Imagine a procurement lead at OpenAI staring at a quarterly invoice from NVIDIA, wondering if the company’s valuation is actually just a giant hedge fund for Jensen Huang’s leather jacket collection. The numbers are absurd. The wait times are worse. For years, the strategy has been to simply throw more H100s at every problem, but eventually, the math stops working. When your burn rate starts to look like a small nation’s GDP, you stop asking “can we do this” and start asking “how do we stop paying the tax.” It is a desperate realization that no amount of VC funding can outpace the margins of a hardware monopoly.
That’s why the move to custom silicon finally happened. According to TechCrunch, OpenAI has teamed up with Broadcom to build its first dedicated chip. It’s the classic play: stop paying the NVIDIA tax and start owning the stack. (I suspect the board practically begged Sam to do this). It’s a move that signals a shift from the “software-only” lab to a full-stack infrastructure company, whether they want to admit it or not. They aren’t just building models anymore; they are building the very furnace those models are cooked in.
Here is the problem: partnering with Broadcom isn’t “independence.” It’s just swapping one dependency for another. Broadcom is the plumbing of the data center world; they provide the IP and the design assistance, but they aren’t giving away the secrets for free. It’s like a chef deciding to grow their own organic vegetables to save money, only to realize they have to pay a specialized irrigation consultant $50k a month just to keep the plants alive. You’ve removed the middleman, but you’ve added a specialized architect who knows exactly how much you’re desperate. You aren’t escaping the vendor lock-in; you’re just moving to a different neighborhood with a different set of landlords.
Let’s talk about the real-world friction. Designing a chip isn’t like updating a Python library or tweaking a prompt. A single mistake in the tape-out process can burn through tens of millions of dollars and six months of development time before you even know if the silicon actually boots. Then there’s the software side. CUDA is a moat for a reason. Moving workloads from NVIDIA’s ecosystem to a custom ASIC means rewriting kernels and fighting with compilers that probably don’t have a decent documentation page. Does anyone actually believe the transition will be seamless? The amount of engineering hours required to make a custom chip actually perform in a production environment is often underestimated by people who only look at the TFLOPS on a slide.
This move is logically sound but strategically desperate. OpenAI is terrified of the compute ceiling. If they can’t optimize the hardware to the specific needs of their future models, they’ll hit a wall where adding more GPUs yields diminishing returns. Owning the silicon is the only way to squeeze out that last 20% of efficiency that differentiates a product from a research project. However, it’s a massive gamble on the idea that the model architectures of tomorrow will be stable enough to bake into hardware today. If the industry shifts toward a completely different attention mechanism or a non-transformer architecture next year, they’ve just built a very expensive monument to an obsolete idea. Or maybe not—maybe they’ve already seen the next architecture and this chip is designed for it.
It’s a necessary evil.
By Q1 2027, we’ll see the first real-world benchmarks showing that these chips are actually slower for general training but significantly more efficient for the specific inference patterns of the O-series models. The goal isn’t to beat the H200 at everything; it’s to stop the bleeding on the balance sheet and reduce the latency of the feedback loop. If they pull it off, the “compute moat” becomes a real thing rather than a slide-deck buzzword. If they don’t, they’ve just spent a few billion dollars to build a very expensive paperweight.
