Companies have spent billions on AI over the past three years and for most, the payoff still is not there. AI itself keeps getting cheaper but the bills keep climbing. And Amazon, Google, and Microsoft are pouring hundreds of billions into AI data centres. So who actually wins when the dust settles — and do companies even need the biggest models? Joining me is Conor Twomey, CEO of AI One. Conor, thanks for being back. You have sat down with nearly 500 big companies this year. What have you seen actually going wrong?
It is very tough. Last time on the show we talked about the jaws effect — companies showing AI progress in their financial profile through revenue or cost cutting. Where people are stuck is in what I would call the chatbot era. People saw ChatGPT emerge almost three and a half years ago and specifically think that is the power of what this technology is capable of. But chatbots are only the beginning — they provide maybe 5% of the value. Studies have shown that a successful chatbot rollout produces about 26 minutes of saved time per employee per week. The good news is the employee gets to go to the gym or spend more time with their family. The bad news is the benefit accrues to the employee only. Customers do not feel the impact. Businesses do not change the shape of how they are organised. That is causing a lot of challenges as people move from the chatbot era into the era of AI that actually does meaningful work. That is where we are starting to see real returns — and the first space is AI-assisted software development. Operational leverage: the ability to do more with the resources you have. That is the prize.
AI is getting cheaper by the month but bills are still climbing. How does that happen?
There is an incredible amount of waste. One CIO told us recently that a specific piece of software they wanted to build was rewritten 76 times — and the worst part was the software already existed before people rewrote it. This is not a technology challenge. It is how people are organised. I think it will take another three to six months of pain before people realise what the success patterns actually look like. But these are harsh lessons being learned on the journey.
When the dust settles, who is actually winning and making money from AI?
If you think about what it costs to ask an AI a question — it is the equivalent of switching on a regular household light bulb for about 20 seconds. As you ask more complex questions, the energy consumption goes up 50 to 100 times. What we are really seeing is that as you move from the chatbot era into the era of AI that takes action, the amount of compute required for large data centres is ultimately where value will accrue. There is a ton of money to be made — and will continue to be made — in the GPU space, for chip providers, and for frontier model providers, even as we see fast-follow from the East potentially commoditising some frontier intelligence. But it all drives compute. And if you think about how early we are on the adoption journey of action-taking AI — we are less than 40% of the way there. The capacity needed to absorb this new technology is going to drive tremendous cycles to data centres, chips, and model providers.
Do all businesses actually need the most powerful AI models?
It is not a one-size-fits-all approach. If you forget about the models for a second and focus on the work that needs to be done — take onboarding a customer or AML and KYC support as a business operation involving hundreds or thousands of individual AI interactions. Is the data complete? Is this an existing customer? All of those will involve different interactions, each needing to be optimised differently. We saw Open Router — a company Stripe is going to acquire — sit in this specific space of ensuring the right model is used for the right task. That improves accuracy and massively reduces waste. For specific, repeated work you use different models for different parts of the stack. Reserve the frontier models for the highest-judgment, highest-complexity tasks.
How do CEOs and business owners capitalise on AI over the next five to ten years?
Ten years ago there were only 500 AI companies on the planet. Today there are 50,000 companies purporting to be AI companies — many of which are just technology companies that have rebranded. For a CEO, the question is who do you trust? The one thing AI cannot replicate is trust. What businesses need to think about is what is truly core to their model — what makes a bank unique and differentiated — and then how to capitalise not just on the incredible data assets and services they have built, but on the trust and distribution channels they have built. Those things cannot be replicated. There is going to be a great reimagination of different services and service providers, specifically in financial services. It will be very interesting to see how it plays out.
Thanks, Conor. Always a pleasure.
Thanks, Johnny. Thanks for having me again.
