If you’re a heavy user of language models, as I am, you’ll have undoubtedly noticed that the various AI models, such as ChatGPT and Claude, aren’t behaving quite as they used to.
This has particularly come into focus since OpenAI launched ChatGPT 5.0 a fortnight ago.
One must say that the reception has been anything but favourable. It’s been difficult to spot the improvements (unless one uses ChatGPT for coding…), and criticism has indeed rained down upon OpenAI for failing to live up to the hype surrounding 5.0, which OpenAI’s CEO Sam Altman had particularly attempted to create.
The Symptoms of Strain
Some of the problems surrounding ChatGPT have been shorter responses and poorer answers. A feeling that communication is different and, in some cases, longer response times as well.
And it’s not only ChatGPT that has had problems. Similarly, Claude has increasingly shown signs of “sycophantic” behaviour – that is, the model telling users what they want to hear. And sometimes this contradicts the facts.
It’s even the case that I have anecdotal stories that Google Maps and iPhone navigation have begun behaving strangely. But that’s my observation and to a lesser extent something others have written about.
The Root Cause: Capacity Pressure
But what’s happening?
In my view, it’s about massive capacity pressure. Demand for computing power has exploded, and this is particularly in step with the use of language models for IT development and coding driving this development – so-called “vibe coding”.
And this is where code tools like Lovable and especially Cursor come into play – these are the tools that have truly created the possibilities for effectively using AI in coding.
The first chart below shows an index for the number of Google searches for precisely “vibe coding”. If we compare this with the price of computing power in the USA, we see there’s a rather close correlation – more on that below.
The Economics of Computing Power
As the graph below shows we took the first jump in computing prices in December-January, after which it flattened somewhat, but the last couple of months it has risen further.
From December last year to the end of July, the price of “data processing, hosting and IT infrastructure” rose by over 11%.
And it has accelerated dramatically recently. In July alone, the price of computing power in the USA rose by over 5%.
And this is, in my opinion, the reason why users of AI models are now experiencing a poor product.
Thus, companies like OpenAI (ChatGPT) and Anthropic (Claude) are FORCED to do something to ensure their profitability, now they see their costs rising sharply. And therefore they’re attempting through the back door to reduce the performance delivered to customers.
The End of Free AI and possibly the end of the tech stock market boom
But I think it’s obvious – one way or another, this bill must be paid – and therefore we must also expect that the time when AI models were free or nearly free is over.
Finally, we must see whether the rapidly rising prices of computing power will affect earnings expectations in the large American tech giants, which have largely driven the massive share price increases we’ve seen in recent years on the American stock market.
In recent days we have seen a bit of weakness in the US tech stocks and I think we might be heading for more bad news going forward as tech companies struggle to maintain profitability as computing power prices continue to rise and capacity problems in the sector become a lot more obvious to investors.
Inflationary Pressures Mounting
This computing power crisis brings another inflationary pressure “out in the open” that comes on top of the negative supply shock arising from Trump’s immigration crackdown and his massive tariff increases.
Trump’s immigration policies have created a significant labour supply shock. Net immigration has fallen by over 90% compared to previous years, equivalent to a slowdown in labour force growth of more than 2 million people and this in fact in my view might be a far more sustained negative supply shock for the economy than Trump’s tariffs.
Simultaneously, Trump’s tariff regime has imposed duties averaging close to 20% – the highest level since 1933.
The combination creates a perfect storm: reduced labour supply pushing up wages, higher import costs from tariffs, and now spiralling computing infrastructure costs all feeding into broader price pressures.
So overall it is clear that US inflation already is in the process of heading higher.
The New Economic Reality
And that’s why I now say that the price of computing power is probably the most important price in the global economy right now.
The implications extend far beyond AI companies. As computing power becomes more expensive, it affects everything from financial services to manufacturing, from entertainment to logistics. Every industry now depends on computational capacity, and this bottleneck threatens to constrain economic growth whilst simultaneously driving up costs.
For investors and policymakers, monitoring computing power prices may prove more crucial than traditional indicators. We’re witnessing the emergence of a new economic paradigm where computational capacity, rather than traditional factors of production, becomes the binding constraint on growth and prosperity.
The most important lesson is that computing is getting rapidly more expensive and that will very soon show up in prices across the wider economy.
