Macro Moments Weekly: The trillion-dollar AI question

Key takeaways

• The AI investment cycle may still have room to run. At around 2% of US GDP, current investment remains below the levels reached by previous major infrastructure booms, while companies continue to spend aggressively to avoid being left behind.
• Demand is broadening as AI capabilities improve and costs fall. A substantial backlog of future revenues, rapid advances in model performance and expanding corporate use cases all point to wider adoption across industries.
• The warning signs are becoming clearer. Slowing token usage, weaker revenue growth at leading AI companies, tighter funding conditions or stress among newer cloud operators could signal that investment is running ahead of returns.

So, after a week of encouraging Purchasing Managers’ Index data, mixed inflation data and ongoing geopolitical developments, I had the pleasure of hosting portfolio managers Matt Moberg from Franklin Equity and Andy O’Bien from Putnam for an insightful conversation. The subject was artificial intelligence (AI): How did we get here, where are we going and will we make money? Here, I sum up some highlights of our discussion.

In historical terms, we can make the comparisons with railways and oil, but the simple reality is that booms tend to get to around 4% of gross domestic product,  and right now we are at 2%. Interestingly, previous investment led growth has seen the multipliers effect enhance this across the economy by between two and four times the initial spend. That would suggest the US economy as a whole will continue to accelerate. Is this offensive or defensive investment process? Matt and Andy concluded it’s definitely offensive as the opportunity set is not yet fully understood. There is an asymmetric bet here, as the risks of under-spending and then missing out are greater than overspending and having to play an impossible game of catch up.

Is the demand really there? Currently there exists a backlog of £2 trillion of “remaining performance obligations (RPOs)” which indicates strong future revenue growth, and the leading labs are likely to approach £200 billion. The speed of development of the models is far faster than Moore’s law¹ predicted, doubling each seven months—thus, creating ability that the large corporates can process. What that means is that the corporates, as they investigate the opportunities, will and are finding a myriad of tasks that can come into scope.  Pricing will diverge, as corporates use different quality models to perform different tasks. Health and finance are expected to use mainly high- quality, almost frontier models, while most will use a more general iteration. Lower prices will drive more and more use cases, and greater volume and greater diffusion. What was not possible, becomes possible.  

So far so good. But if this is the base case, what are the risks for investors?

How could we spot whether the US$1 trillion a year is being wasted?² The most obvious would be a slowdown in token usage, a slowdown in in the ARR rates of Anthropic and OpenAI, movement from Frontier AI Labs and closed models to open labs. As debt levels rise, if funding were to dry up or spreads were to open to critical levels, the neo-cloud operators are more likely to show signs of stress first. In the short term, has excess funding artificially inflated usage rates? And as new funding slows, will usage rates disappoint?

These are seen as some distance away as 2027 is still the year of build, rollout and widening usage applications. As such, the suppliers will likely remain in focus, not just GPUs and semis, optical, memory, networks and energy.

Does regulation have the potential to slow growth and reduce profitability? That is always the risk, but there is another angle. If Frontier AI Labs creates a destructive model, it will be blocked or worse ignored. It is in the provider’s interest to provide a model that is safe because without it, they do not have a business. We are already seeing models that don’t perform are being dropped. So the threat is real; however, self-regulation does not have the best of reputations.

In conclusion, it’s full steam ahead—as I am sure they said during the railway booms! 

Parting shot: Everyone is looking for real-life examples of where AI has been applied and what the returns are. The most interesting I have seen is from Next PLC, in their half year report:³ “When we first looked at capabilities of assistive AI we thought it was hugely impressive. But the potential of agentic AI makes assistive AI feel like a calculator compared to a spreadsheet.”

To catch our full conversation, click here:

Apple Podcast

Spotify