The Claude Code Problem

0:00Today on the AI Daily Brief, the clawed code problem. Before that in the headlines, is the White House about to take a stake in Intel? The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:32KPMG series you can with AI that I am the host of. It's a seven-part series that is really designed as a primer for those who are still just wrapping their heads around enterprise AI. You can check it out at www.kpmg.us slash AI podcasts. Welcome back to the AI Daily Brief Headlines edition, all the daily AI news you need in around five minutes. Well, here's an interesting one. The Trump administration is in talks to have the US government take a stake in Intel. Bloomberg sources said the deal is intended to shore up Intel's planned factory hub in Ohio. Intel had previously pledged to turn the site into the world's largest chipmaking facility and were provisioned with government subsidies out of the Biden administration's Chips Act to make it happen.

1:11However, slowing sales over recent years has meant the facility made less and less sense and no meaningful progress has been made. Earlier this year, reports stated that the administration was trying to put together a TSMC-led joint venture to take over parts of Intel's operations. In fact, those reports have been popping up all year, but they've always been very murky as to exactly what the state of the negotiations were. Still, it seemed pretty clear that TSMC was being dragged to the table and wasn't particularly keen to take on Intel's problems. But then last week, all of this got even louder when Trump called on Intel CEO Liputin to resign.

1:42He posted, The CEO of Intel is highly conflicted, all caps, obviously, and must resign immediately. There is no other solution to this problem. Now, there were no further details, but the outburst was likely related to allegations leveled by Senator Tom Cotton earlier in the week. Cotton wrote a letter to the Intel board asking about Tan's previous CEO role at a chip design firm that contracted with the Chinese government. In July, that firm agreed to a plea deal with the U.S. government to settle claims that it illegally exported products to China between 2015 and 2021 when Tan was CEO. By the end of last week, it emerged that Tan had been locked in a power struggle with Intel's board.

2:15The Wall Street Journal reported that some board members were pushing for Intel to stop manufacturing their own chips. It seemed that Tan was at odds with Chairman Frank Uri on the issue almost immediately after he was appointed CEO back in March. Uri, a former investment banker, had already drawn up plans to sell off Intel's manufacturing business. Tan, meanwhile, wanted to conduct a multi-billion dollar capital raise to fund a restructuring of the business, kicking off in July. Uri and other board members reportedly blocked the fundraising as they wanted to move more slowly. All of this is to say that the intrigue has been building for months, but there has been very little clarity on how things were going to work out for Intel.

2:47Now it appears that the Trump administration views Intel's manufacturing capabilities as a national security issue. And yet, neither the Intel board nor outside investors seem willing to put up the capital to turn around the loss-making division. The solution of a TSMC joint venture was always messy at best, with the White House insisting they could only take a 49 % share to ensure it wasn't majority foreign-owned. And it was with all of that context that new reports came out suggesting the administration has discussed the prospect of taking a stake in Intel, specifically doing so in a meeting with Tan on Monday.

3:16At this stage, we basically have no details on how a deal would be structured, and the reporting stated that the talks are still in early stages. The WSJ painted the news as the latest in a string of unconventional private sector interventions from Trump, obviously referring, among other things, to the unprecedented 15 % export charges on NVIDIA and AMD products bound for China. Now, if the Wall Street Journal was sort of indicating the irregularity of this engagement, the New York Times was much less subtle in their critique, blaring the headline that Trump has made himself commander-in-chief of the chip industry.

3:43Meanwhile, if you go on Twitter slash X, you will variously find accusations of this being an example of fascism or, on the other hand, socialism, just to give you a sense of how it's all going. Luke Groman had a more measured take, trying to suss out why people are reacting differently to this than to some previous government interventions. He writes, October 2008, the US government invests$250 billion into failing US banks and Wall Street cheers. August 2025, the US government invests into US defense industrial base that is losing to China and that just got outproduced 4-1 in Ukraine by Russia and Wall Street booze.

