3 ms·

Yes, many people believe that, but it doesn't seem to be an evidence-based belief. I've written about this in some detail[0][1] before. But since just linking t

by jsnell 1y ago

Yes, many people believe that, but it doesn't seem to be an evidence-based belief. I've written about this in some detail[0][1] before. But since just linking to one's own writing is a bit gauche and doesn't make for a good discussion, I'll summarize :) 1. There is no point in providing paid APIs at negative margins, since there's no platform power in having a larger paid API share (paid access can't be used for training data, no lock-in effects, no network effects, no customer loyalty, no pricing power on the supply side since Nvidia doesn't give preferential treatment to large customers). Even selling access at break-even makes no sense, since that is just compute you're not using for training, or not selling to other companies desperate for compute. 2. There are 3rd-party providers selling only the compute, not models, who have even less reason to sell at a loss. Their prices are comparable to 1st-party providers. 3. Deepseek published their inference cost structure for R1. According to that data their paid API traffic is very lucrative (their GPU rental costs for inference are under 20% of their standard pricing, i.e. >80% operating margins; and the rental costs would cover power, cooling, depreciation of the capital investment). Insofar as frontier labs are unprofitable, I think it's primarily due to them giving out vast amounts of free access. [0] https://www.snellman.net/blog/archive/2025-06-02-llms-are-cheap/ https://www.snellman.net/blog/archive/2025-06-02-llms-are-ch... [1] https://news.ycombinator.com/item?id=44165521 https://news.ycombinator.com/item?id=44165521

what 1y ago
There are more factors to cost than just the raw compute to provide inference. They can’t just fire everyone and continue to operate while paying just the compute cost. They also can’t stop training new models. The actual cost is much more than the compute for inference.
- brookst 1y ago
  I heart you. Classic fixed / variable cost fallacy: if you look at the steel and plastic in a $200k Ferrari, it’s worth about $10k. They have 95% gross margins! Outrageous! (Nevermind the engine R&D cost, the pre-production molds that fail, the testing and marketing and product placement and…)
- ummonk 1y ago
  That's all the more reason to run at a positive margin though - why shovel money into taking a loss on inference when you need to spend money on R&D?
- jsnell 1y ago
  Yes, there are some additional operating costs, but they're really marginal compared to the cost of the compute. Your suggestion was personnel: Anthropic is reportedly on a run-rate of $3B with O(1k) employees, most of whom aren't directly doing ops. Likewise they also have to pay for non-compute infra, but it is a rounding error. Training is a fixed cost, not a variable cost. My initial comment was on the unit economics, so fixed costs don't matter. But including the full training costs doesn't actually change the math that much as far as I can tell for any of the popular models. E.g. the alleged leaked OpenAI financials for 2024 projected $4B spent on inference, $3B on training. And the inference workloads are currently growing insanely fast, meaning the training gets amortized over a larger volume of inference (e.g. Google showed a graph of their inference volume at Google I/O -- 50x growth in a year, now at 480T tokens / month[0]) [0] https://blog.google/technology/ai/io-2025-keynote/ https://blog.google/technology/ai/io-2025-keynote/
lmeyerov 1y ago
I think you miss 2 big aspects: 1. High volume providers get efficiencies that low volume do not. It comes from both more workload giving more optimization opportunities, and staffing to do better engineering to begin with. The result is break even for lower volume firms is profitable for higher volume, and as high volume is magnitudes more scale, this quickly pays for many people. By being the high-volume API, this game can be played. If they choose not to bother, it is likely because strategic views on opportunity cost, not inability. That's not even the interesting analysis, which is what the real stock value is, or whatever corp structure scheme they're doing nowadays: 2. Growth for growths sake. Uber was exactly this kind of growth-at-all-costs play, going more into debt with every customer and fundraise. My understanding is they were able to tame costs and find side businesses (delivery, ...), with the threat becoming more about category shift of self-driving. By having the channel, they could be the one to monetize as that got figured out better. Whether tokens or something else becomes what is charged for at the profit layers (with breakeven tokens as cost of business), or subsidization ends and competitive pricing dominates, being the user interface to chat and the API interface to devs gives them channel. Historically, it is a lot of hubris to believe channel is worthless, and especially in an era of fast cloning.
- jsnell 1y ago
  > High volume providers get efficiencies that low volume do not But paid-per-token APIs at negative margins do not provide scaling efficiencies! It's just the provider giving away a scarce resource (compute) for nothing tangible in exchange. Whatever you're able to do with that extra scale, you would have been able to do even better if you hadn't served this traffic. In contrast, the other things you can use the compute for have a real upside for some part of the genai improvement flywheel: 1. Compute spent on free users gives you training data, allowing the models to be improved faster. 2. Compute spent on training allows the models to be trained, distilled and fine-tuned faster. (Could be e.g. via longer training runs or by being able to run more experiments.) 3. Compute spent on paid inference with positive margins gives you more financial resources to invest. Why would you intentionally spend your scarce compute on unprofitable inference loads rather than the other three options? > 2. Growth for growths sake. That's fair! It could in theory be a "sell $2 for $1" scenario from the frontier labs that are just trying to pump up their revenue numbers to fund-raise from dumb money who don't think to at least check on the unit economics. OpenAI's latest round certainly seemed to be coming from the dumbest money in the world, which would support that. I have two rebuttals: First, it doesn't explain Google, who a) aren't trying to raise money, b) aren't breaking out genai revenue in their financials, so pumping up those revenue numbers would not help at all. (We don't even know how much of that revenue is reported under Cloud vs. Services, though I'd note that the margins have been improving for both of those segments.) Second, I feel that this hypothetical, even if plausible, is trumped by Deepseek publishing their inference cost structure. The margins they claim for the paid traffic are high by any standard, and they're usually one of the cheaper options at their quality level.

