I’m going to tell you something that might sound crazy. The GPU shortage was the best thing that ever happened to AI.
Everyone was panicking. I had founders calling me at all hours, completely freaking out because they couldn’t get their hands on enough H100s. Their burn rates were skyrocketing, their product roadmaps were slipping, and some were on the verge of collapse. And I get it. When your entire business depends on these little boxes of magic, seeing the supply dry up feels like a death sentence.
But from where I sit, after two exits and over 200 angel investments in companies like Anthropic and OpenAI, I saw something different. I saw a culling. A necessary correction. For too long, the default answer to any AI problem was "throw more GPUs at it." It was a lazy, expensive, and ultimately unsustainable way to build.
The Great GPU Reckoning
I remember one of my portfolio companies—a brilliant team with a fantastic product—was spending almost 70% of their seed round on cloud GPU instances. Seventy percent! It was insane. They were essentially lighting money on fire to train a model that was only marginally better than what they could have achieved with a more optimized approach. They were so focused on the model, on the software, that they completely ignored the hardware stack.
When the shortage hit, they were forced to get smart. They had to actually think about efficiency. They started optimizing their code, exploring different model architectures, and even looking at more specialized hardware. It was painful, for sure. There were a lot of sleepless nights. But in the end, they came out stronger. Their burn rate dropped by 50%, and their inference speed tripled. The crisis forced them to build a better, more resilient company.
This is what I mean when I say the shortage was a blessing. It woke people up. It reminded us that hardware isn't just a commodity you rent from the cloud. It's a fundamental part of the AI stack, and if you ignore it, you're building your house on sand.
The Siren Song of Custom Silicon
The GPU reckoning led directly to the next big wave: the rush to custom silicon. Suddenly, every well-funded startup wanted to design its own AI chip. It became a status symbol, a way to signal to investors that you were a "serious" AI company. And on paper, it makes a ton of sense. Why pay a premium for a general-purpose GPU when you can build a chip that does exactly what you need, only faster and cheaper?
This is a lesson I learned the hard way at my first company, MovieLaLa. We were doing massive-scale video analysis, and the costs of our AWS bill were astronomical. We spent months trying to optimize our algorithms, but we kept hitting a wall. The hardware just wasn't built for our specific workload. We made the crazy decision to build our own ASIC. This was long before companies like Groq or Cerebras made it look easy. It was a brutal, multi-year effort that nearly bankrupted us twice.
I’ll never forget the day our first wafers came back from the fab. We had spent over $10 million, a huge chunk of our venture funding, and we had no idea if they would even work. I remember huddling with my co-founders in our tiny lab, the smell of stale pizza and solder in the air, as we brought up the first chip. The initial tests were a disaster. Nothing worked. It was one of the lowest points in my career. We eventually figured it out—a stupid bug in the power delivery network—but it was a stark reminder that chip design is not for the faint of heart. It’s a high-stakes, unforgiving game.
Today, things are easier. You have open-source instruction sets like RISC-V and a whole ecosystem of design tools. But the fundamental challenge remains: building hardware is hard. It requires a completely different skillset than software, and the iteration cycles are painfully slow. You can’t just push a patch. A mistake in silicon is a multi-million dollar paperweight.
The Next Frontier: The Edge
For all the talk about massive, datacenter-scale AI, I’m convinced the next revolution is happening at the edge. And it’s not just about running smaller models on your phone. It’s about rethinking the entire architecture of AI applications.
I have an investment in a company that’s doing real-time defect detection on a manufacturing line. They were initially sending a video stream to the cloud for processing. The latency was a killer. By the time the model identified a faulty part, it was already ten steps down the line. The cost of data transmission was also a huge problem.
They made the switch to an edge device—a small, ruggedized computer with a specialized AI accelerator right on the factory floor. The results were stunning. Latency dropped to near zero. They were catching defects in milliseconds, saving the company millions in wasted materials. And because the data was processed locally, their cloud bill practically disappeared.
This is the future. We’re moving from a centralized intelligence model to a distributed one. It’s more resilient, more secure, and ultimately, more powerful. But it requires a new way of thinking about hardware. You can’t just stick a power-hungry GPU in a factory. You need low-power, highly efficient chips that can operate in harsh environments. This is where the real innovation is happening right now, in the unglamorous world of industrial and embedded AI.
Quantum's Long Shadow
No conversation about the future of computing is complete without mentioning quantum. I get asked about it all the time. Is it going to make all our classical AI hardware obsolete? My answer is always the same: not anytime soon.
I’ve looked at a lot of quantum computing startups. I’ve even invested in a few. The technology is fascinating, and the long-term potential is mind-boggling. But we are still in the very early innings. The fundamental challenges of building a fault-tolerant quantum computer are immense. We're still figuring out the basic physics, let alone how to manufacture these things at scale.
Think of it this way: classical computing is like building with LEGOs. We have standardized components, reliable manufacturing processes, and a deep understanding of how to put them together. Quantum computing is like trying to build with soap bubbles. It’s incredibly fragile, highly sensitive to its environment, and we’re still trying to figure out how to stop the bubbles from popping.
Will we get there? I think so. But it’s going to take a lot longer than the hype would have you believe. For the next decade, at least, the most significant breakthroughs in AI will continue to be driven by advances in classical hardware—smarter architectures, more efficient designs, and a relentless focus on the interplay between hardware and software.
The Real Counterintuitive Truth
So what’s the counterintuitive truth about AI chip design? It’s that it’s not about the chips. Not really. It’s about the whole system. It’s about the software, the data, the algorithms, and how they all come together on a piece of silicon.
I’ve seen too many companies fail because they had the “best” chip but didn’t understand their users’ problems. I’ve seen others succeed with “inferior” hardware because they built a product that people loved.
My advice is simple: don’t get distracted by the shiny objects. Don’t chase the latest and greatest GPU or pour all your money into a custom chip you don’t need. Start with the problem. Obsess over your users. And then, and only then, think about the hardware you need to solve that problem. The most powerful AI hardware is useless if you’re not building something that matters.
That’s the lesson I’ve learned over and over again, from my own startups to the hundreds of companies I’ve backed. It’s not about having the most powerful hardware. It’s about having the right hardware for the job. And that’s a truth that no amount of hype can change.
Related Investments
Sahin Boydas is an angel investor in these companies mentioned in this article:
Frequently Asked Questions
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.