OpenAI’s Chip Rebellion: Why It’s Breaking Free From Nvidia’s Grip

OpenAI’s Chip Rebellion: Why It’s Breaking Free From Nvidia’s Grip

Here’s the thing about OpenAI’s new deal with Broadcom that actually matters: they’re not just buying different chips (boring), they’re fundamentally reshaping how AI companies think about hardware control. And honestly? This feels like the moment the industry collectively realized it was too dependent on one company’s supply chain. Let me back up. OpenAI just struck a deal to develop custom AI chips optimized specifically for inference—that’s the moment when ChatGPT actually runs and talks to you, as opposed to the brutal training phase where models learn from mountains of data. This isn’t exactly shocking news (companies diversify suppliers, tale as old as time), but the timing and the target reveal something genuinely strategic happening. Here’s the framework for understanding why this matters: Think of AI infrastructure in two phases. Training is the heavyweight championship—it demands absolute raw computational power, which is why Nvidia’s H100 GPUs have basically owned this space. Nvidia crushed it here because of CUDA, their software ecosystem, and years of engineering prowess. OpenAI isn’t ditching Nvidia for training (that would be insane). But inference Broadcom gets this. They’ve built custom silicon for hyperscalers before (Google’s TPUs ring a bell?), and they know how to engineer chips that do one job really, really well. The application-specific integrated circuits (ASICs) they’re building for OpenAI will be optimized to the point that they’re probably more efficient than general-purpose GPUs for inference workloads. Translation: same performance, lower power consumption, massive cost savings when you’re running ChatGPT for millions of users simultaneously. What’s wild is that Sam Altman has been publicly signaling this move for months. He’s talked openly about the need for “more hardware options” and “multiple chip architectures.” This wasn’t a secret—it was a warning to Nvidia that the moat was eroding. And look, I’m not here to bury Nvidia (their training dominance is still absurd), but the inference market is massive, and spreading that load across custom silicon? That’s rational infrastructure thinking. The broader context here: We’re watching the AI industry mature past its “just buy whatever Nvidia has” phase. Supply constraints have been real (remember when everyone was fighting over H100s?), costs are astronomical, and companies with billions at stake need redundancy. OpenAI, Google, Meta—they’re all building custom silicon because relying on one vendor during an arms race is actually reckless. No timeline or financials have been disclosed (because of course not), but this move signals something important: the race isn’t just about who builds the best models anymore. It’s about who controls the entire stack—models, software, and now hardware. That’s where the real competitive moat gets built. Watch this space. When custom inference chips start delivering real cost advantages, we’ll see more of this. And suddenly Nvidia’s dominance looks a little less inevitable. Source: The Wall Street Journal Want more than just the daily AI chaos roundup? I write deeper dives and hot takes on my Substack (because apparently I have Thoughts about where this is all heading): limitededitionjonathan on Substack

Microsoft Brings ‘Vibe Working’ to Office Apps (Because Apparently We’re All Just Vibing Now)

Look, I know another Microsoft Office announcement sounds about as thrilling as watching Excel formulas multiply (which, frankly, is what this is partly about). But Microsoft just launched something called “vibe working” for Excel and Word, and I’m genuinely impressed by what they’re pulling off here. The company is rolling out Agent Mode in Excel and Word today – think of it as Copilot’s older, more capable sibling that actually knows what the hell it’s doing. Instead of those helpful-but-limited suggestions we’re used to, Agent Mode can generate complex spreadsheets and full documents from simple prompts. We’re talking “board-ready presentations” and work that Microsoft’s Sumit Chauhan says is “quite frankly, that a first-year consultant would do, delivered in minutes.” Here’s what’s wild about Agent Mode: it breaks down complex tasks into visible, step-by-step processes using OpenAI’s GPT-5 model. You can literally watch it work through problems in real time, like an automated macro that explains itself. For Excel (where data integrity actually matters), Microsoft has built in tight validation loops and claims a 57.2% accuracy rate on SpreadsheetBench – still behind human accuracy of 71.3%, but ahead of ChatGPT and Claude’s file-handling attempts. The Word version goes beyond the usual “make this sound better” rewrites. Agent Mode turns document creation into what Microsoft calls “vibe writing” – an interactive conversation where Copilot drafts content, suggests refinements, and clarifies what you need as you go. Think collaborative writing, but your writing partner has read the entire internet and never gets tired of your terrible first drafts. But here’s the really interesting move: Microsoft is also launching Office Agent in Copilot chat, powered by Anthropic models (not OpenAI). This thing can create full PowerPoint presentations and Word documents from chat prompts, complete with web research and live slide previews. It’s Microsoft’s answer to the flood of AI document tools trying to eat their lunch. The Anthropic integration is telling – Microsoft is hedging its OpenAI bets while exploring what different model families bring to the table. “We are committed to OpenAI, but we are starting to explore with the model family to understand the strength that different models bring,” Chauhan says. Smart move, considering Anthropic’s models are already powering GitHub Copilat and researcher tools. Agent Mode launches today for Microsoft 365 Copilot customers and Personal/Family subscribers (web versions first, desktop coming soon). Office Agent is U.S.-only for now. And yes, this means the office productivity wars just got a lot more interesting. Read more from Tom Warren at The Verge [https://www.theverge.com/news/787076/microsoft-office-agent-mode-office-agent-anthropic-models] Want more than just the daily AI chaos roundup? I write deeper dives and hot takes on my Substack (because apparently I have Thoughts about where this is all heading): https://substack.com/@limitededitionjonathan

