EPISODE 1939

[0:00:12] GV: Hello, and welcome to SED News. I'm Gregor Vand. 

[0:00:16] SF: And I'm Sean Falconer. 

[0:00:18] GV: And I'm sure as many of you know, this is a different format of SE Daily, released monthly, where we touch on the main tech headlines from mainstream news. We go into a bigger topic in the middle, bit of debate in there, and then go to Hacker News at the end and highlight some of our favorite things that have been posted. As usual, we like to just catch up with each other. Sean and I usually having quite busy schedules these days. You've been traveling, Sean, I think. 

[0:00:47] SF: I feel like every time we start these off, it's like, "Oh, you've been traveling," which it seems to be the case over the last few months. Yes. We kind of schedule these around my travel schedule. But yeah, I was in London this past week for our big Europe conference. It was a great time. Saw a lot of people in this data streaming community. I was part of the keynote. Gave a talk there. It was a busy week, but a fun week. And we launched a lot of new products, my team. It was quite a sprint leading up to the event, but it was nice to see everything come together. 

[0:01:16] GV: Yeah. Nice. Well, I guess on the flip side, I've not been traveling since we last recorded SED News. I think I was in SF when we recorded. We met each other, though. You and I had lunch after that in-person. First time we've met each other. So, that was always fun. I stopped back through Hong Kong on the way back, but that was nothing work-related. That was just purely for my own enjoyment. 

Yeah, I'm back in Singapore, and I'll be here for a few weeks, which is good. I think I do need less travel based on just how much was going on in the tech world right now. My Slack is just a deluge every morning, every time I wake up with the time difference with US. It's an interesting time to be part of all of this, as I keep telling people that are not in our industry, especially. 

Given that, we are in the industry, and we do sit across quite a few things. Let's hit the main headlines. The first one is Apple. We haven't touched on Apple in a little while. Apple sort of at a crossroads. A lot of this comes from Financial Times. Did quite a good deep dive on this. But really, the headline here is: can Apple actually progress beyond - let's just say the iPhone for now. 

I think there are other facets, obviously, to their revenue generation which we can touch on briefly, but they've been this hardware business. They have gone into services. And a lot of the last almost decade at this point is really just keep this thing running at optimum efficiency. 

And actually, an interesting anecdote, which feeds back to something that we did touch on at the time when their head of AI departed. But there wasn't like a lot of information, I guess, around that exactly. But we have a bit more color on that now, which was the fact that the Siri demo basically had been fake when they had said - I think it was at the last WWDC, or like two WWDC's ago, saying, "Hey, this is the new Siri." And I think they actually put some supposed demo of it going on, and turns out that was completely fake. 

And the product hadn't been built yet. And they told their engineers, "You've got 9 weeks to get this out." And it didn't work really. And they actually got a 250 million class action lawsuit against false advertising, which I think that was the nail in the coffin for the head of AI heading out. It's just embarrassing, I mean, this whole thing. But I think it's interesting this color. But it does start to paint a picture of do Apple really know what they're doing with AI? It's kind of interesting. 

[0:03:45] SF: First of all, Salesforce has been doing vaporware demos for decade plus. And everyone kind of knows it and laughs about it, but they seem to be able to get away with it. Maybe it's because who they're selling to is not a consumer, so they don't end up with these class action lawsuits through their false advertising. But yeah, I mean, it's interesting. 

Apple hasn't, to my knowledge, ever had anything really impactful in AI. Obviously, they've had Siri for a long time, one of the first voice interfaces, and so forth that was widely used. But in a lot of ways, voice didn't really take off the way that people maybe perceived it to do. I worked on the Google Assistant at one point. It didn't really take off the way that we thought it was going to, and it didn't end up being sort of this new modality that was widely used. Same with Alexa and stuff. Like a lot of these things became glorified ways of turning on and off lights and checking the weather, and really, really simple sort of low-value tasks. 

But despite that, one of these companies that seems immune to what's happening in the AI world, where so many companies that are public, their stocks are being punished right now, even with like record-breaking quarters, because of the perception of how AI might disrupt them. But you look at like Apple stock, and it's like an all-time high. It's still going really well despite them, from my perception anyway, not really having kind of like an AI play. 

I think it would be tough, though, for them to become suddenly some sort of AI company. If you look at Google, Google always thought of itself as an AI company. Larry Page always considered Google to be an AI company. It's not shocking, even though they were slow-moving compared to OpenAI on really monetizing and doing a lot around large language models. But they're well positioned to be very, very successful there, as we've talked about before. There just feels like a real cultural difference between who Apple is at a company and how they see themselves culturally versus how a company maybe Google or any of the AI native companies see themselves. 

What would it take for Apple to become that? Probably a lot of money spent on talent and maybe acquisitions. But even then, can you actually bridge that gap? And does it make sense? Because it's not just about - there's lots of problems we're solving in AI, but how do you map that to what you're uniquely positioned to do well with is also a big question. It's interesting because Apple's such a very successful company, has been for a long time now. People have a lot of respect for it still. It does really well in the market. It's one of the most valuable companies in the world, but it's kind of hard to see what their next act is. 