4:15Why the difference in reaction? Now, I have to say before we move on that the other hilarious-to-me sub-story is that you'll remember that we talked about earlier in the week situational awareness. The hedge fund from 23-year-old former OpenAI staffer Leopold Aschenbrenner, who had raised a billion and a half with no previous financial experience and was sitting on something like 47 % gains for the year. Ever since that article came out, people have been absolutely harping on the dude for his intel calls. That's right, for his Intel calls. Buko Capital this morning posted a video of Cam 'ron freestyling with a pile of money in his hands and said this is Leopold Aschenbrenner today after being clowned on literally this weekend for being neck-teep in Intel calls.

4:54Wild times, man. Wild times. Now, speaking of U.S.-China chip geopolitics, the Financial Times reports that the delay of DeepSeek's new model has been caused by a failed training run on Huawei chips. The report suggests that Chinese authorities pushed DeepSeq to adopt Huawei's Ascend chips rather than use NVIDIA training clusters for the follow-up to their viral R1 model. Sources said that the startup encountered persistent technical issues during the training process, leading them to ultimately use NVIDIA chips for training and Huawei for inference. These issues say the reports were the main reason the model's launch was delayed back in May.

5:26Industry insiders in China said that the Chinese chips suffer from stability issues, slower interchip connectivity, and inferior software compared to NVIDIA GPUs. Huawei reportedly had a team on site during DeepSeek's training runs, but they still couldn't conduct a successful run on the Ascend chips. Obviously, if accurate, this sheds new light on Chinese authorities' recent push to force domestic firms to use Huawei chips. If the chips aren't capable of completing advanced training runs, then obviously that's going to be a huge problem for the advancement of Chinese AI. It also implies that the Chinese government is pushing their companies to use inferior or non-functional chips in pursuit of nationalist or protectionist goals.

5:59Now the Chinese bots on X are out in force calling this fake news, suggesting if anything to me that it's probably more accurate than not. This is now becoming an issue in financial reporting as well, with the president of Chinese tech giant Tencent saying that they have enough AI chips for training. During their earnings call yesterday, Martin Lau told analysts, We don't really have a definitive answer on the import situation of the U.S. chips. I think there's a lot of discussion between the two governments, and we're waiting to see exactly what comes out of that. But from our perspective, we do have enough chips for training and continuous upgrade of our existing models, and we also have many options for inference chips, and we are also executing a lot of software improvement and upgrade in order to drive efficiency gains in inference so that we can actually put more workload on the same number of chips.

6:37If you've been paying close attention to this show all week, you'll know that there's been a back and forth between the White House finally saying, yes, it's okay to export these H20s, with a 15 % take rate, of course, and then China saying, no, don't use those. So all of this is up in the air. The intrigue continues. But before we move to the main episode, let's bring it back to some fundraising news. One fairly big and kind of unexpected one comes from Cohear. The company announced a half-billion dollar raise at a$6.8 billion valuation. Now, on the one hand, this is a fairly modest jump, at least by crazy AI standards, from the$5.5 billion they achieved during their last round a little over a year ago.

7:10At the same time, the round was oversubscribed, and it seems like Cohear were able to take their pick of big strategic names. In addition to the fresh funding, Cohear added Joel Pinnau as chief AI officer. Pinau left Meta in April, where she was the VP of AI Research and a day-to-day leader at their FAIR research lab. The company also added Francois Chadwick as CFO. Chadwick was involved in Uber for 10 years, including serving as acting CFO in 2017. Now, what makes all this notable is that this indicates that Cohere's second act really has some legs. The Canadian startup was originally founded in 2019 and was initially involved in the Foundation Model and Chatbot competition.

7:45However, over the course of the last year or so, the company pivoted hard towards deployment of on-premise AI for enterprises. The headline of their blog post reinforced that this is the approach, saying Cohear raises$500 million at$6.8 billion valuation to accelerate enterprise efficiency with Agentic AI. The new capital, they write, will enable us to accelerate our efforts to make businesses and governments around the world vastly more efficient, simplifying tedious tasks through Agentic AI solutions. We aim to free up people to spend their time on the interesting, challenging, and human parts of work, all while prioritizing data security.

