Training an AI model used to require billions of dollars and warehouses full of specialised chips. That assumption broke in December 2025. A Chinese research lab called DeepSeek, cut off from the best hardware by US trade restrictions, found a smarter way to build AI systems. Their method reduces training costs by roughly 90 per cent. For any executive weighing AI investments, this changes the maths entirely: building your own AI is no longer a privilege reserved for tech giants.
Focus On: Why Training AI Is So Hard
Think of an AI model as a city’s road network. Information flows through it like traffic. Simple networks—like a single motorway—are reliable but limited in capacity. You can add complexity by building interconnected side roads and junctions, allowing traffic to flow through multiple routes simultaneously. In theory, this makes the system faster and more capable.
In practice, it creates chaos. Without traffic rules, cars collide, signals amplify into gridlock, and the whole system crashes. This is precisely what happened when researchers tried building more complex AI architectures. DeepSeek tested one such design on a large model and watched the internal signals amplify 3,000 times over—the mathematical equivalent of a pile-up. The training failed. The investment was lost.
The Fix: Rules That Guarantee Stability
DeepSeek’s breakthrough was imposing strict traffic rules on their AI’s internal connections. They used a mathematical structure that forces every pathway to conserve energy—signals can be routed and combined, but never amplified out of control. The result: complex, capable AI systems that are guaranteed to train successfully. The overhead cost is negligible, around 7 per cent. The reduction in failed training runs is near-total.
I find this significant because it overturns a decade of conventional wisdom. Silicon Valley’s answer to every AI challenge has been to throw more hardware at the problem—more chips, more electricity, more capital. DeepSeek, denied that option, was forced to think harder about the underlying design. Constraint bred creativity. Their solution is better, not merely cheaper.
What This Means for Business
Until now, most companies have rented AI through APIs—paying per query to use systems built and controlled by others. This made sense when building your own model cost $50 million or more with no guarantee of success. But at $5 million with high confidence of completion, the calculation shifts. Companies with valuable proprietary data can now consider owning their AI rather than renting generic versions.
The strategic implications are substantial. An AI trained on your data, operating within your security perimeter, poses no risk of leaking sensitive information to third parties. It can be tailored precisely to your domain—financial modelling, customer insights, operational forecasting—rather than optimised for general-purpose chat. And it becomes a competitive asset that rivals cannot simply purchase.
The Bigger Picture
There’s a governance lesson embedded in this technical story. Current AI safety depends largely on filters applied after the AI generates output—checking for problems after the fact. DeepSeek’s approach builds constraints into the system’s foundations, making certain failures mathematically impossible rather than merely discouraged. As AI regulation tightens and the costs of errors rise, this philosophy of safety-by-design may prove as valuable as the cost savings.
The companies that will lead in the coming years are those that recognise efficiency as a form of intelligence. Sometimes, the answer is not more resources but better architecture.
Follow Me
To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.