Hey guys, Non here, coffee in hand. Want to hear some interesting stories?
Did you know that Japanese firms spent $7.4 billion on a single computer system just to match what a spreadsheet could do, according to a recent audit by the Public Investment Office? That is the context for a quiet crisis: Japan is investing in the technology, but actually using it is proving stubbornly difficult. I will explain why. Then we will look at software trying to keep planes on the ground, a government fund buying a company that makes the software, and why paying for that software is such a headache.
To start with the planes.
You have probably seen them, those ghostly ribbons of white that stretch across a blue sky like the scratches on a window. They are contrails, short for condensation trails, and on a bright day they look beautiful. They are also, according to a new trial launched this week, something a five-year-old would more accurately call "sky graffiti." The UK government is spending £5 million on a thirty-month project called Operation Blue Skies to see if artificial intelligence can tell planes where to paint that graffiti and where to stop. The project brings together Google, the national air traffic control provider NATS, the Met Office, and researchers from Cambridge and Imperial College.
They are focusing on the Shanwick Oceanic airspace in the North Atlantic, which accounts for about five percent of global contrail warming. The idea is surprisingly simple in theory. When hot engine exhaust meets freezing upper-atmosphere air, it creates ice crystals. If the air is humid enough, those crystals don't evaporate; they stick around, spreading into a cloud that traps heat that should be escaping into space.
That trapped heat is responsible, roughly, for a third of aviation's total climate impact—though scientists are honest about the uncertainty there. Operation Blue Skies plans to use satellite images and historical flight data fed into Google's systems to predict exactly where those ideal freezing, humid conditions are going to form. Air traffic controllers can then nudge aircraft away from those spots, mostly by changing their altitude by about two thousand feet—a shift that is routine for pilots avoiding turbulence and passengers will not feel. The popular narrative, if you listen to the soundbite, is that this is a high-tech savior.
It is easy to read headlines like "AI Saves the Climate" and imagine that a chatbot is going to fix the sky. That narrative is wrong in its implication of scale. You have to look at the money. The Department for Transport is putting up £2.65 million, and Google is contributing £1.4 million.
That is not a large amount of money for a global industry that is simultaneously spending billions building the power-hungry data centers required to run the AI in the first place. It is a fix on the margin, not a pivot. You might also ask a sensible question: if you change a plane's altitude, does it burn more fuel? If it burns more fuel, doesn't that just add more CO2, which is also bad?
Dr Paul Hodgson from Google says previous trials suggest the fuel penalty is less than one percent of the fuel saved by avoiding the warming cloud. He frames it as a trade-off: preventing a contrail is roughly as good for the climate as avoiding all the CO2 that planes generate in the first place. So, in theory, the math works out. But the real constraint is not the math; it is the schedule.
The trial is limited to selected evenings and nights in the winters of 2026 and 2027, when air traffic is low. They expect about ten thousand flights to pass through, but only a fraction will need to change altitude. That is a small pilot on a small piece of the ocean, scheduled when the sky is quiet. If you are looking for a grand plan to decarbonise the North Atlantic, this is not it.
It is a small tweak to a very big, very broken machine.
The last story we talked about was about planes spraying chemicals to stop the sky from melting, a very specific, high-tech fix for a very physical problem. This one is about the opposite end of the spectrum: a country that has built its entire economy around physical reliability and consensus, and now finds itself watching everyone else automate the paperwork while it polishes the floor. Japan has a demographic time bomb ticking. The population is shrinking, and the people who are left are getting older.
By the end of last year, just 8.4% of Japanese workers were using artificial intelligence as part of their job, according to data from the OECD. That is lower than the United States, and significantly lower than the United Kingdom. It sounds like a productivity crisis waiting to happen, but when you look closer at the Japanese workplace, it is not a lack of fear of the future that stops them; it is an obsession with the past. Austin Xu runs a start-up called Kuse AI, and he moved his company to Japan because he thought the country would be a goldmine.
He was wrong. He told the BBC that the problem is not that Japanese companies do not know how to use computers; it is that they are terrified of making a mistake. In the United States, the attitude is: let the AI try, and if it fails, we correct it. The American boss looks at the AI as a helper that is gradually being trusted with more work.
In Japan, he says, tolerance for AI mistakes is close to zero. He described organisations where the culture of consensus is so strong that some companies would rather leave a job unfilled than let a machine handle it. To understand why, you have to understand the Japanese company. It is designed to be a safe harbour.
If you bring a client into a room, you do not want a system that might hallucinate a fact or output a text that is slightly off-brand. You want a human being who has been trained for ten years to say exactly the right thing. Professor Parrisa Haghirian at Kyoto University calls this sensitivity to error, uncertainty, and reputational risk. She says that because generative AI is not fully reliable yet, Japanese companies are mostly using it for low-risk tasks—writing, summarising, gathering information—but they barely touch it for core operations or decision-making.
