Large model price wars enter the era of token wholesale
In October 2026, China's large model market saw a new wave of price cuts. After several vendors slashed inference costs to single-digit yuan per million tokens earlier this year, multiple leading companies this month announced “token wholesale” plans, allowing enterprise customers to lock in lower unit prices through annual prepayment. This model directly spurred a race for the “cheapest token” — some small and mid-sized model providers even quoted aggressive prices of less than 0.5 yuan per million tokens. Notably, this wholesale model goes beyond simple discounts, bundling usage commitments, priority scheduling and customised fine-tuning into a single package. Industry analysts believe that when tokens are wholesaled like water and electricity, model inference shifts from pay-per-use to resource pooling, significantly boosting the utilisation of underlying infrastructure.
AI API hubs: invisible pipes connecting supply and demand
As multiple vendors open their API interfaces, enterprises face a dilemma: the cost-effectiveness of different models varies significantly for specific tasks. This has given rise to a wave of specialised AI API hubs. These hubs aggregate APIs from multiple large model providers, offering a unified access point, intelligent routing and usage management. For example, a leading AI hub platform announced in October that its token hub processed over 1 billion daily calls, helping enterprises reduce average inference costs by 35% through dynamic model switching. Additionally, these hubs provide token brokerage services, allowing customers to obtain more stable connections and transparent billing without directly connecting to original vendors. Such platforms, essentially “model routing + billing aggregation,” are becoming a critical component of enterprise AI infrastructure.
Policy and capital drive AI channel agents as a new track
This month, China's National New Generation Artificial Intelligence Innovation and Development Pilot Zone issued new guidelines, explicitly encouraging the establishment of third-party model service platforms and channel agent systems. Spurred by this, a number of former cloud computing distributors began transitioning into AI channel agents, shifting their role from selling compute to selling “model capabilities + operational services.” These AI large model agents not only help SMEs quickly select and deploy models, but also offer secondary pricing capabilities based on token wholesale agreements. Capital markets are active too: in mid-October, a startup focused on operating an AI API hub completed its Series B round at a valuation exceeding $2 billion, with investors including several industry funds. Industry consensus holds that the channel agent model solves the “last mile” adaptation problem for large models, particularly meaningful for SMEs and traditional industries.
Behind the cheapest token: the changing cost structure
Is the “cheapest token” sustainable? On the technical side, multiple vendors this month unveiled new inference acceleration schemes, including sparse attention, prefill-decode separation architectures, and KV cache optimisations for long-context scenarios. These advancements improve hardware utilisation by more than 20%, providing the foundation for price reductions. On the commercial side, however, token wholesale means vendors must rely on economies of scale to offset R&D investment — a challenge for startups with limited capital. Notably, some AI API hubs have begun offering value-added “token brokerage” services, such as automatically switching between different vendors' APIs to balance cost and latency, or providing prepaid discounts based on historical usage. These innovations are reshaping profit distribution rules across the entire chain.
Outlook: enterprise deployment enters a on-demand scheduling stage
Looking at the month's developments, the AI industry is shifting from a “model competition” to an “infrastructure efficiency competition.” Enterprises are no longer fixated on a single model's benchmark scores, but are more focused on obtaining the most appropriate capabilities at the lowest cost. The increased activity at AI hub token hubs and the expansion of channel agent networks point to a clear trend: future enterprise AI deployment will resemble electricity usage, leveraging smart hubs for multi-source supply and on-demand scheduling. Policy support for open-source ecosystems and third-party service platforms further lowers the barrier to entry. It is foreseeable that with the maturation of “token wholesale” and “AI API hub” models, a number of SMEs will complete their first large model deployment via channel agents before the end of 2026, while the industry competition focus will shift from model capability itself to building more efficient and inclusive delivery pipelines.