OpenAI is Finally Building Its Own Chips - September 5, 2025

OpenAI just dropped $10 billion on a deal with Broadcom to build their own AI chips. About fucking time - NVIDIA's been charging whatever they want for H100s and H200s because they had a monopoly on decent AI hardware.

The plan is simple: design chips specifically for GPT models instead of using NVIDIA's general-purpose GPUs that cost a fortune and aren't even optimized for transformer workloads. If they can pull it off by 2026, they'll cut their inference costs in half and never have to beg NVIDIA for allocation again.

I've Seen This Movie Before (It Doesn't End Well)

Here's what the press releases won't tell you: custom chip projects fail constantly. I watched Intel burn billions on Nervana from 2016-2020, promising AI chips that would crush NVIDIA. Project got cancelled after eating $400M and delivering nothing. The test chips kept throwing PCIE_TRAINING_ERROR and the drivers crashed with DEVICE_NOT_FOUND every time you tried to run anything real.

IBM spent a decade hyping neuromorphic chips that would "think like brains." Saw their demos at conferences for years - always "coming soon." The hardware never shipped because it kept failing power-on self-tests and throwing THERMAL_SHUTDOWN errors during basic operations. Know where those chips are now? Nowhere.

Every chip company promises 2026 delivery. It's always 2028. Then 2030. Then quietly cancelled.

OpenAI has zero hardware experience. These are software people trying to out-engineer NVIDIA's 30-year GPU evolution in two years. You can't just throw Python developers at silicon design problems and expect magic.

But here's why they might not be completely insane: Broadcom actually knows what they're doing. They've built custom silicon for Google's TPUs, Amazon's Trainium chips, and custom AI accelerators for major cloud providers. Their semiconductor expertise spans decades, unlike other failed attempts at custom AI silicon. If anyone can turn OpenAI's wishful thinking into actual working hardware, it's probably them.

Why This Matters (If It Works)

NVIDIA's pricing is basically extortion at this point. A single H100 costs $40,000+ and you need thousands of them to run anything meaningful. That's why every big tech company is building their own chips:

• Google has TPUs that crush GPUs for their workloads
• Amazon's got Trainium and Inferentia
• Apple built M-series chips that embarrassed Intel
• Now OpenAI wants in on the custom silicon game

The Broadcom stock pop shows Wall Street thinks this might actually work. But remember - this is the same market that thought crypto would replace money and that the metaverse mattered.

Every big tech company is building their own chips now because NVIDIA's prices are fucking insane. When your inference costs are eating 60% of your revenue, custom silicon starts looking like the only way out.

If OpenAI's chips work, NVIDIA's AI monopoly is fucked. If they don't, OpenAI just burned $10 billion learning why chip design is harder than it looks.

This OpenAI deal isn't happening in a vacuum. Every major tech company is desperately trying to break free from NVIDIA's monopoly on AI chips. Google has TPUs, Amazon built Trainium, and even Tesla made their own Dojo chips because they got sick of waiting for NVIDIA allocation.

NVIDIA controls over 90% of the AI chip market, which means they can charge whatever they want. And they do - H100s cost $40k+ each when you can actually get them. No wonder everyone's looking for alternatives.

Why Broadcom Might Actually Pull This Off

Unlike OpenAI, Broadcom isn't new to this game. They've been designing custom chips for decades and actually know what they're doing. Here's what they bring:

• 2.5D and 3D packaging - basically stacking chips to make them faster and more efficient
• Custom ASIC experience - they've built specialized chips for Google, Amazon, and other giants
• Manufacturing connections - they know the people at TSMC and Samsung who actually make this stuff
• Interconnect tech - the plumbing that connects multiple chips together without creating a bottleneck

The key difference is Broadcom has successfully shipped custom silicon before. They're not some startup with big dreams and no experience.

The Great AI Chip Gold Rush

Here's the pattern: every company that gets big enough in AI eventually says "fuck it, we'll make our own chips." Why? Because NVIDIA's pricing is insane and their GPUs aren't actually optimized for most AI workloads.

Google's TPUs absolutely destroy GPUs for training large language models. Amazon's Trainium costs a fraction of equivalent GPU compute. Tesla's Dojo is specifically designed for computer vision training.

The problem is NVIDIA's CUDA lock-in. Fifteen years of software development means most AI frameworks are built around CUDA. Switching to custom chips means rewriting tons of code, which is why most companies stick with overpriced GPUs. We spent 6 months porting our training pipeline to AMD's ROCm only to hit random HIP_ERROR_INVALID_VALUE crashes that took another 3 months to debug.

But here's the thing - if you're spending hundreds of millions on chips anyway, the cost savings from custom silicon pay for the software rewrite pretty quickly.

What This Really Means

NVIDIA's monopoly is starting to crack. Not because their chips suck - they're still excellent - but because they're pricing themselves out of the market for anyone with serious scale.

The companies building custom chips aren't trying to compete with NVIDIA on performance. They're optimizing for their specific workloads at a fraction of the cost. And for most AI applications, that's more than good enough.

If OpenAI's chips work even half as well as planned, other AI companies will follow. Nobody wants to keep paying NVIDIA's ransom when alternatives exist.