Odd Lots
Odd Lots

Nov 20, 2026 · 56 min

Listen from 52:39

Listen at 52:39

Open models reshape AI startup economics

How an American AI Company Uses Chinese Open-Weight Models

The episode examines whether cheaper, customizable models can reduce AI operating costs while improving data control and bargaining power.

3 key takeaways
  1. 1Frontier models can accelerate product development, while cheaper open-weight models handle repeatable customer workflows.
  2. 2Fine-tuning creates specialization but introduces uncertainty when new model releases require retraining and redeployment.
  3. 3Chinese models are advancing beyond simple imitation, with real progress in architecture, efficiency, and inference techniques.

Don't miss

Daniel Saedi’s clearest test for Chinese-model progress is practical substitution: whether companies can use them in real workflows without noticeable performance losses.

The brief

Daniel Saedi explains how Minerva combines expensive frontier models with cheaper open-weight systems, using each where its cost and capability make the most business sense.

The case for open-weight models extends beyond price: self-hosting can protect data sovereignty, enable fine-tuning, and give startups leverage against frontier-model vendors.

Fine-tuning turns proprietary customer data into a competitive advantage, but moving that advantage across model releases remains an operational and technical risk.

The conversation also confronts uncensored models, whose removed safeguards could lower the cost of offensive cyber operations while complicating legitimate open-model development.

Saedi argues that Chinese models should be judged by real workflow substitution—not benchmarks alone—as advances in architecture and inference narrow practical performance gaps.

What was said on this episode

36 statements · 19 positive · 10 negative · 2 mixed · 5 neutral

  1. Daniel Saedion Minerva marketing automationPositive4:36

    Minerva uses AI agents to automate marketing campaigns from planning through execution.

    “We are an AI company focused on marketing. So we use models to build agents to automate everything from pre-campaign planning. all the way down to execution.”

    Listen at 4:36

  2. Daniel Saedion Minerva AI task economicsPositive7:03

    Minerva optimizes for acceptable task completion at minimum cost to preserve margin.

    “For us, it's how do I make sure that I get the minimum acceptable task completion for the lowest amount of costs? Because Everything outside of that is our margin.”

    Listen at 7:03

  3. Daniel Saedion Open-source artificial intelligenceNeutral8:24

    True open-source models disclose code, data, and training recipes.

    “Open source means that you're disclosing all of the code, all of the data, the exact recipe that was used to train a model.”

    Listen at 8:24

  4. Daniel Saedion Chinese open-weight modelsPositive8:47

    Chinese labs release downloadable model weights that users can self-host.

    “The Chinese labs are at least saying. OK, you can download this model and you can now host it yourself, essentially.”

    Listen at 8:47

  5. Daniel Saedion Open-weight modelsPositive9:43

    Fine-tuning open models is currently the primary benefit for application companies.

    “I actually think that this is the primary benefit to the application layer right now is the ability to fine tune on top of those open.”

    Listen at 9:43

  6. Daniel Saedion Minerva fine-tuned modelPositive11:40

    Minerva’s fine-tuned model is faster and better for its specialized queries than closed models.

    “And now we have a model that's faster than the closed source models for our use case is better in the sense that it takes complicated queries and actually produces the results that our customers want.”

    Listen at 11:40

  7. Daniel Saedion Open-weight modelsPositive12:21

    Open-weight models can give American application companies a margin-generating advantage.

    “Now you have an advantage that you can start to charge a margin for. And that's really the benefit of the open weight models for American companies.”

    Listen at 12:21

  8. Daniel Saedion Open-weight modelsPositive12:59

    Cost is the most important benefit of open-weight models.

    “I think the most important thing is cost.”

    Listen at 12:59

  9. Daniel Saedion Data sovereignty in AIPositive13:23

    Data sovereignty will become a major issue for companies in 2027.

    “I imagine this is going to be a huge issue in 27 as companies realize they're sending their data out to Anthropic and OpenAI and they're not getting anything in return.”

    Listen at 13:23

  10. Daniel Saedion Data sovereignty and model trainingPositive14:20

    Data sovereignty enables companies to improve models using their own data.

    “what sovereignty gives you is the ability to say, OK, if I'm running Kimmy a bunch of times on my data, maybe I should be training it on my data as well so that it gets better.”

    Listen at 14:20

  11. Daniel Saedion Open-weight modelsPositive14:33

    Owning an open model reduces dependence on a single AI vendor.

    “you're no longer beholden to a single vendor.”

    Listen at 14:33

  12. Daniel Saedion Fine-tuned model weightsNegative18:20

    Fine-tuning does not automatically transfer to a new model release.

    “It doesn't simply transfer over.”

    Listen at 18:20

  13. Daniel Saedion Fine-tuning durabilityMixed18:58

    It is uncertain whether the same data will produce optimal weights in a new model version.

    “whether or not it's going to converge to an optimal set of weights with the same training data set, I don't know.”

    Listen at 18:58

  14. Daniel Saedion Self-hosting open modelsNegative20:36

    Self-hosting fine-tuned models previously required an inference-engineering team.

    “You actually would need a team of inference engineers to be able to do this in the past.”

    Listen at 20:36

  15. Daniel Saedion FireworksNeutral21:19

    Fireworks charges roughly $7–$9 per million input and output tokens for fine-tuning.