[0:06:25] GV: Yeah, I think that's exactly it. I mean, this is it. Apple could absolutely decide to not lean into AI. I think they said something in their own legal filings; they had actually warned that a truly capable AI assistant could make manually downloading apps obsolete. Obviously, it's services. Their app store is still quite a big money driver. If you look at their services breakdown for 2025, it's almost 3x all their Mac sales, for example. This is just still a huge part of their business. 

And I think there was also this interesting quote from this FT article, which was back in 2010 when Jobs said he had acquired Siri, the company behind Siri, and people said, "Are you entering the search business?" And he said, "No, I'm not. I'm getting into the AI business." I think it's one of these sliding door moments is what Apple might have become had Jobs persisted as CEO. Instead, it's gone to this incredibly "well-run" hardware business with supply chain excellence under Tim Cook, who's obviously retiring now or becoming chairman or something to that effect. 

And he's played this very well financially. I put a billion sales per day, a trillion dollars apparently returned to shareholders, 75% margins on services. But yeah, that is not innovation. That's a behemoth of money-making. And their R&D spend, I think, has gone - it's gone from 8% to 2% apparently during the iPhone boom. I mean, okay, 2% of a much bigger cash pile, maybe it's exactly the same amount of money. But on relative terms, you'd expect R&D to be keeping pace from a percentage perspective. 

[0:08:09] SF: Yeah. I mean, there is probably something to - if you're focused on financial optimization, is that the enemy of innovation in some sense. Because I think innovation comes with a cost. You see that right now with the AI companies. We saw it with cloud. You see it in these early markets. You're spending a lot and not necessarily making a lot of money, for eventually reaching some economies of scale and cost optimization in the future. And that might not play well into essentially the Tim Cook playbook for where he's putting his time and resources and so forth. Maybe that's something they have to change. 

I also think that on the AI front, Apple drives a lot of revenue today from app monetization. 30%, I believe it is, that they're taking from app sales and so forth from a creator, which is responsible for a lot of revenue. But if AI disrupts apps eventually, if the entry point, sort of the face of the internet, becomes an AI app that becomes a super app, your gateway portal into, like, everything. And we've talked about this in terms of disruption too. E-commerce, for example. Going directly to an e-commerce website versus just navigating to it through a ChatGPT or something. If that becomes the interface for all applications, then that's going to hurt Apple massively in the long run. 

[0:09:32] GV: Yeah, exactly. I mean, yeah, does AI saas-pocalypse? Well, it's almost here. I guess it's app-pocalypse literally. Where is a market for all these different apps? An app that I use is called Time Shifter. It's a jetlag app. But interestingly, my other half, she, for her own reasons, doesn't want to pay for this app. And she'd rather use Claude just to give her a plan around jet lag. 

Now, I like some of the features that this time shifter app gives me. But I'm sure my other half is not in the minority that it's like, "Actually, you know what? I don't need to pay for this app. And I don't need to give, in that case, Google App Store a little cut of that one. I will happily just use my free Claude account and get a good enough plan instead." There must be tons of these examples that could really eat into Apple's App Store margins. 

[0:10:20] SF: Yeah. 

[0:10:21] GV: Yeah, it's definitely one to watch. 

[0:10:24] SF: I mean, it's been a while since - what was Apple's last hit? 

[0:10:30] GV: It wasn't the Vision Pro. We know that. Yeah.

[0:10:32] SF: I mean, the iPhone? That was 2007. 

[0:10:34] GV: Well, again, if you look at - I like the chart they put in the article, the FT article. The iPhone still looks almost 70% of revenue in 2025. It is just unreal how they've managed to coast on - I say coast. Okay, the iPhone is still, I guess, one of the most advanced handsets out there. But relative to peers, it's much hard to distinguish now between what makes it so excellent. Yeah. Hits. I'm struggling. I'm struggling a bit on hits. 

[0:11:04] SF: I mean, there's minor hits, iPad, and so forth, but nothing's - I mean, you're also comparing it to one of the mega hits of all time, the iPhone. The bar was kind of set high. 

[0:11:13] GV: Yeah. Moving on. I guess one of the other big ones, Google, they had their IO conference just recently. And yeah, I believe this was all the agentic pivot, if you like. I always think of you, Sean, as our resident Google expert. What were you thinking about this? 

[0:11:32] SF: I think one of the things that I like them highlighting was the idea of these agents that are running sort of continuously in the background. Things like a personalized agent that can set up your work in the background 24/7, sort of monitoring different systems. And this is something that I've been talking about and writing about for over a year now, which is - 

[0:11:52] GV: This is the Gemini Spark, is it? 

