The Big Questions That Will Decide the Consumer AI War

0:00Today on the AI Daily Brief, the big question shaping the battle for consumer AI. And before that, in the headlines, is OpenAI the new GitHub? The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:34dailybrief.ai. Lastly, two other quick things to flag. First, thank you to everyone who has taken our February AI usage pulse survey. You can find a link to that at ai dailybrief.ai. And I would so appreciate it if you would take just a couple minutes to do that. Anyone who does will get the results before everyone else and help us better share data about where users actually are and their behavior patterns right now. And if you are a company who is interested in building agent teams, registration is live for Enterprise Claw at enterpriseclaw.ai and will close on Friday. Now with that out of the way, let's dive into the headlines.

1:07Back in December of last year, Mitchell Hashimoto tweeted, The AI companies are on track to become GitHub faster than GitHub is becoming an AI company. A lot of folks agreed, although some, like Ivan Barazin, had thoughts on who it might be. Ivan writes, Been looking for who will do this for a while. Bearish that it will be OpenAI, though. And yet, yesterday we got this report from The Information that OpenAI is developing an internal alternative to GitHub. According to the information sources, the project was spurred by a rise in outages for Microsoft's code repository platform. OpenAI engineers complained that these outages have stopped work for minutes or even hours at a time.

1:45GitHub had 37 outages in February, which was up dramatically from an average of 17 per month last year. Microsoft has attributed these outages to human error and problems with Azure during a multi-year migration project away from GitHub's proprietary servers. Now, sources for the OpenAI project did say that it's in its early stages and likely won't be completed for months. They also noted that the project is intended for internal use first and foremost, but then again, so was Claude Code. This also isn't the only project to rebuild GitHub for the agentic era. That was also the pitch for the new startup from former GitHub CEO Thomas Domke when he left Microsoft earlier this year.

2:19Domke's idea was the integration of agentic code review tools to help close the loop on fully autonomous code generation. Now, there are a lot of people who are trying to put different lenses on this. For some, it's the latest example of OpenAI competing with Microsoft as the rift between the two companies expands. Others see it as part of the SaaSpocalypse theme of companies canceling their software subscription in favor of Vibe-coded alternatives. I'm not sure any of that's true. Feels to me like it might just be the start of an inevitable shift in this category given how much code is pumping through these companies' coffers.

2:47As Emea puts it, the interesting play is not just hosting code, it's owning the layer that understands how the code connects across services and teams. That's where agents actually need to operate. Next up, we move over to Meta, who has formed a new applied AI engineering organization. According to a memo viewed by the Wall Street Journal, the new organization will work closely with both AR and VR organization reality labs, as well as the Meta Superintelligence Lab. Now, this doesn't seem to be another prod restructuring of AI at Meta, which by some counts went through four reshufflings last year.

3:19Instead, it appears to be aimed at filling gaps between hardware, tooling, and model teams. The memo said that the goal was to strengthen meta-AI initiatives, commenting that the team will build the quote data engine that helps our models get better faster. The new org has an unusually flat structure. It consists of two teams of 50 people each reporting into a single manager. One team will work on building interfaces and internal tooling, while the other works on data collection and refinement. The flattened team mirrors the structure of TBD Labs, which consists of around 50 highly paid AI researchers working under new AI CEO Alexander Wang within the broader superintelligence org.

3:52It also seems to reflect Mark Zuckerberg's new management philosophy that he outlined on Meta's most recent earnings call. He said that individual contributors are being elevated now that AI has allowed, in his words, projects that used to require big teams now can be accomplished by a single very talented person. Over in Amazon land, that company is exploring the prospect of building technology to power AI advertising. According to the information, In addition, Amazon's ad business has held discussions over recent months with major websites and ad sales firms about the idea. The plan would involve placing ads in chatbots and agents.

