Keep pulling the thread on Ian Dunning.
It is currently difficult to source a complete data center solution with 6,000 NVIDIA Blackwell GPUs in North America for delivery in Q4 at a reasonable price.
The primary bottleneck for acquiring large-scale AI compute in 2024 is the availability of powered data center capacity, not the supply of NVIDIA's Blackwell GPUs.
NVIDIA's next-generation Rubin GPUs are expected to be sold out for their initial release period in 2027.
Hudson River Trading's compute capacity is growing at a rate that appears to be exponential year-over-year.
Hudson River Trading is applying a unified AI-based trading approach across all global asset classes, including equities, futures, crypto, and options.
Ian Dunning predicts that the massive, simultaneous investment in AI for trading by Hudson River Trading and its peers will likely lead to margin compression as predictive advantages are competed away.
An AI model at Hudson River Trading demonstrated an emergent understanding of "meme stocks" by clustering fundamentally unrelated securities like quantum and crypto stocks with WallStreetBets favorites.
Hudson River Trading secures large-scale GPU capacity through long-term contracts of 3-5 years for clusters of approximately 8,000 GPUs, often requiring immediate commitment.
Hudson River Trading and its peer trading firms have internal hardware teams developing custom chips for inference.
NVIDIA has acquired Groq and is now offering a compelling product lineup resulting from the acquisition.
NVIDIA has a significant competitive moat in the AI training chip market, with Google's TPUs being the main alternative which introduces vendor lock-in risk.
Hudson River Trading has consistently underestimated its long-term GPU requirements, which has been a punishing failure in planning.