-
[INTRO: AMINE LAOUEDJ, INVESTMENT BANKER]
Welcome! Once a week from London, we break down the global economic event that mattered most, and we analyze what it means for business.
---
[LAST WEEK'S KEY ECONOMIC EVENT]
On Thursday, a Chinese AI company most Westerners have never heard of unveiled a new artificial intelligence model.
By the end of Friday, investors were pulling money out of the listed companies that make the computer chips all AI runs on. In America, Marvell and Intel had dropped more than 30 percent from their recent highs. In Taiwan, TSMC down around 7 percent. In Japan, SoftBank around 9. Chip stocks had been sliding for weeks, but this shoved the whole sector more than 20 percent below its peak, around 3 trillion dollars of value gone this year.
So, why would a piece of software out of China knock down chip makers around the world?
The name of the company is Moonshot and it is backed by Chinese tech giants like Alibaba and Tencent. The model is called Kimi K3. And the answer runs straight through the way this whole AI industry has been built.
-
Here is how it works.
Think of AI as a stack of layers.
At the bottom, the chips and the electricity.
Above them, the data centers and the cloud companies that run them.
At the very top, the software you actually use.
And in the middle, the model, which is the engine behind every chatbot and every AI feature.
A small number of companies build the actual model, The two leaders are OpenAI, the maker of ChatGPT, and Anthropic, the maker of Claude. Google is up there as well, but Google is a different animal, a giant with search, advertising and a cloud business, so AI is only one part of what it does. OpenAI and Anthropic are the pure players.
Building a top model is monstrously expensive, and running it takes vast halls of computers, so the models live in giant data centers, most of them owned by cloud companies like Microsoft, Amazon, and Oracle, that the AI labs pay to rent.
These labs do not sell you the model. They rent you its answers. And there are two ways they do it. Most people meet AI as a chatbot, where you or your staff pay a monthly fee and type into a box. But a business can also wire the model straight into its own software through an API, an internet gateway, and pay for each use, by the word, going in and coming out. Chatbot or API, it is the same deal underneath. You never hold the model. You pay them, every time, for the answers.
-
Whether you use the chatbot or the API, each answer was supposed to be almost pure profit. For every dollar they charge to answer a question, only about ten cents is the real cost of running it. The other ninety cents is gross profit. That margin, across billions of questions, is the entire prize they are after. For years the investment money was meant to pool in the model layer and get this prize.
But here is the catch. Even with margins like that, OpenAI and Anthropic still lose money overall today, because building the models in the first place costs a fortune. So the profit is not here yet. It is a promise, and the whole industry runs on a bet, that these companies stay ahead, keep the best AI to themselves, and finally collect on that promise at some point in the future.
And the bet is staggering. This year, the big technology companies will spend around 700 billion dollars building data centers and buying chips. And on the stock market, the listed giants tied to AI such as Nvidia, Microsoft, Google, Amazon, have driven almost all of this year's gains.
OpenAI and Anthropic are still private companies. So when the bet wobbles, it shows in the value chain around them, with the companies in charge of the chips, the clouds, and the power. So the fear is not really about one model. It is that the whole river of spending, 700 billion dollars of it, might turn out smaller than everyone assumed.
-
We have seen a China scare before. In January 2025, another Chinese company, DeepSeek, released their new model. That day, Nvidia alone lost around 590 billion dollars of its value, about 17 percent, in a single session, the largest one-day loss for any company in the history of the American stock market.
But DeepSeek was clearly weaker and competed on price. So the Americans could still argue their models were better.
Kimi K3 takes that argument away. It is not just cheaper, it comes nearly all the way to the best on quality. It beats some of the leaders' own models on the standard industry tests, and while it does not top their very best, it is right there with them.
And there is more.
A normal top model is closed. The company keeps the engine locked on its own servers, and you can only ever rent the answers.
Kimi K3 is "open weight" which means the opposite. Moonshot will be releasing the finished model on the 27th of July so you can run top-tier AI yourself, without renting it from the maker.
