The AI Acceleration Gap

**Nathaniel Whittemore** (0:01)
Today on the AI Daily Brief, we are talking about the AI acceleration gap, what it is, why it matters, and what you should do about it. Before that in the headlines, what we learned from a recent town hall at OpenAI. 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. Over the weekend, Sam Altman announced that on Monday afternoons last evening, they would be hosting what they were calling a town hall for AI builders at OpenAI. In his announcement post, Sam said that this was an experiment and a first pass at a new format. He framed the live stream event as an opportunity to gather feedback as OpenAI begins building their next generation of tools. Ultimately, it's sort of played out as a Q&A about the state of the company and the industry. One of the big points of discussion was the performance of GPT-5-2. Altman acknowledged, for example, that the latest model has a writing style that can be unwieldy and difficult to read. He said, I think we just screwed that up. We will make future versions of GPT-5.x hopefully much better at writing than 4.5 was. Now Altman noted that their focus hadn't been on writing, saying, We did decide, and I think for good reason, to put most of our effort in 5.2 into making it super good at intelligence, reasoning, coding, engineering, that kind of thing. And we have limited bandwidth here and sometimes we focus on one thing and neglect another. Now, of course, rumors suggest that the next model, codenamed Garlick, is weeks or even days away. So for those of you who find GPT-5.2's writing clunky, you presumably won't have to deal with it much longer. Altman also discussed a hiring slowdown at OpenAI. Responding to a question about how AI had changed the interview process, he commented, We are planning to dramatically slow down how quickly we grow because we think we'll be able to do so much more with fewer people. He assured the crowd that this was not a hiring freeze and that headcount reductions are not on the table, but did suggest that AI developments could rapidly shift staffing needs over the short term. What I think we shouldn't do and what I hope other companies won't do either is hire super aggressively then realize all of a sudden AI can do a lot of stuff and you need fewer people and have to have some sort of very uncomfortable conversation. I think the right approach for us will be to hire more slowly but keep hiring. In other comments, Altman said that he expects the cost of AI to continue to hyper-deflate, forecasting that OpenAI will be able to deliver quote, GPT-52 level intelligence by the end of 2027 for at least 100 times less. Reflecting something that we've talked about a bunch on this show, Altman said that another big goal of 2026 is to push quote, super hard on memory and personalization. Altman said that he's personally ready to give ChatGPT complete access to his computer and internet history, allowing it to quote, just know everything. And while he acknowledged that security and privacy were still major concerns, he said that the company will focus on building a system that has quote, such a deep understanding of the complex rules and interactions of my life, that it knows what to use when and what to expose where. As part of that goal, Altman explained that login with ChatGPT is coming soon, which will in the short term enable token budgets to be shared across various apps, with the long term vision being allowing portable memory to function across different AI products. We even got some little hints about their hardware plans, with Altman saying that the vision is now a collaborative multiplayer experience. He framed it as five people gathered around a table with what he called a little robot to help the group do better, and highlighting the massive shift that we have talked about extensively and is in fact the theme of the main episode today. Altman commented that, quote, What it means to be an engineer is going to super change. He said there will probably be far more people creating far more value getting computers to do what they want. He noted at the same time, however, that demand for software seems not to be slowing down at all. So, what to think about this overall? There were quite a few snarky responses to this on Twitter slash X. Some people didn't like that comments were off on the livestream. Others thought that the vibes and the atmosphere were just really weird. Some thought that Altman himself looked kind of tired and run down, which others interpreted as OpenAI's competitive struggles, but which also could easily be explained by being father to an infant. Overall, what I would say is this. I think it kind of doesn't matter if their A wasn't all that much revealed on this, and B, people have critiques about the vibes. I think that if OpenAI regularizes this, it could actually be really valuable and build a lot of trust. Having a predictable regular place to have these sort of conversations could go a long way to making things that need to be explained, not always feel like there are PR response. So in that regard, I think it was successful and they should do more of it. Now another discussion from this weekend around OpenAI had to do with their advertising plans. The information reports that pricing sheets are starting to circulate with a premium price tag for OpenAI ads. OpenAI appears to be selling on a CPM basis with an offering at $60 CPMs or $60 per thousand views, which is around three times the cost of placing an average ad on, for example, Metis platforms. The only data available during the early stage will be total ad views and clicks, which is obviously a lot less information than they're going to get from other advertisers, but presumably that will change soon. OpenAI does pledge not to sell personal data to advertisers, and they appear to be taking that stance to the extremes, and the premium pricing, at least at the beginning, probably won't be a deterrent to the early advertisers that are clamoring to get on the platform. Studies have generally shown that AI users have high intent relative to other types of internet users, which could end up easily justifying that premium over other digital ad units. Now, advertising isn't the only place where OpenAI plans to charge a hefty premium. FinTech reporter Simon Taylor recently noted that Shopify merchants are being charged a 4% fee for sales conducted through ChatGPT, which is a fee on top of existing Shopify charges. Shopify CEO Toby Lutke filled in some further details, commenting, This is ChatGPT charging 4% and we collect the fees on their behalf. Everyone gets a free trial that starts after the first sales. Not saying that's good or bad, ads definitely cost more for most. Taylor acknowledged that this is pretty close to fees from buy-now-pay-later services and added, For what it's worth, I think 4% is very defensible if conversion is there. Moving over to chips today, Microsoft has unveiled the second generation of their in-house AI chip, taking aim at custom silicon from Google and Amazon. Called the Maya 200, Microsoft claims the chip is the, quote, most performant first-party silicon from any hyperscaler. The chip is optimized for inference and Microsoft says it's the most efficient silicon in their fleet, outperforming the next best by 30% on a performance-per-dollar basis. The accelerator was built using TSMC's latest 3nm process. Google is also using the 3nm process for their seventh generation TPUs, but NVIDIA chose to stick with 4nm manufacturing on their latest generation Blackwell chips. The chip also features enough memory to easily run the latest models with plenty of headroom for the next generation. Now, whenever a new chip is released, the immediate chatter is all about whether this will end NVIDIA's dominance of the industry. But Tom's Hardware pointed out that comparisons to NVIDIA's leading chips are a little spurious as they do very different things. Microsoft's hardware won't be available for outside sales, so the Maya 200 will only be able to move the needle internally. And Tom's also pointed out that the Blackwell 300 Ultra flagship vastly outperforms on raw power and of course integrates with NVIDIA's highly developed software stack. But the Maya 200 does beat the Blackwell chip in efficiency, operating at nearly half the total power draw. Ultimately, with the Maya 200 though, Microsoft is staking their claim as a player in the custom silicon race. Staying in Chitlan but moving over a bit, NVIDIA has invested a further $2 billion into CoreWeave to kickstart the deployment of AI factories. Now NVIDIA CEO Jensen Huang has been discussing the concept of AI factories for the past year. The language describes the idea that data centers will need to be deployed on a much greater scale to supply the AI tokens that will drive economic outcomes in the future. Essentially, it reframes data centers from large cloud storage and compute providers to the producers of the core commodity of the AI age. With the next leg of their coreweave partnership, NVIDIA will support a scaling up of coreweave's infrastructure. The goal is to deploy 5 gigawatts in capacity by 2030, with NVIDIA using its financial might to help procure land and power for the rollout. NVIDIA already owned a 6.6 stake in coreweave, so this new investment brings their ownership to around 10%.