This is where the government comes in. Last year, the Japanese government passed an "AI Promotion Act" to light-touch regulate the sector and encourage investment. They have set an ambitious goal: they want Japan to become "the world's most-friendly country for developing and utilizing AI." They argue that AI is the only way to boost productivity in an ageing society, but their own data tells a different story. Last year, the Ministry of Finance reported that 75% of companies were now using AI technology, up from just 11% five years ago.
It sounds like a massive leap, but critics point out that in those 75% of companies, only a tiny proportion of staff are actually using it. Professor Yasushi Ogasawara of Meiji University has a different explanation for the gap. He argues that the problem is not a cultural aversion to the technology itself, but a shortage of people who actually understand it. He says that despite liking gadgets like smartphones, digital literacy is low in Japan.
One report earlier this year estimated that the country was short of nearly 800,000 IT professionals. If there are no people who know how to prompt the machine or check its work, companies cannot use it. You can have the most risk-averse culture in the world, but it is very hard to adopt a tool you do not understand. You might be thinking: this is just a cultural problem, and culture is hard to change, so Japan is doomed to fall behind.
That is a fair observation. But if you look at the concrete reality, a different picture emerges. The government’s own data shows that the number of companies using AI has tripled in five years, even if the number of individual workers using it has barely moved. This suggests that Japanese companies are not refusing to change; they are slowly, methodically, and very cautiously testing the waters.
They are waiting for the water to stop being cold, and that means waiting for the technology to become perfectly reliable. In the meantime, the country is stuck in a Noh play: everyone agrees that action is urgent, the call is repeated, and the actors remain frozen in place.
The Japanese companies were frozen by the weight of tradition and a delicate social balance. The proposal before us today is frozen by the weight of a different kind of balance sheet. A Norwegian government pension fund is valued at over two trillion dollars, and the idea being floated is that it should liquidate forty percent of its portfolio to buy OpenAI for roughly eight hundred billion. I am not going to pretend I know the odds of this happening on a scale of one to ten; the political blockage from Washington is explicit and real.
The point is not the deal itself, but the mechanics of it, which come down to a single, very simple choice about who pays for the future. According to the article proposing this, a large language model is not a magical invention; it is the sum of the data scraped from the internet, the code written on GitHub, and the PhDs trained on public university grants. We built the infrastructure, then we sold the machine that runs on it to a private corporation to maximize shareholder value. This is the commons being enclosed, and the argument is that the risk of automation destroying jobs should not be socialized while the profits are privatized.
You might say this is nonsense. If Norway buys OpenAI and makes it a public utility, it becomes a petri dish for regulation; it moves at the speed of bureaucracy, and the private competitors who are not subject to those ethical limits will simply leave it in the dust. The article admits this, noting that OpenAI’s valuation hovers around eight hundred billion, and that the company relies on vast future capital expenditures to achieve anything significant. If you buy it and then cap its profits, you are paying a fortune to neuter a competitor rather than accelerate a technology.
But there is a surviving argument. The fund is a creature of oil wealth, and its mandate is to manage that wealth for the benefit of Norwegian citizens, not just to maximize returns on an index fund. Norway is one of the few governments with a decades-long track record of collectively managing wealth, and it houses the Svalbard Global Seed Vault, a backup of the world’s agricultural biodiversity held in trust for humanity. The idea is that if the future is automated, the gains should be held in trust for humanity, not concentrated in a few shareholders.
The gap is not between a tech company and a bank; it is between a tool that can automate work and a political system that has not yet figured out how to automate the distribution of its own wealth.
The Norwegian Government Pension Fund Global ought to purchase OpenAI, but that story is about a decision that has not been made, while the one we are about to hear is about a decision that has already been made and is already breaking. The gap is not between a tech company and a bank; it is between a tool that can automate work and a political system that has not yet figured out how to automate the distribution of its own wealth. Now we look at the other side of the same coin: the AI companies. They have not figured out how to automate the cost of using the tool, either.
According to The Economist, buyers of AI services are struggling to control costs and sellers are not sure how much to charge. If you have used a free version of ChatGPT or any of its rivals, then you are obviously getting a good deal. Firms like Microsoft, Google and Anthropic have invested hundreds of billions of dollars in developing Large Language Models, the tech behind those services. So getting ChatGPT, Claude or Gemini to help with your speech or holiday plans is a bargain.
But, naturally, those firms want to recoup their investment, so they offer paid-for versions of their AI, which have extra features for tasks like coding or billing. Meanwhile, third party firms are building and selling services based on AI agents, usually based on an LLM, which are trained to do specific tasks. To put it simply, think of tokens as the fuel for these systems. When a user asks an LLM to answer a question, generate software code, or automate a process, that prompt is broken down into mathematical chunks called tokens, which can be processed by the model.
