ENGLISH · ORIGINAL
Reasoning models introduced in 2024 demonstrated that allocating more computational resources and giving models more time to process during inference directly leads to higher-quality intelligence. In addition, compute infrastructure with end-to-end, cluster-level coherence through tight integration across software and hardware systems enables more efficient, stable, and higher-fidelity training and inference at scale—ultimately enhancing model intelligence and performance. Within inference, we expect computationally-intensive reasoning, agentic, and multi-modal workloads will continue to grow as a portion of overall usage. We therefore believe operators with superior model-to-compute integration—the ability to efficiently support and allocate compute across both training and inference workloads—are best positioned to win the AI race. Self-Reinforcing Network Effects Among Lower Cost Per Token, Model Quality, and User Adoption. AI systems are ultimately constrained or differentiated by the cost, speed, and scale at which they can generate and process tokens. A "token" represents the fundamental unit of data consumed and produced by modern AI models. This is because lower cost per token enables more frequent model training, larger and more sophisticated models, longer chains of processing for reasoning and agentic workloads, and significantly higher inference volumes at economically viable prices.
中文翻译
算力规模与"模型—算力"整合:AI 竞赛的胜负手
2024 年出现的"推理模型"(reasoning models)证明了一件事:给模型越多算力、让模型在"推理"时有更多时间思考,AI 输出的智能水平就越高。同样重要的是——一套"从软件到硬件深度整合、整个集群端到端协调"的算力基础设施,能让大规模训练和推理都更高效、更稳定、更"忠于原意"(high-fidelity),最终让模型更聪明、表现更好。在推理这一端,我们预计算力消耗巨大的推理任务、Agent 任务、多模态任务占整体用量的比例会越来越大。所以我们认为:谁能最有效地把"模型"和"算力"整合起来——也就是既能把算力同时服务好训练和推理——谁就最有可能赢得这场 AI 竞赛。
"单 token 成本越低、模型质量越好、用户越多"——这是一个自我强化的循环。 AI 系统到底强不强,最终由它"生成和处理 token"时的成本、速度、规模来决定。"Token"是现代 AI 模型"吃进去、吐出来"的最小数据单位。这个事很重要——因为单 token 成本越低,就能做到:
训练更频繁;
训练更大、更复杂的模型;
让推理和 Agent 任务做更长的"思考链";
在还能赚钱的价格水平下,让推理量大幅放大。
💰 算力成本 = token 成本
想参与互动?打开法量小程序体验完整功能
非常赞同“模型-算力整合”才是核心壁垒的观点。单纯堆算力就像造跑车却堵在乡间小路,而软硬协同的端到端优化才是高速公路。想请教一下,这种整合能力对于初创团队来说,是否意味着必须自研芯片或深度绑定特定云厂商,还是存在通过软件调度实现弯道超车的可能?
这个分析很到位,特别是“推理模型”对算力需求的结构性改变,以及“模型-算力整合”作为护城河的观点。不过想请教一下,这种整合能力对初创公司来说是不是越来越难追赶?毕竟OpenAI和SpaceX这类巨头在软硬件协同上的积累太深了。另外,单token成本降低带来的自我强化循环,会不会最终导致算力资源进一步向头部集中?
这个分析很扎实,尤其是“模型-算力整合”这个点,很多人只盯着算力规模,忽略了软件栈和集群协同的重要性。想问一下,目前SpaceX在推理侧的算力调度策略更偏向于自研芯片还是依赖现有生态?另外,token成本下降带来的用户增长和模型质量提升,会不会反而加剧头部集中效应?
这个分析很到位,尤其是“模型—算力整合”这个视角,比单纯堆算力更关键。但想请教一个问题:在这种自我强化的循环里,小公司或开源社区有没有可能通过算法创新或更高效的架构,在token成本上形成局部突破,还是说最终还是会输给拥有完整软硬件栈的巨头?
非常赞同对“模型-算力整合”的强调。算力规模固然重要,但软硬件的端到端协同才能把效率压榨到极致。关于token成本降低带来的网络效应,想请教一下:在推理任务占比越来越高的趋势下,你们认为算力分配上训练和推理的最优比例大概会如何演变?