Skip to content
Rush Commerce
AI & Automation3 min read

Open-weight models: 6.1% adoption, billions in deals

Ramp puts open-source model platforms at 6.1% of AI-using businesses while Nvidia and Stripe buy the layer. Why you probably should not self-host yet.

The money going into open-weight models and the usage of them are not close to the same size. Ramp's spend data puts open-source model platforms at 6.1% of businesses that buy AI. Meanwhile Nvidia is reportedly paying $13 billion for Hugging Face. Somebody is buying a floor that almost nobody is standing on yet.

What actually happened

The Ramp AI Index published August 12 reports that adoption of model-serving platforms offering open-source and Chinese-developed models reached 6.1% of AI-using businesses, up 0.2 points month over month. Ramp's read is that this is not a switch — existing OpenAI and Anthropic customers are not leaving. It is high-volume spenders adding open weights next to the proprietary models they already run.

TechCrunch's August 28 piece stacks the deal flow against that number: Nvidia's reported $13 billion for Hugging Face, a $6 billion agreement with Poolside, Stripe's acquisition of OpenRouter. It also cites Jellyfish measuring open-weight use at about 2% of software engineers.

So: two percent of engineers, six percent of businesses, and tens of billions in acquisitions. That is not a demand signal. It is a bet on where inference cost lands in three years.

Why open-weight adoption matters for your business

At your volume, self-hosting is a cost increase. Open weights get cheap when you are burning enough tokens to keep GPUs saturated. Below that line you are paying for idle capacity, an ops rotation, and an upgrade treadmill to save cents on a bill that is already small. The 6.1% figure is heavily weighted toward advanced spenders for a reason.

The acquisitions still tell you something. When the chip vendor buys the model registry and the payments company buys the router, they are pricing a future where switching models is routine and the margin sits in routing and serving. Every one of those deals assumes portability becomes normal. Build like that is true.

Keep the exit cheap instead of taking it early. Route through an abstraction you control, keep prompts and evals in your repo, and mirror any weights you actually depend on. Then self-hosting becomes a decision you make on a spreadsheet when your volume justifies it — not a rebuild.

Key takeaways

  • Ramp: open-source model platforms are at 6.1% of AI-using businesses, up 0.2 points month over month
  • Jellyfish, cited by TechCrunch, measures open-weight use at roughly 2% of software engineers
  • Nvidia reportedly agreed to $13B for Hugging Face and $6B with Poolside; Stripe bought OpenRouter
  • Ramp says growth is additive from heavy spenders, not defection from OpenAI or Anthropic
  • Self-hosting pays off at sustained volume, not at small-business token counts
  • Preserve the option: abstraction layer you own, prompts and evals in your repo, mirrored weights

Self-hosting a model is a decision, not a default — and most teams asking us about it would spend more, not less. We size the actual token math before anyone provisions a GPU, then build the routing layer so the choice stays open. Run the numbers on your AI spend, or ask us whether self-hosting is worth it at your volume.

Sources: Ramp AI Index, TechCrunch.

  • #open-weights
  • #self-hosting
  • #inference-cost
  • #ai-spend
  • #model-portability
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.