Everyone talks about AI models, but nobody talks about the brutal reality of the hardware that runs them. I get it. Software is sexy. Models that can write poetry or generate photorealistic images are easy to get excited about. But behind every single one of those models is a mountain of silicon, and that mountain is where fortunes are made and lost.
I’m Sahin Boydas. You might know me from RemoteTeam, which we sold to Gusto, or from my first startup, MovieLaLa. Or maybe you’ve seen my name on the cap table of companies like Anthropic, OpenAI, or Scale AI. I’ve spent the last twelve years in the trenches of Silicon Valley, first as a founder and now as an investor. I’ve seen firsthand how the right AI hardware can be a company’s secret weapon, and how the wrong hardware strategy can kill it before it even gets off the ground.
I’m writing this because I’m tired of the high-level, hand-wavy conversations about AI infrastructure. I want to share the hard-won lessons and contrarian insights I wish I had when I started. This is the unfiltered truth about what it really takes to build and scale in the age of AI.
The Great GPU Lie
Let's start with the elephant in the room: the GPU shortage. For the past few years, it feels like the entire industry has been in a desperate scramble for NVIDIA chips. I’ve had founders call me in a panic, telling me their Series A funding was contingent on securing a cluster of H100s. It’s madness. We’ve created a situation where a single company has a chokehold on the entire AI ecosystem.
I remember one of my portfolio companies, a brilliant team working on protein folding simulations, was on the verge of a breakthrough. They had the algorithm, they had the data, but they were stuck. They spent six months on a waiting list for GPUs. Six months! In startup years, that’s an eternity. They almost died. We had to pull every string I had, call in every favor, just to get them the hardware they needed to survive. They’re thriving now, but it was a terrifyingly close call.
Here’s the thing everyone is missing: the GPU shortage isn’t the disease, it’s a symptom. The real disease is our industry’s monolithic thinking. We’ve become so dependent on a single type of architecture that we’ve forgotten how to innovate our way out of problems. We’ve optimized for convenience, and now we’re paying the price in fragility.
Founders need to think differently. Are you sure you need the biggest, most expensive GPU on the market? Have you explored more specialized hardware? Have you optimized your models to run on less powerful chips? The answer isn't always to throw more hardware at the problem. Sometimes, the answer is to be smarter about the hardware you already have.
The Custom Silicon Minefield
This brings me to my next point: custom silicon. The idea is seductive. If off-the-shelf hardware is the problem, why not build your own? Companies like Google with their TPUs and Amazon with Inferentia have shown that it can be done. They’ve built custom chips that are perfectly tailored to their workloads, giving them a massive performance and cost advantage.
It sounds great, right? But I’m here to tell you that the road to custom silicon is paved with the bodies of dead startups. It is an incredibly difficult, expensive, and risky path. You’re not just designing a chip; you’re building a team, a supply chain, and a software stack to support it. It’s a multi-year, nine-figure commitment. I’ve seen more than one company with a brilliant chip design fail because they underestimated the complexity of the software.
I was an early advisor to a startup—I won’t name them—that was building a chip for natural language processing. The architecture was revolutionary. On paper, it was 10x faster and 20x more power-efficient than anything on the market. They raised a huge seed round. But they got bogged down in the compiler. They couldn’t get the software to a point where developers could actually use the chip. After three years and $50 million, they had a very expensive piece of silicon that did nothing. They quietly folded.
My advice to founders is to be brutally honest with yourselves. Do you have the expertise, the capital, and the stomach for a multi-year slog with a high probability of failure? For 99% of companies, the answer is no. You are far better off focusing on your core product and finding clever ways to use existing hardware. Don’t let the siren song of custom silicon lure your startup onto the rocks.
The Edge is Sharper Than You Think
So if massive GPU clusters are a trap and custom silicon is a minefield, where should we be looking? For me, the answer is clear: the edge. For years, the trend in AI has been towards bigger and bigger models running in massive, centralized data centers. I think that’s about to change.