by jsnell 1y ago

what 1y ago

There are more factors to cost than just the raw compute to provide inference. They can’t just fire everyone and continue to operate while paying just the compute cost. They also can’t stop training new models. The actual cost is much more than the compute for inference.

brookst 1y ago
I heart you. Classic fixed / variable cost fallacy: if you look at the steel and plastic in a $200k Ferrari, it’s worth about $10k. They have 95% gross margins! Outrageous! (Nevermind the engine R&D cost, the pre-production molds that fail, the testing and marketing and product placement and…)
ummonk 1y ago
That's all the more reason to run at a positive margin though - why shovel money into taking a loss on inference when you need to spend money on R&D?
jsnell 1y ago
Yes, there are some additional operating costs, but they're really marginal compared to the cost of the compute. Your suggestion was personnel: Anthropic is reportedly on a run-rate of $3B with O(1k) employees, most of whom aren't directly doing ops. Likewise they also have to pay for non-compute infra, but it is a rounding error. Training is a fixed cost, not a variable cost. My initial comment was on the unit economics, so fixed costs don't matter. But including the full training costs doesn't actually change the math that much as far as I can tell for any of the popular models. E.g. the alleged leaked OpenAI financials for 2024 projected $4B spent on inference, $3B on training. And the inference workloads are currently growing insanely fast, meaning the training gets amortized over a larger volume of inference (e.g. Google showed a graph of their inference volume at Google I/O -- 50x growth in a year, now at 480T tokens / month[0]) [0] https://blog.google/technology/ai/io-2025-keynote/ https://blog.google/technology/ai/io-2025-keynote/

lmeyerov 1y ago

I think you miss 2 big aspects: 1. High volume providers get efficiencies that low volume do not. It comes from both more workload giving more optimization opportunities, and staffing to do better engineering to begin with. The result is break even for lower volume firms is profitable for higher volume, and as high volume is magnitudes more scale, this quickly pays for many people. By being the high-volume API, this game can be played. If they choose not to bother, it is likely because strategic views on opportunity cost, not inability. That's not even the interesting analysis, which is what the real stock value is, or whatever corp structure scheme they're doing nowadays: 2. Growth for growths sake. Uber was exactly this kind of growth-at-all-costs play, going more into debt with every customer and fundraise. My understanding is they were able to tame costs and find side businesses (delivery, ...), with the threat becoming more about category shift of self-driving. By having the channel, they could be the one to monetize as that got figured out better. Whether tokens or something else becomes what is charged for at the profit layers (with breakeven tokens as cost of business), or subsidization ends and competitive pricing dominates, being the user interface to chat and the API interface to devs gives them channel. Historically, it is a lot of hubris to believe channel is worthless, and especially in an era of fast cloning.

jsnell 1y ago
> High volume providers get efficiencies that low volume do not But paid-per-token APIs at negative margins do not provide scaling efficiencies! It's just the provider giving away a scarce resource (compute) for nothing tangible in exchange. Whatever you're able to do with that extra scale, you would have been able to do even better if you hadn't served this traffic. In contrast, the other things you can use the compute for have a real upside for some part of the genai improvement flywheel: 1. Compute spent on free users gives you training data, allowing the models to be improved faster. 2. Compute spent on training allows the models to be trained, distilled and fine-tuned faster. (Could be e.g. via longer training runs or by being able to run more experiments.) 3. Compute spent on paid inference with positive margins gives you more financial resources to invest. Why would you intentionally spend your scarce compute on unprofitable inference loads rather than the other three options? > 2. Growth for growths sake. That's fair! It could in theory be a "sell $2 for $1" scenario from the frontier labs that are just trying to pump up their revenue numbers to fund-raise from dumb money who don't think to at least check on the unit economics. OpenAI's latest round certainly seemed to be coming from the dumbest money in the world, which would support that. I have two rebuttals: First, it doesn't explain Google, who a) aren't trying to raise money, b) aren't breaking out genai revenue in their financials, so pumping up those revenue numbers would not help at all. (We don't even know how much of that revenue is reported under Cloud vs. Services, though I'd note that the margins have been improving for both of those segments.) Second, I feel that this hypothetical, even if plausible, is trumped by Deepseek publishing their inference cost structure. The margins they claim for the paid traffic are high by any standard, and they're usually one of the cheaper options at their quality level.