[0:11:55] SF: Yeah. Yeah. And this is something I've been talking about for quite some time, of we've kind of limited ourselves in terms of thinking of agents as these purely chat systems. I ask the agent to go do something, and it goes works for a bit, comes back, and gives me an answer, and so forth. But there's lots of situations where you don't really want an agent to have to wait around for you to prompt it to go do the work. You kind of want it to just be able to do the work because it knows something's going on. I'm glad that they're sort of highlighting that. And hopefully, that's something that will help educate the world that an agent doesn't just have to be locked behind chat. 

[0:12:30] GV: Yeah, just jumped. There's another headline that I wanted to pull out here. Actually, it was via PC Gamer, but it was that DuckDuckGo traffic has grown after Google dropping - sorry. When I say dropping, I mean producing their AI mode. It's hard to drop a product or dropping a product. I don't know. But Google have kind of defaulted to AI mode, and you have to kind of go looking now for the classic, if you like, Google search, which was already under a lot of stress from people saying this is just too ad-heavy. It doesn't do what it's supposed to do anymore. And DuckDuckGo, which is this alternative search engine, has actually seen a 28% increase in visits following the default to AI mode for Google, which that's pretty fascinating, I think. 

[0:13:20] SF: Yeah. I wonder how long that will last. Is this like the gut reaction to change? It kind of reminds me of when Facebook first introduced the newsfeed, and everybody was up in arms about this way back. Annoyed by it, that they kind of rolled it out without telling anybody. And suddenly, there's newsfeed. And people were upset about it, and then they kind of got over it. And now, basically, the newsfeed became social media. 

I do wonder, is this kind of a blip reaction and temporary growth to DuckDuckGo, or is this something more meaningful that will continue to gain traction? But I think that there is definitely some truth to this obstacle course of what search has become. You go, and you do a search now on Google. And you have your AI mode response, which, if you click on, takes you into a chat experience. But if you scroll past that, then you're looking at ads. And if you scroll past that, you get the organic results. It's been a long time since I've been on the second page of a Google result. 

A year ago, maybe, it wasn't unheard of for me to go multiple levels deep on a Google search result, especially if I was looking for something obscure. And part of that is I'm doing a lot less Google searches these days because I go and I ask AI for more meaningful things, and it can just give me the answer. But it's hard to see where's Google going to go with this. You kind of either lean into it, and you're just like, "Hey, we're a full AI chat experience now, like a ChatGPT." Or is this really what users want from search? 

[0:14:53] GV: Yeah, exactly. I think we did cover this in a chunkier section a few months ago in SED News, sort of where we're looking at where will search go. We're getting a bit of signal then on this one, which is when Google tries to default to, I guess, moving much closer to a - call it like a ChatGPT experience, where the bulk of the result returned is sort of an AI response but with some context. 

For example, today, I just searched for a specific company through the AI mode just to see what happened. And you get it almost like a Wikipedia-type answer, which is just isn't what I'm looking for. And I'm like, "No, I want to see -" if I'm looking in Google, let's just say I am looking for, "Oh, what are the latest Reddit posts on it?" Just give me a list. Or I want to see any news results, give me a list. I'd also just like to see what the website is, and I'll go there myself. That's why I'm using search. It is interesting how Google are just caught in this interesting spot right now where I don't think they really know actually which direction they should go in. 

[0:15:54] SF: Yeah. Well, it's tough because they are in this place where they're obviously trying to stay ahead of ChatGPT and perplexity, these types of applications. But at the same time, they're taking away from what the core value of Google search was in the first place. And then it's this push-pull, where they need to try to protect future revenue, but they're doing at the cost of the current user experience. It's kind of hard to know exactly what the end strategy is going to be. 

[0:16:26] GV: Yeah. And then for our final headline this week, this is via TechCrunch. Remote, which is a company that I think quite a few listeners might be aware of and may even be recipients of their payroll through Remote, they've been growing pretty dramatically. They are a 7-year-old company based in Amsterdam. I didn't actually know that until I read this article. I thought they were American. But they've surpassed 300 million in ARR, which is pretty exceptional. I know, compared to, say, an AI builder platform that says it's done 200 million in ARR in a year or something, but I think this is a slightly more durable business. I think people would agree. 

But what is interesting is if we look at what's been reported is their per employee revenue basically. But it's just the fact that they claim that they've pushed through this 300 million ARR, become cash flow positive, but they have not been hiring any more people. Of course, people could say this, but it sounds - this is CO saying that their headcount has stayed flat to get to this place. 

At least in theory, that proves that this is possible. I think we're also got to - certainly, if I've scat ahead to my own Hacker News highlights, it is going to be a bit of this coming in, like AI and people. Again, what does that mean? It should be more time spent at work, less time spent at work. At least I think you and I showed we're both seeing roughly the same thing in industry right now, which is more time spent at work. But okay, maybe headcount does stay flat. But that headcount is spending more time at work.

[0:17:59] SF: Yeah. You keep it flat, but double the efficiency or double the output by having everyone work 80 hours. 

