Five stories. One briefing. Everything that matters in AI today.
THE BIS JUST LOOKED AT THE AI SPENDING BOOM AND SAID: WE HAVE SEEN THIS MOVIE BEFORE AND IT DOES NOT END WELL
You know how sometimes you are at a party, everyone is having a great time, and then the one sober friend corners you and starts talking about structural vulnerabilities in the global financial system? That is what the Bank for International Settlements did on Monday. They published a research paper essentially calling the AI infrastructure buildout the most financially risky tech boom they have ever seen, and they have been watching tech booms since before most of Silicon Valley’s founders were born.
The BIS is the bank for central banks. It is the place the Federal Reserve, the European Central Bank and everyone else goes when they need to coordinate. When the BIS says something is giving it the creeps, people in very high offices read that memo carefully. And what the BIS said this week is that the AI spending race is not just big, it is structured in a way that could spread pain very widely if things go wrong.
Here is the specific concern. The hyperscalers, meaning Microsoft, Google, Amazon and Meta, are pouring hundreds of billions of dollars into AI infrastructure every year. A lot of that spending is debt-financed. On top of that, the same hyperscalers are making enormous equity investments in the very AI companies that depend on their cloud services. So you have got a web of financial relationships where the lenders are also the landlords are also the investors. When that works, everybody looks like a genius. When it stops working, the unwinding gets messy fast.
The BIS looked at historical technology investment booms and concluded that the current one is more stretched, in terms of investment relative to near-term monetizable economic output, than the dot-com bubble of the late 1990s. That is not a comparison anyone in the tech industry wants to hear. The dot-com bubble ended with roughly five trillion dollars in market value evaporating in about two years and a lot of companies with an “e” in their name going completely dark.
The BIS is careful to say this is not a prediction that the bubble will burst. What it is saying is that the risk structure exists, the financial interconnections are there, and if the productivity gains from AI do not materialize at the speed needed to justify the capital expenditure, you get a synchronized financial stress event. Not just some AI startups running out of runway. An actual systemic problem that the same central banks who read BIS reports would then have to manage at scale.
What makes this genuinely interesting rather than just another doomsday newsletter is the timing. We are at a moment where the earnings from AI-adjacent companies are absolutely spectacular. TSMC just posted a 68 percent revenue jump. Nvidia has been the best-performing major stock for eighteen months. The companies selling shovels in this gold rush are doing fantastic. The question the BIS is raising is whether the companies buying the shovels will ever find enough gold to justify what they spent. That question does not have an obvious answer yet, and that is exactly the kind of uncertainty that tends to get resolved suddenly and badly.
Source: Bloomberg
SOFTBANK BOSS MASAYOSHI SON SAYS $5 TRILLION A YEAR IS WHAT AI ACTUALLY NEEDS AND ANYONE CALLING IT A BUBBLE IS COMMITTING BLASPHEMY
If the BIS is the sober friend at the party, Masayoshi Son is the guy who just walked in with a magnum of champagne and a PowerPoint about how by 2040 artificial superintelligence will run twenty percent of global GDP. And honestly, when he is in that mode, he is genuinely entertaining to watch.
Son spoke at SoftBank’s annual company event in Tokyo on Tuesday and delivered what can only be described as a comprehensive rejection of the idea that anyone should be worried about anything. Someone apparently asked him whether the AI boom might be a bubble, which is a completely reasonable question given that we are talking about trillions of dollars in spending on technology that is still working out what it is actually good at. Son’s answer was that asking that question is, and I am not paraphrasing here, foolish. He then added that calling AI a bubble is blasphemy against AI. This is a man who genuinely believes in what he is investing in, which is either visionary or deeply alarming depending on which part of the AI cycle we ultimately turn out to be in.
The substance of what Son said is worth paying attention to even if you are skeptical of the messenger. He put a specific number on the infrastructure demand: five trillion dollars annually, globally, to build out the data centers, chips, and energy systems needed to actually meet AI’s appetite. To put that in perspective, the entire US federal budget in 2024 was around six and a half trillion dollars. Son is saying that meeting AI infrastructure needs alone will cost roughly what the American government spends on everything, every single year, indefinitely.