Huawei Plans to Double AI Chip Production as Nvidia Stumbles in China

Look, I know another chip story sounds like more tech industry inside baseball, but this one’s actually wild when you dig into what’s happening. Huawei Technologies is preparing to massively ramp up production of its most advanced AI chips over the next year—we’re talking about doubling output of their flagship Ascend 910B processors. And the timing? That’s the interesting part. While Nvidia is getting tangled up in geopolitical headwinds (export restrictions, compliance issues, the usual US-China tech drama), Huawei is essentially saying “hold our beer” and going full throttle on domestic AI silicon. Thing is, this isn’t just about making more chips—it’s about winning customers in what Bloomberg calls “the world’s biggest semiconductor market.” Here’s what makes this fascinating from a technical standpoint: The Ascend 910B isn’t some budget knockoff. We’re talking about chips that can genuinely compete with high-end GPUs for AI training workloads. Huawei has been quietly building this capability for years (remember, they’ve been dealing with US restrictions since 2019), and now they’re ready to scale production significantly. The broader context here is that China’s AI companies have been desperate for alternatives to Nvidia’s H100s and A100s. With export controls making it increasingly difficult to get the latest US chips, there’s been this massive pent-up demand for domestic alternatives. Huawei is basically positioned to fill that void—and they know it. What’s particularly smart about Huawei’s approach is the timing. As Nvidia navigates compliance requirements and export restrictions that slow down their China business, Huawei gets to swoop in with locally-produced chips that Chinese companies can actually buy without worrying about geopolitical complications. It’s like being the only restaurant open when everyone else is dealing with supply chain issues. The ripple effects could be huge. If Huawei can actually deliver on this production ramp (and that’s a big if—chip manufacturing is notoriously difficult to scale), we’re looking at a genuine alternative ecosystem for AI development in China. That means Chinese AI companies won’t be as dependent on US technology, which fundamentally changes the competitive landscape. Of course, there are still questions about performance parity and ecosystem support (CUDA is hard to replace), but the mere fact that a viable alternative exists puts pressure on everyone. Competition drives innovation, and having two major players fighting for the world’s largest AI chip market? That’s going to accelerate development on both sides. This is one of those stories where the technical development (doubling chip production capacity) intersects with geopolitics in ways that could reshape how AI infrastructure gets built globally. Worth watching closely. Read more from Yuan Gao at Bloomberg [https://www.bloomberg.com/news/articles/2025-09-29/huawei-to-double-output-of-top-ai-chip-as-nvidia-wavers-in-china] Want more than just the daily AI chaos roundup? I write deeper dives and hot takes on my Substack (because apparently I have Thoughts about where this is all heading): https://substack.com/@limitededitionjonathan

