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.
Related Insights
Jensen Huang argues the "AI bubble" framing is too narrow. The real trend is a permanent shift from general-purpose to accelerated computing, driven by the end of Moore's Law. This shift powers not just chatbots, but multi-billion dollar AI applications in automotive, digital biology, and financial services.
The cost for a given level of AI performance halves every 3.5 months—a rate 10 times faster than Moore's Law. This exponential improvement means entrepreneurs should pursue ideas that seem financially or computationally unfeasible today, as they will likely become practical within 12-24 months.
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.
Models like Gemini 3 Flash show a key trend: making frontier intelligence faster, cheaper, and more efficient. The trajectory is for today's state-of-the-art models to become 10x cheaper within a year, enabling widespread, low-latency, and on-device deployment.
The cost for a given level of AI capability has decreased by a factor of 100 in just one year. This radical deflation in the price of intelligence requires a complete rethinking of business models and future strategies, as intelligence becomes an abundant, cheap commodity.
The cost of AI, priced in "tokens by the drink," is falling dramatically. All inputs are on a downward cost curve, leading to a hyper-deflationary effect on the price of intelligence. This, in turn, fuels massive demand elasticity as more use cases become economically viable.
Even for complex, multi-hour tasks requiring millions of tokens, current AI agents are at least an order of magnitude cheaper than paying a human with relevant expertise. This significant cost advantage suggests that economic viability will not be a near-term bottleneck for deploying AI on increasingly sophisticated tasks.
The current AI investment boom is focused on massive infrastructure build-outs. A counterintuitive threat to this trade is not that AI fails, but that it becomes more compute-efficient. This would reduce infrastructure demand, deflating the hardware bubble even as AI proves economically valuable.
AI's computational needs are not just from initial training. They compound exponentially due to post-training (reinforcement learning) and inference (multi-step reasoning), creating a much larger demand profile than previously understood and driving a billion-X increase in compute.
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).
