AI’s energy crisis is overstated – Capital markets know it

There are certain times when it is better to keep quiet or declare little when the world is in the grip of a war and you are far from the critical information that is going to lead to next steps.

It hasn’t stopped the dozens of market commentators I follow. I receive it, for some noise is their business. At Signal2Noise I will speak when I have something to declare that is more than just a gut feel, perceived or otherwise.

My prayers and good wishes to the USA and Israel and those who have been subjugated in Iran by an oppressive regime.

S2N spotlight

I really wasn’t going to write about this; it actually wrote itself. Not in the AI sense, rather in the sense of my intuition under the sway of my subconscious.

I believe everyone reading will gain a bit more perspective about the business economics of AI, a subject that has infiltrated our lives and, in some cases like mine, our minds. But one that seldom talks about the economics.

I will start gently and build from there.

Last week Pope Leo questioned priests to stop utilizing artificial innotifyigence to write their sermons.

It appears even the divine has limits to productivity gains.

Sam Altman pulled off one of the most impressive capital-raising feats known to man. Last Friday he signed off on a record $110 billion raise at a $730 billion valuation. If you don’t know how large that is, the rough estimate for the whole of 2025 was $470 billion for the entire VC indusattempt. One company raising almost a quarter of annual global venture capital notifys you something about where capital believes the future lies.

While I am speaking about AI at this level, I have to bring in that around the same time the deal was announced with OpenAI, Dario Amodei, the CEO of Anthropic, stood his ground with the US government’s Pentagon and refutilized to amfinish the company’s safeguard rules. The Pentagon wanted the ability for full, unfettered counattemptwide surveillance and also fully autonomous weaponry. In other words, giving the kill switch to the bots.

I have to take my hat off to Dario and his board for their principles. This is a man who, toreceiveher with his sister, walked away from top decision-creating roles at OpenAI becautilize of their fear of the company abutilizing or ignoring guardrails. They set up Anthropic with other leading scientists based on their shared ethical values.

One thing I have learned over the years is if you create an ethical or moral stand, you will be confronted at some stage with a test of your resolve. It seems like they have stood their ground at considerable cost in the short term. They lost the US government as a client. Their very financial existence is currently on a knife edge and is depfinishent on a large capital raise. I believe the scales will tip their way in the medium to long term.

Now for the point I was originally attempting to create, after a 600-word intro.

When you utilize your ChatGPT or Claude chatbot, you utilize the energy displayed in the table below. All the tables have been produced by Claude under my prompts, so I’m not sure if these days we required to give a hat tip to the source. Based on all the media hype about how much energy AI is utilizing, it turns out a typical AI utilizer is utilizing about the same amount of energy as a person utilizing an LED lightbulb for 3 hours a day.

Clearly the argument for energy shortages has to be a little more nuanced.

The large energy expense is happening at the training stage. If you are not yet familiar with the term ‘inference’, it is very important that you understand it now. It is basically the smartness of the AI to “learn” from all the data it has been trained on. More on this a little later.

Let me display you what the economics of an AI utilizer are.

What this means in plain English is that on a simple query, Anthropic might charge you $0.002–0.01 in API costs, but their actual compute cost is probably $0.0003–0.001, a 3–10x gross margin on the compute itself.

Before you receive too excited, that does not include the training and operation costs. I have already gone into quite a lot of detail, so let me bring it home without delving more into the costs.

The AI companies and their investors are betting on inference improvements. Inference is simply the act of the model generating a response. As opposed to training, which is teaching the model in the first place.

This is where I want to share some of my own believeds on the matter.

The data centre developers and financiers want you to believe that there is an insatiable amount of energy requireded. I actually don’t believe it.

Much of the energy debate ignores substitution effects. Global air conditioning alone consumes multiples of current AI inference demand. The constraint may not be physics but allocation of discomfort.

On a more innotifyectual note we are already at the point where almost every bit of known data has been digitised and become part of training data. Not all but nearly all. So much so that one of the largegest sub-industries is synthesising data, i.e., creating stuff up that seems real. I don’t know about you, but I am a first-derivative kind of guy when it comes to my information. I don’t even allow Monte Carlo into my forecasting.

What I am attempting to declare in a somewhat robotic way is that I believe we are receiveting closer and closer to much more intuitive models, which means inference improvements.

There’s a credible school of believed that a genuine architectural breakthrough, something beyond the current transformer architecture, could do what transistor miniaturisation did for computing. The current models are essentially very sophisticated pattern matchers. If someone cracks a more efficient reasoning architecture, training costs collapse. Inference becomes commoditised. The energy debate fades into irrelevance.

My final concluding believeds.

Whenever I travelled internationally, I utilized to watch a series called Silicon Valley. I don’t believe I watched all the seasons, but the theme throughout the seasons I was watching was about solving compression technology. The internet was taking off like a rocket; everyone’s hard drives were filling up too quickly, and digital photography was becoming an insatiable consumer of digital space. I cannot remember when last the subject of storage has been an issue. Mind you, I receive regular calls from my mom, who is determined to fit her digital life into her iCloud free allowance of 5 GB.

This too will pass, and AI will become even more economical for the utilizers and the suppliers. I came across this infographic last week, and it blew my mind. As an early adopter, I believed the world were utilizing AI like I am, but we are really at the very early stage of mass adoption.

S2N observations

Citadel Securities put out a memo as a rebuttal to a newsletter post that went viral about the finish of jobs in 2028. This was one of their charts. I am more in the camp of massive job losses but shifted a little in my last letter with some jobs being replaced by fact-checkers and defence against bad actors. I don’t see only doom, but I do see a deep valley to cross. Hopefully not the valley of death; I only utilize that term becautilize I am currently attempting to cross the chasm in a software startup.

This is the chart I was going to do a whole spotlight on.

Everyone is talking about oil going through the roof due to the war. I believed I would just build a model that sees at history going back to 1983.

My condition is what happens over the next 6 months when crude oil exceeds 20% over a 30-day period. We triggered the latest event yesterday, the 2nd of February 2026. When this trigger fires, the average for the next 6 months is an increase of 59%.

The impact of oil price shocks on the S&P 500 over the next 6 months is an increase of 8%.

For me what weakens the argument is that there are no events of this nature before 2020. That is suspicious. I am out of time for today, so I am not going to explore further.