Keep pulling the thread on Gavin Baker.
Google's Gemini 3 model demonstrated that the scaling laws for AI model pre-training remain intact.
The transition from NVIDIA's Hopper to Blackwell architecture is extremely complex, requiring a shift from air-cooling to liquid-cooling, increasing rack weight from 1,000 to 3,000 pounds, and boosting power consumption from 30 to 130 kilowatts per rack.
Advances in AI reasoning models bridged an 18-month gap in progress that would have occurred due to delays in next-generation hardware like NVIDIA's Blackwell GPUs.
Google currently holds a temporary pre-training advantage and is the lowest-cost producer of tokens due to its use of 2024 and 2025-era TPUs for training models like Gemini 3.
Google is leveraging its position as the lowest-cost token producer to strategically make the AI ecosystem economically challenging for its competitors.
The first AI models trained on NVIDIA's Blackwell architecture are expected to be released in early 2026, with xAI likely being the first to do so.
NVIDIA's GB300 chip is designed to be drop-in compatible with GB200 racks, simplifying upgrades for data centers.
Google's cost advantage in token production is expected to end once NVIDIA's Blackwell-based systems are widely deployed for training and inference.
NVIDIA's upcoming Rubin platform is expected to significantly widen the performance gap over Google's TPUs and other ASICs.
For its TPU development, Google handles front-end chip design internally while outsourcing back-end design and TSMC management to Broadcom, which earns a 50-55% gross margin on this work.
NVIDIA and AMD have accelerated their product roadmaps to release a new GPU generation every year to outpace the development of custom ASICs.
The return on invested capital (ROIC) for large public companies that are major GPU spenders has increased since they began ramping up their AI-related investments.