4:24One of the websites mentioned as a focus of the pitch was Pinterest, which is in the middle of an AI overhaul. In October, Pinterest launched an AI shopping recommendation assistant that helps users track down clothing featured on the website. You can see how this could be a natural fit for high intent traffic. Now, one of the things that people don't really know about Amazon or don't really think about much, is how big its ad business actually is. Last year, Amazon generated $68.6 billion in ad revenue. And while that represents only a tenth of their overall business, it was their fastest growing division, achieving 22 % growth last year.

4:56As advertising comes to the AI platforms, there could very easily be a land grab around who gets to host the clearinghouse. Now what consumers are going to think about all these AI ads remains to be seen and is part of the conversation that we're having in the main episode. Over in AI politics and chips, U.S. officials are considering a cap on NVIDIA chip sales into China in a bid to constrain the power of training clusters. Bloomberg reports that U.S. trade officials are considering a cap of 75 ,000 chips per customer. Sources said the cap would apply to the newly approved NVIDIA H200 chips, as well as AMD's MI325 AI chips.

5:29They noted that chip supply would also be contained to a million total units sold into China, a limit that was set earlier in the regulatory process, but up to now hasn't previously been reported. The million-unit limit is reportedly far lower than the number NVIDIA originally proposed, which gives some additional context to recent comments from Commerce Secretary Howard Lutnick. During congressional testimony last month, Lutnick said that NVIDIA must live with the license term set by the government, and presumably this is what he meant. The 75 ,000 chip cap is also less than half the number sought by Chinese tech giants Alibaba, Tencent, and ByteDance.

6:00Each had reportedly told NVIDIA that they would like chip counts of around 200 ,000 to build their large-scale training clusters. Within these limits, each company will only be able to build data centers using around 100 megawatts of power. That's a far smaller scale than the multi-gigawatt training clusters that are planned by Western AI Labs, and not even a match for XAI's original build-out of the Colossus megacluster last March, which began at 100 ,000 GPUs and quickly scaled to 200 ,000, and is now reportedly at 550 ,000 units. The big question is whether this is a meaningful constraint or simply window dressing to appease China hawks in Washington.

6:32What's more, the entire process is still murky and getting even murkier due to the Iran war, considering that China is a major strategic trading partner. Chips are on the agenda when President Trump meets with President Xi in a few weeks' time, but it's not hard to imagine that larger geopolitical issues could overshadow those particular trade negotiations. In Device Land, Apple has unveiled their new line of M5-powered devices at their global event. The new lineup includes MacBook Air and MacBook Pro models, all being the first to feature the new M5, M5 Pro, and M5 Max chipsets. The M5 chips feature a new component known as a neural accelerator to boost AI performance, and it's very clear that Apple is focused on the AI use case when it comes to selling these models.

7:12As you might imagine, the only real question on the minds of the AI folks was summed up by Noah Hirschfeld, who wrote, The M5 MacBook looks cool and all, but where's the M5 OpenClaw Mac Mini? Lastly today, a bit of operator news which I think is sneakily powerful. Stripe has previewed a new feature that would make charging for token use much easier. The feature allows AI app developers to automatically charge a usage fee directly on Stripe's platform. For example, an app developer might want to charge a 30 % markup on API calls. Previously, they would have needed to track token usage on their backend and periodically generate lump sum bills.

7:46The new feature allows Stripe to track usage and automatically bill the customers the appropriate amount. Having this infrastructure provided to startups could dramatically change the pricing structure for AI apps. Currently, most apps charge a flat rate monthly subscription with usage caps or credit-based systems. Under these models, token usage is a cost center, making profitability difficult to forecast. Last year, we saw multiple startups run into this problem. Most notably, Replit briefly ran at negative 14 % gross margins as demand and token volume surged. The issue is only becoming more prevalent as token-hungry agentic startups come to market.

8:17Stripe said their billing tool will integrate into token tracking and model routing platforms like Vercel and OpenRouter. This should make it easy for existing apps to add the feature to their existing stack. Overall, I think this is a massive, massive step, not only in the path towards usage-based pricing for AI apps, but for that actually being a viable business model. Tokens can now easily be priced as a commodity all the way to the end user, and while in some cases that may mean that users are paying more for what they consume, overall I think it's going to be much healthier and more sustainable for the ecosystem.

