3 ms·

Several top level comments are attacks on the author’s credibility that do not engage the substance of the piece at all. I think the fundamental problem for pe

Several top level comments are attacks on the author’s credibility that do not engage the substance of the piece at all. I think the fundamental problem for people on all sides of the various debates surrounding AI is that “AI is a powerful, transformative technology” and “Anthropic and OpenAI are both doomed, and this poses risks to the broader economy” are compatible statements of fact. Just because you believe (1) does not justify dismissing (2) out of hand. (2) is the point of this article. OAI and Anthropic are spending by far the most money of anyone in the space, as the article rightly notes, but they have no path to becoming profitable, meaning they cannot occupy that position forever. The other entities that rely on their spending to support their own margins - in this case the major cloud providers - are vulnerable to revenue collapse if OAI and Anthropic fail. The premise most would disagree with is that the labs have no path to profitability. Two points support this: demand for inference is functionally infinite, or at least is so great that it is not meaningful to discuss its limits; and the labs are profitable on inference and are only taking losses to compete with one another. Some would extend this further and say that once the tech is good enough it will be able to drastically reduce their costs by some combination of speeding up research and creating efficiencies to reduce compute spend. These are valid criticisms. But “AI has gotten better since he started saying ‘AI bad’” is not a reason to ignore the fact that major cloud computing providers are taking on massive new debt while becoming increasingly dependent on only two customers who face meaningful margin pressures. Unless OAI and Anthropic can find a durable moat and a means to exert pricing power, this is a serious issue going forward. That is true whether we wind up with a machine god (although we might have bigger problems in that case) or if we plateau at current capabilities.

rich_sasha 2mo ago
I largely agree with this take. What gives me disquiet is that I thought more or less exactly this about Uber, and was proven comprehensively wrong. Despite burning money like a furnace on an app for taxis, it seems that anyone who invested privately made a handsome return, and they are at a profitable steady state. Is AI the same?
- arctic-true 2mo ago
  The cash burn is not the problem. You can’t build a big business without it. The difference is network effects. With ride sharing apps, you need a lot of drivers and riders collected on one platform. It is nearly always better, for both drivers and riders, to switch to a larger platform. Thus, it was worthwhile to spend the money to become the biggest fish. Once this was accomplished, Uber could raise prices because switching to a smaller rival would mean less availability, and thus less utility (for riders) or less earnings (for drivers). With AI, by contrast - at least in its current state - there is no benefit to be gained from using the same model provider as somebody else. Switching is trivial for most use cases. Since they can’t capture consumers using network effects, the labs only have the levers of price and quality to pull to acquire and retain customers. To pull the price lever, they have to reduce their revenues; to pull the quality lever, they have to increase their expenditures. Indeed, they are sowing the seeds of their own demise by making inference cheaper and more efficient: since they can’t exercise pricing pressure, efficiency gains will be passed on to the consumer, which is unsustainable if your GPU debt is priced based on yesterday’s efficiency expectations.
  - rich_sasha 2mo ago
    This only works up to a point. Plenty of different makers of phones for example all making fungible phones at ever lower prices, yet managing to eke out a profit. I can see a world where OpenAI, Anthropic, Chinese companies corner the market, make themselves indispensable and start charging market rates, while also getting better and more cost efficient.
    - arctic-true 2mo ago
      Here again you have network effects, though. I personally have an iPhone because it’s what all my friends and family use and there are communication functions that are much easier to engage with if you all have the same sort of device. It’s also not trivial to switch, you need to, at minimum, go to the store or wait for something to get delivered - to say nothing of the wasted money from buying multiple phones. If I have unused OpenAI tokens I can burn them on side quests or something. If you have an unused iPhone you need to find a way to sell/return it or eat the cost. There’s also potential violations of your contract with your mobile carrier, etc. If I’m using ChatGPT and I decide I want to use DeepSeek instead, I am only a couple of keystrokes away from doing it, and that’s if I have never used DeepSeek before. As for your “corner the market” scenario, it’s possible, but unlikely. It is too easy to enter; even if you somehow got all of the major players to commit to growing their margins - and somehow manage not to violate the antitrust laws in the process - a newcomer could spoil the party far easier than it could in an industry like mobile phones (where you need tons of components, manufacturing capacity, network relationships, etc.) or ride sharing (where you need a large user base to justify your existence).
lowbloodsugar 2mo ago
TFA admits that Amazon's AI revenue is only 59% from Anthropic and OpenAI, with 41% coming from actual customers using bedrock. That seems like they are in a winning position. I've run Qwen on Bedrock. Big models on Bedrock do outperform the models I can fit on my single RTX PRO 6000.

rich_sasha 2mo ago

I largely agree with this take. What gives me disquiet is that I thought more or less exactly this about Uber, and was proven comprehensively wrong. Despite burning money like a furnace on an app for taxis, it seems that anyone who invested privately made a handsome return, and they are at a profitable steady state. Is AI the same?

arctic-true 2mo ago
The cash burn is not the problem. You can’t build a big business without it. The difference is network effects. With ride sharing apps, you need a lot of drivers and riders collected on one platform. It is nearly always better, for both drivers and riders, to switch to a larger platform. Thus, it was worthwhile to spend the money to become the biggest fish. Once this was accomplished, Uber could raise prices because switching to a smaller rival would mean less availability, and thus less utility (for riders) or less earnings (for drivers). With AI, by contrast - at least in its current state - there is no benefit to be gained from using the same model provider as somebody else. Switching is trivial for most use cases. Since they can’t capture consumers using network effects, the labs only have the levers of price and quality to pull to acquire and retain customers. To pull the price lever, they have to reduce their revenues; to pull the quality lever, they have to increase their expenditures. Indeed, they are sowing the seeds of their own demise by making inference cheaper and more efficient: since they can’t exercise pricing pressure, efficiency gains will be passed on to the consumer, which is unsustainable if your GPU debt is priced based on yesterday’s efficiency expectations.
- rich_sasha 2mo ago
  This only works up to a point. Plenty of different makers of phones for example all making fungible phones at ever lower prices, yet managing to eke out a profit. I can see a world where OpenAI, Anthropic, Chinese companies corner the market, make themselves indispensable and start charging market rates, while also getting better and more cost efficient.
  - arctic-true 2mo ago
    Here again you have network effects, though. I personally have an iPhone because it’s what all my friends and family use and there are communication functions that are much easier to engage with if you all have the same sort of device. It’s also not trivial to switch, you need to, at minimum, go to the store or wait for something to get delivered - to say nothing of the wasted money from buying multiple phones. If I have unused OpenAI tokens I can burn them on side quests or something. If you have an unused iPhone you need to find a way to sell/return it or eat the cost. There’s also potential violations of your contract with your mobile carrier, etc. If I’m using ChatGPT and I decide I want to use DeepSeek instead, I am only a couple of keystrokes away from doing it, and that’s if I have never used DeepSeek before. As for your “corner the market” scenario, it’s possible, but unlikely. It is too easy to enter; even if you somehow got all of the major players to commit to growing their margins - and somehow manage not to violate the antitrust laws in the process - a newcomer could spoil the party far easier than it could in an industry like mobile phones (where you need tons of components, manufacturing capacity, network relationships, etc.) or ride sharing (where you need a large user base to justify your existence).

lowbloodsugar 2mo ago

TFA admits that Amazon's AI revenue is only 59% from Anthropic and OpenAI, with 41% coming from actual customers using bedrock. That seems like they are in a winning position. I've run Qwen on Bedrock. Big models on Bedrock do outperform the models I can fit on my single RTX PRO 6000.