Concise Summary简洁概述
DeepSeek's lack of coding subscriptions, multimodal models, or eval harnesses looks like strategic negligence — but the author argues it's a deliberate bet on infrastructure-level innovation over product polish.
A chain of attention-mechanism and memory innovations (MLA/DSA, Engram, mHC) drastically shrinks KV cache and shifts compute burden from scarce HBM to abundant Chinese-made NAND and LPDDR memory.
DeepSeek 没有编程订阅、没有多模态模型、也没有评测框架，看起来像是战略性的疏忽——但作者认为这其实是一场刻意押注基础设施层创新、而非打磨产品的豪赌。
一系列注意力机制与内存创新（MLA/DSA、Engram、mHC）大幅压缩 KV 缓存，把算力负担从稀缺的 HBM 转移到中国厂商产能充足的 NAND 与 LPDDR 内存上。
Infographic信息图
KV cache math as the crux
KV 缓存数学是关键
At 1M context, DeepSeek V4 needs ~5.48GB HBM vs GLM5's 60GB and Qwen3-235B's 89GB — a >10x compression via MLA/DSA-style attention. This single number underlies every downstream claim about cheap long-context agents.
在 100 万上下文下，DeepSeek V4 仅需约 5.48GB 显存，而 GLM5 要 60GB、Qwen3-235B 要 89GB——通过 MLA/DSA 类注意力机制实现十倍以上压缩。这一个数字撑起了后面所有关于廉价长上下文智能体的论断。
Offload cache to SSD, not HBM
缓存转存 SSD 而非依赖显存
Because compressed KV cache is small, it becomes cheap to offload to SSD and reload fast instead of recomputing or hoarding HBM. This directly benefits China's NAND leader YMTC by creating a new high-value market for flash storage.
由于压缩后的 KV 缓存体积极小，把它转存到固态硬盘再快速重新加载，比重新计算或占用显存划算得多。这直接为长江存储（YMTC）这样的 NAND 闪存厂商创造出全新的高价值市场。
Engram: trade memory for compute
Engram：以内存换算力
Engram upgrades N-gram embeddings into O(1) hashed lookups, giving transformers a retrieval shortcut instead of forcing them to 'compute' recall. The cost is a huge embedding table — which conveniently needs LPDDR memory, not scarce HBM.
Engram 把经典 N-gram 嵌入升级为基于哈希的 O(1) 查找，让 Transformer 拥有检索捷径而不必靠算力硬算“回忆”。代价是需要巨大的嵌入表——而这恰好可以用 LPDDR 内存承载，无需依赖稀缺的 HBM。
Weight streaming favors Chinese memory makers
权重流式加载利好中国存储厂商
DeepSeek's MoE architecture with 4-bit weights and many experts is naturally suited to streaming weights from LPDDR into HBM on demand, easing HBM bottlenecks. China's CXMT is only about half a generation behind in LPDDR speed, making this gap closeable.
DeepSeek 采用大量专家、支持 4 位权重的混合专家架构，天然适合把权重从 LPDDR 按需流式传输到 HBM，从而缓解显存瓶颈。中国的长鑫存储（CXMT）在 LPDDR 速度上仅落后国际顶尖水平约半代，差距是可以追平的。
Detailed Summary详细解读
The article's opening move is to treat DeepSeek's absent product lines — no coding subscription, no multimodal, no eval harness — not as gaps but as signals of where its attention isn't going. This is a classic 'strategy by omission' read: what a company refuses to build often says more than what it ships. The author uses this to set up a contrarian thesis against the natural read (that DeepSeek is simply behind).
The technical inventory (MoE, GRPO, RLVR, MTP, zero-bubble pipelines, MLA/DSA/CSA/HCA attention, Engram, mHC) is presented as a coherent research program rather than a grab-bag. Each item independently reduces either training cost, inference memory, or stability at scale — the throughline is systematically attacking the resource constraints (compute, memory, stability) that would otherwise cap a Chinese lab without frontier HBM access.
The KV cache calculation is the article's evidentiary centerpiece: a concrete, checkable number (5.48GB vs 60-89GB at 1M tokens) that the reader can reproduce via a public calculator. This grounds an otherwise speculative geopolitical thesis in something falsifiable, which is what elevates the piece above pure conjecture — though note the comparison models (GLM5, Qwen3) are chosen by the author, not exhaustive.
The hardware-substitution logic — small KV cache enables SSD offload, which favors NAND (YMTC); MoE weight streaming favors LPDDR (CXMT) — is the article's most novel contribution. It connects a software/algorithmic choice directly to which memory technologies benefit, turning a model architecture story into an industrial-policy story without requiring any explicit coordination between DeepSeek and chipmakers.
The piece is explicitly framed as an X (Twitter) thread rewritten into essay form, written by an enthusiastic outside analyst (not a DeepSeek insider). The 'hero's journey' framing and confident tone should be read as advocacy — persuasive synthesis of publicly available technical papers and one calculator demo, not an insider strategy document or DeepSeek's own stated roadmap.