[0:18:05] GV: Yeah. Exactly. Yeah. I wonder if there's, like, a sort of almost like group dynamics at play where people see other colleagues "achieving more", and it just sort of keeps - when I say raising the bar, I'm not sure I want to say that in a positive sense. It just raises the expectation, I think, is what I'm actually meaning. Raises the expectation of output. 

[0:18:26] SF: There's probably some of that. If you're on a team and everybody's working extra hours, and you're the one person who isn't, then it could potentially create some issues because your teammates might be like, "Well, why am I working 12 hours a day and Gregor's getting away with less?" It can create - 

[0:18:45] GV: If only that was true. 

[0:18:48] SF: It's some social problems within the company. But I also think that there's also an addictive component to a lot of these AI tools. And I experienced it myself, like using Claude Code or whatever your agentic engineering tool of choice is. It is addictive. It's like a swap machine. You're putting your tokens in. It's churning away. You're getting some tokens out. And you keep trying to get it to do more. And it feels like, "Hey, if I'm not using my compute cycles right now to cycle through tokens and generate code or generate something, then I'm just wasting all this compute sitting there idle on my computer." You want it running all the time. 

And it's taxing, but it's a different type of taxing than being in full focus mode in locked-in coding. What used to sort of be a requirement to really be efficiently generating code, now it's become somewhat of like an asynchronous process. And it's kind of changed the mode of behavior. And maybe that leads to you being able to do this for longer sustained periods than what we've been able to do before as well. 

There's a novelty factor. There's a lot of things I think going on that leads to the increase in people sort of spending time. And then, also, it's like a highly competitive market. People are getting laid off. There's pressure on companies to churn out more. It's got to come from somewhere. And then you have AI companies that seem to be moving incredibly fast and are very competitive against each other. Valuations are super high. There's lots of pressure all around. I think that creates some of this as well. 

[0:20:17] GV: Yeah, absolutely. And investors, I'm not entirely sure of remotes set up if they're fully VC-backed or etc. But I'm sure to get to 300 million ARR, I assume they do have investment. And they probably have investors looking quite closely at what does seem to be a bit of a new Northstar that's per employee, revenue per employee. 

I think I've mentioned this before. I still find this funny because of running a service business for 10 years. And I would always look at revenue per employee. Literally, you're effectively selling time of a human to another business. And so you have to look at revenue per employee and then compare that to, say, one of our competitors who has 5x the number of employees. But if I did the maths on their reported revenue, I was like, "Oh, we're making a lot more revenue per employee. I would rather be doing that than hiring 50 more people." But it's now become very trendy in tech to be reporting that. 

And actually, gone are the low-interest years of boasting about just how many employees you have and how much you've grown the headcount. And now it's almost the opposite. It's like boasting about keeping it flat or reducing. 

[0:21:21] SF: Yeah. I mean, that becomes the new sort of north star, I guess. 

[0:21:25] GV: Yeah. I think, for many companies, this is all just the classic tech noise that comes out in the news. And I think every company just has to make their own assessment. They shouldn't be like sheep following, just saying, "Oh, well, everyone's not hiring. So we shouldn't either," or whatever. Every company should make their own judgment on that. That's a fairly utopian view from me. 

I think just rounding out with what you were saying, Sean, about just the actual filling of time that an employee has. It is, I guess, Parkinson's laws. The time available. Whatever time is allotted will kind of always get filled. And I think we're just seeing that 100%. Where, despite the efficiencies of the tooling and enabling us to generate the output in theory faster and potentially "better", it just means we're being expected to output even more. And also, when we talk about input, everything you output, at least in where I sit, everything that I output has to taken in by somebody else in the team somewhere. 

And so we've got this debate going on right now of, like, "Well, where's the line between you can't just keep churning out documents expecting a human to keep reading them? Or is it like AI versus AI on the input and the output, or the output and the input?" 

[0:22:39] SF: Yeah, you write a few short points, and then use AI to explode it into a document. 

[0:22:44] GV: Yeah, exactly. 

[0:22:45] SF: And then you give it to somebody, and then that person puts it in AI to summarize it back into points so they can digest it. 

[0:22:51] GV: Yep. Definitely have seen that. I think that's a good place to leave that. Very interesting. We'll sort of follow along with how other - I like that this was remote and not just yet another talking about meta or something. I think it's interesting to follow these companies that a lot of them are powering - I see a lot of the same names pop up. Power a lot of other startups these days. And remote definitely has played a big part in that. Yeah, very interesting. 

Moving to our main topic. We did touch a little bit on it in the last 10 minutes, but it really is just, "Okay. Well, business models of especially coding agents." And this is based on Simon Willison's blog post. He put out something. I think it was yesterday. And for those that aren't familiar, Simon Willison, you'll see his posts pop up on Hacker News quite frequently. Not always long, but just very detailed dives into, especially at the moment, how certain models operate. He runs a bunch of experiments against them and so on. 

