Bankless
@Bankless 发文指出,AI 行业正在经历一次关键转向:市场焦点不再只是训练更强模型,而是如何支撑越来越庞大的「推理」需求。随着 agent、大规模自动化工作流与 AI 产品爆发,「推理」正在成为新的核心资源。
@cerebras IPO 被 20 倍超额认购,也被视为市场押注「推理」时代正式到来的信号。
文章认为,
@AskVenice 正处于这一趋势的核心受益位置。Venice 通过 $VVV 与 $DIEM 两种代币,将 AI 推理能力转化为链上可交易资源:$DIEM 持有者可以每日获得 API Credit,而 $VVV 则承担平台入口与上游资产角色。包括
@dphnAI 的 $POD 在内,一批围绕「去中心化推理」的项目也开始获得市场关注,其核心逻辑都是将持续性的 AI 推理需求映射为长期现金流。
文章同时提到,Anthropic 为应对 Claude 使用量暴涨,已包下拥有 22 万块 NVIDIA GPU 的 Colossus 1 数据中心全部算力资源。相比一次性的模型训练成本,推理属于持续性消耗,并且会随着 agent 调用 agent 而指数级增长。JP Morgan 甚至预计,未来推理市场规模可能达到训练 的 10 至 50 倍。
@Bankless 认为,AI × Crypto 下一阶段的竞争重点,可能不再是谁拥有最大的模型,而是谁能够掌握未来 AI 世界里的推理基础设施与流动性入口。「推理」正在成为 2026 年 AI 市场最重要的新关键词之一。
Venice’s privacy + AI angle made the VVV trade easy to understand. But there’s another, larger AI shift it's set to capture: the move from training to inference.
Training gets the attention but inference costs are recurring and ever-growing, scaling alongside users and agents.
That is becoming one of AI’s biggest bottlenecks:
> Cerebras’ IPO landed 20x oversubscribed as investors piled in for exposure to its fast inference chips
> Anthropic unexpectedly cut usage limits, reports of
Claude being degraded spread everywhere, then Anthropic disclosed usage had GROWN 80x more than expected
> So, they then partnered with SpaceXAI to use the ENTIRE compute capacity of the Colossus 1 data center
> Now, Claude usage limits are rising again, and free API credits are coming in June
That is the broader tailwind behind Venice: its ecosystem is a gateway to endless inference.
VVV gives access to Venice Pro + can be staked to get perpetual inference via DIEM. Projects are already stockpiling DIEM to service inference for their own platforms, agents, and users.
Then there’s Dolphin, which built the default Venice Uncensored model and runs a distributed inference network using idle consumer GPUs. Its token POD can be used for inference payments and xPOD stakers get daily inference allocations across the network’s models.
David Christopher expects we’ll see more projects like this, serving decentralized inference rather than training.
もっと見る