我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
从DeepSeek83天永久降价的崩塌猜测其过度追求缓存命中的策略失败
视频
音频
原始脚本
83天永久低价的崩塌，从硬件架构猜想，解析 Deepseek 前缀缓存战略的结构性失败。 注文中关于昇腾集群内存池化带来的长 TTLKV 缓存的推论。 属于工程视角的推演猜想，并非厂商公开披露的官方事实，但是可以完整解释这条商业策略从成立到破产的完整逻辑链。 一、事件的表层叙事与被忽略的底层矛盾。 2026年5月底，Deepseek 正式宣布 V4 Pro 的优惠价转为永久定价。 其中缓存命中输入定价低至0.025元每百万 token 在全行业独树一帜。 依靠超长生命周期的前缀 kv 缓存。 打造出一套 append only 长会话最优解的开发者生态引导。 该承诺仅仅维持83天，就被峰谷阶梯涨价方案推翻。 高峰时段缓存命中价格最高上涨超过11倍。 市场主流解读大多停留在调用量激增、商业化诉求、算力扩容滞后等表层理由。 但很少有人去追问一个核心问题。 为什么只有 Deepseek 愿意投入极高工程代价去维护一套长 TTL 高复用目标的全局前缀缓存？而 Anthropic OpenAI Copilot 全都刻意回避这条技术路线。 答案可以分为两个层面，业务范式陷阱以及硬件架构带来的最初误判。 一、业务层面，平台与开发者双双落入缓存思维陷阱。 Deepseek这套低价缓存机制本质上是一种平台主动释放的战术红利。 而非健康的长期架构范式。 从平台的设想来看，只要引导开发者采用线性追加式，append only 的长对话，大量相同会话前缀可以反复复用 kv 矩阵。 平台侧，Prefill 算力开销下降，硬件资源利用率提升。 再把一部分收益以极低的缓存单价返还开发者，形成正向循环。 这套设想带来了双向的思维枷锁。 对，开发者，我们长期被这套红利束缚，形成了输入 token 几乎免费。 优化重心只放在输出 token 的惯性，架构设计下意识迁就缓存命中，不敢做上下文摘要、记忆重排、召回片段插入。 为了保住缓存，牺牲了 Agent 记忆编排的自由度。 而现在整个行业已经形成新共识，Claude Hermes Copilot 都选择持续主动压缩上下文。 把短时工作窗口稳定锁在200500K舒适区间，放弃缓存收益，换取模型推理稳定性，减少无效轮次与无效输出。 缓存只能作为副产品，绝对不能作为架构约束。 对 Deepseek 自身，他把这套私有特性当成了核心竞争力，把业务增长的赌注压在了大量长连接，pand only 绘画带来高跨绘画，绘画内 kv 复用率这个前提之上。 而这个前提最终并没有在真实流量中兑现。 真实的 Agent 流量高度碎片化，绝大多数长 kv 缓存都是单会话私有，跨用户共享的公共前缀极少，写入量大，二次复用少。 平台付出存储调度 IO 开销，却拿不到预期的资源摊薄收益。 二、硬件视角的猜想，升腾集群全域内存池化。 是这条策略最初成立的先决条件。 Deepseek 同时运行着异构算力集群，同时存在 NVIDIA HGX 集群与华为昇腾超算集群。 两套硬件的内存模型天然决定了常驻留 kv 缓存的成本天差地别。 英伟达 nvlink 七二 gb 两百集群的设计目标。 是极致算力吞吐。 它的高速 HBM 显存是极其昂贵的稀缺资源。 硬件设计的首要目标是拉高并发，快速周转资源。 在英伟达集群上，把大量 kv cache 长期驻留于 hbm 中，等于长期锁定珍贵的高速显存资源，拖低整体可承载并发数。 即便做 kv cache 卸载到主机内存，也只能作为短时应急策略，厂商不会以长 ttl 缓存作为顶层设计目标。 这也是海外厂商全部只做短生命周期缓存数分钟级别的根本原因。 长时间占用高速资源，在经济上完全不合算。 而昇腾 Cloud Matrix 超算集群拥有一套完全不同的硬件能力，通过 UB 互联总线实现 NPU CPU 全域统一编址，内存池化。 高速 hbm 仅作为 npu 本地高速缓存，海量的 kv 矩阵可以驻留在全局 cpu dram 内存池之中，不需要立刻落盘到 ssd 冷存储。 数据搬运的链路开销远低于反复读写磁盘。 这就是我们猜想的核心原点。 在项目前期流量很低的时候，这套内存池化架构可以以很低的边际成本。 维持一个很长的 kv 缓存、TTL 缓存条目可以长期留在全域内存池，不需要频繁序列化、落盘、再加载。 平台此时认为自己找到了独有的硬件红利，可以靠这套能力打出差异化竞争。 在流量规模小的阶段，硬件调度游刃有余，平台与开发者双方都拿到收益。 这一个判断带来了最初的战略误判。 团队将小流量下的硬件红利当成了可以支撑亿万级 Token 流量的长期商业模式。 当调用量爆发之后，矛盾立刻爆发。 即便拥有全域内存池，物理内存的总量依然存在硬上限。 海量碎片化、低复用率的私有会话 kv 缓存不断占用全局内存池，挤占新请求可用的资源。 调度压力、内存碎片、跨节点数据路由开销持续上涨，集群的并发吞吐开始被缓存管理本身拖累。 此时平台只剩下两条路，一是大规模驱逐缓存，缩短 TTL 但是这种后台静默改动会直接摧毁开发者信任。 二是用价格杠杆抑制长缓存场景的流量，也就是后来推出的峰谷涨价策略。 这里需要澄清，升腾架构并不是做不到高并发推理。 而是常驻留 kv 缓存和高并发是一对互斥目标。 想要维持大量 kv 条目长期驻留，就必须预留巨大的内存池空间。 这部分资源不能用来调度新的推理任务。 集群整体并发上限被硬性压低，这是资源分配的零和博弈，和算力本身的峰值性能无关。 三、涨价并不是故事的终点。 行业范式已经发生不可逆的转向。 Deepseek 现在的调价本质上是在用价格信号引导开发者行为，不再鼓励无脑的超长 append only 会话。 虽然平台没有直接修改缓存的 TTL 策略，但是高昂的高峰时段缓存价格会倒逼开发者重新思考上下文管控策略。 长远来看。 未来开发者的架构选择会和全球主流范式对齐。 短时工作上下文主动压缩，记忆检索与绘画上下文解耦，不再依靠平台私有缓存作为成本支柱。 这场行业实验留给开发者最重要的教训有两点。
修正脚本