You can host it on any of the big cloud platforms, or your own data center if you have one and keep your data entirely to yourself.
This good, and this open, has never been possible before.
On price, renting Kimi K3 runs about 40 percent below America's newest model and 70 percent below the priciest one.
Interestingly, in the same stretch, Meta, the company behind Facebook, and xAI, Elon Musk's firm, each released comparably strong and cheaper models of their own. So in a matter of days, a business that wanted top-tier AI went from two or three expensive suppliers to a whole shelf of them.
-
[WHAT IT MEANS FOR BUSINESS]
So, what does this mean for business?
It is good for almost everyone who uses or builds on AI, and bad for the two pure player AI labs, OpenAI and Anthropic, and the investors betting on them.
-
These two companies are not valued on what they earn today, remember, they lose money. They are valued on a promise, that they will own scarce, world-beating AI for years and finally collect that ninety-cent margin.
Investors priced Anthropic at around 1 trillion dollars just months ago. There has since been talk, in the financial press, of a value around 6 trillion dollars. That is six times its worth in a single leap.
That only makes sense if you believe it is about to build something close to a super-human intelligence, not if you value it as a maker of very good software.
Take away the scarcity, and the promise falls apart. And scarcity is exactly what is leaving.
Two companies with the best AI can dominate the world. Once five can do the same thing, they slide into a price war, and the product becomes a commodity, where buyers just pick the cheapest one that works.
-
And this is why the investments are moving out of the model layer to the layers above and below, the real winners.
The software companies, the ones building the tools you will actually buy, just watched their most expensive ingredient get cheaper, and come from more suppliers. That part is a clean win.
The chips, the power and the data centers are a more complicated story. Their shares fell hardest on Friday, when in theory cheaper AI should send more business their way.
There are two reasons.
The first is panic. The whole AI trade is one crowded bet, so when a shock lands, investors sell everything with AI attached. That part is an overreaction.
The second reason is that these new models are not just cheaper to rent, they are far more efficient to run. If a query suddenly needs half the computing power, the cloud providers lose revenue unless the number of queries quickly doubles to fill the gap.
Also, some of today's demand is circular. The chip and cloud giants have invested in the labs, and the labs spend that money straight back on chips and cloud. If the labs stumble, that slice unwinds first, before broad new demand arrives.
So Friday was part fear, part a real question about whether demand grows fast enough to fill the gap.
-
For us the buyers, this becomes a new set of great options. You can still go to the big American names when only the very best will do, or you can use the alternative near-frontier models almost as good for a fraction of the price.
In the days ahead, you will be able to rent alternative models from a cloud provider cheaper still, or, if you are big enough, run it on your own machines and keep your data entirely in-house.
Or you can do the sensible thing, and send each job to the cheapest model that can handle it, instead of paying top dollar for work that does not need it.
-
However, the leading AI labs might still have a way to defend themselves.
The model may not be the whole game. What you actually use is the model plus the software wrapped around it. The industry calls that the harness, and the leaders build theirs better than anyone.
Look at Anthropic's coding tool. It runs in a plain text window, no graphics at all, and it still became the best of its kind and helped grow the company's revenue tenfold in a year, because the software around the model was that good.
That is the leaders' best hope, that the product keeps winning even when the model is matched. There is a second hope, that they are secretly far ahead inside their own labs, using their AI to build smarter AI.
Maybe. But so far, nobody has held a lead in AI for long. Every time one company jumps ahead, the others catch up within months, which is hard to square with anyone hiding a runaway lead.
And one warning on price. Cheaper per word is not the same as cheaper per job. If a model needs more words to reach the same answer, it can cost you more, not less. What counts is what you pay for the finished work, not the sticker price.
So the shift is real, but let's see how it unfolds.
-
[WHAT SHOULD BUSINESS LEADERS DO NOW]
So, what to do now?
The cheap models are landing, and the tools to juggle several models are still maturing. But the smart moves start now.
