Together AI token volume surges 13,000x as enterprises swap closed models

The massive migration to open-weight models signals a structural pivot in the AI stack toward infrastructure that prioritizes data control and IP protection. For streaming and video platforms, this enables the deployment of high-volume agentic workloads—such as automated metadata tagging or personalized content recommendations—without the uncapped cost exposure of proprietary APIs. As enterprises decouple intelligence from specific vendors, competition shifts from raw model capability to the efficiency of the inference layer. Watch for whether closed-source providers respond by slashing API pricing to defend their volume share as open-weight models reach benchmark parity.

The shift toward open-weight models is reshaping global token consumption patterns. Per the Vercel AI Gateway Production Index (July 2026), open-weight models grew their share of total processed tokens from 11% in April to 29% in June 2026. This volume surge is increasingly dominated by efficient Chinese architectures; data from OpenRouter and Andreessen Horowitz (July 2026) shows that Chinese open-weight models such as DeepSeek V3 and Xiaomi’s MiMo V2 Pro now account for over 45% of total tokens on major aggregators, a steep climb from under 2% a year prior.

Simultaneously, enterprise leaders are expressing heightened anxiety over 'data communism' in closed systems. Per Forbes (July 2026), Palantir CEO Alex Karp sparked market debate by arguing that frontier labs 'steal business alpha' by training on sensitive enterprise data. This sentiment has formalized into a 'Sovereign AI' movement, highlighted at the RAISE Summit in Paris (July 2026). At the event, European policymakers and French President Emmanuel Macron urged independence from Silicon Valley's closed ecosystems to ensure local data residency and long-term IP security.

Infrastructure providers are responding to this demand by introducing more predictable capacity models. Together AI launched its 'Provisioned Throughput' service in July 2026, offering reserved inference capacity with a 99% uptime SLA to bridge the gap between serverless convenience and dedicated hardware. According to Startuphub.ai (July 2026), this allows enterprises to run frontier-quality open models at scale with costs potentially 90% lower than high-end proprietary models like Claude Opus 4.8, further accelerating the volume migration away from closed APIs.