readonrisk

a16z (Andreessen Horowitz) · Oct 14, 2026 · 59 min

Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert

Nathan Lambert argues that open models will remain a tier behind closed frontier labs because of compute, data and revenue-flow advantages, but fast-following is very doable and open models fill valuable enterprise niches. He is skeptical that distillation explains Chinese progress, expects no fast RSI takeoff because scaling needs exponential compute for linear gains, and stresses that the RL-environment and data industry is the key driver of capability gains.

Watch on YouTube ↗More from a16z (Andreessen Horowitz)

Where they stand

AI

Bullish

AI progress will keep compounding with heavy spend, driven by data and efficiency, without a fast RSI takeoff, and with many winners across closed and open models.

Medium conviction · years

Key takeaways

  1. Model progress comes down to compute, data and talent; closed labs (OpenAI, Anthropic, Google) have the most, and Chinese labs match on talent but lack compute. 0:00
  2. Open models will form a tier one step behind the best closed models, serving markets with private enterprise data and lower margins; fast following will remain very doable. 5:28
  3. Lambert doubts distillation is a major factor in Chinese lab progress given the public post-training literature, though he sees big error bars and thinks cracking down would cause more harm than good. 10:55
  4. Open model economy takes longer to build because information and revenue do not flow back to builders, but companies ultimately want owned weights and a positive ROI. 19:44
  5. Frontier progress outside math and code comes largely from buying high-quality data and RL environments; the data industry is early, with quality and opsec problems and reward hacking, and will take years to mature. 25:12
  6. Model serving costs are expected to fall multiplicatively (2x every 8 months), raising a question of Jevons paradox versus shrinking revenues. 26:19
  7. No fast takeoff expected: scaling laws need exponential compute for linear gains, so frontier labs can't pull away indefinitely; their edge is cheap inference and serving. 55:09
  8. Greater worry is the near-nonexistent societal preparation, such as cybersecurity, for capabilities already arriving. 57:21

Calls and positions

  • Open models will settle into a near-frontier tier a step behind the best closed models, which command the best margins.
    “there's the near frontier models which is a step behind this. And those will be open”
    NeutralPrediction: open models a step behind closed frontierNathan Lambert, Oct 14, 2026Watch at 5:28#OpenAI#Anthropic
  • Model prices will fall about 2x every 8 months for the next 3 years, reaching about 1% of cost in roughly 4 years.
    “reduce in price 2x every 8 months for the next 3 years”
    NeutralPrediction: 2x price drop every 8 months (next 3 years)Nathan Lambert, Oct 14, 2026Watch at 26:19#OpenAI
  • Open models will be fine on harnesses because labs were pressured to support many harnesses; open labs will copy new multi-agent coordination systems after frontier labs pioneer them.
    “I think it'll be okay for open because I thought years ago that say the closed labs would be like Apple”
    BullishPrediction: open labs follow on agent swarmsNathan Lambert, Oct 14, 2026Watch at 50:48#OpenAI#Anthropic

Views expressed

  • Chinese labs have equal talent to US labs but are far more compute-starved, which is the biggest difference.
    “to say that China is anything but equal in talent is going to you're going to have a hard time”
    NeutralNathan Lambert, Oct 14, 2026Watch at 8:44
  • Evidence from the public literature makes it hard to believe distillation is a major factor in current reasoning post-training.
    “it's very hard for me to be like distillation is a major factor”
    NeutralNathan Lambert, Oct 14, 2026Watch at 10:55
  • Taking action against distillation would cause more harm than good because Chinese labs will keep doing it.
    “you're going to cause more harm than good if you do anything about distillation”
    NeutralNathan Lambert, Oct 14, 2026Watch at 15:54
  • Open model economy takes much longer to build than closed labs' integrated stack, but ends with customers owning weights with no terms and conditions.
    “they have real ownership, which is model weights, no terms and conditions that they can run on their hardware”
    NeutralNathan Lambert, Oct 14, 2026Watch at 20:50
  • Much frontier progress in non-verifiable domains comes from buying high-quality data and building worlds for models to hill climb on, not magic generalization.
    “a lot of the frontier progress in domains that are not fully verifiable is by buying data that is high quality”
    BullishNathan Lambert, Oct 14, 2026Watch at 25:12#OpenAI
  • The RL environment boom is leading to cut corners and reward hacking, but the labs are improving fast and quality will rise over a few years.
    “the data industry has a lot of low-hanging fruit on improving quality and improving just like opsec”
    BullishNathan Lambert, Oct 14, 2026Watch at 30:40
  • Frontier labs will win on serving the cheapest intelligence per watt through more inference infrastructure and higher margins.
    “I'm very confident in the frontier labs being able to serve the cheapest intelligence per per watt”
    BullishNathan Lambert, Oct 14, 2026Watch at 52:59#OpenAI#Anthropic
  • A fast takeoff where models build themselves is unlikely because scaling laws require exponential compute for linear performance gains.
    “The all the scaling laws are exponential increase in compute for linear increase”
    NeutralNathan Lambert, Oct 14, 2026Watch at 55:09
  • New pre-training entrants are viable because a relevant model does not need $10 billion; the hard part is making a model good enough to support the next fundraise.
    “I'm like not betting against new entrants because like like it's not that you need 10 billion dollars”
    BullishNathan Lambert, Oct 14, 2026Watch at 45:52
  • Preparation for AI-driven risks like cybersecurity is far below what is needed, which worries him more than RSI.
    “the pretty much non-existent preparation of cross public cybersecurity”
    BearishNathan Lambert, Oct 14, 2026Watch at 57:21#AI and cybersecurity

Topics covered

More from a16z (Andreessen Horowitz)

Summary generated by AI from the video's captionsand checked against it; quotes are verbatim. Processed Oct 25, 2026.

Read on Risk summarises what other people said in public videos and podcasts. Summaries are generated by AI and checked by code against the transcript, but can still be wrong; follow the source links before relying on anything. This is not investment advice and nothing here is a recommendation to buy or sell. Disclaimer ·How this works