Open Weights Is a CapEx Decision

As of Q4 2026, this is true. Who knows how things will change in the near future.

The week of October 5, I was putting together a model-choice briefing and wanted a slide on open weights, since it’s the hot topic in the industry right now. My AI chief of staff suggested making it a backup slide and leaving costs off. I overruled it, which is a sentence I never expected to write. If you leave costs off, there’s nothing to say. The question everyone is actually asking is whether to use open-weights models because they’re just as good and cheaper than the closed ones.

That question has a hidden variable, and it isn’t only quality. Open models have gotten good, but on the charts I was looking at they still trail the frontier. The variable that decides the purchase is volume.

The chart nobody draws

Most of the open-versus-closed debate runs on benchmark charts: one model here, another there, a bar for price per million tokens. What’s missing is a volume axis. Cheap per token only matters once you’re pushing a lot of tokens.

So I asked for three lanes on one chart:

• Closed frontier models on flat-rate seats. You pay per person, per month.
• Open weights hosted by someone else, billed per token. You pay for what you use.
• Open weights on your own GPUs. You buy the hardware up front, and that’s the CapEx lane.

Each lane has a different shape. Seats grow with headcount. Per-token grows with usage. Self-hosting is a big step up front and then mostly flat until you need the next rack.

The self-hosted step is bigger than people assume. In August I figured out loud that running a frontier-size open model would take 16 Nvidia B200s, which comes to about a million dollars once you add the rack and everything else. That’s the floor, because the model is that big, and I figured it covers about 60 heavy users. For a large company, fine. For a small team, that’s a lot of money spent so you can say you own your stack.

The cache-read tax

There’s a second cost the benchmark charts miss. Flat-rate seats quietly absorb heavy cached usage. Coding agents reread the same context over and over, and on a seat you don’t feel it. Per-token hosts do discount cache reads, sometimes steeply, but even at the cache discount an agent rereading the same context all day runs up a bill a seat never shows you. That’s a big part of why per-token hosting costs more in practice than the sticker price suggests.

Where I was wrong

Here’s my method showing, warts included. My gut said that above 60 or 70 developers the CapEx starts to look good. I said it with some confidence.

Then we ran the actual usage numbers. Self-hosting never beat flat-rate frontier seats at any team size we modeled. It beat only the per-token hosted lane, and only at a much smaller team size than I’d guessed. My instinct had the direction right and the size wrong, and the honest answer was stronger than my claim.

Halfway through, I changed my own axis. Headcount was the wrong variable. What matters is token volume: the average tokens per developer times the number of developers. Two teams of 50 can sit on opposite sides of the line if one uses AI like autocomplete and the other runs agents all day.

Open weights still looks like a quality fight on social media. On a cost chart with a volume axis, it reads as a capital decision, and I’d guess most companies are too small to make it.

Why this will age like milk

I’m not attached to any particular model. In August, Kimi K3 led the open-weights pack on the Artificial Analysis index we were using. By early October it was third, behind Xiaomi’s MiMo-V2.6-Pro and Z.ai’s GLM-5.3. Hardware prices move too, and so does seat pricing. The crossover point will shift, maybe a lot, maybe soon.

So don’t take my conclusion. Take the method:

1. Plot your own monthly token spend, not your headcount.
2. Put the three lanes on the same chart.
3. Count cache reads honestly.
4. Redo the math when prices change, because they will.

Where could this break? If open models get much smaller at the same quality, the hardware floor drops and the CapEx lane opens up to small teams. If seat vendors keep tightening the caps they’ve already put on heavy users, the flat-rate advantage goes away. Either could happen next quarter. If one does, I’ll redraw the chart, and I expect to be wrong about something on it.

For now, at our size, it doesn’t justify the cost. We’re staying on the closed frontier lane. What does your token volume say?