83天永久低价的崩塌，从硬件架构猜想，解析 Deepseek 前缀缓存战略的结构性失败。 注：文中关于昇腾集群内存池化带来的长 TTLKV 缓存的推论。 属于工程视角的推演猜想，并非厂商公开披露的官方事实，但是可以完整解释这条商业策略从成立到破产的完整逻辑链。 一、事件的表层叙事与被忽略的底层矛盾。 2026年5月底，Deepseek 正式宣布 V4 Pro 的优惠价转为永久定价。 其中缓存命中输入定价低至0.025元每百万 token 在全行业独树一帜。 依靠超长生命周期的前缀 kv 缓存。 打造出一套 append only 长会话最优解的开发者生态引导。 该承诺仅仅维持83天，就被峰谷阶梯涨价方案推翻。 高峰时段缓存命中价格最高上涨超过11倍。 市场主流解读大多停留在调用量激增、商业化诉求、算力扩容滞后等表层理由。 但很少有人去追问一个核心问题。 为什么只有 Deepseek 愿意投入极高工程代价去维护一套长 TTL 高复用目标的全局前缀缓存？而 Anthropic OpenAI Copilot 全都刻意回避这条技术路线。 答案可以分为两个层面，业务范式陷阱以及硬件架构带来的最初误判。 一、业务层面，平台与开发者双双落入缓存思维陷阱。 Deepseek这套低价缓存机制本质上是一种平台主动释放的战术红利。 而非健康的长期架构范式。 从平台的设想来看，只要引导开发者采用线性追加式，append only 的长对话，大量相同会话前缀可以反复复用 kv 矩阵。 平台侧，Prefill 算力开销下降，硬件资源利用率提升。 再把一部分收益以极低的缓存单价返还开发者，形成正向循环。 这套设想带来了双向的思维枷锁。 对开发者，我们长期被这套红利束缚，形成了输入 token 几乎免费，优化重心只放在输出 token 的惯性，架构设计下意识迁就缓存命中，不敢做上下文摘要、记忆重排、召回片段插入。 为了保住缓存，牺牲了 Agent 记忆编排的自由度。 而现在整个行业已经形成新共识，Claude Hermes Copilot 都选择持续主动压缩上下文。 把短时工作窗口稳定锁在200-500K舒适区间，放弃缓存收益，换取模型推理稳定性，减少无效轮次与无效输出。 缓存只能作为副产品，绝对不能作为架构约束。 对 Deepseek 自身，他把这套私有特性当成了核心竞争力，把业务增长的赌注压在了大量长连接，append only 会话带来高跨会话、会话内 kv 复用率这个前提之上。 而这个前提最终并没有在真实流量中兑现。 真实的 Agent 流量高度碎片化，绝大多数长 kv 缓存都是单会话私有，跨用户共享的公共前缀极少，写入量大，二次复用少。 平台付出存储调度 IO 开销，却拿不到预期的资源摊薄收益。 二、硬件视角的猜想，昇腾集群全域内存池化。 是这条策略最初成立的先决条件。 Deepseek 同时运行着异构算力集群，同时存在 NVIDIA HGX 集群与华为昇腾超算集群。 两套硬件的内存模型天然决定了常驻留 kv 缓存的成本天差地别。 英伟达 nvlink 72GB HGX 集群的设计目标。 是极致算力吞吐。 它的高速 HBM 显存是极其昂贵的稀缺资源。 硬件设计的首要目标是拉高并发，快速周转资源。 在英伟达集群上，把大量 kv cache 长期驻留于 hbm 中，等于长期锁定珍贵的高速显存资源，拖低整体可承载并发数。 即便做 kv cache 卸载到主机内存，也只能作为短时应急策略，厂商不会以长 TTL 缓存作为顶层设计目标。 这也是海外厂商全部只做短生命周期缓存数分钟级别的根本原因。 长时间占用高速资源，在经济上完全不合算。 而昇腾 Cloud Matrix 超算集群拥有一套完全不同的硬件能力，通过 UB 互联总线实现 NPU CPU 全域统一编址，内存池化。 高速 hbm 仅作为 npu 本地高速缓存，海量的 kv 矩阵可以驻留在全局 cpu dram 内存池之中，不需要立刻落盘到 ssd 冷存储。 数据搬运的链路开销远低于反复读写磁盘。 这就是我们猜想的核心原点。 在项目前期流量很低的时候，这套内存池化架构可以以很低的边际成本。 维持一个很长的 kv 缓存、TTL 缓存条目可以长期留在全域内存池，不需要频繁序列化、落盘、再加载。 平台此时认为自己找到了独有的硬件红利，可以靠这套能力打出差异化竞争。 在流量规模小的阶段，硬件调度游刃有余，平台与开发者双方都拿到收益。 这一个判断带来了最初的战略误判。 团队将小流量下的硬件红利当成了可以支撑亿万级 Token 流量的长期商业模式。 当调用量爆发之后，矛盾立刻爆发。 即便拥有全域内存池，物理内存的总量依然存在硬上限。 海量碎片化、低复用率的私有会话 kv 缓存不断占用全局内存池，挤占新请求可用的资源。 调度压力、内存碎片、跨节点数据路由开销持续上涨，集群的并发吞吐开始被缓存管理本身拖累。 此时平台只剩下两条路，一是大规模驱逐缓存，缩短 TTL 但是这种后台静默改动会直接摧毁开发者信任。 二是用价格杠杆抑制长缓存场景的流量，也就是后来推出的峰谷涨价策略。 这里需要澄清，昇腾架构并不是做不到高并发推理。 而是常驻留 kv 缓存和高并发是一对互斥目标。 想要维持大量 kv 条目长期驻留，就必须预留巨大的内存池空间。 这部分资源不能用来调度新的推理任务。 集群整体并发上限被硬性压低，这是资源分配的零和博弈，和算力本身的峰值性能无关。 三、涨价并不是故事的终点。 行业范式已经发生不可逆的转向。 Deepseek 现在的调价本质上是在用价格信号引导开发者行为，不再鼓励无脑的超长 append only 会话。 虽然平台没有直接修改缓存的 TTL 策略，但是高昂的高峰时段缓存价格会倒逼开发者重新思考上下文管控策略。 长远来看。 未来开发者的架构选择会和全球主流范式对齐。 短时工作上下文主动压缩，记忆检索与会话上下文解耦，不再依靠平台私有缓存作为成本支柱。 这场行业实验留给开发者最重要的教训有两点。
英文翻译
The Collapse of 83-Day Permanent Low Pricing: Analyzing the Structural Failure of Deepseek's Prefix Caching Strategy Based on Hardware Architecture Conjecture Note: The inferences in this article regarding the long TTL KV cache enabled by Ascend cluster memory pooling are deductive conjectures from an engineering perspective, not official facts publicly disclosed by the manufacturer. However, they can fully explain the complete logical chain of this business strategy from its inception to its collapse. 