But he declared yesterday, I think Anthropic and OpenAI have found product market fit. And this is based on the idea that subscriptions that certainly the consumer pays has absolutely no correlation with, A, the cost to the business. But why is their product market fit? It's actually because enterprise is paying so much money for the use of these models. 

We have touched on this in the past a little bit, sort of saying this all looks a bit strange. How can a $20 a month Claude subscription in any way be covering its costs? And why does my - if I look at sort of our work subscription, those numbers look very different. And Simon Willison saying he's got the $100 a month max plan from Anthropic and the $100 a month pro from OpenAI. 

He used a token usage tool over the past 30 days, and he'd found he'd consumed 2,000 worth of credits or tokens. Just to do the math there, that's 2,000-ish worth of usage for 200. However, enterprise are paying far, far over that. And that's the business model. 

[0:25:02] SF: Yeah. I mean, essentially you get people hooked on this. It's almost like a premium model. Or you could think of it, it's open source or something like that where people, in their own projects and on their own free time, they can use something that's less expensive. But then it's a way to get network effects across the industry because anybody can try it. And then once they try it, because they have product market fit, they're hooked. And people legitimately feel like they can't do their jobs now without these tools. I feel it myself. I depend on this so much now. Companies are willing to pay a lot for the enterprise licenses and because their employees are pushing for it. 

And then additionally had the market dynamics we were talking about, there's just a lot of pressure on companies to try to generate more efficiencies and show that they're more AI-forward, and whatever it happens to be. I think there's multiple things going on where you have both a bottoms up effect of people clamoring for this because they're using it themselves, are paying for it with these cheaper licenses, and they're pushing their companies for it. But then the companies are also sort of top-down want these things. It's a very interesting dynamic, which of course creates rocket fuel for all these companies. 

And I think in a lot of ways, I don't think this was necessarily expected that this was going to be where the model companies found their true revenue driver. For a long time, a long is relative in this world, but not that long ago, it was all about models. It was all about competing for the best model. And now that's really changed to competing for sort of the best agentic harness. The true moat has a lot to do with sort of the context management of the information environment around the model rather than just the base model. 

And we've even found in our own company that we can get similar performance from some of the lower-cost models available from, say, Anthropic versus the top-tier models. In a lot of ways, the underlying models have kind of been at a performant-level good enough for almost a year since about November of 2025, I'd say, was when they kind of became good enough for tool differentiation. But the biggest change that's happened has been really the management of that information environment around coding tasks, which is completely transformed the developer experience over the last half a year or whatever. That's kind a very different competition landscape than it being about bigger model wins, which is interesting. I don't think it's something that anybody necessarily saw coming. 

I also think that the other interesting bit about this is if you thought about like OpenAI and their strategy initially was maybe, "Hey, we're going to own a sort of -" become the face of the internet, the way that Google became the front door of the internet. That's a much different business model. Google is a volume-based business at least from search and ads and stuff like that. They get a ton of searches, they got a ton of users. And they make a little bit of money off of every user, and that leads to like a gigantic business. 

But if the way that I think there's 900 million users of ChatGPT and only a single-digit percentage of them pay, that's an okay business. But that's not trillion-dollar valuation business. And I think where they're now seeing this is more of the B2B play that Anthropic, I think, really was figured that out even before OpenAI did. 

Now I think the comparison, at least from like a business model perspective, is not Anthropic, OpenAI are the new Google. They're more like the new like Salesforce or SaaS. You're selling seats and licenses. You're doing B2B enterprise sales. That's a completely different motion and completely different muscle, which is also interesting. It's kind of a deviation from maybe where people expected the market to go. 

[0:28:53] GV: Yeah. There was a lot of popped up in many news outlets of the Uber CTO saying he'd maxed out the full-year AI budget in just a few months, in 2026. That's definitely done the rounds, I think, especially in enterprise. CFOs, CTOs saying, "Exactly. Look, we've got to be super careful. This is Uber saying that they've rinsed their budget." 

I think the subtext there is I don't think Uber would be saying that if they were worried that's what had happened. I think they're kind of saying it more like this is only positive that actually we just made a miscalculation on how valuable these tools are and we happen to have spent what we predicted, but that's fine. We're just going to keep paying these companies. And I think that's where especially - we might touch on it a bit, but the IPO is probably coming up with Anthropic and SpaceX, which doesn't include Grok, for example. 

The way that Simon Willison put it sort of I think was very helpful. He says this isn't an AI failure story. It's just that a budget set in 2025, which is what happens at companies like an Uber of that size. The budget for '26 is set far in advance in '25. Basically, failed to predict how indispensable these tools will become. 

And I think just also tied into this is the model selection. And to your point, Sean, that actually the bit I'm curious about is that the models, as you say, the choice is less of a factor now. What are the splitting hairs on what's the quality of output, let's just say on Opus versus Sonnet? Of course, a lot of people say Opus is far and away better. And maybe there are some specific tasks that yes, it is. But Sonic can still get you very, very, very far as it should be able to. 

