I'm Afraid to Take a Vacation

Some friends invited me on a trip to Europe. France, and a few countries in Eastern Europe.

I've always wanted to go to France a second time.

It was very enticing. Good people, a real break, the kind of trip you say yes to without thinking.

I didn't say yes. I kept thinking about our GPUs.

Eventually I decided: you know what, I might as well just stay here.

The Bill Arrives Anyway

We launched our AI effort at the end of June. It's been a few months.

We committed to these GPUs. Committed means the bill arrives every day, whether we use them or not.

Idle compute isn't free compute. It's the same bill with nothing to show for it.

In my fintech days, I couldn't have imagined burning this much money in a single day. When I think back to our biggest expenses then, it's compliance and marketing. Those felt big at the time. But they were bills. This is a meter that never stops.

And the technology underneath keeps changing fast. There's still a lot of infrastructure to build. Managing AI has quietly become a full-time job.

I want to be clear about one thing. I feel lucky to have this problem. Very few people get to run real compute as a founder-controlled, independent company. I take that seriously.

But taking it seriously is exactly why I couldn't leave.

Busy Is Not the Same as Useful

The easy fix for idle GPUs is to give them something to do.

I think that's a vanity metric.

You can keep every GPU busy and accomplish nothing. Filler work gives you activity to point to. It doesn't tell you whether the activity was worth doing. And I'd still be responsible for what it accomplished.

Filler is actually worse than idle. Idle makes me uncomfortable. It keeps the question in front of me: is there something better we should be running? Filler makes the discomfort go away without ever answering the question.

The times I've felt good about our compute, it was answering something.

When we evaluated Cyber 1.0, our cybersecurity fine-tune, the runs took a while. We wanted to know whether the model was actually good, or whether it was just a joke.

The answer was somewhere in between. We trained on a small number of samples, so it wasn't a big jump from the base model. But it proved a fine-tune could push a base model toward cybersecurity. It showed us what's possible, and what's not.

We also tried a different transformer architecture. That one failed. I wrote about it on the SimpleDirect blog.

One worked a little. One didn't work at all. Both were good uses of compute, because both answered a question we actually had.

That's my test now. Not "are the GPUs running?" But "can I say what they're for?"

GPUs Are a Currency

I'll be honest. I've doubted this bet at times.

I'm more committed to it now than ever.

I believe GPUs are becoming a currency. Not a replacement for dollars. Something different: a currency that can produce lasting value.

But only if its owners know how to spend it.

That's what changed how I see the idle problem. Idle compute is currency sitting unspent. Filler is currency spent on nothing. Neither is the goal. The goal is spending it on work that lasts.

Own or Be Owned

Here's the part that's uncomfortable to write.

I talk about Own or Be Owned constantly. Own your AI. Own your infrastructure. I say I'm the asset, not the company. I say time and freedom matter more to me than anything else.

Then my friends were planning a trip to France, and I was thinking about GPU utilization.

I own the compute. Right now, it owns my calendar.

And going wouldn't have fixed it. I'd have been in Paris checking in, wondering if the machines were doing anything worthwhile, wondering if I needed to step in. That's not a vacation. That's working remotely with a nicer view.

It wouldn't be fair to my friends either. They invited me. Not the half of me that's still thinking about the company.

What Has to Change

Being afraid to leave isn't a sign of commitment. It's a sign the system depends on me.

That's on me to fix. Not by working harder. By making decisions I don't need to be present for.

Someone other than me owns what runs on our compute.

A written list of work we've already agreed is worth doing, so nobody has to ask me.

A rule for idle time. If nothing meaningful is queued, it stays idle. No filler.

None of this is complicated. I just haven't done it yet.

I'm grateful for the position we're in, and I want to spend it well. I also want to be the kind of founder who can get on a plane.

I stayed home this time.

France will still be there. Next time, I want to go and actually be there.