How an American AI Company Uses Chinese Open-Weight Models
How an American AI Company Uses Chinese Open-Weight Models
43 minutes ago•Odd Lots•Bloomberg
Podcast48 min 23 sec
Listen to Episode
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights
  • Consider NVIDIA (NVDA) as a long-term way to gain exposure to growing AI-computing demand, but treat demand and returns as uncertain; the alleged Reflection AI acquisition is unconfirmed and is not a reliable near-term catalyst.
  • Favor AI application businesses that can use lower-cost models to automate specific workflows, while assessing customer value, cost per completed task, data quality, and ongoing model-maintenance costs.
  • Coinbase (COIN) and the discussion of digital assets offer no actionable investment signal; no price targets or specific public-stock recommendations were provided.
Detailed Analysis

NVIDIA (NVDA)

  • The guest described NVIDIA as backing Reflection AI, a U.S. open-weight model developer. The transcript says Reflection AI was allegedly being acquired by NVIDIA; this was presented as unconfirmed, not as a completed transaction.
  • Reflection AI’s model was described as several months behind the leading frontier models, despite access to a large cluster of NVIDIA GPUs. The guest used this as an example of the resources needed to develop competitive models.
  • The discussion also pointed to potentially large long-term demand for compute: the guest estimated that the economy could eventually need far more AI capacity if knowledge workers use AI extensively.

Takeaways

  • The discussion supports a potential long-term demand theme for AI computing infrastructure, including GPUs, but it does not establish that demand will translate into a particular company’s returns.
  • The guest also emphasized uncertainty over near-term compute availability and costs. He said his own company could access the compute it needed, while noting that industry narratives about shortages and overbuilding shift quickly.
  • Treat the Reflection AI acquisition reference as an allegation, not a confirmed catalyst. No price target or investment recommendation was given.

AI Models and Application Software

  • The central theme was whether lower-cost open-weight models can compete with closed frontier models in business applications.
  • Minerva, the guest’s marketing AI company, uses frontier models such as Claude and other leading models during product development, but considers lower-cost models for repeatable tasks where adequate results are sufficient.
  • The guest said fine-tuning open-weight models can help application companies specialize a general model for specific workflows. He gave examples of Harvey and Legora in legal technology and Minerva in marketing.
  • He argued that companies with good proprietary or customer-provided data may be able to deliver useful results at lower cost. He also described model costs as only part of the calculation: companies must account for setup, hosting, ongoing operations, and the work required to keep fine-tuned models current.
  • The guest said fine-tuning does not automatically carry over when a model provider releases a new version, creating uncertainty and ongoing costs.
  • The episode discussed DeepSeek and other Chinese open-weight models as relatively inexpensive options, while noting debate over how much their performance reflects distillation from Western models.
  • The hosts said they would “place [their] bets on the cheaper models winning out,” while acknowledging uncertainty and noting that closed-model providers also offer cheaper versions.

Takeaways

  • The opportunity discussed is in AI applications that can use lower-cost models to automate well-defined tasks while retaining enough quality to satisfy customers.
  • For evaluating this theme, the discussion suggests focusing on cost per completed task, customer value, data quality, and the expense of maintaining fine-tuned models—not just token prices or benchmark scores.
  • The episode did not identify a specific public stock tied to this application-layer opportunity. Minerva, Harvey, Legora, and Reflection AI were discussed as companies, not as publicly traded investments.

AI Inference, Hosting, and Data Companies

  • The guest named Fireworks, Modal, and Thinking Machines as companies that can help businesses host or serve models without building all the required infrastructure themselves.
  • He said the availability of these services has made fine-tuning and deploying models easier for smaller companies. He also claimed that Fireworks and Modal were each at roughly $1 billion in annual recurring revenue (ARR); this was his statement during the interview.
  • The discussion emphasized that model performance depends heavily on high-quality training data and evaluation methods. The guest said companies such as Mercor, Surge, and Micro1 are involved in acquiring or supplying data.
  • He said some company data sets were being purchased for $250,000–$500,000 and cited substantial demand for useful business data.
  • He described data quality and carefully graded examples as a potential source of advantage, while warning that insufficiently useful data could produce noise rather than better models.

Takeaways

  • The discussion points to potential opportunities in AI infrastructure services and data providers, as businesses seek help with model deployment, fine-tuning, and training data.
  • It also suggests that infrastructure becoming easier to access does not eliminate competitive challenges: the quality of a company’s data and its ability to use that data effectively may matter more than simply having access to a model.
  • The named providers were discussed as operating businesses, but the transcript does not establish that they are publicly traded or offer a specific investment recommendation.

Coinbase (COIN)

  • Coinbase was mentioned briefly when the hosts referred to CEO Brian Armstrong discussing the company’s use of AI models.
  • The episode did not discuss Coinbase’s crypto business, financial results, valuation, or a specific AI strategy in enough detail to form an investment thesis.

Takeaways

  • The mention offers no specific bullish or bearish signal for Coinbase shares.

Digital Assets: Tokenization, Stablecoins, and Virtual Assets

  • The opening promotional segment raised the question of whether regulation can keep pace with—and help accelerate—adoption of virtual assets, tokenization, and stablecoins.
  • A separate promotional segment discussed institutional participation in digital assets and the role of regulation, but these topics were not part of the main interview.

Takeaways

  • The transcript offers only a broad regulatory and adoption theme, not analysis of a particular cryptocurrency, token, company, or investment opportunity.

Ask about this postAnswers are grounded in this post's content.
Episode Description
This week, we got news that the Trump administration is reportedly considering a requirement for US companies to publicly disclose their use of Chinese AI models. Any effort to discourage US businesses from using Chinese open-source could be a big deal for both the AI industry and the market, where investors are still debating whether customers will ultimately be willing to pay for pricier homegrown frontier models versus cheaper alternatives. Everyone remembers last year's 'DeepSeek moment' for AI, when a Chinese company released a cheaper model that seemed nearly as good as cutting-edge US ones. So why do businesses actually use open-source? And what are the considerations that take into account in doing so? In this episode, we talk to Daniel Saedi, co-founder of Minerva, to better understand how an AI native company chooses its models. He breaks down how exactly his company mixes and matches models for specific use cases, the risk of abliterated models, and the issue of data sovereignty. Read more: India Startup Sells DeepSeek Access While Pursuing Sovereign AI Anthropic Raises $107.3 Billion to Top SpaceX’s Record Debut Only http://Bloomberg.com subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at  bloomberg.com/subscriptions/oddlots Subscribe to the Odd Lots Newsletter Join the conversation: discord.gg/oddlots See omnystudio.com/listener for privacy information.
About Odd Lots
Odd Lots

Odd Lots

By Bloomberg

<p>Bloomberg's Joe Weisenthal and Tracy Alloway explore the most interesting topics in finance, markets and economics. Join the conversation every Monday and Thursday.</p>