8:15This represents a security-first category of enterprise AI that is simply not being met by repurposed consumer models. I actually think that Cohear have found a really important and valuable niche here. The foundation model companies do not have the time, bandwidth, or appetite, frankly, to go in and do the sort of customization work and implementation work that enterprises require in many circumstances. At the same time, the GSIs who are doing that work don't have anywhere near the type of talent and actual technical capacity to go do that in any meaningful way as well. Enter, cohere, and there's probably room for a couple other players in that space as well, but congrats to them.

8:51It is no mean feat to transition a company at that stage this successfully. Lastly today, it appears that Cognition AI is also raising a big new round. The Wall Street Journal reports that the startup has raised$500 million at a$9.8 billion valuation. Reportedly, Founders Fund is leading the round, which more than doubles the company's valuation from back in March. Now, Cognition most recently made a splash by acquiring the remains of Windsurf after Google Aqua hired their founder. And between that IP and this new round of money, Cognition is now armed with a ton of fresh capital to make a serious run to disrupting the still nascent AI coding space.

9:23Now, speaking of the nascent AI coding space, that is the subject of our main episode. So let's turn to that now. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Next. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task.

9:54Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native STLC into their org. Blitzy is providing a limited-time, 30-day free proof-of-concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press Book Demo to learn how Blitzy transforms your STLC from AI-assisted to AI-native.

10:29That's blitzy.com. As a founder, you're moving fast towards product market fit, your next round, or your first big enterprise deal. But with AI accelerating how quickly startups build and ship, security expectations are higher earlier than ever. Getting security and compliance right can unlock growth or stall it if you wait too long. With deep integrations and automated workflows built for fast-moving teams, Vanta gets you audit-ready fast and keeps you secure with continuous monitoring as your models, infra, and customers evolve. Fast-growing customers like Langchain, Writer, and Cursor trusted Vanta to build a scalable foundation from the start.

11:06And look, as someone who lives in the world of enterprise procurement, I love how Vanta makes it easy to get compliance right. The last thing you need when you're trying to win that big deal is to have it scuttled by something that Vanta has solved for over 10 ,000 companies. Go to vanta.com slash NLW to save$1 ,000 today through the Vanta for Startups program and join over 10 ,000 ambitious companies already scaling with Vanta. That's V-A-N-T-A dot com slash NLW to save$1 ,000 for a limited time. If you are a regular listener, you will have heard about Superintelligent's agent readiness audits at this point.

11:41But I wanted to tell you today about the full suite of agent readiness products that go beyond just the initial readiness report. Over the last six months, Superintelligent has built out an entire agent planning suite. We help you move from discovery to planning to implementation. After you've completed your agent readiness audits, we help you double-click on your most important use cases with what we call our use case planning reports. These reports are going to help you understand what sort of technical preparation you need to do to be ready for a use case, what challenges you might face in implementation, and whether you should be thinking about building, buying, partnering, or some combination.

12:16After that, you can even get a spec document in what we call our technical blueprint that gives either your developers or the developers of the partner you work with what they need to build exactly the agent that you're looking for. If you want to learn more about Super Intelligence Agent Planning Suite, we built a custom GPT to answer your questions. Just go to bit.ly slash super super agent. That's bit.ly slash super super agent, all one word. And if you have any questions, the agent can even help you book an appointment with our team. Welcome back to the AI Daily Brief. There has been an increasingly loud conversation happening in AI world about the business model that underlines AI coding.

12:57Investor Chris Pike recently called this the cursor problem, or more specifically, he titled a blog post published as a Google Doc, as he does. But in conversations with my friend Sean, better known as Swix from Latent Space, he thought a more apt name for this was the Claude Code problem, and I tend to agree. Now, just to give you a little bit of an example, Riley Brown really sums it up without even meaning to. when he tweeted late last night, I hate how good 4.1 Opus is. So expensive, but so good. It just one-shots full apps if you tell it what you want, which is harder than it sounds. Now, Riley should know, not only is he one of the most prominent creators around Vibe Coding, they just announced that their mobile Vibe Coding app raised over 9 million bucks in a seed, meaning they've spent a lot of time trying to figure out how to get the best performance out of these different models.

