Open-weight AI models now run 54% of token volume on Vercel’s AI Gateway—up from 28% in June—driven by DeepSeek and Chinese labs. For European SMBs, the math is brutal: cost-efficient Chinese models are finally viable production alternatives to expensive US platforms, forcing an urgent recalculation of vendor lock-in and total cost of ownership.
Open-weight models hit 54% of token volume on Vercel’s AI Gateway this week, surpassing proprietary alternatives for the first time. DeepSeek alone now handles roughly a quarter of all gateway tokens—more than twice Google’s share. This isn’t a niche shift. It’s a fundamental restructuring of how production AI gets routed, and it hits your vendor economics harder than most headlines acknowledge.
The speed is what matters. In April, open-weight was 11% of volume. By June, 29%. Now, 54%. Meanwhile, Anthropic alone still captures 61–65% of all spending on the gateway despite processing only 30% of tokens. The gap is the whole story.
What does that gap mean? Companies are now running the majority of their token work on models that cost roughly one-tenth as much as proprietary alternatives, while keeping the high-stakes, highest-precision work on expensive frontier systems. That’s not indecision. That’s ruthless workload routing discipline.
The cost structure actually works for European SMBs
Imagine a 50-person logistics software firm in Stuttgart that built its AI strategy around Claude Opus for everything. Every internal process automation, every customer-facing summarization, every document classification ran through Anthropic’s API. Costs were predictable—per-token billing—but they added up quickly.
Now that same firm can run 70% of that workload on DeepSeek V4 Flash at about one-tenth the token cost, route the 20% that actually needs deep reasoning or rare edge-case handling to Claude, and keep the remaining 10% as a failover buffer. The total AI bill doesn’t drop to a tenth—because not everything can move—but it easily halves. For a 50-person company, that’s real cash.
Chinese labs trained these models under export control constraints. No H100 clusters; they worked with H800 chips and built sparse mixture-of-experts architectures that activate only a subset of parameters per token. DeepSeek trained its V3 model for roughly $5.6 million. US frontier labs routinely spend tens to hundreds of millions on a single model. That efficiency shows up in the API pricing: DeepSeek slashed V4-Pro prices by 75% in May. Chinese firm 01.ai prices inference at roughly 14 cents per million tokens—the lowest in the world.
The result? Chinese models now dominate cost-per-intelligence benchmarks, and they’re no longer fringe bets. Z.ai’s GLM 5.2, released in June, hit the top Vercel models within two weeks—growing 50x in daily token volume—all at roughly one-fifth the cost of Anthropic’s Opus 4.8.
What this actually changes for your procurement
First, vendor lock-in just became expensive. If your SMB is still on a single proprietary platform for all AI workloads, you’re now leaving money on the table every month. The economics don’t require you to abandon closed models—they require you to be intentional about which workloads land where.
Second, you have negotiating leverage you didn’t have six months ago. When open-weight was 11% of production, US vendors could afford to shrug. Now that it’s 54%, and growing faster than anyone predicted, they can’t. If your contract comes up for renewal, or if you’re evaluating a new AI vendor strategy, the existence of a credible, cheaper alternative changes the conversation entirely.
Third, and this matters more for European compliance: open-weight models run on your infrastructure mean your data stays in your jurisdiction. Depending on your industry—financial services, healthcare, legal—that’s not a nice-to-have. It’s a requirement. Proprietary APIs force data outside your borders; open-weights don’t.
But—and this is critical—the open-weight win is not universal. Closed models still capture the majority of spending for a reason: on the hardest reasoning tasks, multimodal work, and agentic coding at the edge of what’s possible, frontier models still hold an edge. You’re not replacing Opus on a mission-critical agent or on work where a mistake is expensive. You’re replacing it on the 60–70% of work where it’s overkill, and that’s where the money is.
Here’s what I’d push on: Don’t let this data sit as a curiosity. The moment open-weight hit 50% on Vercel, a threshold crossed. Every European SMB should spend an afternoon auditing which of your AI workloads actually require frontier capability, and which are just there because they were always there. That audit pays for itself the first month.
The vendor strategy that wins in 2026 isn’t “all open” or “all proprietary.” It’s routing discipline: cheap models for volume, expensive models for precision, and enough structural flexibility that you can shift workloads as new models arrive and prices collapse.
FAQ
Do I have to host open-weight models myself?
No. You can run them through managed hosting (modal, together.ai, lambda) which gives you the cost advantages without the operational overhead. You rent compute rather than licensing. It’s still cheaper than proprietary APIs for high volume, and you avoid the infrastructure hassle.
But will open-weight models stay available?
For now, yes—but there’s risk. China’s Ministry of Commerce is reportedly considering export controls on advanced Chinese AI models, similar to US moves on Anthropic. If restrictions land, weight downloads could become licensed rather than free. For European SMBs, the practical answer is: treat open-weight as a cost-reduction tool today, not a permanent replacement for proprietary. Keep your sourcing diverse.
How do I know if a workload can move to open-weight?
Start with high-volume, low-stakes work: document summarization, classification, routine extraction. These run well below frontier capability and pay for themselves immediately. Avoid moving work that requires rare reasoning, novel problem-solving, or where a mistake costs you. Test on a pilot workload first, measure latency and accuracy, then expand only if both are acceptable.
Will US vendors cut prices to compete?
They’re already moving. But the gap is hard to close. A frontier US model reflects years of scale, R&D, and integration investment. A Chinese open-weight model reflects efficiency and export constraints. The US can’t match the Chinese price without breaking their model economics. Instead, expect them to differentiate on capability (which they do have), integration (which they own), and enterprise relationships (which matter). For SMBs, that’s not enough.
Look at Alibaba’s Qwen strategy—they’re deliberately feeding the open-weight ecosystem as a way to build adoption and lock you into their broader stack. That’s not altruism. It’s a long-term vendor play. The difference is: at least with open-weight, you hold the weights and you’re not locked in by license. With proprietary, you’re locked in by architecture.
What to do Monday morning
1. Run a workload audit. Which of your AI applications are hitting expensive proprietary models? Which could run just as well on open-weight? Track token volume and cost per task.
2. Test one high-volume, low-risk workload on an open-weight alternative. Use a managed host (not self-hosted) to avoid infrastructure overhead. Measure cost, latency, and accuracy.
3. Renegotiate any vendor contracts due in the next six months. The fact that 54% of production workloads now run on open-weight changes your leverage. Use it.
4. Don’t abandon proprietary entirely. Keep closed models for mission-critical, high-stakes reasoning work. The hybrid strategy is where the real savings are.
I write daily on AI strategy and procurement decisions at RDZ Media on LinkedIn. If you’re building procurement strategy or evaluating AI infrastructure, I’d rather you push back on my takes than take them at face value. Come argue.
0 Comments