''' I’m going to tell you something that might sound crazy. The software-is-eating-the-world mantra? It’s missing the bigger picture. For the last decade, we’ve all been obsessed with models, algorithms, and the elegant dance of code. But behind the curtain, the real kingmaker, the silent giant calling the shots, is the hardware. I’ve seen it. I’ve lived it. And I’ve got the scars to prove it.
My journey into the world of custom silicon wasn’t a strategic pivot or a visionary leap. It was a street fight. It started with a simple, brutal reality: we couldn’t get our hands on enough GPUs. We were building a new AI-powered product at RemoteTeam, and our entire roadmap was grinding to a halt because of a global chip shortage. We were at the mercy of NVIDIA, and it felt like being a peasant begging for scraps from a king.
That frustration, that feeling of being powerless, is what pushed us to do something radical. We decided to build our own AI chip. Everyone told us we were insane. "You’re a software company," they said. "You have no idea what you’re getting into." They were right. We didn’t. But we did it anyway. And the lessons we learned along the way were more valuable than any funding round or press hit. Here are 15 of them.
1. The GPU Shortage Is Not a Bug, It’s a Feature
First, you have to understand that the GPU shortage isn’t a temporary supply chain hiccup. It’s the new normal. The demand for compute is growing exponentially, and the supply is, and will remain, constrained. This isn’t just about NVIDIA’s dominance; it’s about the fundamental physics and economics of chip manufacturing. Building a new fab costs billions and takes years. So, if your business model depends on an infinite supply of cheap GPUs, you’re building on quicksand.
2. Cloud AI Is a Gilded Cage
AWS, Google Cloud, Azure… they sell you a dream of infinite scalability. It’s a beautiful dream, until you get the bill. The margins on cloud AI are brutal. We were spending hundreds of thousands of dollars a month on GPU instances, and our costs were scaling faster than our revenue. It’s a gilded cage. You get the convenience, but you trade away your margins and your control. For us, the cost of cloud AI was the single biggest threat to our profitability.
3. Your Biggest Enemy Is Latency
Everyone focuses on training, but the real battle is in inference. That’s where the user experience is won or lost. And in the world of inference, latency is the enemy. We were building a real-time collaboration tool, and we needed responses in milliseconds, not seconds. The round trip to a cloud data center was killing us. We had to bring the compute closer to the user, and that meant edge AI.
4. Edge AI Is Not a Buzzword, It’s a Necessity
Bringing AI to the edge isn’t just about speed. It’s about privacy, reliability, and cost. When your data doesn’t have to leave the device, you have a much stronger security and privacy story. When your app can run offline, it’s more reliable. And when you’re not paying for every inference call to the cloud, your unit economics start to make a lot more sense. For us, edge AI was the only way to build a sustainable business.
5. Don’t Boil the Ocean
When we first started talking about building our own chip, we had grand visions of a general-purpose AI accelerator that could do everything. That was a mistake. The key to custom silicon is to focus on a narrow set of tasks and optimize the hell out of them. We weren’t trying to build a better GPU than NVIDIA. We were trying to build a chip that was 10x better at running our specific models. That focus was our superpower.
6. The Right Team Is Everything
Building a chip is a completely different ballgame than building software. You need a team of specialists—experts in chip design, verification, and fabrication. We had to go out and recruit a team of absolute rockstars, people who had spent their entire careers building silicon. It was a huge investment, but it was the only way to de-risk the project.
7. The First Rule of Custom Silicon: Don’t Talk About Custom Silicon
We made a conscious decision to keep the project under wraps for as long as possible. We didn’t want to alert the competition, and we didn’t want to set unrealistic expectations with our investors and customers. We called it "Project Chimera," and we treated it like a skunkworks project. That stealth was critical to our success.
8. Your Best Friend Is the FPGA
Before you tape out a chip, you need to be damn sure it’s going to work. That’s where FPGAs (Field-Programmable Gate Arrays) come in. We used FPGAs to emulate our chip design and run our models in a real-world environment. It was a slow and painful process, but it allowed us to find and fix a ton of bugs before we committed to silicon.
9. The Second Rule of Custom Silicon: Don’t Run Out of Money
Building a chip is expensive. Really expensive. We had to raise a dedicated round of funding just for Project Chimera. And even then, we were constantly looking for ways to save money. We used open-source tools, we negotiated hard with our vendors, and we were incredibly disciplined about our spending. We knew we only had one shot at this, and we couldn’t afford to screw it up.
10. The Tape-Out Is Just the Beginning
Getting to tape-out is a huge milestone, but it’s not the end of the journey. It’s the beginning of a whole new world of pain. You have to bring up the chip, test it, and integrate it into your product. It’s a long and grueling process, and it’s where most custom silicon projects fail.
11. The Software Is Harder Than the Hardware
This was the biggest surprise for us. We thought building the chip would be the hard part, but it turned out that building the software stack to run on the chip was even harder. We had to build a compiler, a driver, and a whole new set of libraries. It was a massive undertaking, and it required a completely different set of skills than we had in-house.
12. The Payoff Is Worth It
After all the pain and suffering, was it worth it? Absolutely. Our custom chip gave us a 10x improvement in performance and a 100x improvement in cost. It became our single biggest competitive advantage. It allowed us to build a product that was faster, more reliable, and more profitable than anything else on the market.
13. Don’t Believe the Hype
There’s a lot of hype around AI right now, and it’s easy to get caught up in it. But the reality is that most of the "breakthroughs" you read about are just incremental improvements on existing technology. The real breakthroughs are happening at the hardware level. That’s where the next generation of AI companies will be built.
14. Quantum Computing Is a Mirage
I get asked about quantum computing all the time. My answer is always the same: it’s a mirage. It’s a fascinating area of research, but it’s not going to have a meaningful impact on the AI industry for at least another decade. Don’t get distracted by the shiny new toy. Focus on the here and now.
15. The Future Is Custom
The era of general-purpose computing is over. The future is custom. As AI models become more specialized, the hardware that runs them will need to become more specialized too. The companies that understand this, the companies that are willing to invest in custom silicon, are the ones that will win in the long run. It’s not an easy path, but it’s the only path that leads to true, sustainable differentiation.
Building our own AI chip was the hardest thing I’ve ever done in my career. But it was also the most rewarding. It taught me that the biggest risks often lead to the biggest rewards. And it taught me that sometimes, the only way to win is to change the game. '''
Frequently Asked Questions
How were these items selected?
Each item on this list comes from direct experience, either from building my own companies or from patterns I've observed across the 200+ startups I've invested in. I prioritize practical, actionable items over theoretical concepts.
How do I know which items apply to my situation?
Start by honestly assessing where your biggest bottleneck is right now. The items that address that specific constraint will give you the highest return on your time and energy.
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.
Can I implement all of these at once?
I'd strongly recommend against it. Pick the 2-3 items that resonate most with your current situation and focus there. Trying to do everything simultaneously is a recipe for doing nothing well.