The American Startup Taking On China’s AI Models

Anthropic

Across the shows: 77 speakers in other episodes, 93 bullish and 56 bearish calls, past 30 days

• Misha Laskin · Reflection AIBullish

  Anthropic will do just fine despite cheaper open-weight models, in a positive-sum rental and ownership market.

  Laskin says Anthropic and OpenAI are incredible companies that have scaled at an unprecedented rate and keep discovering breakthroughs. He expects a positive-sum world with open-model companies, builders and distributors alongside successful closed-model companies. Rental and ownership markets can coexist, as owning is growing without everyone owning a house. He thinks the closed labs will do just fine.

  “Yeah. I think they're gonna do just fine. These are incredible companies that have scaled at an unprecedented rate and keep discovering new breakthroughs in AI.”at 52:50

OpenAI

Across the shows: 27 speakers in other episodes, 12 bullish and 21 bearish calls, past 30 days

• Misha Laskin · Reflection AIBullish

  OpenAI will keep open-sourcing only lesser models, and will still do just fine.

  Laskin says OpenAI has open-sourced models before, and that when open source is not a company's core strategy, it tends to release a lesser version of what it actually monetizes. He cites OpenAI's GPT-OSS as similar to Google's Gemma, which is a different class from its internal models. He does not expect that strategy to change, since OpenAI has a lot of opportunity chipping away at its current business. He thinks OpenAI will do fine.

  “I think that when something is not the core strategy of the company, you tend to not pay as much attention to it, and you tend to open source maybe like a lesser version of the thing that you're actually kind of monetizing.”at 53:40

Open-source AI models

• Misha Laskin · Reflection AI

  Open models are free, but the compute, serving, agents and fine-tuning around them are paid.

  Open models can be free under permissive licenses. Beam, his company's model, is Apache 2.0. Enterprises still have to buy compute, software to serve models, and agentic or customisation workflows, often through a partner. He describes the model as an engine that is useless without a car. Revenue comes from the systems around the model, including inference serving, packaged agents and fine-tuning services for enterprises.

  “But just because the model is free, doesn't mean that everything around it is free. If you are a customer or an enterprise, and you want to run open models, well, you need to get the compute from somewhere.”at 14:03

• Misha Laskin · Reflection AI

  Open-weight models will likely win majority consumption, as open Linux and Android did after closed operating systems.

  Foundational technology often starts closed and is then followed by an open version that takes most consumption. Microsoft's closed DOS and Windows built a large business, but Linux came after and now runs over 90% of the world's servers (he thinks over 95%). Android followed Apple's closed mobile OS and now runs most of the world's phones. The internet also began closed inside government and universities before becoming open. He argues AI is following the same path, so open models should eventually take most of the market.

  “Operating systems started closed, and Linux came after... And now over 90% of the world's servers, I think the number is even higher, over 95%, run on some variant of an open source operating system, most often Linux.”at 25:49

• Misha Laskin · Reflection AI

  Chinese open-weight models are a gift to the ecosystem, not a harmful AI dumping campaign against the West.

  Laskin rejects the premise that Chinese open models are AI dumping that harms Western markets. He argues that great open models are a gift regardless of origin. When closed labs moved into coding and tall-poppied promising US startups, those startups moved onto Chinese open models, customized them and stood on their own. Taking away open source over geopolitical worries would hurt the American innovation ecosystem. He says a world with no good open models would be darker. He also says the 2025 DeepSeek market scare was irrational.

  “The fact that there are great open models out there, period, doesn't matter where they come from, is a gift to us all.”at 42:05

• Misha Laskin · Reflection AI

  Wants a Western open-model ecosystem rivalling China's, since China's strength comes from many competing companies.

  Laskin says China's open-model strength comes from a whole ecosystem of companies, not one. Leapfrogging among them keeps great models coming out of China. He hopes Reflection becomes a standalone leader in the open-model space, and that a comparable set of open-model builders emerges in the US this year or next. He wants a West-based ecosystem of such companies.

  “We hope to see an ecosystem of open model builders come out of The States this year, next year, that rivals... That is comparable, right, in quality to the Chinese open model ecosystem.”at 55:11

US–China AI & chip race

• Misha Laskin · Reflection AI

  If one country dominates open models, the rest of the world depends on its infrastructure; America must lead open AI.

  Open technology becomes the foundation that everyone adopts, and it brings along the infrastructure and software of the nation that originated it. Western open models began significantly lagging China after DeepSeek V3 and the wave of strong Chinese open models that followed. If only one country produces open models, the rest of the world becomes dependent on that country's models and infrastructure. He argues America needs to be competitive and have the best open source models, and that its allies and domestic users should build on American open intelligence.

  “If it's only one country that's producing open models, then the rest of the world kind of become dependent on those models, and on the infrastructure of that country. And it's extremely important for America, as a country, to be competitive.”at 10:19

• Misha Laskin · Reflection AI

  Chinese labs likely keep their domestic market captive and give models away abroad to undercut Western margins.

