20VC: 70% of Neolabs Will Die | There Will be a $100BN US Open-Source Model | Data is a Trillion $ Market | Governments Cannot Regulate Models: It is Too Late | The Cyber Attacks to Come Will be Insane with Anastasios Angelopoulos @ Arena

Summary

Anastasios Angelopoulos frames Arena as a real-world evaluation layer for AI, using user interaction and agent traces instead of static benchmarks to measure performance, hallucinations, steerability, and actual work completion. He argues that Chinese open-source models have materially improved and, in some cases, surpassed leading American closed models on important tasks like front-end coding, challenging the U.S. narrative that these systems are mostly the result of distillation. A major thesis is that enterprises will increasingly want AI sovereignty: owning their stack, fine-tuning on proprietary data, and relying on data moats rather than software alone. He sees data as a scaling complement to model growth and predicts a very large market by 2030, while also warning that model routing, AI security, and guardian models will become essential as agents proliferate. The conversation also covers Neolab selection pressure, inference margin compression, model-provider encroachment into applications, and the rise of AI-enabled fake applicants and cyber threats.

Notes

• Arena is positioned as a real-world AI evaluation platform, not a static-benchmark shop.
• The company measures factuality, steerability, hallucinations, errors, preference, and task completion in production-like usage.
• KimiK3 is described as beating the best closed American models on an important task subset.
• Angelopoulos says Kimi beat all American models, including Fable, on some tasks.
• He argues distillation is only part of the story behind Chinese model progress.
• OpenRouter is said to distort inference visibility because of its fee-based business model and open-source-heavy usage.
• Enterprises are expected to want AI sovereignty and end-to-end ownership of their AI supply chain.
• He says software alone is no longer a moat because it can be generated too quickly.
• Data moats become central to durable businesses in his view.
• He predicts at least one massive American-first open-source company could become a multihundred-billion or trillion-dollar asset.
• Open-source monetization may come through revenue share with inference providers.
• He sees AI modernization as a large services market, not just software distribution.
• A sustainable western stack may be FDE plus open American models.
• He characterizes Thinking Machines / Inkling as effectively a version-zero model built quickly.
• The U.S. chip ecosystem is described as a key tailwind, with China hardware-constrained.
• Export controls may push China to build a separate hardware ecosystem.
• He notes China already restricts U.S. models domestically.
• A locally hosted model does not eliminate jailbreak or exfiltration risk.
• He thinks restrictions on Chinese open models in the U.S. are likely, but timing is uncertain.
• Closed-source AI would mean less choice, higher costs, and less competition.
• Arena thinks it needs at least three major model providers to stay healthy.
• Enterprise buyers are said to be nervous about both frontier labs and Chinese open-source models.
• A Fortune 50 customer reportedly asked to replace Qwen with an American model.
• He lists several American open-source contenders and says Gemma is efficient on the performance-cost curve.
• Routing models is technically hard because query difficulty and model performance vary by use case.
• A routing layer must also track model performance and onboard new releases quickly.
• He expects inference prices to fall over time as markets get more efficient.
• Public-margin disclosure could improve buyer negotiating leverage.
• He says Anthropic has the strongest incentives to IPO first.
• If open models outperform frontier models broadly, that could hurt Anthropic's IPO case.
• He says the OpenAI/Hugging Face incident showed model breakout risk and the need for open-source defense.
• He argues businesses need guardian models that monitor agent traces and make safety judgments.
• He rejects a central government body deciding when models can be released.
• He prefers outcome-based regulation and fines over pre-approval bureaucracy.
• He says AI-enabled fake candidates are already passing interviews at Arena.
• He does not know whether the threat is espionage, cybercrime, or nation-state activity.
• Arena is considering in-person onboarding and identity verification to reduce fake applicants.
• He says the hiring market for elite AI talent is brutally competitive.
• Top researchers with deep expertise and many citations are expensive.
• He says there are at least 75 Neolabs, and roughly two thirds will be worth nothing or sold for parts.
• Winning Neolabs need strong strategy and a real business model, not just model quality.
• He says a $10B valuation needs 25x-30x revenue expansion to justify a $100B outcome.
• That implies roughly $4B of revenue in two to three years.
• He uses Mistral and ElevenLabs as examples of companies that may do well.
• Data is framed as a scaling complement to model growth, not a commodity.
• He says frontier labs spend roughly 10% to 20% of GPU spend on data.
• He predicts the data market could reach at least $100B by 2030, potentially $1T.
• Data businesses may expand into enterprise model-building and routing.
• He says every business will need its own AI model and its own data moat.
• Arena says agent evaluations are driven by a large consumer product with 30M+ monthly visitors.
• Those users create an organic flywheel for agentic evaluations based on real work.
• Evaluation is described as the biggest bottleneck to AI deployment because value is hard to define.
• Arena extracts organic performance measurements from agentic traces.
• Arena says it is above a $100M annualized revenue run rate.
• He distinguishes productive businesses from GMV-style token or GPU reselling models.
• A terminal-value lens matters: hosting GPUs alone should not command much more than electricity-cost economics.
• He says model providers moving into apps is a real threat to application vendors.
• He specifically flags Harvey and Legora as exposed to model-app competition.
• More operations-heavy businesses may prefer large labs and system-integrator style deployments.
• He ends by saying AI in medicine could eradicate diseases one by one, but biology needs fast feedback loops.
• The missing piece in biology and medicine is data infrastructure, not compute.