But especially in enterprise, it's very interesting. I don't think a lot of people understand the default settings on their Claude Code. Or even when they go and just use Claude interface, there'll be some nontechnical users that are effectively given access to the enterprise account. And they're just running everything through Opus and not really paying attention. Maybe the average employee doing just some knowledge work through Opus is racking up, I don't know, $200 a month of usage. But we've got developers who basically want to say, "Look, I'm only going to use the best model. Because why would I use anything less?" And then racking up thousands and thousands and thousands per themselves. 

And I think companies, they're not saying, "Don't do this. Don't use this." But they're definitely starting to say, "Hang on, we got to keep looking at this because we want to make sure that at least use a cheaper model if you can." But I'm not fully clear on what the best way to address that is, because you're kind of saying if you think you can figure out that you'll still do as good work with a cheaper model, then go ahead. But if not, it's fine. Use the expensive model. 

[0:31:39] SF: Now you're seeing more. And I actually talked to one of them today. It's like more startups that are focused on dynamic cost optimization, where they're routing prompts to the correct or less expensive LLMs based on understanding sort of which models are good at certain tasks. Or looking at the history of interactions and so forth within the company. I think there'll be more and more. There's going to be a lot of cost optimization stuff. But I feel like at some point, especially if the company with OpenAI and Anthropics are staged to become public, at some point, someone's going to pay for this. 

Basically, right now, we're in its early innovation mode. We talked about companies are under pressure. They're not kind of worried about the token cost. They're more worried about the public perception that they're a dinosaur and they're not adopting these AI tools. But suddenly, all your employees are having a $25,000 token bill per month. What's that mean to your business? You're essentially getting like some sort of AI compute tax on every head that you hire within the company that you have to factor in. 

It's almost like in the US, factoring in like, "Okay. Well, we fully load an employee. We got to pay for insurance. We got these other expenses that we have to -" now you're going to have like a line item that's like what is the token usage of a senior engineer versus, I don't know, some other role. Product marketing or something like that. You're going to have to be thinking about that as a company as well to factor this stuff in. 

I think at some point, there's going to be some impact, ultimate impact to either these companies are going to have to charge more under public. Similar to what we saw with Uber or Lyft for a while for a long time. The rideshares were sort of VC subsidized. And at some point, there's no free lunches. Someone's going to have to pay for this. And then that's going to also lead to I think a lot more cost optimization, which might start with sort of independent players and then eventually become some of these companies. 

We saw the same thing with warehouses as well. For the early days, when Snowflake was growing like crazy, people weren't really worried about sort of their warehouse bill as much. But then when the market dynamics changed and things became a lot less about growth at all costs and more around cost optimization, and margin optimization, and stuff like that, then suddenly people were started being very concerned about their gigantic Snowflake bill and other similar products. 

So then a bunch of companies and consulting services spun up around cost optimization. And eventually, Snowflake, in order to reduce the risk of churning people, built a lot of that tooling functionality actually directly into snowflake so that people had more controls of how spend gets allocated and stuff. I could certainly see things like that coming as well. But those are all like signs of a more mature market. And we're just in early, early days. So people are just kind of spending blindly. 

[0:34:34] GV: And I guess tying it back to the headline about remote saying we kept headcount flat, and it's because of AI. They obviously haven't - well, I say they obviously haven't. But we didn't see any breakout from them of, "But what are they spending on AI?" If you're spending - I'm just picking numbers out here. But let's say as a business, you have 100 employees, and then you're spending 10K a month on these tools. Well, okay, what's that? That's at least three to four potential employees you might have had otherwise. A bit less, maybe. Two or three. It depends on the company. But that probably still makes a ton of sense. 

Don't employ two to three people. Do put that kind of money into tooling for the other 100. Surely that makes sense. But as you say, Sean, we definitely haven't reached the real inflection point or the next stage in the journey, which is when do these costs start to come under scrutiny and when the cost optimization come in. 

With cloud, cloud is still the one that there's a ton of tools out there helping teams optimize their cloud bills on AWS, etc. But if you look at the numbers, AWS still just keeps growing revenue. And okay, people are much better at optimizing costs. But for example, at Supabase, we have a team of two to three who their whole job is to optimize costs on AWS. And I know that, I guess, we're a little bit unusual in that respect that that makes a lot of sense for us. But equally, that's just symptomatic of the problem. You need to have three full-time employees if you use all of AWS to make sure you're not spending money where you shouldn't need to spend it. 

[0:36:10] SF: Yeah, I think companies probably reach a similar spot with model and token cost, or use and stuff like that. What do you think all this might mean for the open-weight models? The Llamas of the world essentially. If the real value prop or moat is sort of this like agentic harness and the way you manage the context and information environment, and not necessarily just like the model, then what's the play for the open-weight models? Do they have to go and find a similar sort of business to go after? Or can they survive just as a model alone? 

