Ant = Anthropic, not the Chinese Ant. Good write up. There seems to be an inconsistency in Anthropic's justification of their spend on model training. According to them, there is a feed back loop of model training to become more efficient (with statements indicating that 90% of the company's code is AI-written), while at the same time the R&D expenses (which covers the training of new models) make an enormous jump. That is a bit weird. If you expect exponential growth, the exponent matters, not the base. Maybe Anthropic thinks that gobbling up all existing hardware via e.g. leasing allows them to iterate faster, thus increasing the exponent. But to me it seems this mainly allows for increasing the base (bigger models). I would expect that improving your models means that they become smaller and more efficient, not that they grow in parameter size. So it seems like Anthropic thinks that they need to brute force intelligence by creating increasingly heavier models (which indeed might suit their business model), whereas from a customer point of view increasingly leaner models are more 'fit' in llm evolution.
The article raises some concerns about the CSP revenue (which could be gamed, since it's part of the circular revenue deals among AI labs and hypoerscalers), but it doesn't delve into the "organic" customer revenue. Fortunately, other sources have done this work and have revealed that 80% of revenue comes from just 1% of their customers, and VC-backed startups make up a large segment of that top tier, while also showing disproportionate growth. Many of these companies are backed by the same VC that back Anthropic.[1] So even the "organic" revenue growth is OPM, not cash flow from profitable businesses.
Ant = Anthropic, not the Chinese Ant. Good write up. There seems to be an inconsistency in Anthropic's justification of their spend on model training. According to them, there is a feed back loop of model training to become more efficient (with statements indicating that 90% of the company's code is AI-written), while at the same time the R&D expenses (which covers the training of new models) make an enormous jump. That is a bit weird. If you expect exponential growth, the exponent matters, not the base. Maybe Anthropic thinks that gobbling up all existing hardware via e.g. leasing allows them to iterate faster, thus increasing the exponent. But to me it seems this mainly allows for increasing the base (bigger models). I would expect that improving your models means that they become smaller and more efficient, not that they grow in parameter size. So it seems like Anthropic thinks that they need to brute force intelligence by creating increasingly heavier models (which indeed might suit their business model), whereas from a customer point of view increasingly leaner models are more 'fit' in llm evolution.
The article raises some concerns about the CSP revenue (which could be gamed, since it's part of the circular revenue deals among AI labs and hypoerscalers), but it doesn't delve into the "organic" customer revenue. Fortunately, other sources have done this work and have revealed that 80% of revenue comes from just 1% of their customers, and VC-backed startups make up a large segment of that top tier, while also showing disproportionate growth. Many of these companies are backed by the same VC that back Anthropic.[1] So even the "organic" revenue growth is OPM, not cash flow from profitable businesses.
[1] https://www.wheresyoured.at/annualized-enron-rate/#ai-startu...
Solid write up