1. Surface Narrative of the Incident and the Overlooked Underlying Contradiction At the end of May 2026, Deepseek officially announced that the preferential price of V4 Pro would be converted to a permanent price. Among them, the pricing for cache-hit inputs is as low as 0.025 yuan per million tokens, which is unique in the entire industry. Relying on prefix KV cache with an ultra-long lifecycle, it has built a developer ecosystem orientation that takes append-only long sessions as the optimal solution. This promise only lasted 83 days before being overturned by the peak-valley tiered price increase plan. During peak hours, the cache hit price increased by more than 11 times at maximum. Most mainstream market interpretations mostly stay at surface-level reasons such as surging query volume, commercialization demands, and delayed computing capacity expansion. But few people have pressed for a core question. Why is only Deepseek willing to pay extremely high engineering costs to maintain a global prefix caching system targeting long TTL and high reuse, while Anthropic, OpenAI and Copilot all deliberately avoid this technical route? The answer can be divided into two levels: the business paradigm trap and the initial misjudgment brought by hardware architecture. 1. At the business level: Both the platform and developers have fallen into the caching thinking trap Deepseek's low-price caching mechanism is essentially a tactical dividend actively released by the platform, rather than a healthy long-term architectural paradigm. From the platform's assumption, as long as developers are guided to adopt linearly appended, append-only long conversations, a large number of identical session prefixes can reuse KV matrices repeatedly. On the platform side, Prefill computing overhead decreases, and hardware resource utilization increases. Part of the benefit is then returned to developers at an extremely low cache unit price, forming a positive cycle. This assumption has brought two-way thinking shackles. For developers, we have long been bound by this dividend, forming the inertia that input tokens are almost free and optimization focus only falls on output tokens. We subconsciously accommodate cache hits in architectural design, and dare not implement context summarization, memory rearrangement, or insertion of recalled fragments. To retain cache validity, we sacrificed the freedom of Agent memory orchestration. Now the entire industry has reached a new consensus: Claude, Hermes and Copilot all choose to continuously and proactively compress context. They stably lock the short-term working window in the comfortable range of 200-500K, give up cache benefits in exchange for model inference stability, and reduce invalid rounds and invalid