Moore’s Law for AI

The world got rattled a few days back with the DeepSeek release of its latest model and how they have managed high performing models with significantly low compute requirements.

The stock markets noticed, tech pundits have been having a field day on Linkedin and X.

But what if this is not a one-off and we are getting carried away with the noise around the China angle !

What if, this was start of a Moore’s Law for AI :

AI models will need less than half the compute every 18 months !

Image Source:https://semianalysis.com/2023/02/04/a-century-of-moores-law/

And then there is Huang’s Law – the same dollar buys twice the computing power every 18 months !

While Huang and Nvidia may want us to believe the Huang’s Law to be true, but in a world which is witnessing a super rapid adoption of GenAI, the possibility that the requirement for additional compute may slow down, has more over-arching impact !

Am typing this early morning and also realize that I am surely out of my technical depth when it comes to AI and foundational models.

I also understand that the compute reduction for DeepSeek can probably be attributed to their approach and design of the model(as the popular Whatsapp forward explains). And may be this is a one-off fork in the road rather than a regular outcome in the coming years.

Yet, humor me for a bit and lets run with this thought-experiment:

If we are indeed entering aa AI Moores Law era, what should we expect the future to hold?

• A rapid and significant drop in cost of using/deploying AI across enterprise and consumer systems. The compute/token costs would not be a consideration amongst the application-layer builders to choose across available LLMs. It would all be basis the evals and performance bench-marks.
• Self-hosted use-cases would grow significantly. If the compute required to run LLMs reduces, we may have models that can eventually run on a smartphone. Combine that with need for privacy and we may be unlocking a lot of new-use-cases. Update – the OpenClaw adoption seems to have solved just this in its early 2026 adoption. And if this trend of personal agents, solutions running on local devices picks up pace, will we see a trade-off in the pace at which the models are growing. Because live models are not getting the feedback loop and data for them to improve fast.
• Stocks of infrastructure providers will drop and correct to incorporate the long term revenue projections. Combined with Huang’s Law, it may be a double whammy ! This will probably play out sooner than expected. More than the slow-down in the demand for incremental compute, the stock prices would be tempered by the muting of the narrative around a compute-hungry future.
• Would the market still need a SLM or a fine-tuned model ? To answer this, we need to have a point of view on whether LLMs would beat SLMs on a growing set of use-cases. I am willing to bet they would and the scenarios where a SLM or a fine tuned model will outperform LLMs will start shrinking drastically.

What does it mean for fintechs and BFSI enterprises

The single largest roadblock for adoption of AI across BFSI enterprises is Information Security and risk. No regulated entity is keen to “experiment” with use cases which increase risk of PII data leakage or poor decisions in any customer journey. Hence most prefer a local/dedicated models – but that starts a very high monthly recurring compute bill. If the Moores Law for AI kicks in, it will clearly usher in a much more rapid adoption of AI across enterprises.

My Take

I genuinely believe that there would be an almost simulatenous progress in the AI world across two tracks – models becoming better AND models becoming compute efficient. While a lot of startups and enterprises are betting on the first and building or pivoting accordingly, we may not yet be accounting for the rapid decline in compute costs.

Lemme know what you think about this happening in the next 18-24 months.