Thanks for all of your recent posts! Your clear-headed approach is very refreshing to read.
I am curious about your conclusion. In terms of API-based billing, your analysis suggests that per-token costs may not be so heavily subsidized. But for subscription-based billing, there certainly seems to be a mismatch (and maybe the mechanism of that mismatch is debatable - is it due to a subsidy, or something else?). The SemiAnalysis post you mentioned is evidence of this, and I have my own anecdotal evidence (e.g. I spent >$1000 of tokens with a $100 subscription last month according to the ccusage tool).
If this mismatch is not due to a subsidy, what else do you think might cause it? If it is a subsidy, do you think the providers will not want to increase prices in the future? I want to think about this in a rational way, but it's hard for me to imagine for-profit companies avoiding the temptation to do this!
In fairness, there are certainly other examples of profitable companies subsidizing a product to make money elsewhere. And, I might be ignorant of the overall economics - maybe individual subscriptions are a small piece of the pie, so they don't mind subsidizing them, compared to enterprise API billing.
Still, as the use of LLMs starts to feel more and more necessary, it is hard to escape the feeling that I'm getting locked into a subscription that someone else has a lot of power over!
I looked at hosting Kimi K3.1 Pro as well for a Danish only sovereign inference and worked out it would be about 10M dkk ($1.5M) and you would need a commitment of around 200.000dkk (~$30k) to make it commercially viable e.g. token speed reasonable for multiple users.
Its plainly not the Uber model that the "tokens are subsidised" position implies, different economics entirely - supply constrained (not enough GPUs), not demand (taxi rides).
Before the Anthropic IPO I tried to get top line numbers by taking the published API cost, reported revenue, and published data-centers commitments. I assumed a roughly 50/50 split between inference and training as that is the most efficient way to use GPUs, and came up with a figure looking like a margin of around 80% on tokens. This also gets better over time, as serving gets cheaper for the same model, and API prices stay the same.
But recent reporting is saying its more like 70% training, 30% on inference
Now we have the IPO documents, we can take the reported figures:
Scenario 1: all compute is inference (100%) (worst case)
6M 2026: ($16,286 − $8,945) / $16,286 ≈ 45%
Scenario 2: 50% of compute spend is inference (my guess)
So bottom line they have likely spent $4.5B to make $16B (amazing), and have not spent the $100B in the bank from the money they raised (eg. it wasn't used to subsidise tokens) but then have commitments to spend $800B in compute as intelligence increases as log of GPU flops, so they need 100x more to be 10x better. Crazy economics.
I think that many of the subsidies people talk about are in the form of the subscriptions you mention, but there are also a lot hidden behind the creative accounting that nvidia and the big hyperscalers are involved in. Beyond that, there's enormous public infrastructure support being thrown at them.
The most worrying thing to me is the risk exposure being offloaded to retirement accounts via index funds; without recent rule changes, this wouldn't be allowed to happen for such newly listed companies. The folks running the frontier labs are in a desperate rush to trade equity for cash; if they don't think it's sustainable, I'm not any more optimistic. No one is getting any serious money back in 3 years if the bottom falls out, and current interest rates make treasurey bonds pretty attractive compared to what these companies are issuing.
The whole idea behind a system of tokens is to hide costs. Providing use costs and markups transplantly, even without factoring in training costs, is not in anyone's best interest. It's why Chuck E. Cheese use tokens and carnivals use tickets, because that cheap stuffed bear is clearly quite expensive in any actual currency.
More visible pricing, with markup, is the only way I see out of the current environment that doesn't end with a collapse. Because model use doesn't benefit from economies of scale, there isn't a road to sustainability (to say nothing of profitability) without government interventions of unprecedented scale.
Thanks for all of your recent posts! Your clear-headed approach is very refreshing to read.
I am curious about your conclusion. In terms of API-based billing, your analysis suggests that per-token costs may not be so heavily subsidized. But for subscription-based billing, there certainly seems to be a mismatch (and maybe the mechanism of that mismatch is debatable - is it due to a subsidy, or something else?). The SemiAnalysis post you mentioned is evidence of this, and I have my own anecdotal evidence (e.g. I spent >$1000 of tokens with a $100 subscription last month according to the ccusage tool).
If this mismatch is not due to a subsidy, what else do you think might cause it? If it is a subsidy, do you think the providers will not want to increase prices in the future? I want to think about this in a rational way, but it's hard for me to imagine for-profit companies avoiding the temptation to do this!
In fairness, there are certainly other examples of profitable companies subsidizing a product to make money elsewhere. And, I might be ignorant of the overall economics - maybe individual subscriptions are a small piece of the pie, so they don't mind subsidizing them, compared to enterprise API billing.
Still, as the use of LLMs starts to feel more and more necessary, it is hard to escape the feeling that I'm getting locked into a subscription that someone else has a lot of power over!
I looked at hosting Kimi K3.1 Pro as well for a Danish only sovereign inference and worked out it would be about 10M dkk ($1.5M) and you would need a commitment of around 200.000dkk (~$30k) to make it commercially viable e.g. token speed reasonable for multiple users.
Its plainly not the Uber model that the "tokens are subsidised" position implies, different economics entirely - supply constrained (not enough GPUs), not demand (taxi rides).
Before the Anthropic IPO I tried to get top line numbers by taking the published API cost, reported revenue, and published data-centers commitments. I assumed a roughly 50/50 split between inference and training as that is the most efficient way to use GPUs, and came up with a figure looking like a margin of around 80% on tokens. This also gets better over time, as serving gets cheaper for the same model, and API prices stay the same.
But recent reporting is saying its more like 70% training, 30% on inference
Now we have the IPO documents, we can take the reported figures:
Scenario 1: all compute is inference (100%) (worst case)
6M 2026: ($16,286 − $8,945) / $16,286 ≈ 45%
Scenario 2: 50% of compute spend is inference (my guess)
6M 2026: inference cost ≈ $4,473M, so margin ≈ ($16,286 − $4,473) / $16,286 ≈ 73%
Scenario 3: 30% of compute is inference (70% training if as reported)
6M 2026: inference cost ≈ $2,684M, so margin ≈ ($16,286 − $2,684) / $16,286 ≈ 84%
So bottom line they have likely spent $4.5B to make $16B (amazing), and have not spent the $100B in the bank from the money they raised (eg. it wasn't used to subsidise tokens) but then have commitments to spend $800B in compute as intelligence increases as log of GPU flops, so they need 100x more to be 10x better. Crazy economics.
I think that many of the subsidies people talk about are in the form of the subscriptions you mention, but there are also a lot hidden behind the creative accounting that nvidia and the big hyperscalers are involved in. Beyond that, there's enormous public infrastructure support being thrown at them.
The most worrying thing to me is the risk exposure being offloaded to retirement accounts via index funds; without recent rule changes, this wouldn't be allowed to happen for such newly listed companies. The folks running the frontier labs are in a desperate rush to trade equity for cash; if they don't think it's sustainable, I'm not any more optimistic. No one is getting any serious money back in 3 years if the bottom falls out, and current interest rates make treasurey bonds pretty attractive compared to what these companies are issuing.
The whole idea behind a system of tokens is to hide costs. Providing use costs and markups transplantly, even without factoring in training costs, is not in anyone's best interest. It's why Chuck E. Cheese use tokens and carnivals use tickets, because that cheap stuffed bear is clearly quite expensive in any actual currency.
More visible pricing, with markup, is the only way I see out of the current environment that doesn't end with a collapse. Because model use doesn't benefit from economies of scale, there isn't a road to sustainability (to say nothing of profitability) without government interventions of unprecedented scale.