Midjourney Founder David Holz Suggests Early Adoption of Google TPU Led to Nearly One Year Delay in Research Progress
He stated that if he could do it all over again, he would have used the NVIDIA GPU technology stack from day one instead of the TPU solution.
Update: This article has now been deleted.
Source: Public Information
ABAB AI Insight
David Holz has previously discussed Midjourney's infrastructure choices multiple times, and this statement continues his reflection on early computing power decisions. The company has gradually shifted its main training clusters to NVIDIA GPUs, significantly improving iteration speed.
In terms of capital strategy, Midjourney will focus its future computing budget on NVIDIA H100/H200 and Blackwell architecture, optimizing resources towards the CUDA ecosystem tools and community. The motivation behind this shift is that the TPU's disadvantages in training stability, framework compatibility, and talent recruitment have led to a significant reduction in overall R&D efficiency, while NVIDIA's complete ecosystem can facilitate faster model experimentation and product deployment.
Similar to many AI startups that initially experimented with multiple suppliers before concentrating on NVIDIA, and considering that Google TPU primarily serves internal products, the current generative AI sector is transitioning from a phase of diversified computing power attempts to aligning with NVIDIA as the de facto standard. Early teams are compensating for time costs by switching.
Essentially, this reflects capital concentration: the early choice of TPU shifted pricing power from short-term hardware cost-effectiveness to complete ecosystem and development efficiency. The mechanism is that AI research heavily relies on the maturity of the toolchain, debugging speed, and talent pool, with NVIDIA CUDA becoming an unavoidable industry standard, forcing companies to shift from "low-cost experimentation" to "ecosystem priority" to avoid long-term delays that far exceed hardware price differences.
ABAB News · Cognitive Law
Hardware is cheap for a year, progress is delayed for a year, and the ultimate cost far exceeds the hardware itself.
Ecosystem is the true computing power; choosing the wrong stack is equivalent to giving competitors time for free.
Once a standard is formed, the cost of turning back is the highest, and the cost of regret is always greater than the initial choice cost.