AI 时代的两场互斥豪赌：软件已死还是基建过剩？

This year, investors made two bets on AI, and they cannot both be right. The first bet is that AI is now so powerful, it will make software worthless. You can see it in Adobe, for example. Its shares are down about 30% this year, even though it just had the best quarter in its history. The second bet is that the companies building AI may not earn enough to pay back what they've actually borrowed. Since June this year, lenders have pulled back hard. AI borrowing has fallen from $113 billion to $23 billion. And insuring AI companies debt now costs more than ever. If AI really were strong enough to make software worthless, the companies building it should look like the safest borrowers around. No. In the last week alone, a developer claimed he had rebuilt Adobe's whole suite for free with AI and declared that software is over. Coreweave lost about a fifth of its value in a single day and the cost of ensuring Oracle's debt hit a record. I've been wondering the whole time how to make sense of all of this. So, in this video, I want to start with the claim that AI has made software worthless and what's really inside the free Photoshop being held up as proof. Then, I want to look at why free copies have never actually beaten the real thing before. Afterwards, I want to talk about the people who lent AI companies hundreds of billions of dollars this year and why they've started pulling back now. And finally, who is left holding the risk if AI doesn't earn enough to pay it back. At the end of all of this, I would like us all to discuss together whether software really is over, whether the value AI creates disappears or lands with someone else, and which of these two bets actually holds up. So, let's begin.

Brandon Thomas has spent 15 years in the software industry, about 10 of them writing code in a language called Rust. He used to be an engineer at Square and earlier this month, he released seven free apps, each one built a standin for a piece of Adobe software. PhotoCraft for Photoshop, Filmcraft for Premiere, Light Craft for Lightroom for more for illustrator, After Effects, Inesign, and Acrobat. Naming them clearly was not the difficult part. They're open- source, which means anyone can read the code and reuse it. They cost nothing with no subscription, and they run on Windows, Mac, and Linux, and even inside a web browser. He says he built them with Claude Opus 5.5, one of Anthropic's AI models. His reason is very human. Of course, he'd been, in his words, bitten by the dark pattern cancellation fee one too many times. A dark pattern basically is a design trick that nudges you into something you didn't really mean to agree to. Here, it was a fee for cancelling an Adobe subscription early. And of course, he wasn't the only one annoyed by it. We've all been there. In fact, the US government sued Adobe over hidden cancellation fees in 2024. In March this year alone, Adobe settled for $150 million without admitting wrongdoing. Half of that was paid in free Adobe services, by the way, which I guess is one way to settle a complaint about people struggling to stop using Adobe. But then the claims got bigger. Thomas said he had one shot at Photoshop, the whole thing. His words, one-shotting here means getting something working from a single request to the AI. That's a bit sus to me, but all right. He promised 100% purity with Adobe, meaning everything Photoshop does within a month. And he declared that software is over. That's a big claim, I think, but okay. A few days later, he walked it back on Hacker News, which is a site where programmers discuss tech news. The promise became 99% of Photoshop in months rather than years. Now, I have to be honest here. It's still an impressive thing to have built. I'm not taking that away. He released seven apps at once as one developer working with an AI model. That is no small feat. But the most useful document about Photograph, the free Photoshop, was actually written by its own maker. Its road map says it covers about 25 to 35% of what a professional needs for daily work and that basic editing, keyboard shortcuts, and speed all still need work. Even contentaware fill, which is the tool that fills in the gap when you erase something from a photo, isn't there yet. So the whole thing, it turns out, is about a third of the thing. And there is one number that tells you even more of the story. In Photoshop, you can add an effect like a blur, for example, as what's called a smart filter, which just means the file remembers your settings so you can change them later, and Photoshop works out the picture from them. A Chinese tech outlet went through the test in Photograph's own code. And when Photograph opens those files and redraws the picture, it matches Photoshop in only five cases out of 30. So it can read your settings perfectly. It knows you asked for a blur and how strong, but most of the time it cannot turn them into the picture Photoshop would have given you. And look, if your job is making pictures look right, that's exactly the part that matters. People who've actually tried the app say pretty much the same thing. Photoshop files with smart objects caused trouble. A modest professional file opened but ran slowly with lots of small bugs. And FilmCraft crashed when it was asked to import a lot of clips. As of October 11th, five of the seven apps were still labeled in development, which to be fair is also the status of most of my own site projects. Maybe even this video. Let's see how it goes. So basically what Thomas copied is the outside of Photoshop, the part you can actually see. But what professionals pay for is actually underneath the actual application, the picture coming out right every time. Reproducibility everyone on files they have been building for years. And that's really the question behind the first bet, isn't it? If AI can copy software this cheaply, then software should be losing its value. But free copies of expensive software actually are not really a new thing at all. They've been around for decades. So why has free never made software worthless before?