13:41Now, I have touched on this particular problem in an episode sometime in the last week, but it was specifically in the context of this post from Antonio Garcia Martinez, who, as I mentioned then, I like and think is very smart, but basically my argument was that he was drawing a too lazy comparison. He wrote, Every tech bubble is initially pumped by some extraneous source of liquidity poured into unsustainable growth. Web2 consumer use paid for VC ads to pump MAUs. Crypto used retail tokens to pump user rewards to inflate usage and hence the token. AI is using VC and inflated equity to subsidize compute costs and inflate consumer usage.

14:13Now, he was talking broadly not just about AI coding, but the post that he was referring to is one which we'll read in a minute and was specifically talking about AI coding. Now, my argument was not that there wasn't a misalignment in the business model where people were being charged less than the cost of the actual services that they were using. My argument was that there was a fundamental difference in the demand profile of these industries. The implication of his comparison is that if you took away the subsidy, people would not use nearly as much of the service. I think that that's not true with coding.

14:43In fact, I think that the problem is that it turns out that the appetite and demand for AI coding is effectively unlimited. It is not just a one-to-one transition from people who are already coding. It is unlocking and opening a new market which I believe will be as ubiquitous as word processing in the future. But I'm getting a little bit ahead of myself here. The thing we're going to explore today is this Claude Code problem, the mismatch between what people are paying for AI coding and what it's actually costing companies, with an eye to understanding what it says about where we are in the AI cycle and just how worried we should be about that mismatch.

15:17The conversation started last week when we got a string of reporting suggesting that AI coding startups, as violently high growth as they were, were actually facing some financial problems. The information got their hands on Replit's financials, which showed their gross margins had fallen from 36 % in February to negative 14 % in April. They bounced back a little bit since then, and CEO Amjad Massad insisted that unit economics are not a problem. The information reported that similar issues were happening at Lovable, while TechCrunch added that Cursor and Windsurf were also facing cost pressures.

15:45Nicolas Charrière, the founder of Mocha, which is a vibe-coding startup and backend hosting solution, commented in the article that, quote, margins on all of the cogen products are either neutral or negative. They're absolutely abysmal. Now, at core, the issue here is, in fact, not even the margin on paid users. Now, many people think about this issue as vibe coding apps subsidizing tokens for their heaviest users. At the end of July, for example, when Anthropic rolled out new weekly rate limits, they pointed to a very small number of users who would be impacted. They pointed to their biggest power users as causing the problem, saying, Some of the biggest Claude Code fans are running it continuously in the background 24-7.

16:23These uses are remarkable and we want to enable them, but a few outlying cases are very costly to support. For example, one user consumed tens of thousands in model usage on a$200 plan. And when this story came out, one of the things that we discussed on this show was how part of the challenge when it comes specifically to AI coding is that it's sort of a the more you use it, the more you use it situation. Where especially with the rise of background agents, not only are people accomplishing more with AI coding tools, they're doing so in a way that consumes just a radically higher amount of tokens.

16:54This was, it appeared, running headlong into the business model. However, there's a whole separate dimension of this, which I think is actually wildly under-discussed, which is not about the power users, but about the fact that the very small number of paid users have to subsidize all of the free usage. And when only a tiny fraction of people are paying for a tool, that puts a lot of pressure on those paid users. So let's talk about how Pace Capital investor Chris Pike thinks about the problem. He basically argues that there are two separate things that founders need to be thinking about, product market fit and business model product fit.

17:28Product market fit is the one we all know. He defines it as users repeatedly choosing your product. Business model product fit, on the other hand, is defined by the extraction of value being sustainably in excess and proportional to the cost of delivering value. Now, before we get into the rest of his explanation about this as relating to the coding tools, it's important to note that this is not a phenomenon that's limited to the AI era. In April, for example, The Atlantic wrote about how millennials got cheap Ubers, cheap on-demand delivery. I can certainly validate this. I was there in San Francisco as this was all happening.

