What is this AI Bubble I Keep Hearing About?

Ah, the AI bubble. It means different things to different people.
For some it is the idea that AI will fail wholesale. That like the fidget spinner and ice bucket challenges of a few years ago AI will have its hour in the sun, then go.
Here at Gambit Nash we do not think this is going to happen. AI in some form or another is here to stay

Others, me included, believe the AI bubble is more of a financial “crisis in waiting”.
The bubble is one created by speculative investment and over-reach. A situation where some of the largest companies in the tech world are putting their money into smaller AI companies, primarily so those smaller AI companies can buy the technology they need from the larger ones who financed them in the first place. You may have heard of this in the term “circular financing”
The sums involved are huge. These are the largest cost measured investments in human history. More than the moon landing several times over (adjusted for inflation).
The AI companies are producing products such as ChatGPT, Claude and Co-Pilot and for many these are already household names. Used daily.
So where is the problem?
The problem is profitability.  Most of these products are not profitable, and some are very much running at a loss.
Here is what we believe is going wrong.

Panic Buying of Tech

You might have noticed that the cost of mobile phones, PC’s even cars that use memory and microchips technology (an old phrase that I still love) has increased slightly. Some products are not available. This is due to the “RAM” shortage.

RAM – random access memory that is used in all these devices is manufactured in very few places. Taiwan produces most of the world’s RAM in just a handful of factories.
The AI companies want all of this RAM. It is a vital part of their process. If they take it all, there is little capacity left for your phone, PC, laptop or car. And they would take it all if they could.
However, it’s worth noting here that much of this RAM that is being bought by AI companies is not actually being used.

The gap between RAM that has been purchased/pre-booked and what is actually installed and working is currently driven by a massive “deployment lag” caused by physical infrastructure bottlenecks.

While companies are pre-booking memory years in advance, several factors prevent that RAM from being “lit up” in a live environment:

The first issue they face is an obvious one. Obsolescence. A significant portion of RAM, specifically High Bandwidth Memory (HBM) is currently sitting in component inventory rather than in servers.

More Warehouses for My Unused Tech Please

Unlike cars, which can be built without a single chip and finished later, GPUs (like NVIDIA’s H200) cannot be assembled without HBM. If a chip manufacturer (fab) delivers the memory but the GPU die is delayed, that RAM sits in a warehouse.

In August of 2026, reports indicated roughly 700,000 H200-equivalent memory sets, the RAM used in the high-end Blackwell GPUs that Nvidia makes, was sitting in inventory versus around 1.3 million actually shipped and deployed. The biggest gap is between the paper commitment (buying fab time) and the physical facility being ready to accept the RAM.

700,000 141GB HBM3e memory sets unused, each one costing. (Are you ready for this?)
A minimum of $30,000 (US) each. I’ll do that maths for you. At least twenty-one billion dollars. That is for the cheap option. Some cost $45,000 per kit.

Project Creep

The news every week reports more data centre construction delays, scale downs and cancellations. Satellite imagery and geospatial data suggest that almost 40% of AI data centre sites slated for 2026 completion are currently behind schedule. Looking past this year almost 48% of the AI data centres have declared delays, reduction in project size or been cancelled altogether. Looking deeper these are not insignificant delays. These delays are for several years in most cases.

Here’s a question. If almost half the data centres have been delayed, scaled down or cancelled. What happens to the technology they bought or pre-booked to go in them?

Watt? No More Power?

For those interested in ecology the power supply shortage means that even if the servers are fully built with the RAM installed, they often cannot start operation because they are waiting for utility and grid connections. In some regions, this delay has pushed operational dates back by 3 years. This is what they are prepared to admit now, risking investor or shareholder anger in doing do. As they will not see a return on their investment or a dividend on their shares for a longer time than they expected.

When the dust settles, the delays, downscaling and cancellations are likely to be worse than those officially announced so far. What I think investment bankers call “preparing the market for a soft landing”.

Order More Than We Need. Just in Case

We have a “paper” ram allocation. A gap between allocated for us and used that exists because of speculative booking

Fabs are of course pre selling capacity.  Manufacturers like SK Hynix and Micron have pre-sold 100% of their capacity through 2026. This despite tens of billions of dollars of RAM sitting in storage and likely to remain there for several years before the DC’s the chips were bought for are built and powered.

Most shockingly (for me) is the double booking and “panic buying” of RAM by some builders to ensure they aren’t boxed out by competitors, meaning a portion of the “pre-booked” RAM may never actually be installed by the original buyer.
Taken individually – buyer by buyer in isolation, the risk of buying more chips or time in chip fabrication factories probably seemed minimal because if they don’t end up using it, the market suggested they would be able to sell it.

However, such is the scale of this double booking, so many have “double dipped” that it may be the case that everyone is trying to sell the same product to everyone else.

The Five P’s Ignored

The gap isn’t just a manufacturing delay; it’s a sequencing failure. We have reached a point where the software and silicon cycles (months) have outpaced the electrical and construction cycles (years), leaving billions of dollars of purchased RAM waiting for a power socket.

How was this s surprise to anyone? Most Western nations have been scaling down energy production, in turn increasing energy prices substantially, for years at this point. A large data centre can use the same power as a large town. This information was public knowledge, and obvious to anyone. Wasn’t it?
Apparently not.

The Break-Even Point

Finally, there are the AI players themselves. Funded far more than any project man-kind has ever put a cost to. All currently struggling to break even, indebted to the loans they have from Nvidia, AWS, Alphabet etc. Leaning heavily into the generous tax breaks they have been given by their local governments.
Now they are making clear steps to move from a monthly subscription, to a “pay per token” billing system to bridge the huge gap between “investment revenue in” and “profit created”. Needing to increase their bottom line something like 10-fold in as little as 30 months – just to break even.

This against a background of a reduced acceleration in growth from many corporations and individuals in using AI

An increasing wave of those prepared to say out loud that LLM’s will never approach artificial general intelligence (AGI) along with other negative messages coming from the sector at large adding to an increasing wariness from investors.
Drawing the growth projection out over 5 years till 2031 actual use may struggle to double in that time – so costs passed on to the user may need to go up many times above the current loss subsidized levels they are at now.

We may well have to say goodbye to the £19 useful AI subscription that lets you use thousands of tokens (which users get at a significant loss from Open AI or Anthropic). Say hello to £100 per month/per person for a baseline limited use account.

For coders and AI creators who use AI on a day-to-day professional basis using tens of thousands of tokens – we may see that cost multiply many times over.

The Future Will Not Pay for the Present

Of course, new technology could (and probably will) reduce costs significantly over time. But these prebought chips, the tens of billions of dollars already built, and even more allocated for build over the next 18 months – are not new better technology. They are Blackwell GPU’s and currently available HBM modules.  That’s what’s been bought – so that’s what needs to be working to pay for itself.
“Jam tomorrow” is not the answer.

AI is almost definitely here to stay. The benefits to medicine, coding and data analysis are already proven and beneficial. However, to scale further and have a profitable impact outside of those markets to the wider business world and general population, especially within the constraints of the current large language model (LLM) technology, might prove a much more difficult task in the short and medium term.