[0:36:45] GV: Yeah. Well, I guess to plug an episode, I spoke to Benny at Fireworks, fireworks.ai. They run open weight models. And e.g., they run an open-weight model for cursor. Obviously, as people that I think that know use Cursor, it's not all their options when you ask it to do something. There's an auto mode, which, yes, does run on the sort of tuned open-weight models. And then you can of course choose your specific non-open-weight models. I think it's going to be a really interesting space, because I think when more companies realize they could probably get the same outputs, but then rather than paying a SaaS bill to just straight to Anthropic, what if you set up your own - again, call it like cloud instance on something like a Fireworks, where you say, "Hey, we know that this open-weight model actually would be the best allrounder for the kind of work we do. And we would end up spending 20x less over a year if you guys just run it for us." I think that's a very interesting space. 

[0:37:51] SF: Yeah. Yeah. And the other thing too is the total TAM of the space is gigantic. There's going to probably be some model winner or a collection of model winners that kind of own the open-weight model inference market. 

[0:38:07] GV: Yeah. And at least from what I understand at present day, I think the thing that's interesting about open-weight models is that just the tense - especially the bigger more general ones, the majority are kind of coming out of China, basically. Llama, yes, is a stand out from US. But how much more effort does Meta put into that? I'm not totally sure. It's definitely a big play. And all of these models coming from China, which underpin a lot of some of the offerings from, say, Fireworks. And we actually talk about that on the episode. I think there's just a lot to keep an eye on there. And maybe, actually that's a topic for a future SED News, where we sort of look at how the open-weight models are doing competing. 

Hopefully, thanks to Simon Willison for his very good article on that. I think it's very present, as we say, looking at IPOs coming up. We expect to see probably the S1 document, which is what gets produced before an IPO to attract investors. We expect to see that from Anthropic probably in the next few weeks, maybe months. 

Why is that interesting? Because that should have numbers break down. What is the actual revenue coming in from - maybe they won't break it down precisely, enterprise versus consumer. But we'll get to start to see at least what is the revenue coming in from usage. And what are the costs involved with running such services? That'll be the SpaceX one. We don't have time to dig into that one today. But the problem with that one is it's a lot of financial chicanery. It's a bit harder to dig into that one at least from the AI perspective of like what does it actually cost to run. And how much revenue are they making? 

Just a sidebar is apparently the Starlink is actually the most profitable piece of SpaceX. Or the only profitable piece of SpaceX is actually Starlink. That's already an indicator of that one. But yeah, definitely don't have time dig into that crazy S one today, but maybe another time. 

It's that time of the show. Hacker news highlights. Some of our favorites that have popped up. Oh, I see Doom. Doom's in there again. I love it. 

[0:40:15] SF: Yeah. Yeah. I couldn't resist. You always got to highlight Doom on everything. But yeah, this was running Doom on a travel router with touch. Essentially, it's called a GL.Net's Slate 7 Pro travel router. It's like a networking device with a 2.8 in touchscreen. Apparently, it runs some version of Linux. And it has root SSH access. To get Doom on there I guess was relatively straightforward. 

But if you dig in the comments, there's some pretty funny ones. Someone claims that they actually did this like 10 months previously. But I think one was there's kind of like an earnest thread about whether humanity should have picked a less violent game for this tradition of porting this one game across every type of device. Should it be Doom, or should it be something that's maybe a little less silent? 

[0:41:05] GV: Maybe someone needs to kick off like a Mario or something. Mario on something. 

[0:41:10] SF: Yeah, like the original Mario. 

[0:41:12] GV: Yeah, exactly. This didn't get very high on Hacker News, but it did the rounds, and it's related. There's now something called Doombench. It's like can your stack run Doom? It's a performance analyzer, as it sounds. Can your stack run Doom? Check that out as well if you're thinking of trying to port Doom onto something. But yeah, I love the travel router with the little screen. That's very fun. 

My first one is this was at the top of Hacker News before we were recording today. It's literally just called 'Can We Have the Day Off'. And I thought this was a very timely. It's quite a short, little article blog post by somebody called Mike. And this was posted by a user code MLSU. It's basically just sort of, "Well, if we're all getting 10x productivity from AI, can I just take Friday off now? Is that okay?" I think probably a lot of us are feeling that at times. And this person also talks about they're paying $6,000 a month for childcare in California. Can I not just go into the office for four days and save some money on my childcare as well? Yeah, I think I like these little kind of zeitgeist articles that do actually make it up where I think it's usually a lot of what we're thinking in our own lives, then someone just says it. 

[0:42:30] SF: Yeah. I think we talked about this throughout the episode. But I think there's probably a lot of people feeling like they're all running sprints right now. And how long can you sustain the sprint? 

[0:42:40] GV: Yeah. A marathon of sprints right now. Exactly. 

[0:42:42] SF: Yeah. Exactly. Yeah, it's a never-ending sprint. The other one I had was YouTube is automatically labeling AI-generated videos now. Before, I think things were basically like an honor system. A creator could say that was AI-generated, and now they switching to automatic. they're going to detect sort of photo realistic AI-generated content, auto apply labels, even if the creators don't disclose it. 

