OpenAI Just Built Its Own Chip
Here's the thing about Nvidia's dominance in AI hardware: it was never going to last forever. Not because Nvidia did anything wrong—they've been phenomenal—but because no single company should hold that much leverage over an entire industry's infrastructure.
OpenAI and Broadcom just proved that point on June 24, 2026, when they unveiled Jalapeño. It's OpenAI's first custom-built AI chip, though calling it just a "chip" undersells what they've actually created. It's an Intelligence Processor designed from the ground up based on OpenAI's deep understanding of how models like ChatGPT actually work.
The engineering samples are already running ML workloads, including GPT-5.3-Codex-Spark, at target speed and power levels. Early results show significantly better performance with less power than leading AI chips on the market. That's not a marginal improvement—that's the kind of gain that makes investors sit up and pay attention.
But here's what really matters: this isn't a one-off experiment. OpenAI is building a multi-generation computing platform, and Jalapeño is just the first generation. They're planning to deploy at gigawatt scale with data center partners, which means we're talking about real infrastructure, not a lab curiosity.
Broadcom and Celestica helped transform the design into a production-ready chip, with Broadcom's Tomahawk networking technology playing a key role in making this scale possible. The partnership model here is interesting—OpenAI brings the AI workload expertise, Broadcom brings the chip design and manufacturing muscle, and Celestica handles system integration. It's a clean division of labor that could become the template for future custom silicon projects.
The Custom Silicon Trend Is Real
OpenAI isn't alone in this. Google's been building TPUs for years. Apple transitioned from Intel to its own silicon and it was a masterclass in what custom hardware can do. SpaceX is building chips too, though their use cases are more specialized.
The goal here isn't a clean break from Nvidia. It's a hedge against single-supplier risk, and that's an important distinction. These companies aren't trying to eliminate Nvidia from their stacks—they're trying to ensure they have options.
Think about it: when you're spending billions on AI infrastructure, betting everything on one supplier is reckless. Nvidia has dominated the AI chip market for years, and they deserve credit for that. But the era of total dependence is ending, and OpenAI's Jalapeño announcement is just one data point in a much larger trend.
Custom silicon offers three things that off-the-shelf chips can't match: more control over the design, hardware tuned to specific workloads, and performance gains that come from optimizing for exactly what you need rather than trying to be everything to everyone.
Apple's transition from Intel is the poster child here. They didn't just swap one chip for another—they redesigned their entire computing architecture around what their software actually needed. The result was better performance, better battery life, and more control over their hardware roadmap. OpenAI is trying to do the same thing with AI workloads.
What This Means for Nvidia
Let's be clear: Nvidia isn't going anywhere. They're too big, too well-established, and their CUDA ecosystem is too deeply embedded in the industry. But they're facing a new reality.
Companies that were once loyal Nvidia customers are now building their own chips. That doesn't mean they'll stop buying Nvidia products—it means they have leverage. They can negotiate better pricing, demand better support, and ensure they're not at the mercy of a single supplier's roadmap.
This is actually healthy for the industry. Competition drives innovation, and having multiple players in the custom silicon space means more experimentation, more breakthroughs, and better options for everyone.
Nvidia's response will be interesting to watch. They'll likely double down on their strengths—training workloads, the CUDA ecosystem, their relationships with cloud providers. But they'll also need to adapt to a world where their biggest customers are also their competitors in chip design.
The shift isn't about replacing Nvidia. It's about creating a more resilient, more competitive hardware ecosystem that can serve the diverse needs of AI companies better than any single supplier could.
The Bigger Picture: AI Loops and Autonomous Systems
The Equity podcast discussion touched on something that's often overlooked in these chip debates: the relationship between custom hardware and AI loops.
AI loops—systems where AI models operate in feedback cycles, making autonomous decisions and continuously learning from their outputs—are becoming more common. And they need hardware that's optimized for exactly that kind of workload.
General-purpose chips are fine for training large models. But inference at scale, especially when you're running continuous loops with tight latency requirements? That's where custom silicon shines. You can design for the specific patterns of your inference workloads, optimize data movement, and reduce energy consumption in ways that off-the-shelf chips simply can't match.
This is why companies like OpenAI are investing in custom hardware. They're not just trying to save money on chips—they're trying to build infrastructure that's perfectly matched to how their AI systems actually operate.
The implications extend beyond just OpenAI. Any company running autonomous AI systems—whether it's for robotics, autonomous vehicles, or industrial automation—will benefit from hardware that's tuned to their specific loop architectures.
Deals Worth Watching
The podcast also covered a few deals of the week that are worth keeping an eye on, though the specific details weren't fully fleshed out in the article format. The broader trend is clear: investment in AI infrastructure, custom silicon development, and the companies enabling it all continues to accelerate.
What's interesting is how the narrative has shifted. A few years ago, building your own AI chip was seen as a vanity project for well-funded labs. Now it's becoming standard practice, and the companies that get it right will have a significant competitive advantage.
The OpenAI-Broadcom partnership is just the beginning. Watch for more announcements like this throughout 2026 and beyond, especially as the demand for AI inference capacity continues to grow exponentially.
What Comes Next
Jalapeño is generation one. OpenAI has a multi-generation platform strategy, which means we should expect to see iterative improvements and new capabilities as they refine their approach.
The gigawatt-scale deployment plans suggest they're thinking about this in terms of real infrastructure, not just technical proofs of concept. That's a significant step up from most custom chip projects I've seen, which tend to stay in the lab for years.
The broader industry impact will be felt over the next few years. As more companies build custom silicon, we'll see: better performance for specific workloads, more competition in the AI hardware space, and a more resilient supply chain that's less vulnerable to single points of failure.
Nvidia will adapt. They always do. But the landscape is changing, and companies that were once purely customers are now becoming competitors in chip design. That's a shift that will reshape the entire AI hardware ecosystem.
The era of Nvidia dominance isn't over. But it's evolving, and OpenAI's Jalapeño is a clear signal of where things are heading.
Related Reading
- Beyond the GPU: The 'Jalapeño' ASIC and the Future of Inference Infrastructure — A deeper technical look at how OpenAI and Broadcom engineered Jalapeño for LLM inference scaling.
- The CapEx Conundrum: Markets Retreat as Wall Street Realizes AI's Long Horizon — Why the economics of AI infrastructure spending are under scrutiny as investors demand faster returns.
- Scaling AI: Why Data Infrastructure is the Real Bottleneck — How data delivery, not compute, is the hidden constraint holding back production AI systems.