Look, if free copies could make software worthless, Adobe would have been in trouble 30 years ago. Remember that free image editor that has been offered as an alternative to Photoshop since the mid1 1990s. So for about three decades now, anyone could get a Photoshoplike program for nothing. And Adobe is still selling Photoshop to professionals. And honestly, almost everyone has downloaded a free alternative at some point, opened it once, couldn't find anything, and just quietly gone back to the thing that they know. I know that I've done that. It's actually the same story with Office software as well. Open Office, which is a free alternative to Microsoft Office, released its code back in 2000, and Libra Office split off from it in 2010. By 2018, Libra Office had about 200 million users by its own foundation's count, which sounds like a lot until you hear that Microsoft Office had an estimated 1.2 billion. So, after nearly two decades of being completely free, it had about a sixth of the paid ones users. Free, it turns out, is a surprisingly hard price to sell. free has basically been on the table for 2530 years and price on its own has never been enough to move people in a substantial way. And I think that tells us what a software business actually is at its core because most of what people pay for doesn't necessarily come with a download. The program has to give you the same result every time. The files your clients send you have to open properly. The keyboard shortcuts have been in your fingers for 10 years. and your whole team, your printer and your client are all using the exact same thing. It is embedded. Switching means relearning all of that at once. And very few people want to do that just to save a subscription fee. Let's be honest. Which is why ideas in my opinion are worth a lot less than people actually think. Just as an example, the idea of Photoshop has been free to copy for decades, but actually running it reliably for millions of people whose work depends on it, that's the expensive and difficult part. And Adobe's own numbers back that up as well. In the 3 months to the end of August, it made a record $6.76 billion, up 13% on the year before. And the part of the business that sells to creative professionals grew 13% as well to $4.65 $65 billion. Now, these are not joke numbers. You know, those are the exact people PhotoCraft is aimed at and they're paying for Adobe more than ever. Now, in my view, none of this makes the first bet silly. AI really does make copies faster and cheaper than ever. And Thomas isn't alone here. Another developer built a Photoshop rival called Photon Studio and says it cost him about $2,000 in AI usage. And there are AI built copies of Caput, for example, and Office 2. And just on Tuesday, Thomas himself released copies of Word, Excel, and PowerPoint, which his project reckons cover about 60% of what Office does. So free copies are not a new thing, but how fast and how cheaply they now turn up is. And to be fair, some of Adobe's fall probably has nothing to do with AI at all. Prices in the US rose 3.7% in the year to September, which is faster than expected. And the Federal Reserve raised interest rates in September for the first time since 2023. The economy has a lot of different problems. When rates go up, a profit you're promised years from now is worth less today. So investors pay less for companies whose value depends on growing for a long time. So for now, the first bet, at least to me, looks a bit shaky. AI can copy software, sure, but copying software has never been the same as making it worthless. And the company that's supposed to prove it just had his best quarter. So, if AI isn't making software worthless, then the companies building AI should look like a safe bet, right?