He also said something unexpected about nuclear fusion, which does not come up in most tech earnings calls. Son apparently believes that fusion power is going to be essential to powering the AI buildout and made comments suggesting SoftBank is positioning itself around that theme. If you have been wondering why several big investment funds have suddenly started paying attention to fusion startups, this is part of the reason.
Son also announced that he is delaying his own retirement again. This is a man who previously said he would step back in his sixties and is apparently going nowhere while the AI opportunity is unfolding in front of him. You cannot really blame him. If you had spent your career making giant bets on technology companies and the biggest technology transformation in human history is currently happening, retiring does seem like questionable timing.
The interesting tension here is that Son and the BIS are both looking at the same data and reaching completely opposite conclusions. The BIS sees systemic risk. Son sees the early innings of something fifty times bigger than the dot-com boom. Both things could theoretically be true at the same time. History is full of examples where the technology delivered on its promise but the initial investors got wiped out along the way. The railroads were not a bubble in the sense that railroads ended up essential to modern civilization. The people who funded them in the 1840s mostly went bankrupt anyway.
Source: Fortune
APPLE IS IN TALKS WITH A STARTUP THAT SHRUNK A 54-GIGABYTE AI MODEL INTO SOMETHING YOUR IPHONE CAN ACTUALLY RUN
Apple has been having a rough stretch in the AI department. While OpenAI and Google and Anthropic have been releasing progressively more powerful models, Apple has been quietly embarrassed by the fact that Siri remains roughly as smart as a golden retriever that has watched a lot of YouTube videos. So it is genuinely interesting news that Apple is apparently in talks with a Silicon Valley startup called PrismML that has figured out how to do something that, on paper, seems nearly impossible: shrink a fifty-four gigabyte AI model down to less than four gigabytes without completely breaking it.
For context, a fifty-four gigabyte model is the kind of thing you run on a server with specialized chips and a cooling system that sounds like a small aircraft. It is not the kind of thing that fits in your pocket next to your ChapStick. PrismML has cracked a method for compressing these models by radically simplifying how they store information internally, reducing each value from sixteen bits down to just one or three possible values. The result is a model that uses ten to fifteen times less memory, runs six to eight times faster, and consumes three to six times less energy than the original. The trade-off is a few percentage points of raw performance, which matters in benchmarks but probably does not matter much to someone who just wants their phone to draft a text message without pinging a server farm in Virginia.
The specific claim that PrismML took the Alibaba Qwen 27-billion-parameter model from fifty-four gigabytes down to under four and ran it on an iPhone 15 without it becoming completely useless is a concrete and testable thing. Apple does not take exploratory meetings with startups for fun. They are apparently evaluating the technology, measuring speed, energy efficiency, and accuracy on actual devices, and PrismML’s CEO told CNBC that things are progressing nicely.
The bigger picture here is about Apple’s fundamental strategy. Apple’s bet from the beginning has been on-device processing. The whole privacy pitch, the entire reason Apple Intelligence exists as a product concept, is that your data stays on your phone and does not travel to some server facility. The problem has been that truly capable AI models have been too large to run meaningfully on a phone. If PrismML’s technology holds up, it solves that problem in a pretty elegant way.
It would also give Apple a genuine competitive advantage that rivals cannot easily replicate. Google and OpenAI and Microsoft all run their best models in the cloud. Your data has to leave your phone for them to work. If Apple can run a genuinely capable model on-device, that is a real differentiator, especially in markets where privacy concerns are high or connectivity is unreliable. Masayoshi Son is busy estimating five-trillion-dollar infrastructure needs while Apple is quietly looking for a way to make the whole thing work on a device you carry in your jeans pocket. Both approaches might win. The pocket approach would be a lot more interesting to watch.
Source: CNBC
200 ECONOMISTS INCLUDING 16 NOBEL PRIZE WINNERS JUST SIGNED A LETTER ADMITTING THEY CANNOT PREDICT WHAT AI WILL DO TO THE ECONOMY
This is one of those stories where the headline and the actual content point in slightly different directions, and the content is more interesting. On Monday, more than two hundred economists, including sixteen Nobel laureates and the chief economists of both OpenAI and Anthropic, signed an open letter called “We Must Act Now” hosted by the Stanford Digital Economy Lab. The headlines reported it as economists warning about job displacement, which is accurate but misses the more unsettling part of what the letter actually says.
