownership percentages, the dangers of gaming metrics, and the need for VCs to play the current market game while maintaining authentic values.
20VC: The Future of Datacentres: What You Need to Know | Why Everyone Gets GPU Depreciation and AI's Energy Costs Wrong | Who Really Makes Money From AI & Why Most Moats Don't Exist with Chase Lochmiller
Chase Lochmiller, CEO of Crusoe Energy, discusses how data centers and AI infrastructure are becoming critical bottlenecks, arguing that energy-distributed computing rather than centralized hubs will power AI's future, and that vertical integration across data centers, GPUs, and inference services creates sustainable competitive advantages in the AI infrastructure space.
20VC: $1BN ARR in 18 Months; The Untold Story of Higgsfield | Spending $4M Per Month on Models | Why Moats in AI are BS | Scaling a Content Team to 150 People with Alex Mashrabov
Alex Mashrabov, founder of Higgsfield, discusses how his AI video generation startup achieved $1 billion ARR in 18 months—faster than Coursera—by focusing on product-market fit, strategic model routing between open-source and proprietary AI models, and building a 400-person team primarily based in Kazakhstan. He shares insights on AI moats being largely meaningless, the $4 million monthly spend on model inference, and his vision for Higgsfield to become bigger than Shopify by owning the distribution layer for AI-generated content.
20VC: Jensen Huang Declares AGI Has Arrived | GPT Astra and Fable 5.1 Accelerate the Model Race | Tesla Launches Cybercabs | Index Pulls Out of Town & Anthropic Pulls From Descartes Acquisition
In this 20VC episode, Rory O'Driscoll and Jason Lemkin discuss major tech news including Jensen Huang's AGI claims, AI agent risks and rule-breaking behavior, Elon's Cybercab launch, venture capital conflicts around AI assistants, and high-growth companies like Wonderful raising at massive valuations with significant secondaries.
20VC: The $100 Billion AI Assistant Race: Town vs Instinct vs GrokBot | We Spend $75K Per Engineer on AI Tools | Why the AI Assistant Market Is Not a Bubble & AI Assistants Will Replace Every App on Your Phone with JD, Founder of Town
JD, founder of Town, discusses how his AI assistant platform operates within email and calendar to automate work tasks, argues that network effects at the agent level will be the true moat in the AI assistant market, and explains why the speed of market development means companies can only learn at human speed while building at machine speed.