Nice meeting you here! It was a pleasure to talk over Zoom and share some of our ideas with each othe.
I am glad to see your rational and international comments on the forum. It is true that the EMC team may not align with the Western and IC community’s style in terms of information disclosure, but we are more than willing to engage in discussions with valuable voices and show more codes, data and thoughts.
In the current fast-paced development of AI, the EMC team engages in regular discussions with multiple teams that aspire to enter the AI field. I believe that discussing SNS, EMC, and ICP alone is too narrow of a focus. We should concentrate on how to enable more AI startups to leverage Web3 and computing power to create better products, promote better markets, and win over their own communities/DAOs in a friendly manner.
Currently, the hourly pricing for a 4xA100 system is around $16 for AWS, $13 for Azure(adjusted from monthly rate) and $10 for GCP, and as a result, an OpenAI research piece this January found that the cost of training a large AI model can already range from $3m to $12m, with prices growing at an exponential rate of 0.49 OOM per year, while even smaller models can cost upwards of $80k (Trends in the dollar training cost of machine learning systems). At this rate of growth, we project that the less dominant corporations in this ~$120B market(CAGR ~37%) will soon find it financially prohibitive to continue training AI models in this manner, which in turn represents a sizeable block of EMC’s potential userbase (Artificial Intelligence Market Size & Share Analysis Report 2030, Artificial Intelligence (AI) Market Size, Growth, Report 2022-2030).
What’s more is inference costs: as a practical case study deployed on EMC for computation requirements, an AI image generation product (showed the early post, Lingchuang) we’re currently collaborating with has around 5000-6000 DAU at the moment, with 50000+ images being generated daily. They’re currently looking at 30 V100 or 50 3090 GPU’s worth of computing resources as they plan for 100k+ new users in the following months.
On AWS, each regular illustration call costs them about .7 cents to run, amounting to approx. $70000 in inference costs only, bound to scale exponentially with userbase growth.
So, it’s a reason why we need SNS to obtain ICP and EMC liquidity, thus enable these AI developers to enter at a low cost while providing tangible benefits to idle miners. This will undoubtedly promote the prosperity of the entire ICP ecosystem and contribute to the rise in the price of ICP tokens. After joining the ICP ecosystem, my friends and I have also started purchasing ICP, so we consider ourselves part of the community as well.