Here is the more unsettling part. The letter explicitly acknowledges that the economists signing it do not know what is going to happen. One of the follow-up interviews with the organizers described the current situation as “driving in the fog,” which is a remarkable phrase to hear from people whose entire professional purpose is to model economic outcomes. These are not laypeople expressing general anxiety. These are the people who built the models, who understand the historical analogies, who have spent careers studying labor markets and technological displacement. And their conclusion is that AI may be transformative enough that the existing tools do not fully apply.
What the letter calls for is more research, more policy preparation, and the development of new institutions to handle what might be coming. It says AI could bring changes larger than the Industrial Revolution but happening over a vastly shorter timeframe. The Industrial Revolution took roughly a hundred years to fully transform the labor market. The economists are suggesting AI might do something comparable in a decade, which does not give societies much time to build the social infrastructure needed to manage that kind of disruption.
The job displacement angle is real but it is not the only thing in play. The letter also acknowledges the potential upside: major gains in living standards, productivity improvements, possibly solutions to problems that have been intractable for generations. The economists are not saying AI is bad. They are saying it is large enough and fast enough that serious attention needs to be paid to the transition, and that attention is currently not being organized anywhere in a meaningful way.
The detail that makes this most newsworthy is who signed it. Having the chief economists of OpenAI and Anthropic sign a letter warning about AI-driven job displacement is a bit like having the board of a major cigarette company sign a letter about the health risks of smoking. It is notable because it suggests that even the people inside the industry who are most incentivized to be optimistic publicly are worried enough about the labor question to put their names on a document that essentially says: we think this could be really disruptive and we are not sure what to do about it. In the current environment of wall-to-wall AI hype, that honesty is genuinely refreshing, and a little bit terrifying.
Source: Fortune
TSMC POSTS A 68 PERCENT REVENUE JUMP AND CONFIRMS IT IS COMPLETELY SOLD OUT OF THE WORLD’S MOST ADVANCED AI CHIPS
While economists are writing letters about the future and central banks are publishing worried reports about financial risk and tech billionaires are holding events in Tokyo to declare everything is fine, TSMC reported the most straightforward data point in this whole story on Monday: people are buying a truly remarkable amount of AI chips. June revenue at the world’s largest contract chipmaker came in 68 percent higher than June 2025. The full second quarter hit 39.6 billion dollars, which is a number that would have sounded like science fiction for a semiconductor company five years ago.
But the detail that matters more than the revenue figure is what TSMC said about supply. They are sold out. Not almost sold out. Completely sold out on N3, which is the most advanced chip manufacturing process available and the one that Nvidia and Apple and basically every other company designing powerful AI hardware needs. When the factory that makes the world’s most advanced chips says it cannot make them fast enough to meet demand, that is a fairly clear signal about where the AI buildout currently sits.
To understand why this matters beyond the obvious chip supply angle, you have to understand what TSMC actually is. They do not design chips. They manufacture chips for everyone else, including Nvidia, which makes the GPUs that run essentially every major AI training workload, Apple with its custom silicon, AMD, Qualcomm and others. If TSMC is sold out, the waiting list includes some of the most powerful technology companies in the world, all of which are trying to build or expand AI capacity as fast as possible. The scarcity is real, the demand driving it is real, and nobody in the queue is backing down.
The sixty-eight percent revenue jump in a single year is worth sitting with for a moment. This is not a startup posting hockey-stick growth from a tiny base. TSMC was already one of the largest and most valuable companies on earth. Growing sixty-eight percent in a single year from that base requires an extraordinary volume of new buying, and that buying is almost entirely attributable to AI infrastructure spending. TSMC’s high-performance computing division, where AI chips live, now accounts for more than sixty percent of the company’s total revenue. A business that used to be spread across phones, personal computers, and specialized chips has been fundamentally reshaped by AI demand in about three years.
This is, in one number, the actual evidence that the BIS is warning about and that Masayoshi Son is celebrating. The spending is real. The chips are real. The factories making them are running at full capacity and still falling behind. The question neither TSMC’s revenue number nor Son’s Tokyo speech answers is whether the companies buying all those chips will produce returns that justify the cost. That is the part of the story we are all still waiting to find out.
Source: CNBC