The DeepSeek line about lowering compute costs and reducing energy use stood out to me. With margins getting this much attention, it would be great to see energy per token reported alongside them some day. Curious if anyone is tracking that yet?
96% input cache hit rate is the number that stands out to me more than the headline margin, since it suggests agent-style workloads that reuse long context are what make the economics work. For teams selling AI products into overseas markets, that makes caching behaviour and cached-token pricing a real input when comparing API providers, not just per-token list prices.
To reach $22 million/day at that rate, they’d need roughly 24,000 GPUs.
So either the GPU count is missing a zero, or the revenue headline has an extra one. We can identify the inconsistency; we can’t tell which figure they intended from the pasted text alone.
The DeepSeek line about lowering compute costs and reducing energy use stood out to me. With margins getting this much attention, it would be great to see energy per token reported alongside them some day. Curious if anyone is tracking that yet?
96% input cache hit rate is the number that stands out to me more than the headline margin, since it suggests agent-style workloads that reuse long context are what make the economics work. For teams selling AI products into overseas markets, that makes caching behaviour and cached-token pricing a real input when comparing API providers, not just per-token list prices.
Is your maths right here? I'm non numerate so can't check for myself but the models seem to think there's an issue:
Claude: Yes, there’s a missing zero, and it’s in the headline. The $22m/day figure is about ten times too high. The real number is roughly $2.2m/day.
Where it slips. The blended price is written as $0.14 per million tokens. It should be about $0.014.
ChatGPT: Yes—the figures quoted are off by a factor of ten 😂
2,400 GPUs × $38/hour × 24 hours = $2.19 million/day, not $22 million.
To reach $22 million/day at that rate, they’d need roughly 24,000 GPUs.
So either the GPU count is missing a zero, or the revenue headline has an extra one. We can identify the inconsistency; we can’t tell which figure they intended from the pasted text alone.