Open-weight models win hearts, minds, and wallets

Open-weight models’ share of Vercel AI Gateway spending more than quadrupled from 5% in June to nearly 25% in September, while they processed over 70% of tokens; gateway spending was about 13x December 2025 levels, but average token prices were 47% lower. Chinese models gained usage, though more than two-thirds of September spending on them went to US providers, versus under 10% in December 2025. China proposed adding 40 technology items, including 23 AI-related items, to its export control list, raising uncertainty for US businesses hosting or relying on Chinese models.

Analysis

The investable signal is not simply rising open-model adoption: it is a shift in bargaining power. Falling token prices alongside much higher gateway activity suggest demand is responding to cheaper inference, but may also cap revenue per unit and put pressure on proprietary-model pricing. Gateways and application developers gain leverage from easy model switching; model providers face weaker lock-in and a greater need to differentiate on quality, reliability, or tooling. The Vercel sample is selective, however, and cannot establish market-wide share or provider economics.

The China exposure is asymmetric. US businesses hosting Chinese models may capture near-term usage without owning the underlying model, yet their revenues could become vulnerable to licensing restrictions imposed by Beijing as well as US policy. For Alibaba (BABA), Qwen’s reported threshold relevance is a contingent policy risk, not evidence that a specific model or Alibaba revenue stream is currently restricted. No direct financial sensitivity is provided.

Near term, policy headlines can raise uncertainty around US access to Chinese models. Over 1–3 months, the key catalysts are the final export-control scope, implementation and licensing terms, and evidence of customer migration. Over 6–18 months, continued model portability could structurally shift value toward routing, hosting, and application layers while compressing model-provider pricing. The contrarian point: team adoption can look dramatic while spending remains concentrated elsewhere; token share and paid-team penetration are not revenue share. Treat the gateway data as a directional competitive signal, not a basis for sizing company-level exposure.

Key Decisions for Investors

• No standalone BABA trade on this evidence. Keep Qwen-related exposure on watch; verify whether any proposed controls explicitly cover Qwen, whether licenses are curtailed, and whether Alibaba discloses material overseas usage or revenue exposure.
• For the next 1–3 months, monitor US-hosted inference and routing providers as potential relative beneficiaries of model switching, but require evidence that gateway growth converts to retained revenue and attractive unit economics before initiating a position.
• Avoid paying up for model-provider growth on token or team counts alone. Track spend share, realized pricing, and retention: continued falling prices without offsetting paid usage or revenue growth would falsify the bullish inference-demand thesis.
• Reassess promptly if China publishes binding restrictions affecting model licensing, or US rules limit hosting or use; those events could reverse near-term usage gains for US intermediaries. Conversely, clear licensing protections and continued customer migration to open models would weaken the policy-risk case.