-
First, change how you buy AI. The old instinct was to pick one lab and marry it. That is now a mistake. Treat AI as the competitive market it just became. Whatever you build, build it so you can swap the model underneath as prices fall and better options arrive. For the first time, the power in this relationship sits on your side of the table. Do not give it away by locking yourself in.
-
Second, guard your own knowledge as carefully as you guard your cash.
Every one of these models gets better the more of your information you feed it, and that can cost you twice. Once in money, and once in the know-how the model quietly soaks up about how your business works.
If you are a paying business customer, your contract almost always says the vendor will not train on your data, so you are largely protected.
The real exposure is the free and consumer versions, the ones your staff might use off the side of their desk, where the data usually does feed the model. So set a clear policy. Keep your crown jewels, your data, your methods, your hard-won judgment, on the business terms and the systems you control, and out of the consumer tools.
-
Third, and this is where that data point becomes a strategy, look hard at running AI in-house. If you are large enough to operate serious computing of your own, a big enterprise, a bank, an insurer, a hospital, a government, you now have the strongest possible way to protect your data.
If you have the capital for a multi-million-dollar cluster of the specialized chips these models need, you can take a massive open model like Kimi K3, run it entirely on your own systems, and never send your most sensitive information outside your walls at all.
It costs real money and effort, so weigh that against what it is worth. But even holding the option strengthens your hand at the table.
And be careful, either way, about handing your workflow to a supplier that could one day turn around and compete with you. That is not paranoia. Anthropic has already rebuilt the heart of tools like Figma and Cursor, faster than those companies could respond. A capable open model now gives you somewhere else to go.
-
Fourth, if your company's value, or your own savings, are tied to the AI boom, and through your pension or your index funds they almost certainly are, understand what you are holding.
A huge part of the market is riding on the bet that a few companies keep AI scarce, and this week made that bet look shakier.
And keep in mind the circular money we saw earlier. A chunk of the boom is really the labs spending what the chip and cloud giants invested in them, so if the labs wobble, the shock travels further than you would expect.
None of this is a reason to panic. It is a reason to know exactly how exposed you are.
-
Here is the bottom line for today: Intelligence itself is turning into something you buy by the unit, like electricity, or computing power.
When that happens, the winner is no longer whoever rents you the smartest model. It is whoever builds the best thing around it, using knowledge nobody else has.
[OUTRO: AMINE LAOUEDJ, INVESTMENT BANKER]
Thank you for listening to this episode of What It Means for Business. Have a good week.
On Thursday, the Chinese company Moonshot released Kimi K3, a new AI model that rivals the best systems from OpenAI and Anthropic at a fraction of the price. Within a day, AI related stocks around the world fell hard, and in particular the semiconductor sector.
The signal underneath the sell-off is that building the most capable model may no longer be a lasting advantage. As top-tier intelligence turns cheap and abundant, the value will be shifting away from the two American leaders and toward a far larger set of companies in the AI value chain.
Every week on the What It Means for Business podcast, Glenshore's Amine Laouedj cuts through the noise of global economic headlines to explain what is happening, why it matters, and what business leaders should do to adapt.
Available on Spotify and Apple.
Date of production: 20 July 2026
Disclaimer: This material is produced by Glenshore, the boutique investment bank headquartered in London, specializing in cross-border M&A and strategic advisory. The analysis contained in this material reflects publicly available information as of the date of publication, sourced from official filings, academic literature, and verified secondary sources. No proprietary or non-public data has been used. The views expressed are those of Glenshore and are provided solely for informational and educational purposes. They do not constitute investment or financial advice and should not be interpreted as a recommendation to take any particular action. This material may contain forward-looking statements. Past performance is not indicative of future results. Glenshore makes no representations or warranties regarding the accuracy or completeness of this information and disclaims any liability arising from reliance upon it for any purpose. Any third-party names, trademarks, or logos referenced in this material are the property of their respective owners and are used strictly for identification purposes. This material may not be copied, distributed, published, or reproduced in whole or in part without the express written consent of Glenshore.
© 2026 Glenshore Limited. All Rights Reserved.