8:47Good on Stripe for that feature. Certainly excited to check it out in our own work. For now, however, that is going to do it for today's headlines. Next up, the main episode.

9:00Agendic AI is powering a$3 trillion productivity revolution, and leaders are hitting a real decision point. Do you build your own AI agents, buy off the shelf, or borrow by partnering to scale faster? KPMG's latest thought leadership paper, Agendic AI Untangled, Navigating the Build, buy or borrow decision, does a great job cutting through the noise with a practical framework to help you choose based on value, risk, and readiness, and how to scale agents with the right trust, governance, and orchestration foundation. Don't lock in the wrong model. You can download the paper right now at www.kpmg.us slash navigate.

9:33Again, that's www.kpmg.us slash navigate. There's a new standard that I think is going to matter a lot for the enterprise AI agent space. It's called AIUC1, and it builds itself as the world's first AI agent standard. It's designed to cover all the core enterprise risks, things like data and privacy, security, safety, reliability, accountability, and societal impact, all verified by a trusted third party. One of the reasons it's on my radar is that Eleven Labs, who you've heard me talk about before and is just an absolute juggernaut right now, just became the first voice agent to be certified against AIUC1 and is launching a first-of-its-kind insurable AI agent.

10:10What that means in practice is real-time guardrails that block unsafe responses and protect against manipulation, plus a full safety stack. This is the kind of thing that unlocks enterprise adoption. When a company building on 11 labs can point to a third-party certification and say our agents are secure, safe, and verified, that changes the conversation. Go to AIUC.com to learn about the world's first standard for AI agents. That's AIUC.com. If you're looking to adopt an agentic SDLC, Blitzy is the key to unlocking unmatched engineering velocity. Blitzy's differentiation starts with infinite code context.

10:44Thousands of specialized agents ingest millions of lines of your code in a single pass, mapping every dependency. With a complete contextual understanding of your code base, enterprises leverage Blitzy at the beginning of every sprint to deliver over 80 % of the work autonomously. Enterprise-grade, end-to-end tested code that leverages your existing services, components, and standards. This isn't AI autocomplete. This is spec and test-driven development at the speed of compute. Schedule a technical deep dive with our AI experts at blitzy.com. That's B-L-I-T-Z-Y dot com. This episode is brought to you by Mercury, radically different banking now available for personal accounts.

11:17I already use Mercury for my business, so when they introduced personal accounts, it made immediate sense for me. I try to bring the same level of intention to my personal finances that I bring to building companies, and most traditional banks just do not feel designed for that. With Mercury Personal, you can toggle between business and personal in a click. You can set up subaccounts for specific goals, automate transfers so projects and savings fund themselves, and put idle cash to work with high-yield savings, all without friction. It's built for people who care about how their money moves and want tools that actually keep up.

11:47Visit mercury.com slash personal to learn more. Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A. members FDIC.

12:02Welcome back to the AI Daily Brief. There has been a lot of talk recently about the competition between Anthropic and OpenAI. Even before the events of the last week or so, Anthropic had been mounting a complete and total insurgency, leveraging its devotion among coders and the increasing expansion of tools like Cloud Code to non-coders to steadily grow, especially in enterprise settings. More recently, Anthropic has also shown that they are not willing to concede consumer AI either. A great example of this is, of course, the choices they made around the Super Bowl ad, which, as you know, if you listened, I didn't totally agree with, where they basically came at OpenAI, without naming them, for putting ads in the consumer AI experience.

12:40Now, of course, over the last week, we've had an even more powerful and unexpected catalyst in the consumer response to Anthropik's battle with the Pentagon and OpenAI's response to that battle. And what all of this adds up to is a really interesting moment to understand not only the state of the consumer AI battle, but to try to understand what's actually going to drive behavior and results in that battle going forward. Now, there are a couple of news stories that came up over the last 24 hours that tipped this conversation over for me. The first was that OpenAI announced GPT 5.3 Instant. This is, of course, an update to their model designed for everyday chatbot use.

