The Big Questions That Will Decide the Consumer AI War

**Nathaniel Whittemore** (0:00)
Today 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.
Welcome back to the AI Daily Brief headlines edition, all the daily AI news you need in around five minutes. Back 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 complain that these outages have stopped work for minutes or even hours at a time. GitHub 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. Domke'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 SaaS-pocalypse theme of companies cancelling their software subscription in favor of vibe-coded alternatives. I'm not sure any of that's true. It 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. As Ameya 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 broad restructuring of AI at Meta, which by some counts went through four reshufflings last year. Instead, 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. It 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, 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. One 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. 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. As 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, US officials are considering a cap on NVIDIA chip sales into China in a bid to constrain the power of training clusters. Bloomberg reports that US 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. They 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. Each 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 buildout 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. What'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.