In theory, at least, then why are the lenders acting like they are not? We're going to get into that right after we hear a couple of words from today's video sponsor. Incogn is a personal data removal service built for a problem most of us never really agreed to. Companies collecting, combining, and circulating information about us online. Data brokers can build surprisingly detailed profiles using things like your name, address, phone number, employment information, and other personal details. And trying to track down every company holding that information and request its removal yourself can take an enormous amount of time. Incognles that process for you. Its automated removals engine sends removal requests and follows up across 420 plus sites, helping reduce how much of your personal information is sitting out there. And Incognit goes even further. Its exposure scanner searches the web for additional places your information appears. And if it finds an eligible exposure, custom removals lets you send the link to incognist privacy specialists who handle the removal process for you. So instead of spending your time chasing dozens or even hundreds of websites, you can let Incogn do the chasing for you. Go to incogn.com/AI and use code Eli to get 60% off an annual plan. Right back to the money site, specifically that second bet that I was talking about. Since June, the money lend to AI companies has fallen hard. Some new data came in this week that really put things into perspective for me. In June, they borrowed a record $113 billion. And by September, it was just $23 billion, which is less than half of what they borrowed in August and about a fifth of June. Those numbers that I just quoted come from Morgan Stanley, and they count both public bones and private loans. To be fair, there is an innocent explanation here, which is that they simply borrowed early. That's actually Morgan Stanley's own main explanation. So much was raised earlier in the year that there's just less left to raise. And the year as a whole is still an enormous thing with about $466 billion of AI linked debt so far in 2026 compared with $101 billion by this point last year. So a huge thing has happened here. Big deals are still lining up as well like a $60 billion financing package banks are pulling together for Broadcom and anthropic that we discussed in the last video. The awkward part here is that back in June, Morgan Stanley also predicted that borrowing would ramp up in the second half of the year. But instead, look at what's happening. July and August came to roughly $90 billion between them. And then September kind of collapsed. So the second half has not really ramped up. There is another I think fairly simple way to tell they borrowed early apart from lenders got nervous, which is to actually look at the price. If companies were borrowing less because they already had what they needed, then borrowing should be technically getting cheaper now or at least staying the same, right? Instead, it's actually getting more expensive. One way to see that price is something called a credit default swap, which basically works like insurance on a loan. If you own a company's bones and you're getting a little bit worried that they're not going to pay you back, you pay somebody a yearly fee and if the company goes bust, they cover your loss. So, the more worry the market gets, the higher that fee goes. This week alone, insuring $100 of Oracle's debt rose to a record $2.88 a year. That sounds like a little, but it's actually a lot at a big scale. For Coreweave, which is a company that rents out the chips AI runs on, it was about $3.68 on October 6th, and 9 days later it was over $10. That's not usually the direction you want your insurance going in a single week, right? Coreweave is losing money and owes nearly $50 billion and his bones due in 2032 now trade at 77 cents on the dollar. In other words, if you're holding $100 IOU from Coree, you can only sell it for about $77 because buyers are not sure that the full $100 is ever coming back. On Wednesday, its shares fell 15% and then another 6% after hours. And it's not only the weaker companies that we're talking about here. This is a huge deal and it's widespread. The cost of ensuring Amazon, Alphabet, Meta, and Nvidia's debt also hit records this week, although from much lower levels. Rising interest rates are part of this, of course, because they make all borrowing more expensive. But this insurance fee sits on top of interest rates. It's the extra you pay for one particular company's risk. So when it climbs, lenders are basically telling you something about that company specifically. Make of that what you will. I'm not going to give financial advice. I never will. But I'm also gonna say that you can see the same nervousness around Oracle's huge data center project in New Mexico called Project Jupiter. After problems getting power to the site, Oracle sent a notice that would let it delay payments if the site isn't up and running by 2028, although it says the project remains on our plan schedule, their words. Still, the loans tied to it, about $18 billion of them, were quoted at around 90 cents on the dollar in September. And by October 10th, they had slipped to about 85. And if that was not enough, even the people selling the chips sound uneasy. On Monday, Broadcom's chief executive, Hawk Tan, said, "A lot of investors have no clue what they're investing in." And that investing in AI was, I'm quoting here, almost like they're in a casino, which is a brave thing to say, I think, when Broadcom is one of the two companies in that $60 billion financing package that I was mentioning earlier. So lenders are lending less and charging more at the same time which points to real caution, not just companies that have already borrowed enough. Just to make it simpler, in other words, the second bet is very much on the people with the money are starting to doubt that AI will earn enough to pay it back.

