Comma.ai just dropped a blog post that's making every founder who's ever winced at an AWS bill feel personally attacked: they built and run their own $5 million data center. In their office. And they claim it's saved them $20 million compared to cloud.

Before you dismiss this as "cool but not for me," hear them out. Because the underlying lesson isn't about buying racks of GPUs — it's about a trap most startups sleepwalk into.

The Cloud Trap Nobody Warns You About

George Hotz, Comma's founder, puts it bluntly: "Cloud companies generally make onboarding very easy, and offboarding very difficult. If you are not vigilant you will sleepwalk into a situation of high cloud costs and no way out."

Sound familiar? It should. This is the same playbook we've seen from every platform that's ever eaten a startup alive. Cheap or free to start, increasingly expensive once you're dependent, and migration costs that make staying feel like the "rational" choice.

The cloud providers aren't stupid. They know that once your data lives on their servers, once your engineers are trained on their APIs, once your infrastructure-as-code assumes their services exist — you're locked in. Not legally, but practically.

The Real Cost of "Don't Worry About Infrastructure"

Here's what VCs and accelerators don't tell you when they say "just use AWS and focus on your product": you're not renting compute. You're renting optionality. And you're paying a premium for it.

Comma's math is brutal:

  • Their data center cost: ~$5 million total
  • Equivalent cloud spend: $25+ million (and counting)
  • Power costs: $540K/year (high, but predictable)
  • Result: 600 GPUs, 4 petabytes of storage, complete control

Now, does every startup need 600 GPUs? Obviously not. But every startup does need to ask: at what point does renting stop making sense?

The Control Argument

Comma makes a point that resonates beyond just cost savings: "If your business relies on compute, and you run that compute in the cloud, you are putting a lot of trust in your cloud provider."

They're not wrong. We've seen cloud providers:

  • Change pricing with minimal notice
  • Deprecate services that startups built their entire stack around
  • Introduce competing products using insights from hosted workloads
  • Suspend accounts over billing disputes or vague TOS violations

When your entire company runs on someone else's infrastructure, you're one policy change away from an existential crisis.

What This Actually Means For You

Let's be clear: this isn't "everyone should build a data center." That's absurd. Most startups don't need 450 kilowatts of compute power.

But the principle scales down. Ask yourself:

1. What's your cloud concentration risk? If AWS/GCP/Azure raised prices 50% tomorrow, what happens to your runway? Your margins? Your ability to raise?

2. Do you know your actual compute costs? Not the blended average your finance person reports. The real, per-unit economics of your core workloads. Many founders are shocked when they actually calculate this.

3. What's your exit strategy? Not from the company — from your cloud provider. Could you move if you needed to? How long would it take? What would break?

The Uncomfortable Math

Comma pays 40 cents per kWh in San Diego — triple the national average — and they're still saving money versus cloud. That tells you something about cloud margins.

The standard advice is "cloud when you're small, own when you're big." But Comma isn't that big. They're a self-driving car company with hundreds of employees, not a FAANG. They just did the math early and made a different choice.

More startups should at least run that math. You might not build a data center. But you might discover that a few colocated servers, or a hybrid setup, or even just reserved instances instead of on-demand, could extend your runway by months.

The Bottom Line

The cloud is a tool, not a religion. Comma.ai's blog post is a reminder that "best practices" often mean "best for the vendor's practices."

You probably won't build your own data center. But you should absolutely understand what you're paying for — and whether the convenience premium is actually worth it at your scale.

Because "we'll optimize infrastructure later" is how startups end up spending more on AWS than on payroll. And by then, the switching costs make "later" feel like "never."