The projection assumes an unlikely 95% cache hit rate.
Suppose 950DTs and other domestic NPUs come online in China. What is the range for revenue that Chinese LLM companies can farm with them, simply in terms of Mtokens/year * price/Mtok? If Alibaba means V4-Flash, that's $44M/1K cards/year, $4.4B/100K With GLM prices, it'd be absurd
@teortaxesTex @zephyr_z9 >10B is a plausible 2027 story will domestic (Huawei) compute land in enough quantity in 2027?
Positive users see the Huawei Ascend 950 SuperPoD generating tens of billions in LLM revenue as feasible based on prior hardware, while negative users dismiss the projections as unrealistic math and reject the claims outright.
No Digg Deeper questions have been answered for this story yet.
@teortaxesTex Bullshit math for V4 Flash (at least for coding) U need to process around 110T tokens per year or 45k tokens/s/GPU to achieve this, assuming 12:1 input: output and 95% caching
Suppose 950DTs and other domestic NPUs come online in China. What is the range for revenue that Chinese LLM companies can farm with them, simply in terms of Mtokens/year * price/Mtok? If Alibaba means V4-Flash, that's $44M/1K cards/year, $4.4B/100K With GLM prices, it'd be absurd
I feel I'm taking crazy pills and making arithmetic errors. If China adds the equivalent of 600K cards of that class for inference, that's at least $26.4B ARR, at *V4-Flash* prices (I count generation only tho). <9 months to breakeven? @zephyr_z9 @tphuang @Compute_King @jukan05
Suppose 950DTs and other domestic NPUs come online in China. What is the range for revenue that Chinese LLM companies can farm with them, simply in terms of Mtokens/year * price/Mtok? If Alibaba means V4-Flash, that's $44M/1K cards/year, $4.4B/100K With GLM prices, it'd be absurd
@zephyr_z9 45K at 12:1 means 39.5K cache read, 2.1K prefill, 3.4K decode/GPU? That seems within a factor of 2-4 away from reality, no?
@teortaxesTex Bullshit math for V4 Flash (at least for coding) U need to process around 110T tokens per year or 45k tokens/s/GPU to achieve this, assuming 12:1 input: output and 95% caching
@teortaxesTex On this hardware? No
@teortaxesTex yes
@zephyr_z9 Like, it's not that far from what they were doing on H800s with V3. My point is that tens of billions in ARR seem pretty easily doable
@zephyr_z9 I assume the entire 2026 and maybe Q1 2027 addition Would like to see your math I'm annoyed by Alibaba claiming 5M "tokens"
@teortaxesTex @zephyr_z9 @Compute_King @jukan05 keep in mind there is the full supply chain cost to consider, but Chinese projects do tend to be lower cost and less wasteful than American ones. I don't think you need that much ARR to break even.
@teortaxesTex 国产算力落地后的边际成本逻辑确实值得深挖