Really tried to avoid the underlying discussion and stick to letting Nathan L know that it was a distracting assumption on his part, and since it was irrelevant to the article, probably better left out. He might as well have wandered into tabs VS spaces or brace styles.
Right. Like totally out of left field and not even the consensus opinion these days in the consumer products space. This is the value traditional editors brought to journalism.
"Where closed labs are an oligopoly, open model builders and users will be far more diverse and numerous."
This is exactly why intelligent model routing is the most critical skill right now. We will always need the frontier models for heavy logic, but a product's actual scalability relies entirely on routing daily noise to the open ecosystem.
I don't like the vitriole from some commenters, but it seems actually ignorant of the smartphone product landscape since circa 2019 to just insist that iPhones are categorically better than every Android phone...they aren't even categorically more expensive.
Such a wildly distracting and ignorant analogy to use.
As someone who uses both daily, one is not inherently better than the other. They both have their pros and cons. Seems strange to use the analogy in an otherwise decent article.
Nathan identifies a crucial mechanism: as long as open models keep pushing the boundary of “good enough” upward, they will continue to compress the monetizable space of closed models. The capability gap may remain, but the commercial gap will narrow. At the same time, many enterprise tasks do not require the smartest model. They require models that are low-cost, deployable, controllable, customizable, and easy to integrate into workflows. The advantage of open models is that they can enter real industrial systems faster and become part of internal enterprise processes, software tools, industry applications, and edge deployment.
However, the real moat in the future may not be a single model. It may be the system combination of model + data + toolchain + workflow + agent framework + distribution + enterprise trust. Even if closed models are caught up by open models in some capabilities, they may still preserve pricing power through product ecosystems, enterprise integration, platform access, and high-reliability services.
Ultimately, open models and closed models appear to be a contest between two technical paths. At a deeper level, they represent two ways of organizing national AI capacity. The United States is closer to a model of frontier labs + proprietary APIs + capital expenditure + platform rent. China is closer to a model of open diffusion + low-cost deployment + industrial adoption + supply-chain integration. The former pursues the strongest models and the highest rents. The latter pursues faster diffusion and broader industrial embedding.
This is also what I argued in my previous essay: the U.S.-China AI technology war has become an independent competitive system organized around models, data, computing power, application ecosystems, industrial diffusion, and institutional capacity. The United States may continue to hold advantages in frontier models. But if China can use open models, low-cost inference, industrial scenarios, and engineering diffusion to turn AI faster into system capacity across manufacturing, energy, transportation, e-commerce, content, robotics, and enterprise software, then the center of competition will shift from who has the strongest model to who can turn AI into a broader infrastructure of productivity.
Hi Nathan, your "different exponentials" split is the one the PC era already ran.
A dozen companies assembled PCs from the same off-the-shelf parts and competed the price down to commodity margins. That's your open side. IBM lost more than a billion dollars on PCs in 1998. The two layers a buyer could not swap, Intel's processor and Microsoft's operating system, kept the profit. In 2004 those two earned over $15 billion in net income, while Dell, HP, and IBM together made about $2.5 billion from PCs.
Now the model layer is splitting the same way. The open side is served by many companies, every layer has an alternative, and the price falls to commodity. The profit pools where the bundle cannot be copied.
The open ecosystem can still be the bigger market in total, the way PC sales dwarfed Intel and Microsoft's revenue while those two kept the profit.
"In 5-10 years I expect both OpenAI and Anthropic to be valued in the $2-10T". --- more like 5 to 10 months the way these markets are trading. Both would probably be at 2T if traded publicly. Bigger question is how long with they be able to raise capital from the circular ecosystem of funding ( ie nvdia, micron sk , Samsung invest in companies like anthropic and core weave, who then buy chips ) and how much capital will they need in 3-5 years. ...
The "iphone vs android" comment really took me out of the article.
sounds like you still have an android?