outputs. Caching can only be a by-product, and must never be used as an architectural constraint. For Deepseek itself, it regards this private feature as its core competitiveness, and bets its business growth on the premise that a large number of long-connected, append-only sessions will bring high cross-session and intra-session KV reuse rates. And this premise was ultimately not fulfilled in real traffic. Real Agent traffic is highly fragmented: the vast majority of long KV caches are private to single sessions, public prefixes shared across users are extremely rare, the write volume is large, and secondary reuse is low. The platform pays storage scheduling IO overhead, but cannot get the expected resource dilution benefit. 2. Conjecture from the hardware perspective: Global memory pooling of Ascend clusters is the prerequisite for the initial establishment of this strategy. Deepseek operates heterogeneous computing clusters at the same time, with both NVIDIA HGX clusters and Huawei Ascend supercomputing clusters. The memory models of the two sets of hardware naturally determine that the cost of resident KV cache varies greatly. The design goal of NVIDIA's nvlink 72GB HGX cluster is extreme computing throughput. Its high-speed HBM video memory is an extremely expensive and scarce resource. The primary goal of the hardware design is to increase concurrency and quickly turn over resources. On NVIDIA clusters, keeping a large amount of KV cache resident in HBM for a long time is equivalent to locking precious high-speed video memory resources for a long time, dragging down the overall number of concurrent connections that can be carried. Even if KV cache is offloaded to host memory, it can only be used as a short-term emergency strategy, and manufacturers will not take long TTL caching as a top-level design goal. This is also the fundamental reason why all overseas vendors only implement short-lifecycle caching at the minute level. Occupying high-speed resources for a long time is completely uneconomical. In contrast, the Ascend Cloud Matrix supercomputing cluster has a completely different set of hardware capabilities: it realizes global unified addressing for NPU and CPU and memory pooling through the UB interconnection bus. High-speed HBM only serves as NPU local high-speed cache, and massive KV matrices can reside in the global CPU DRAM memory pool, without needing to be immediately written to SSD cold storage. The link overhead of data handling is much lower than repeated reading and writing to disks. This is the core origin of our conjecture. When traffic is very low in the early stage of the project, this memory pooling architecture can maintain very long TTL for KV cache at very low marginal cost. Cache entries can stay in the global memory pool