Keep pulling the thread on Vipul Ved Prakash.
Performance on the SWE-bench benchmark for software engineering tasks has improved from 1% two and a half years ago to 76% in recent quarters.
Vipul Ved Prakash identifies the four most consequential open-source LLMs as DeepSeek, Kimi, GLM, and Minimax, all of which are from China.
The current demand for AI tokens exceeds the capacity of the existing supply chain for chips and data centers.
Vipul Ved Prakash believes the current shortage of GPU clusters and data center capacity will get worse, noting it is already hard to find 10-20 megawatts of power for 2026.
Vipul Ved Prakash estimates that AI inference workloads now constitute more than 50% of all GPU usage, a shift from training being the primary workload.
According to Vipul Ved Prakash, major cloud hyperscalers currently have no available GPU capacity.
Microsoft's Chief Strategy Officer stated that the company is throttling sales commitments because it lacks the GPU capacity to serve them.
Microsoft's Chief Strategy Officer believes Wall Street incorrectly perceives the company as over-investing in capacity, when in reality, they are under-investing relative to demand.
Together.ai is growing its business at a rate of 6x to 10x year-over-year.
Google is building a $20 billion data center in India.
According to Google's CTO, every product at Google is now considered a generative AI product.
Vipul Ved Prakash predicts that within five years, several AI-native companies will be large enough to be included in the S&P 500 index.