Apple’s Secret AI Weapon: The Internal Chatbot That Should Be Public

Look, I know another AI chatbot announcement sounds about as exciting as watching paint dry (we’ve had roughly 847 of them this year), but this one’s different. According to Bloomberg’s Mark Gurman, Apple has been quietly testing its next-generation Siri through an internal ChatGPT-style chatbot called “Veritas” — and honestly, it sounds kind of amazing. Here’s what makes this interesting: Instead of the usual corporate approach of endless closed-door testing, Apple built what’s essentially their own version of ChatGPT for employees to actually use. We’re talking back-and-forth conversations, the ability to dig deeper into topics, and crucially — the kind of functionality that Siri desperately needs. Employees can search through personal data and perform in-app actions like photo editing, which sounds suspiciously like what we were promised when Apple Intelligence was first announced. The thing is, Apple’s AI struggles aren’t exactly a secret at this point. The company has delayed the next-gen Siri multiple times, and Apple Intelligence launched to what can generously be called “tepid” reception. Meanwhile, every other tech company is shipping AI assistants that can actually hold a conversation without making you want to throw your phone out the window. But here’s where it gets frustrating: Gurman reports that Apple has no plans to release Veritas to consumers. Instead, they’re likely going to lean on Google’s Gemini for AI-powered search. Which feels backwards, right? You’ve built this internal tool that apparently works well enough for your employees to test new Siri features with, but regular users get… nothing? Think about the framework here: Apple has created a testing environment that lets them rapidly develop and collect feedback on AI features. That’s exactly the kind of iterative approach that made ChatGPT and other conversational AI successful. The difference is OpenAI, Anthropic, and Google let millions of users participate in that feedback loop. Apple is keeping it locked to their own employees. This feels like a missed opportunity on multiple levels. First, Apple could actually compete in the AI assistant space instead of just licensing someone else’s technology. Second, they’d get the kind of real-world usage data that makes these systems better. And third (this might be the most important part), it would give Apple Intelligence some actual credibility instead of the current situation where Siri still can’t reliably set multiple timers. The irony here is that Apple traditionally excels at taking complex technology and making it accessible to regular people. But with AI, they’re taking the opposite approach — building sophisticated tools for internal use while consumers get a watered-down experience that relies on external partnerships. Maybe Apple will surprise us and eventually release some version of Veritas publicly. But given their track record with AI announcements (remember when we were supposed to get the “new” Siri by now?), I’m not holding my breath. In the meantime, the rest of us will keep using ChatGPT while Apple employees get to play with what sounds like a genuinely useful AI assistant. Sources: The Verge [https://www.theverge.com/news/787046/apples-veritas-siri-ai-chatbot] Want more than just the daily AI chaos roundup? I write deeper dives and hot takes on my Substack (because apparently I have Thoughts about where this is all heading): https://substack.com/@limitededitionjonathan

Gemini Robotics 1.5: Google DeepMind Just Cracked the Code on Agentic Robots

Look, I know another AI model announcement sounds boring (trust me, I’ve written about 47 of them this month), but Google DeepMind just dropped something that actually made me sit up and pay attention. Their new Gemini Robotics 1.5 isn’t just another incremental upgrade—it’s a completely different approach to making robots that can think, plan, and adapt like actual agents in the real world. Here’s what’s wild: instead of trying to cram everything into one massive model (which, let’s be honest, has been the industry’s default approach), DeepMind split embodied intelligence into two specialized models. The ERVLA stack pairs Gemini Robotics-ER 1.5 for high-level reasoning with Gemini Robotics 1.5 for low-level motor control. Think of it like giving a robot both a strategic brain and muscle memory that can actually talk to each other. The “embodied reasoning” model (ER) handles the big picture stuff—spatial understanding, planning multiple steps ahead, figuring out if a task is actually working or failing, and even tool use. Meanwhile, the visuomotor learning agent (VLA) manages the precise hand-eye coordination needed to actually manipulate objects. The genius part? They can transfer skills between completely different robot platforms without starting from scratch. What does this look like in practice? These robots can now receive a high-level instruction like “prepare this workspace for the next task” and break it down into concrete steps: assess what’s currently there, determine what needs to move where, grab the right tools, and execute the plan while monitoring progress. If something goes wrong (like a tool slips or an object isn’t where expected), the reasoning model can replan on the fly. The technical breakthrough here is in the bidirectional communication between the two models. Previous approaches either had rigid, pre-programmed behaviors or tried to learn everything end-to-end (which works great in simulation but falls apart when you meet real-world complexity). This stack lets robots maintain both flexible high-level reasoning and precise low-level control. Here’s the framework for understanding why this matters: we’re moving from “task-specific robots” to “contextually intelligent agents.” Instead of programming a robot to do one thing really well, you can give it general capabilities and let it figure out how to apply them to novel situations. That’s the difference between a really good assembly line worker and someone who can walk into any workspace and immediately start being useful. The implications are pretty staggering when you think about it. Manufacturing environments that need flexible reconfiguration, household robots that can adapt to different homes and tasks, research assistants in labs that can understand experimental protocols—we’re talking about robots that can actually collaborate with humans rather than just following pre-written scripts. DeepMind demonstrated the system working across different robot embodiments, which solves one of the biggest practical problems in robotics: the fact that every robot design requires starting over with training. Now you can develop skills on one platform and transfer them to others, which could dramatically accelerate deployment timelines. This feels like one of those moments where we look back and say “that’s when robots stopped being fancy automation and started being actual agents.” The combination of spatial reasoning, dynamic planning, and transferable skills wrapped in a system that can actually explain what it’s doing? That’s not just an incremental improvement—that’s a fundamental shift in what’s possible. Read more from MarkTechPost [https://www.marktechpost.com/2025/09/28/gemini-robotics-1-5-deepminds-er%e2%86%94vla-stack-brings-agentic-robots-to-the-real-world/] Want more than just the daily AI chaos roundup? I write deeper dives and hot takes on my Substack (because apparently I have Thoughts about where this is all heading): https://substack.com/@limitededitionjonathan