for a long time, without frequent serialization, disk writing, and reloading. At this stage, the platform believed that it had found unique hardware dividends and could use this capability to achieve differentiated competition. At the stage of small traffic scale, hardware scheduling is more than sufficient, and both the platform and developers get benefits. This judgment led to the initial strategic misjudgment. The team mistook the hardware dividends under small traffic for a long-term business model that can support hundreds of millions of tokens of traffic. After the query volume exploded, contradictions broke out immediately. Even with a global memory pool, there is still a hard upper limit on the total amount of physical memory. Massive fragmented, low-reuse private session KV caches continue to occupy the global memory pool, crowding out available resources for new requests. Scheduling pressure, memory fragmentation, and cross-node data routing overhead continue to rise, and the cluster's concurrent throughput begins to be dragged down by cache management itself. At this point, the platform only has two options: first, evict a large number of caches and shorten TTL, but this silent change in the background will directly destroy developer trust. Second, use price leverage to curb traffic in long-cache scenarios, which is the peak-valley price increase strategy launched later. It needs to be clarified here that the Ascend architecture is not incapable of high-concurrency inference. Instead, resident KV caching and high concurrency are a pair of mutually exclusive goals. To maintain a large number of KV entries resident for a long time, a huge memory pool space must be reserved. This part of resources cannot be used to schedule new inference tasks. The overall concurrency upper limit of the cluster is forcibly lowered. This is a zero-sum game of resource allocation and has nothing to do with the peak performance of computing power itself. 3. Price increase is not the end of the story The industry paradigm has undergone an irreversible shift. Deepseek's current price adjustment is essentially using price signals to guide developer behavior, and no longer encourages mindless ultra-long append-only sessions. Although the platform has not directly modified the TTL strategy of caching, the high cache price during peak hours will force developers to rethink their context management strategies. In the long run, developers' architectural choices in the future will align with the global mainstream paradigm: proactively compress short working context, decouple memory retrieval from session context, and no longer rely on the platform's private caching as a cost pillar. This industry experiment has left two most important lessons for developers.
back to top