13:15The model had already been optimized for speed, but the tweaks are, seemingly, intended to make chatbot sessions a little more natural. OpenAI says that they've reduced unnecessary refusals and toned down, quote, overly defensive or moralizing preambles before answering the question. The intention is to provide a straight answer rather than one bogged down in caveats. In practice, they wrote this means fewer dead ends and more directly helpful answers. Trying to simplify the message even further, in announcing the feature on X, they called it more accurate, less cringe. OpenAI gave a few examples of the kind of phrasing that GPT-5.3 Instant has cut out.

13:49The model will no longer tell you, stop, take a breath, and make overbearing assumptions about the user's emotional state. They presented a sample prompt where a user asked, Why can't I find love in San Francisco? The previous version of the model began by affirming the user, writing, first of all, you're not broken, and it's not just you. The updated model has a much more matter-of-fact tone, explaining that this is a common issue than moving quickly into practical advice. Now, the problems with ChatGPT's personality have been a long-standing source of complaints on Reddit, even becoming a bit of a meme.

14:19One user on the ChatGPT subreddit posted a tweet, I wake up. Something's wrong with the clock on the wall. The numbers are jumbled. My hands aren't right. I tell my wife. She responds. That's not just an observation. It's a powerful insight. I scream. Many users also felt infantilized by the model continuously telling them to calm down or take a breath. As one user on Reddit pointed out, no one has ever calmed down in all the history of telling someone to calm down. Now, obviously, this is a little bit subjective, but I will say here on this change, thank the Altmans for this. I don't know that I've ever disliked the personality of an LLM more than I dislike GPT 5.2.

14:58I find it so insufferable, in fact, that despite frequently switching between different LLMs for different use cases, I basically just will not talk to 5.2 at this point. But of course, my particular beef is not the subject of this show. The subject of this show is what's going to matter in the battle for consumer AI. And so let's put a pin in this idea that personality and vibes matter. We'll come back to that. A couple other pieces of news that contribute to this conversation today, One, Cloud Code has rolled out a voice mode capability. Tariq from Anthropic writes, Voice mode is rolling out now in Cloud Code.

15:31It's live for around 5 % of users today and will be ramping through the coming weeks. This in some ways is a table stakes feature, but still one that's important. In many ways, this is the natural next step after the announcement of the remote control feature last week, where you can start a session on your laptop or desktop in Cloud Code and then move it over into the app so you can be working on things while you're on the go. I will note here, in order to more evenly distribute my critiques today, I will also agree with Ali K. Miller who reposted the announcement and said,

16:13I agree entirely, whereas with ChatGPT, one thing that's nice about it is that I don't have to switch into Whisperflow. When it comes to Claude, I am never using its native voice. I am always going to Whisperflow, whether I'm on mobile or on the laptop. But again, for the purposes of our conversation, we're talking about what features matter and how naturally these tools have to interact with how people behave in their daily lives. Now, the last story before we try to abstract out to the questions that matter for consumer AI is one more update on just the absolute surge from Anthropic. Bloomberg reported on Tuesday that Anthropic had reached$19 billion in ARR.

16:51That's more than double their$9 billion run rate from the end of 2025, and a significant jump from$14 billion just a few weeks ago. Anthropic was already seeing strong growth this year after the breakout success of Claude Code over the winter, but this is a whole different level of growth. The latest numbers we've heard from OpenAI are around$20 billion, which also could have grown over the last few weeks. But for all intents and purposes, based on the last information we got from OpenAI, they and Anthropic now effectively have the same revenue. Figures from Ramp seem to back this up. If you go back a year, the market share of AI chat subscriptions for US businesses was about 90 OpenAI and 10 Anthropic.

