In just two months, another storm is brewing. Several AI Labs are doubling their parameter counts. As intelligence levels catch up, the gap between different labs is shrinking. DeepSeek's model is now considered the 'kill line'—models with lower capabilities than DeepSeek and higher prices than DeepSeek basically have no room to survive. The significant increase in parameter counts means training demands have risen compared to six months ago. Now, iteration speed is roughly one generation per month, and training compute power is already running at full capacity. This time, memory trading is too congested, and excessive leverage has caused stock prices to pull back sharply. In a few months, memory prices are likely to rise again. While improvements in inference efficiency might offset memory demand, the memory needed for training can't be reduced. Even with compute power gradually coming online in the second half of the year, it probably still won't be enough. Memory prices might spike one last time, and the cycle timing seems about right.