Arena is the central AI evaluation platform with organic flywheel from 30M monthly knowledge-worker users; past $100M ARR growing fast; building guardian models for agent safety.
Arena past $100M ARR with 30M+ monthly users, evaluation is bottleneck for AI deployment
Arena is one of largest consumer AI apps (30M+ monthly visitors, mostly knowledge workers); organic flywheel of real-world usage builds agentic evaluations; every business needs e…
Anastasios: Kimi K3 beat US closed models, violating Chinese distillation narrative
Kimi K3 (27T params) beat best US closed models on frontend coding tasks, proving Chinese labs are innovating beyond distillation; shifts scientific hegemony narrative and open so…
Thinking Machines could become $100B+ American open-source AI champion
Only 6 months into current model development after restructuring, already #1 US open-source model (#10 globally); strong proponent of their FDE + open model strategy for enterpris…
Anastasios: Thinking Machines will become $100B+ US open source champion
Strong proponent of Thinking Machines; believes US will produce at least one multi-hundred-billion to trillion dollar open source company; Inkling v0 achieved #1 US open source in…
Mistral executing well on open-source + services model, likely public company
Mistral's strategy of open-source model as lead gen for fine-tuning/services is viable; AI modernization market (helping enterprises retool) will be massive over next 10 years.
Anastasios: Mistral pursuing sustainable open source + services model
Mistral is executing well on open source business model combining revenue share with inference providers and deployed engineer services for enterprise AI modernization.
Etched innovating in US chip ecosystem, critical for national security
US chip ecosystem (Nvidia/TSMC/Etched) is a national security necessity; export controls may incentivize Chinese hardware competition; Etched is an example of new US chip innovati…
Anastasios: Etched innovating in US chip ecosystem, national security priority
Etched is a new US chip company innovating to keep Nvidia/TSMC ecosystem ahead; US government should protect and grow this semiconductor supply chain as national security necessit…
Sam Altman's lobbying power may drive US restrictions on Chinese open models
Sam Altman is the best politician in tech and will coalesce investors to lobby for restrictions on Chinese models to protect US labs; OpenAI not ready to IPO this year due to inte…
Nvidia benefits from open source proliferation and enterprise AI adoption
Open source models drive more training and fine-tuning on GPUs, increasing Nvidia's TAM; Jensen's pro-open-source letter is both patriotic and self-serving; enterprise AI adoption…
Anastasios: Enterprises demanding removal of Qwen from US stacks due to sovereignty
Fortune 50 enterprises are actively requiring US vendors to replace Qwen with American models due to regulatory/sovereignty concerns, creating tailwind for US open source alternat…
Anastasios & Harry: Poolside among top 5 US open source contenders
Poolside listed as one of five American open source contenders (along with RC, Reflection, Mistral, Thinking Machines, Google, Nvidia); Harry was impressed by their Laguna model.
Anastasios: Gemma models efficient on Pareto curves, strong open source contender
Google's Gemma models show strong performance-per-dollar efficiency on Arena leaderboards, making Google a significant open source player alongside Nvidia.
Anastasios: Anthropic generating FCF, best positioned for IPO this year
Anthropic has hockey-stick revenue growth, generating free cash flow, and appears better prepared for IPO than OpenAI (possibly as soon as October); high inference margins will fa…
Anthropic generating free cash flow, positioned for IPO as early as October
Anthropic has disgustingly high inference margins and is generating free cash flow, making it better prepared for public markets than OpenAI; public listing will force margin tran…
Anastasios: OpenAI not ready for IPO, faces open source disruption risk
OpenAI has internal discussions suggesting it's not ready to IPO this year; if open models beat Opus 5 or GPT-5 across categories, it would be a major business risk for their publ…
75+ Neolabs exist, 2/3 will be worth zero; sustainable business model now required
Market has become P&L driven; creating a model is old news; need hypergrowth revenue ($4B+ in 2-3 years to justify $10B valuation); zero-revenue $3B valuations rely on team acquih…
Data providers are scaling complements to AI, could reach $100B+ valuations by 2030
Data is a scaling complement to model training (more/bigger models = more data needed); durable need until AGI; companies spend 10-20% of GPU budget on data; revenue concentration…
Harry & Anastasios: Mercor generational revenue ramp, raising at $20B
Mercor is a generational revenue ramp company in data labeling; Harry is an investor; Anastasios believes data providers could reach hundreds of billions in value as 3-5% of front…
Fireworks at mid-30s margins, inference layer faces terminal value questions
Inference providers reselling tokens/GPUs have tough terminal value - why pay more than electricity cost? Margin structure must support public company; Fireworks at 30% margins vs…
Anastasios & Harry: Harvey CEO admits model labs are biggest competitive threat
Frontier model providers (OpenAI, Anthropic) will move up application layer to avoid commoditization, directly threatening vertical AI apps like Harvey in legal; Harvey CEO confir…
Salesforce has strong AI strategy, SaaS apocalypse overstated for entrenched platforms
Salesforce and ServiceNow have deep entrenchment, network effects, and data moats that are hard to replicate; model providers moving up application layer is a risk for vertical ap…
Anastasios: Salesforce will thrive, SaaS apocalypse overstated due to entrenchment
Salesforce has strong AI strategy and deep enterprise entrenchment (data, network effects) making it resilient; SaaS disruption is overstated because replicating incumbent moats i…
Anastasios: Nvidia leads $10T race, enterprise AI will 10x demand
Nvidia has the best chip ecosystem and benefits from open source proliferation driving more training/fine-tuning; export controls protect its moat; enterprise adoption will reliab…
Korean memory stocks down 40%, market realizing overinflation without demand destruction
SK Hynix and Samsung crashed ~40% on realization of overinflation; no destabilizing factor in AI demand suggests this is valuation reset not demand destruction; bonuses paid befor…
ServiceNow incredibly difficult to displace due to enterprise entrenchment
Deep integration and stickiness in enterprise workflows makes ServiceNow resilient to AI disruption; entrenchment is key differentiator vs less integrated SaaS.