One credit investor even told the Financial Times that he is unsure where the value is really acrewing in the system and that's what makes the two bets so strange. side by side. In my eyes, if AI really was taking value away from software companies like Adobe, that value should in theory be flowing to the AI companies and their lenders should actually be getting calmer, not more nervous. No. So, if the money is leaving Adobe and it isn't landing with the people building AI, then where is this money actually going? Well, the most obvious place would be the companies that make the AI models, right? Antropic told investors its revenue rose 45% in the third quarter to $16.8 billion. Beautiful number. Up from 11.6 billion in the second quarter and under $5 billion in the first. So that is real money arriving somewhere. The catch is that the growth is slowing from more than doubling in one quarter to 45% in the next. And that slowdown is part of what unsettled markets this week. I guess only an AI can growing 45% in three months count as a disappointment. But let's just ignore that for a second. Funnily enough, the free apps from earlier fit this picture as well. By the way, they cost nothing in theory, but the business behind them sells access to AI image and video models, and that's what pays for development. So, the free Photoshop is in the end paid for by selling AI. The second possibility that I'm thinking of is that the value doesn't land with any company whatsoever and just disappears into lower prices for the people using the software. The free apps cost nothing. And even Adobe says it's growing through a premium strategy, meaning a free tier it hopes people will one day upgrade from. And it now counts over a billion monthly users off the back of that. If that's where the value goes, users win and both bets lose a little bit because Adobe earns less per person and no AI company earns it instead. I guess all of us win. The third possibility is the actual uncomfortable one. The value lands with the model makers or with users and the companies stuck in the middle are left holding all of the risk. Those are the companies who borrow to build data centers and now rent out the computing power. And just as an example, because this is exactly what's happened, an analyst at SNBC called this week's fall a business modeldriven sell-off. Mind you, I have to be honest here. The demand itself is a real thing. The chipmaker TSMC posted its sixth record quarter in a row on Thursday morning with profits up 69%. That's a huge number. So, the question lenders are actually asking is in fact a narrower one. It's basically whether renting those chips out earns enough to pay back what was borrowed to build around them.

Which leaves a stranger question underneath all three. And honestly, it's probably the one that I keep coming back to. What is something worth once copying it costs almost nothing? Somebody built a Photoshop rival for about $2,000 of AI usage. If the code can be copied for that money, then whatever Adobe is worth must live somewhere else in trust, in habit, and years of people's work saved in its files. Those things are real, but they're very hard to put a price on. And if most of the value of cheap code ends up with the people using it, that might be good for almost everyone except the people who actually lend money expecting it to land somewhere else. Which brings us all the way back to the first bets that I mentioned at the start of this video. The first bet said AI is strong enough to make software worthless. From everything in this video, it is strong enough to make software cheap to copy, which is actually a completely different claim. Adobe just had its best quarter. The clone covers about a third of a professional's work by its own count, and free alternatives have been around for 30 years without breaking it. The second bet was that AI may not earn enough to repay its debts and lenders are definitely acting like it, lending less and charging more with the companies in the middle feeling it the first. But then again, the demand for chips is real and AI revenue is still strong. So honestly, I don't think either bet holds up on its own. And I think they can actually both be partly right. AI can make code cheap without making anyone rich. If the value mostly ends up as lower prices for the people using the software, yay us, Adobe loses a little, the AI companies don't gain as much as they hoped and the lender's worry turns out to be about exactly that. And personally, that's the direction I'd like things to go because technology is supposed to make useful things cheaper. Whether the people who borrow to build it can live with that is a separate question and is theirs to answer. As for where the value lands, I think that's still open. Some of it might be arriving at the model maker slowly, and some of it might actually be leaking out to everyone using the software quietly. Truth is usually somewhere in the middle. So, it's probably a bit of both. Big claims sell a lot. You all know this. And there is definitely a new one every single week. Software was declared over last week, and since then, Adobe's professional subscription are still growing. Brandon Thomas has shipped three more apps and I'm still here making these videos. So maybe software isn't over. Maybe what's over is the idea that writing the code was ever the valuable part. Honestly, software is over isn't even the biggest claim going around about AI. The biggest one yet is of course that AI is going to kill us all. So, in this video that I'm linking here on your screen, I went through the actual research on that very question to see what the evidence really supports and was just a very confident headline. If you want to know where that one lands, that's the video that I would watch.