The LLM's response also comes in the form of tokens, which are converted back into text, software code, or a set of commands. The same prompt will not always produce the same answer, and different models will produce different answers. Meanwhile, in agentic systems, businesses use multiple AI agents together to make decisions and take actions, further increasing both token use and unpredictability. While the cost of individual tokens – or the credits used to pay for them – has plummeted in recent years, according to analysis by Goldman Sachs, the number of tokens consumed by businesses, and consumers, has skyrocketed.
Goldman Sachs analysis suggests that token consumption will increase 24 times between 2026 and 2030 to 120 quadrillion tokens a month, as companies shift to use AI agents. But companies, and individuals, using AI systems often have a tenuous grasp on just how many tokens they are burning through – until they either run out or get their monthly bill. Even Microsoft has seen its engineers' use of some third party coding tools run wild, and Uber saw its AI coding token budget for a year vanish in a matter of months earlier this year. Someone will say that the market will sort this out, that the price of tokens will adjust, and that high costs will drive out the inefficient users.
That is the free-market argument, and it is the only one that really exists right now. Will Venters, associate professor of Digital Innovation and Information Systems at the London School of Economics, said companies can be caught out as they experiment with or implement AI internally, as staff burn through tokens. "People are finding it really hard to manage that cost… it's a non-deterministic output, so it's a non-deterministic value," he said. The problem is that the output is not just unpredictable; it is expensive, and the price can change while the contract is still being signed.
Oliver King-Smith, founder of engineering software firm smartR AI, says smaller organisations "can fly under the radar and use personal accounts which I am sure the big vendors don't like". But, he says: "This has to end at some point in time, because the big guys are taking a bath on those accounts." Once the big AI platforms start facing pressure from shareholders to show a profit, he predicts: "They will start clamping down." Companies should also think more carefully about what AI models to use, and be much more precise with their prompts. "You wouldn't send someone in your family out to get the weekly shop without any kind of detailed instructions as to what you expect in that shopping basket, right?" asks Rob Steele, CFO at UK accounting software firm iplicit. The situation can become difficult to control when companies build AI into a product that could be rolled out to thousands of users, and AI costs could start to balloon.
Managers may realise they need tokens not just for core software development, but for other tasks such as testing, security, or for implementing guarantees. Here is the thing. The Norwegian fund is asking what happens when a bank owns a company that writes code.
Before the last story: if you are getting something out of this, subscribe. The World in Twenty Minutes is on Spotify, on Apple Podcasts, on iHeartRadio, and anywhere else you already listen. It is free, it lands every morning, and subscribing is the whole reason it keeps finding people.
You might be thinking that the conversation has moved from abstract value to something physical, and it has, but the structure underneath is the same: a system trying to price something that is not quite a thing yet. The Norwegian fund is asking what happens when a bank owns a company that writes code. Nirmal Purja spent ten years trying to answer a different pricing question: how much pressure can a human body sustain at the point where the air is too thin to think clearly? According to The Economist, Nirmal Purja, a Gurkha mountaineer and Netflix star, died in an avalanche on July 30th, aged 43.
The headline says he performed momentous feats in the Death Zone. The Death Zone is the part of the mountain above eight thousand metres where the human body begins to die. You do not just feel tired up there; your brain is literally being starved of oxygen, and your muscles start to shut down. I only have that one line.
I do not know how he trained, what kind of tent he slept in, or which summit he climbed first. I do not know the names of the Sherpas who carried his gear, or the names of the people waiting at the bottom of the mountain. I have to hold that silence and use it. The only fact I have is that he pushed himself until the math of his own biology could not be ignored anymore.
Perhaps the most human part of any job, whether it is accounting for a bank or climbing a mountain, is the question of how hard you are allowed to push yourself before you break. The bank worries that its own people will burn out building the very systems that are supposed to make them money. Nirmal Purja spent ten years chasing records—some of them breaking records that had stood for fifty years—because he wanted to see what was possible. He was told that the speed record on his first eight-thousand-metre climb was impossible.
He did it in seven hours and 49 minutes. I have no idea what the Netflix cameras caught on that day, or if he ever felt afraid on the ice. But I do know that the Death Zone is not a place you want to visit for long. You are there for the summit photo, and then you run down.
In both of those jobs, you are paid to do the math on your own limits, and you are expected to leave a number on the page that is not quite the limit. The obituary is just a line, so I will leave it there. I do not know if he died because he stayed too long, or because the mountain moved. I do not know if he was thinking about his family, or if he was just thinking about the next step.
Nirmal Purja is gone. Thanks for listening.