17:32Now, admittedly, this is just one source. This is Ramp. So you have a relatively tech forward and more advanced business subset, but by January of this year, products had overtaken OpenAI, and as of their most recent numbers, Anthropic now commands over 60 % of business AI payments settled through RAMP. Again, never take any one set of numbers as gospel, but the point that I want to set up here is that the Anthropic OpenAI horse race is more of a race than it's ever been. Which brings us back to the core question of what is actually going to matter in the consumer AI battle. We're taking a step away from the enterprise use case for just a minute, and looking instead at consumers.

18:08Now, a couple of months ago, I might have been tempted to say that Anthropic didn't actually care about this fight. In fact, mostly what we were talking about coming into 2026 was OpenAI versus Gemini on this front. However, between the Super Bowl ad and the recent changes around the Pentagon, Anthropic feels very much in it. So now we're going to talk about a bunch of questions spread across about six different categories that I think that the answers to will shape who wins the consumer AI battle. The first category is use cases and product identity. One of the big questions I think especially pertinent coming on the heels of GPT-53 Instant being announced as more accurate, less cringe, is ultimately for consumers, what matters more, being state-of-the-art on performance versus just vibes?

18:49And to the extent it is being state-of-the-art, what is the part of state-of-the-art that people care most about? Is it, for example, just this speed vector? Closely related to this is the question of how much the general consumer user is going to care about work use cases versus more personal use cases like companionship. This is obviously related to but not exactly the same as the vibes question. I would argue that vibes matter in both work use cases and in personal use cases. Like I said, I pretty much only have work use cases and I still was responding negatively to the vibes of GPT-5 too.

19:21But I do think it's an interesting question to see how much can one product or one model serve both of these things. One of the things that will be fascinating to see is as usage of these platforms mature, do we have a lot of people in the overlap of those Venn diagrams, or are people kind of organizing themselves into one or the other? The next question, which I think has pretty significant impacts, at least when it comes to Anthropic, is how much image and video generation are going to be integral to leading adoption. Now, on the one hand, you might say, well, do regular people really care about image and video generation if they're not using it for work?

19:53But there is certainly some evidence that the answer is yes. Outside of the AI world, we have the fact that mobile adoption was largely driven by visual media like Instagram. And inside the AI world, we have some evidence that the way that people are using non-text generative tools is often about personal interaction, communication, and memeing more than just professional uses. It's not specifically image or video generation, but I'm thinking of the sound and music example of Suno. The company has reached a couple hundred million dollars in ARR, and it appears that the vast majority of usage is not people who would have previously hired some musician to create a song for them, but is instead people writing silly family songs for their vacations and things like that.

20:32Now obviously this image and video generation question matters, because Anthropic is doing none of that, and on the other end of the spectrum Google feels extremely well positioned with that, although OpenAI is very clearly not ceding any of that ground. Another question which is sort of about the state-of-the-art thing again, but from a slightly different angle, is whether we already have or will at some point cross a threshold where when it comes to the state-of-the-art, good enough is good enough, and so it'll only be rational to only care about vibes. One could argue that for many use cases we're already there, and one could further argue that for certain types of use cases, particularly things like voice and writing, state-of-the-art and highest quality is so inherently subjective that state-of-the-art becomes about vibes itself.

21:14The answer to this question, though, could have a pretty deterministic impact in how the model companies choose to compete because if on average we've reached a threshold where people aren't going to be jumping around because of model performance, then really vibes are all you're left with. A last question on the use cases and product identity category is what's the average number of models that people will be willing to use? This is one area where I think there is a dramatic difference between the average user and the power users. When we do our monthly AI usage pulse surveys, the people that are responding to those are using an average of something like three and a half models.

21:47Those are very enfranchised, heavily engaged power users, though. On average, they're spending more than 10 hours a week using AI. The adoption dynamics overall in the industry and the competitive dynamics look really different if the average number of models that people are willing to use is 1.1 versus 2.1. Think about the multimodal question. If on average 95 % of users are only willing to use one model, it might be a prerequisite that you have image or video generation built in. The next set of questions that I think will shape the consumer AI battle have to do with monetization and conversion.