17:57And none of the stuff that I had access to back then is available today for anywhere near the same price. And they made the explicit connection between this and cheap AI tokens now. The question is, of course, what happens when the subsidy ends? And that's really a lot of what Chris is talking to. Chris writes, Cursor has relied on a subscription model that historically allowed for unlimited use. That's a fixed revenue variable cost setup. Without actual discipline, pricing, segmentation, caps, and exclusions, this type of model drifts into the same ditch that killed MoviePass, Oyster, and forced ClassPass to retire unlimited.

18:26The pathologies rhyme. First, cohorts invert. Your most profitable users churn because they use the product the least and get the least value. Oftentimes, they can get better value from a competitor who prices them more accurately or with less breakage. The remaining users are those who extract more value than they pay. Over time, older cohorts morph towards deeply negative gross margins. Then top line masks rot. New, larger cohorts can briefly offset the drag, hiding the deterioration in early cohorts. Revenue grows, margin quality quietly decays. Chris says that the lesson is if there's an operational path to positive margins and future pricing power, temporary subsidies can be a bridge.

19:01He also argues that venture capital is precisely the right instrument to facilitate these kinds of companies. The question he has is, does this bridge exist for Cursor? And the challenge he points out is that Cursor doesn't control two critical parts of its cost of goods sold. It doesn't control model performance frontier, i.e. what quality model its users will demand, and it doesn't control model input and output pricing, which is what Cursor pays to OpenAI, Anthropic, etc. He continues, If Cursor steps down to cheaper, weaker models, the users who care about performance will notice and churn.

19:32Those who can tolerate weaker models can get them cheaper elsewhere. If it stays at the frontier while keeping prices flat, the variable real cost to service their heaviest users will explode. In an effort to combat this, Cursor has been forced to raise prices and institute usage caps leading to user outrage and churn. Any time Unlimited shows up in a variable cost business, product market fit becomes a permanently open question. Are users here for the product or for the subsidy? Would they still use as much or at all at true marginal cost? Until Cursor prices consumption in proportion to cost, it cannot know.

20:02So that's the discussion that Chris brought up, which I think is a really interesting and important one. There are a few things confounding this, though. There is an entire additional dimension to the cursor problem or the Claude Code problem that was not present with these other examples. And that is that, one, the quality of the goods sold is rising dramatically, and two, the cost of the goods sold is coming down precipitously. The cost of inference with AI and just the general token cost has come down at a rate that absolutely no one anticipated. The problem and the reason we're still having this conversation is that demand just grows even faster.

20:34However, the other part of this is really interesting to me, which is this question of model quality. I think Chris is right to point out that Cursor and any other company in this space doesn't control model performance frontier or what quality of model users demand. And so far, what's clear, at least according to studies like Menlo Ventures' mid-year LLM market update, is that people are not switching between models because of price considerations. They are entirely focused on getting better performance. In other words, right now, all indications suggest that as much as users might complain and hem and haw and squeak and squawk on Twitter in the short term, they're going to pay what anthropic charges.

21:09Now, part of what makes that interesting, though, is that we are only just on the other side of where these models are good enough to actually be in production workflows for a lot of coding tasks. In fact, for some coding tasks, they're still not good enough. If, and this is of course a big if, the rate of performance continues to increase. I wonder what the situation will be in a year. Right now, it's very clear that at least the core base of users that exist right now for these coding tools want the highest performing models, period, even if they're more expensive. In a year, when today's state-of-the-art models are actually a bit older and incredibly cheap in comparison to whatever the state-of-the-art is then, will people who are using this for incredibly large enterprise-grade workloads be willing to use the models that are today state-of-the-art, but in the future will not be state-of-the-art?

21:55Or will they always just want whatever the newest model is? I, of course, can't answer that with any sort of assuredness, but I do think that just using today as a snapshot in time doesn't give us a full picture because of the fact that we have only just hit this frontier where these models are actually good enough. Now that we are operating entirely in the context of all of these models being fairly high performance, how is that going to change in a year's time? In the meantime, what I think is for sure is that we are going to see lots and lots of pricing experiments. We've already seen shifts among some of the leading competitors right now.