    “I'm going to pay around $7 to $9 per million input and million output tokens”

    Listen at 21:19

  16. Daniel Saedion Open-model fine-tuning economicsPositive21:36

    The economics of fine-tuning open models have improved dramatically.

    “the economics have collapsed.”

    Listen at 21:36

  17. Daniel Saedion AI training dataPositive23:40

    High-quality data is currently the main source of model-training advantage.

    “The alpha, per se, as you would say, is in the data right now.”

    Listen at 23:40

  18. Daniel Saedion Marketing dataPositive24:58

    Marketing data is relatively plentiful for Minerva’s use case.

    “And so for us, it's actually quite easy because it's quite plentiful.”

    Listen at 24:58

  19. Daniel Saedion Construction-domain training dataNegative25:45

    Construction-domain training data is difficult to obtain.

    “And so it's really hard to find that data.”

    Listen at 25:45

  20. Daniel Saedion MicroOneNeutral26:20

    MicroOne pays $250,000–$500,000 for a complete company dataset.

    “they're spending anywhere from $250,000 to $500,000 per complete data set from a company.”

    Listen at 26:20

  21. Daniel Saedion Open-weight modelsNegative27:36

    Uncensored open-weight frontier models pose the biggest risk of open models in public.

    “that's really the biggest risk, I think, personally, to having open-weight frontier models out in the wild.”

    Listen at 27:36

  22. Daniel Saedion Obliterated DeepSeek modelPositive28:16

    The obliterated DeepSeek model costs about ten times less than Sonnet and Fable.

    “It's about 10x cheaper than Sonnet, which is about 10x cheaper than Fable.”

    Listen at 28:16

  23. Daniel Saedion Obliterated AI modelsNegative28:24

    Obliterated models can be instructed to perform unrestricted tasks.

    “And you can tell it to do anything.”

    Listen at 28:24

  24. Daniel Saedion Obliterated open-weight modelsNegative28:51

    Removing safeguards could enable many capable models to conduct offensive hacking.

    “I can just turn off a few parameters in this model. And now suddenly I have this PhD level. You know, Elliot Alderson, if you watched Mr. Robot, I have a thousand of these Mr. Robots now running in the cloud and I can tell them to hack anything.”

    Listen at 28:51

  25. Daniel Saedion open modelsNegative29:43

    Policy should prevent terrorists from downloading powerful open models.

    “you have to create a world where people can't – you can't have the terrorists just downloading this.”

    Listen at 29:43

  26. Daniel Saedion Chinese AI labsNeutral32:46

    Distillation is not the sole explanation for Chinese labs’ model progress.

    “the distillation is not the pure reason that the Chinese labs are where they are.”

    Listen at 32:46

  27. Daniel Saedion Chinese AI model distillationNeutral33:20

    Distillation contributes to the performance of Chinese AI models.

    “Distillation is is real. And there's lots of evidence for it.”

    Listen at 33:20

  28. Daniel Saedion Chinese AI model distillationMixed34:13

    Distillation helps Chinese models but does not explain all their quality.

    “Is distillation responsible for all of the reason why they're good models? Probably not, but it's definitely helping them out for sure.”

    Listen at 34:13

  29. Daniel Saedion AI model benchmarksNegative38:17

    AI benchmarks are becoming less meaningful for measuring model progress.

    “the benchmarks are starting to become meaningless”

    Listen at 38:17

  30. Daniel Saedion Kimi 3.2 in coding workflowsPositive38:31

    Replacing a frontier model in daily coding workflows would demonstrate Chinese models have caught up.

    “if I can take Kimi 3.2 and start replacing it as my daily driver and how I'm coding, okay, now I know that things have caught up.”

    Listen at 38:31

  31. Daniel Saedion AI application operating costsPositive40:18

    Businesses with fixed revenue should minimize variable operating costs.

    “when you have a fixed revenue coming in and a variable cost, you want to reduce that variable cost as much as you can.”

    Listen at 40:18

  32. Daniel Saedion open-source modelsPositive41:02

    Simple repetitive tasks can use very cheap open-source models.

    “for some use cases, if it's a highly repetitive task that's quite simple, we might just use a very cheap open source model.”

    Listen at 41:02

  33. Daniel Saedion AI compute capacity per knowledge workerPositive45:11

    Within four years, compute capacity may reach roughly one B300 per knowledge worker.

    “It won't be a surprise that we have kind of one B300 worth of. computational capacity for every knowledge working employee in the economy.”

    Listen at 45:11

  34. Daniel Saedion AI compute capacityNegative45:39

    The economy will remain substantially short of compute for widespread AI adoption.

    “in the long term, we're still massively short the amount of compute we need to actually be an AI forward society and labor force.”

    Listen at 45:39

  35. Joe Weisenthalon AI modelsNegative46:55

    Lagging-edge models may be uneconomical when engineer time is scarce.

    “it may not be cheap to use a lagging edge model because, you know, like you just want to.”

    Listen at 46:55

  36. Tracy Allowayon Cheaper AI modelsPositive47:33

    Cheaper AI models are more likely to win over time.

    “I would place my bets on the cheaper models winning out, but I don't know.”

    Listen at 47:33

Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.

Weekly email

State of the Pod, every Monday

The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.

Listen to the full episode and explore every guest, topic, and moment on PodLume.

Open models reshape AI startup economics · PodLume