It makes a lot of sense to me. I think we talked a lot about this, particularly with social media. If all the social media is generated by AI, and then you're using AI to kind of optimally show it to somebody and stuff like that. At some point, who's this for? At least I don't think they're changing the serving of it. They'll still serve you AI-generated content, but they're at least letting you know that it's AI-generated content. And I'm sure maybe there'll be more controls or something like that over time. But at some point, you're going to end up with more content, not just on YouTube, but more content that exists on the internet that's been generated by AI than generated by humans. And that may or may not be majorly problematic. 

[0:43:50] GV: Yeah, that's interesting. I would like to see that on Instagram. I don't spend a ton of time on Instagram these days, but it is still where I probably off any network where I go to just see what various people I do know are up to, as well as just random people I follow. But the AI thing is annoying, because I see a bunch of stuff. And I do have to kind of go to the comments to check if people say, "Obviously, this is AI." 

I'm into bird watching, for example, and there was a period when all these sort of AI birds started appearing. And I'm like, "That's not a real bird." Just voting it down. I would just like a little label, and I can just filter out anything within certain topics that is AI. That would be nice. 

[0:44:28] SF: Yeah, I know Dr. Christian Hubicki, who's been a guest on the podcast a couple times. And I don't know if he's still doing this, but for a while on Twitter, he would debunk fake robotics videos. Because people would always show some robot doing something ridiculous and stuff like that. And he would post his takes on it like of why this is like a fake video and things like that. 

[0:44:49] GV: Yeah, especially robotics. I guess you sort of get into the physics of it as well. There's no robot that could possibly do this or something to that - 

[0:44:56] SF: Yeah, exactly. 

[0:44:57] GV: Yeah. Cool. My second one is I feel it's like quite vanilla, but I quite like it. It's um SimCity 3000 in 4K. Someone that has uh put together a nice port of what it sounds like, SimCity 3000. And you go to GOG, which has game exe's that you can grab. And you grab that, and you do a bunch of tweaks that this person has put together. And then it runs in 4K, which is super nice. You run it on your nice wide screens, etc. 

Interestingly, I always like this genre, but I ended up playing what's a game called Cities: Skylines. I've never quite understood why it's not City Skylines. But I believe it's like a Finnish developer, so it ended up being called Cities: Skylines. But it was an excellent sort of - The first one wasn't 4K. I believe there's now a second, Cities: Skylines 2, which is 4K. But then it is nice if we can go back and play these classics that we're so used to, but just with uprated graphics. 

A sort of sidebar, I did end up downloading Metal Gear Solid for the PS4, the original. And they actually have two modes. You can run it in original graphics mode and improved graphics mode. And I've got to say, I actually almost prefer the original. It's like this kind of fun, blocky experience that reminds you of how PlayStation used to be. Each to their own on whether they want to uprate, see their classics uprated or just play it as they were. 

[0:46:22] SF: Yeah. Yeah. 

[0:46:23] GV: So then, yeah, just any thoughts for what we might see across the next month when we check in next? 

[0:46:29] SF: The only thing that jumps to mind is - and I think maybe next month is probably too fast. But I do think that we're going to see a lot more around token cost optimization. I think maybe you look ahead 6 months, I think that's where we'll start to see that, where companies will become either more concerned about it. Or we're going to see a lot more sort of startups coming on the scene that will help you sort of manage your token cost and token bill that you're now spending on Claude. 

Because, at some point, can you really justify, I don't know, a $25,000 or whatever it is per month bill per head of everybody in your company? Maybe, maybe not. I guess it depends on what the efficiency gain and the ROI is on that. But there's a pretty high bar for getting ROI from everybody at that level. 

[0:47:16] GV: Yeah, I guess sort of related, but maybe I'm taking a slightly alternate take. And I probably ahead of my time at this one, but S1 an S1 from Anthropic, I think we might see that. And if we do see it, I'm going to stick my neck out and say people will be very surprised positively at how much money they're making and say that this is like the next Microsoft or even the next Nvidia to some degree, sort of. Yes. The R&D costs have been massive. But this is clearly going to be where all enterprise money ends up for the next 10 years. 

[0:47:48] SF: Mm-hmm. Yeah. 

[0:47:49] GV: Yeah. Let's see what happens. As usual, thank you everybody for tuning in. We've had a few notes from various people that SED News is what they look out for on a monthly basis. So, we really appreciate that. Both Sean and I, I think, have had separate anecdotes from people we've met. So, anything you would like to see covered, always do just drop us a message on various socials. And you can find SE Daily on LinkedIn, X, and all the usual places. Do give us feedback. We love it. 

[0:48:16] SF: Yeah, absolutely. Feedbacks are welcome. 

[0:48:18] GV: Yeah. Cool. All right. Well, until next time. We'll catch everybody on next month's SED News. 

[0:48:23] SF: Thanks everyone.

[END]

SED 1939		Transcript

	(c) 2026 Software Engineering Daily	1