Half the Fortune 500 already run on Hugging Face’s open models. That’s not open source having a moment. That’s open source becoming default infrastructure.
Opus costs roughly 23x more per token — so it wins on revenue while losing on volume.
Hugging Face CEO Clem Delangue has watched the same pattern enough times to name it: companies start on frontier APIs, then costs push them toward open source as they scale. It’s a real number backed by a real mechanism, not a trend piece.
This isn’t a fight between open and closed models, though. It’s a lifecycle. Frontier labs own discovery. Open source increasingly owns production. Those are two different jobs, and conflating them is why most “open source vs. closed” takes, miss what’s actually happening.
The data backs this up better than the sentiment does:
- On Vercel’s AI Gateway, DeepSeek now carries over a third of total token volume.
- Anthropic still captures more than half of total spend, despite raising prices.
- On OpenRouter, DeepSeek V4 Flash processes 5.3 trillion tokens a week against Opus 4.8’s 2 trillion.
- Opus costs roughly 23x more per token — so it wins on revenue while losing on volume.
That’s not a contradiction. That’s the lifecycle working exactly as you’d expect. Mature, cost-sensitive deployments migrate to cheaper open source once the use case is proven. New use cases keep showing up needing a frontier model’s ceiling first. Discovery never stops generating fresh demand for the expensive tier — which is why the “coffee beans to Starbucks” commoditization story keeps getting predicted and keeps not fully arriving.
What actually changed my mind reading this: I’d been treating “open source vs. proprietary” as a market-share question, winner and loser. It’s not. It’s a maturity question, and every company answers it twice — once when they pick a model to start with, and again when they decide it’s time to graduate.
Delangue’s other worry is worth sitting with too: even in an “open” ecosystem, a handful of large companies could end up controlling the infrastructure everyone else depends on. Open weights don’t automatically mean open power.
If you’re choosing a model today, don’t treat the choice as permanent. Plan the migration path before you need it — know what “time to graduate to open source” looks like for your use case before cost forces the decision on you.
— –
Sources
Hugging Face’s CEO on why companies are done renting their AI — TechCrunch
Open source AI matters more than ever, according to Hugging Face’s Clem Delangue (podcast) — TechCrunch
Why the rise of open source AI isn’t hurting Anthropic … yet — TechCrunch