Really tried to avoid the underlying discussion and stick to letting Nathan L know that it was a distracting assumption on his part, and since it was irrelevant to the article, probably better left out. He might as well have wandered into tabs VS spaces or brace styles.
Right. Like totally out of left field and not even the consensus opinion these days in the consumer products space. This is the value traditional editors brought to journalism.
"Where closed labs are an oligopoly, open model builders and users will be far more diverse and numerous."
This is exactly why intelligent model routing is the most critical skill right now. We will always need the frontier models for heavy logic, but a product's actual scalability relies entirely on routing daily noise to the open ecosystem.
I don't like the vitriole from some commenters, but it seems actually ignorant of the smartphone product landscape since circa 2019 to just insist that iPhones are categorically better than every Android phone...they aren't even categorically more expensive.
Such a wildly distracting and ignorant analogy to use.
There’s a very apt analogy between Apple and the closed labs, but I’ll explain it better next time!
apple fanboys sicken me! such a stupid analogy.
doing tech business analysis without acknowledging apple products are better at this point is simply not possible
As someone who uses both daily, one is not inherently better than the other. They both have their pros and cons. Seems strange to use the analogy in an otherwise decent article.
Nathan identifies a crucial mechanism: as long as open models keep pushing the boundary of “good enough” upward, they will continue to compress the monetizable space of closed models. The capability gap may remain, but the commercial gap will narrow. At the same time, many enterprise tasks do not require the smartest model. They require models that are low-cost, deployable, controllable, customizable, and easy to integrate into workflows. The advantage of open models is that they can enter real industrial systems faster and become part of internal enterprise processes, software tools, industry applications, and edge deployment.
However, the real moat in the future may not be a single model. It may be the system combination of model + data + toolchain + workflow + agent framework + distribution + enterprise trust. Even if closed models are caught up by open models in some capabilities, they may still preserve pricing power through product ecosystems, enterprise integration, platform access, and high-reliability services.
Ultimately, open models and closed models appear to be a contest between two technical paths. At a deeper level, they represent two ways of organizing national AI capacity. The United States is closer to a model of frontier labs + proprietary APIs + capital expenditure + platform rent. China is closer to a model of open diffusion + low-cost deployment + industrial adoption + supply-chain integration. The former pursues the strongest models and the highest rents. The latter pursues faster diffusion and broader industrial embedding.
This is also what I argued in my previous essay: the U.S.-China AI technology war has become an independent competitive system organized around models, data, computing power, application ecosystems, industrial diffusion, and institutional capacity. The United States may continue to hold advantages in frontier models. But if China can use open models, low-cost inference, industrial scenarios, and engineering diffusion to turn AI faster into system capacity across manufacturing, energy, transportation, e-commerce, content, robotics, and enterprise software, then the center of competition will shift from who has the strongest model to who can turn AI into a broader infrastructure of productivity.
Hi Nathan, your "different exponentials" split is the one the PC era already ran.
A dozen companies assembled PCs from the same off-the-shelf parts and competed the price down to commodity margins. That's your open side. IBM lost more than a billion dollars on PCs in 1998. The two layers a buyer could not swap, Intel's processor and Microsoft's operating system, kept the profit. In 2004 those two earned over $15 billion in net income, while Dell, HP, and IBM together made about $2.5 billion from PCs.
Now the model layer is splitting the same way. The open side is served by many companies, every layer has an alternative, and the price falls to commodity. The profit pools where the bundle cannot be copied.
The open ecosystem can still be the bigger market in total, the way PC sales dwarfed Intel and Microsoft's revenue while those two kept the profit.
"In 5-10 years I expect both OpenAI and Anthropic to be valued in the $2-10T". --- more like 5 to 10 months the way these markets are trading. Both would probably be at 2T if traded publicly. Bigger question is how long with they be able to raise capital from the circular ecosystem of funding ( ie nvdia, micron sk , Samsung invest in companies like anthropic and core weave, who then buy chips ) and how much capital will they need in 3-5 years. ...