  From the outside, he sees Chinese labs such as DeepSeek, Kimi, GLM and Zhipu serving models mainly through their own APIs, with few third-party distributors of their open models inside China. Their business in China therefore looks similar to Western closed API businesses. Western enterprises are also reluctant to consume Chinese APIs, so the models can't be monetized abroad. Giving the models away abroad can still pay off, because it has geopolitical advantages and pressures the margins of competitors abroad. He hedges that he is judging from the outside and suspects this pattern.

  “I do think it has geopolitical advantages, and as a sort of, one could think about it as a sort of economic incentive to drive down the margins of your competitors abroad.”at 36:02

• Misha Laskin · Reflection AI

  Chinese open models could lock allied nations into Chinese chips and infrastructure, a US national-security risk.

  Laskin argues open models are Trojan horses for the infrastructure they carry. The engine needs a car, and the car is chips. As Chinese companies sell open models into the Middle East, Asia and Brazil, they will also start selling chips for those countries' data centers. Allied governments would become locked into Chinese infrastructure. He concludes the US must have the best open models and chips so the world builds on American technology, and that this is a matter of national security.

  “Chinese companies go and sell open models into the Middle East... they will now also start selling chips in their data centers. And what that means is that other governments, allied nations, will start getting locked into Chinese infrastructure.”at 47:08

AI pricing & unit economics

• Misha Laskin · Reflection AI

  Falling model margins are healthy; intelligence should be a commodity like electricity, yet inference margins stay high.

  He compares intelligence to electricity and water, which are utilities with thin margins that still support large services. Margin compression at the model layer is therefore healthy. Inference margins remain high, even on open models, because compute is scarce. The amount of intelligence per watt keeps rising faster than the ability to build compute, which is a physical buildout. Cheaper intelligence should drive more use, in line with Jevons paradox. He argues intelligence should not be treated as a premium product like an iPhone.

  “I think what's happening in terms of margin compression on the model side is a healthy thing, because intelligence should be viewed as foundational fundamentals, electricity, water, it's commodity.”at 27:46

• Misha Laskin · Reflection AI

  Falling AI costs have consistently raised demand, a Jevons-paradox dynamic that expands the market.

  Laskin says AI is not a fixed pie. When the cost of something falls, demand rises, a Jevons paradox. So far, every time AI costs have been driven down, demand for these intelligence systems has increased. He notes enterprises are cost-conscious right now because AI has been expensive to date. He argues enterprises would spend billions if it yielded a multiple of that in value, but the payoff takes time to show up.

  “in this instance, the demand goes up. It's a sort of Jevons paradox sort of thing. So, when you take AI, so far, every time you've kind of driven down the costs, the demand has gone up for them, for these intelligence systems.”at 43:54

Enterprise AI adoption & AI software

• Misha Laskin · Reflection AI

  The largest consumers of open models will be enterprises, big banks, public sector and sovereign entities.

  Open models carry trust, because the companies building them are seen as strategic to the nation and the West. Enterprises, public-sector bodies and sovereign entities therefore find them more trustworthy. He expects the largest consumption of open models to come from these buyers, including large banks. The business model depends on serving these customers with inference, agents and fine-tuning.

  “where we see the markets for consumption of open models, the biggest markets for consuming them are going to be in enterprise, like the big banks, public sector, sovereign entities.”at 37:36

• Misha Laskin · Reflection AI

  Fortune 500 firms avoid Chinese models over risk and regulation, leaving US enterprises and agencies at a disadvantage.

  Laskin says large US enterprises do not want to use Chinese models because of risk, perception and regulatory concerns. Chinese open models rarely go into production at Fortune 500 companies, though some firms test them. Startups that accept the risk therefore have an edge. Enterprises that want to own their intelligence, rather than rent it, need a good US solution. Public sector agencies and the largest enterprises are at a competitive disadvantage to Chinese companies and to US peers willing to take on that risk.

  “our Fortune 500 enterprises do not wanna use Chinese models for various reasons around risk, perception, regulatory.”at 44:53

AI regulation & safety

• Misha Laskin · Reflection AI

  Open weights bring far more researchers to model safety than the small safety teams at closed labs.

  He estimates that closed labs employ about 5,000 researchers who understand these systems, with perhaps 500 working on safety, a very small community for such powerful technology. Open models let a much wider group of researchers, engineers and developers inspect, align and contribute. He draws on open-source security, where many eyes make bugs shallow and the most hardened software tends to be open. He sees this as how science works.

  “With an open model ecosystem, you're inviting a much larger group of researchers, engineers, developers, anyone who takes an interest can take a look and contribute to these things.”at 23:15

AI infrastructure & data centers

• Misha Laskin · Reflection AI

  The DeepSeek selloff was irrational; compute-linked stocks rebounded as open models raised compute demand.

  When DeepSeek came out, some stocks fell on a market scare. Laskin calls that reaction irrational. Those stocks rebounded and did very well because demand for compute went up. A good open model means more buyers for compute, since anyone can get the model and run it on their own hardware. That creates more diverse demand for compute from startups and scientific users.

  “when DeepSeek came out, there was a market scare, and some stocks went down. But the reality is that even... It was an irrational reaction. Those stocks rebounded and started doing very well because of demand for various things like compute went up.”at 50:08