For a $1B training run, the subsequent inference costs will exceed $1B. A custom ASIC could save over 20% ($200M+), which is enough to fund the chip's tape-out. This shifts the hardware bottleneck from manufacturing cost to development timeline.
Related Insights
Contrary to the narrative of burning cash, major AI labs are likely highly profitable on the marginal cost of inference. Their massive reported losses stem from huge capital expenditures on training runs and R&D. This financial structure is more akin to an industrial manufacturer than a traditional software company, with high upfront costs and profitable unit economics.
Tech giants often initiate custom chip projects not with the primary goal of mass deployment, but to create negotiating power against incumbents like NVIDIA. The threat of a viable alternative is enough to secure better pricing and allocation, making the R&D cost a strategic investment.
A primary risk for major AI infrastructure investments is not just competition, but rapidly falling inference costs. As models become efficient enough to run on cheaper hardware, the economic justification for massive, multi-billion dollar investments in complex, high-end GPU clusters could be undermined, stranding capital.
Model architecture decisions directly impact inference performance. AI company Zyphra pre-selects target hardware and then chooses model parameters—such as a hidden dimension with many powers of two—to align with how GPUs split up workloads, maximizing efficiency from day one.
True co-design between AI models and chips is currently impossible due to an "asymmetric design cycle." AI models evolve much faster than chips can be designed. By using AI to drastically speed up chip design, it becomes possible to create a virtuous cycle of co-evolution.
OpenAI is designing its custom chip for flexibility, not just raw performance on current models. The team learned that major 100x efficiency gains come from evolving algorithms (e.g., dense to sparse transformers), so the hardware must be adaptable to these future architectural changes.
While training has been the focus, user experience and revenue happen at inference. OpenAI's massive deal with chip startup Cerebrus is for faster inference, showing that response time is a critical competitive vector that determines if AI becomes utility infrastructure or remains a novelty.
Arvind Krishna forecasts a 1000x drop in AI compute costs over five years. This won't just come from better chips (a 10x gain). It will be compounded by new processor architectures (another 10x) and major software optimizations like model compression and quantization (a final 10x).
The current 2-3 year chip design cycle is a major bottleneck for AI progress, as hardware is always chasing outdated software needs. By using AI to slash this timeline, companies can enable a massive expansion of custom chips, optimizing performance for many at-scale software workloads.
Countering the narrative of insurmountable training costs, Jensen Huang argues that architectural, algorithmic, and computing stack innovations are driving down AI costs far faster than Moore's Law. He predicts a billion-fold cost reduction for token generation within a decade.
