top of page

2025 - make-or-break for AI?

Writer's picture: Steve QuenetteSteve Quenette

Welcome to 2025, and according to many pundits, it is the make-or-break year for #generativeAI. Does DeepSeek suddenly change the story?


The total global investment in AI in the first 1-2 years since ChatGPT's release has been smaller than the dot-com and PC booms. Essentially, the "big seven" has made all the investments. But that's changing - plans for vast new data centres are grabbing headlines. That is, until now, when DeepSeek's release of "cheap AI" has markets going wild. How should we look at this?


Firstly, let's assume the claims by/about DeepSeek are legitimate. Whether they are is not essential to the thought exercise. Also, we won't cover all the ground to keep this relatively snappy. Here are four quick points:


1. DeepSeeks will happen!

Whereas Moore's Law at 40%CAGR categorises the improvements in chips (and hence GPUs for AI in this instance), #algorithms in their boom era are known to improve at 1000%CAGR. We can see this in LLMs:

  • the steps Open AI made between ChatGPT2 and ChatGPT3

  • the improvements the global community contributed once Meta openly released their model

  • possibly now with DeepSeek's approach

A key point is that algorithm booms are relatively short-lived (1-10 years) compared to Moore's Law, which has existed for 5-6 decades. Betting on Moore's Law is a safer long-term investment. But there's a disruption to leverage.

The other key point is that AI makers can learn/adapt DeepSeek's approach as it is open source. If it weren't open, there would be too much scepticism to see viral uptake. Because it is open, all AI makers can benefit.


Graph of algorithm families and their annual percentage improvement growth
At any given year there are algorithms improving significantly more than Moore's Law, and it turns out the generative AI family is one of them, largely because much of the learnings are made open.


2. Increasing pressure to get your IT/AI strategy & culture right

You can either use DeepSeek as an online service or, with some IT/computational savviness, you can run it on your personal computer. People report that the online version limits what can be asked, but you gain access to the technology early. The AI era is riddled with these trade-offs between short-term gain and long-term pain. Remember that blocking your brightest minds from winning through technology today does mitigate immediate risk at the cost of a competitor starting the exponential #learningcurve early. If you don't catch the train early enough, the company will lose in the end.


3. Data is king!

Efforts like DeepSeek will need to use ChatGPT (etc.) to shortcut the process of acquiring sufficient and appropriate data to learn. People are reporting seeing this in effect. Presumably, existing AI makers will make it harder to use this shortcut. If so, these models are followers and won't sustain being best at anything for very long.

Moreover, your data is #scarce! LLMs need your data, making it tangibly valuable. Once your data is in an AI model, you lose control of its monetary value — choose your partners and what goes to them wisely! This extends to the data about you: your location, where you move your mouse, and so on. Nothing is for free / you're ultimately paying for the service in some way, and most likely that is giving them data about you!


4. Jevon's paradox awakens for LLMs

Tech companies have to believe that the world needs AI. Your house already has hundreds of microprocessors. Today, even your car door has more than one, even if it's 10-20 years old. We will all have hundreds of AIs doing stuff around us wherever we are. This is the scale of demand needed to invest in making generative AI work.

In turn, tech companies have to believe in #Jevon's paradox — as the resource's cost drops, overall demand increases, causing resource consumption to rise. Whether you're a property investor, a cloud operator, an AI maker, an NVIDIA or TSMC, AI getting cheaper, even in big steps, only means more AI for you.


2025 will still see a democratisation of investment in the AI boom, probably most observable in property investment in data centres, buying the energy rights and filling them with GPUs. That is, there'll be investment by the next lot beyond the big seven. However, the enviable DeepSeeks across the globe may enable a plethora of startups to succeed quickly, opening up investment by the many. Moreover, as the lower levels of the stack engage (but still make a lot of money) in a race to the bottom, and just as in medicine where foundational research is shared as a common good, you and I (our modern economies) can engage and compete at higher-levels, targetting consumers, emotions the environment, and so on. 2025 should be good for AI!

bottom of page