The real world is messy and disconnected. A self-driving car can’t wait for a round trip to the cloud to decide whether to brake. A smart factory needs to detect defects in real-time, on the assembly line. A doctor needs a device that can analyze medical images instantly, without an internet connection. All of these applications require powerful, efficient AI that can run on small, low-power devices. This is edge AI.
At MovieLaLa, we were obsessed with recommendation speed. We learned that every 100-millisecond delay in showing a user a relevant movie trailer cost us a measurable drop in engagement. We had to push as much of our recommendation logic as possible onto the device itself. This was a decade ago, long before “edge AI” was a buzzword. We were just trying to build a better product.
Today, the opportunities are a thousand times bigger. I’m seeing incredible companies building AI-powered hardware for agriculture, for retail, for logistics, for healthcare. They’re building smart cameras that can detect crop diseases, handheld scanners that can identify counterfeit goods, and wearable sensors that can predict medical emergencies. This is the future. It’s not about building one giant brain in the cloud; it’s about distributing intelligence everywhere.
Don't Write Off the Data Center Just Yet
Now, does this mean the data center is dead? Absolutely not. Training the next generation of massive AI models will still require enormous amounts of computing power. But the data centers of the future will look very different from the ones we have today.
We are on a collision course with the laws of physics. The power consumption of AI data centers is doubling every few months. Some estimates suggest that by 2030, AI could consume as much electricity as the entire country of Japan. This is not sustainable. We can’t just keep building bigger data centers and hoping for the best.
We need to rethink everything, from the ground up. We need more efficient cooling systems, like liquid immersion. We need faster and more flexible networking, like optical interconnects. And we need to get serious about co-locating data centers with renewable energy sources.
I’m an investor in a company that’s building small, modular nuclear reactors specifically to power data centers. When I first heard the pitch, I thought it was insane. But the more I dug into the numbers, the more I realized they were onto something. It’s a bold, contrarian bet, and it might not pay off. But that’s the kind of thinking we need.
A Word on Quantum
I can’t write an article about the future of AI hardware without mentioning quantum computing. There is a lot of hype around quantum, and a lot of money being invested. My take is simple: it’s a science project.
Yes, the long-term potential is enormous. A fault-tolerant quantum computer could break modern cryptography and revolutionize drug discovery. But we are a long, long way from that. The quantum computers we have today are noisy, error-prone, and can only solve a very limited set of problems. They are fascinating research tools, but they are not a practical solution for running AI workloads.
If you’re a founder, don’t bet your company on quantum. If you’re an investor, be very, very skeptical of any startup that claims to be using quantum computing for AI. The timeline is just too long and the technical hurdles are too high. It’s a distraction from the real, here-and-now problems we need to solve in classical AI hardware.
The Real Work Begins
Building an AI company is hard. Building one that depends on a solid hardware strategy is even harder. You have to be a strategist, a supply chain expert, a hardware architect, and a software developer all at once.
The biggest lesson I’ve learned after all these years is that there are no easy answers. The right hardware strategy for your company will depend on your product, your team, and your budget. Don’t follow the hype. Don’t assume you need the biggest and best of everything. Think from first principles. Be skeptical. Be creative.
The next decade of AI will be defined not just by the models we build, but by the hardware we build them on. The real work is just beginning. It’s a brutal, complicated, and fascinating challenge. And I wouldn’t have it any other way.
Frequently Asked Questions
How long did it take to see results?
Most meaningful business results take 3-6 months to materialize. Anyone promising overnight success is selling something. The companies in my portfolio that grew fastest were the ones that stayed patient and consistent.
Can these results be replicated?
The specific numbers will vary, but the underlying patterns and principles are transferable. The key is understanding the context behind the results, not just copying the tactics. Every company has unique constraints that shape what works.
What would you do differently looking back?
I'd move faster on the things that were working and cut the things that weren't sooner. Most founders, myself included, hold onto failing strategies too long because of sunk cost. Speed of learning is everything.
What was the biggest challenge in this case?
Almost always, the biggest challenge is people and alignment, not technology or strategy. Getting the right team focused on the right problem is harder than any technical challenge I've encountered.