文章开篇先把 DeepSeek 缺失的产品线——没有编程订阅、没有多模态、没有评测框架——不当作短板，而是当作它注意力没有投向哪里的信号。这是一种典型的“靠缺失读战略”的思路：一家公司拒绝做什么，往往比它做了什么更能说明问题。作者借此引出一个反直觉论点，对抗“DeepSeek 只是落后了”这种自然而然的解读。
文章列出的技术清单（MoE、GRPO、RLVR、MTP、零气泡流水线、MLA/DSA/CSA/HCA 注意力机制、Engram、mHC）被呈现为一套连贯的研究纲领，而非零散堆砌。每一项都分别降低了训练成本、推理内存占用或大规模训练的稳定性问题——贯穿始终的主线是系统性攻克那些原本会限制一家缺乏顶级 HBM 资源的中国实验室的资源约束（算力、内存、稳定性）。
KV 缓存计算是文章的证据核心：一个具体、可核实的数字（百万 token 下 5.48GB 对比 60-89GB），读者可以通过公开计算器自行复现。这为一个原本颇具思辨色彩的地缘政治论断提供了可证伪的落地依据，也是这篇文章超越纯粹猜测之处——不过需注意，用于对比的模型（GLM5、Qwen3）是作者自选的，并非穷尽性对比。
硬件替代逻辑——缓存小到可以转存 SSD，利好 NAND（长江存储）；MoE 权重流式加载利好 LPDDR（长鑫存储）——是文章最具原创性的部分。它把一个软件/算法层面的选择，直接同哪些存储技术会因此受益联系起来，把一个模型架构故事变成了一个产业政策故事，而无需假设 DeepSeek 与芯片厂商之间存在任何明确协调。
这篇文章原本是一条 X（推特）长推文改写成的文章体，作者是一位热情的外部分析者，而非 DeepSeek 内部人士。“英雄之旅”式的叙事框架和笃定的语气应被视为一种带说服意图的解读——是对公开技术论文和一次计算器演示的综合性论证，而不是内部战略文件或 DeepSeek 官方公布的路线图。
FAQ常见问答
Is the 10 trillion / 1 trillion valuation figure an official DeepSeek target?10 万亿/1 万亿美元的市值数字是 DeepSeek 官方给出的目标吗？
No — it's the author's own extrapolation from observed technical choices, not a figure from DeepSeek, its founder, or any filing. Treat it as a thesis, not disclosed guidance.
不是——这是作者根据观察到的技术选择推导出的个人估算，并非 DeepSeek、其创始人或任何官方文件给出的数字，应视为一种论断而非公开指引。
Why does smaller KV cache matter so much for agents specifically?为什么更小的 KV 缓存对智能体应用特别重要？
Long-horizon agents hold context (tool outputs, memory) open across many steps; a huge cache forces expensive HBM reservation or recomputation. Shrinking it makes long-running, low-cost agent sessions economically viable.
长程智能体需要在多个步骤中持续保留上下文（工具输出、记忆），庞大的缓存会迫使系统占用昂贵的显存或反复重算。缩小缓存让长时间运行、低成本的智能体会话在经济上变得可行。
Does GLM5 already use similar techniques, undercutting DeepSeek's edge?GLM5 是否已经采用了类似技术，从而削弱了 DeepSeek 的优势？
The article notes GLM5 has adopted DeepSeek's MLA and DSA but not its newest compressed attention — so the gap is real but narrowing, and framed by the author as a moving target, not a permanent moat.
文章提到 GLM5 已经采用了 DeepSeek 的 MLA 和 DSA，但尚未用上最新的压缩注意力技术——差距是真实存在的，但在缩小，作者将其视为一个动态目标而非永久护城河。
Is YMTC/CXMT benefiting a deliberate DeepSeek strategy or an unintended side effect?长江存储/长鑫存储受益，是 DeepSeek 有意为之的战略，还是无心插柳的副作用？
The article doesn't claim direct coordination — it argues the algorithmic choices create market incentives that structurally favor domestic NAND/LPDDR makers, regardless of whether that was an explicit goal.
文章并未声称存在直接协调——它认为，这些算法层面的选择在结构上制造了有利于本土 NAND/LPDDR 厂商的市场激励，无论这是否是明确设定的目标。
What happens to this thesis if HBM shortages ease or Nvidia regains full China access?如果 HBM 短缺缓解，或英伟达重新获得对华全面供货许可，这套论点还成立吗？
The article doesn't address this counterfactual; the whole thesis rests on HBM scarcity as a binding constraint, so easier HBM access would weaken (though not eliminate) the economic case for memory substitution.
文章没有讨论这种反事实情形；整套论点建立在 HBM 稀缺是一个刚性约束的前提上，如果 HBM 获取变得容易，内存替代方案的经济性论证就会被削弱（但不会完全失效）。
In-depth Analysis · Pros & Cons深入解读 · 优缺点
This piece reframes DeepSeek's seemingly odd product choices — no coding subscriptions, no multimodal push, aggressive open-sourcing — as deliberate infrastructure-first moves in a much larger play. It argues DeepSeek is engineering the technical preconditions for a domestic Chinese AI hardware ecosystem worth trillions of dollars.
这篇文章把 DeepSeek 那些看似古怪的产品选择——不做编程订阅、不追多模态、疯狂开源——重新解读为一盘更大棋局里的基础设施先行棋。作者认为，DeepSeek 正在为价值万亿美元级别的中国本土 AI 硬件生态铺设技术前提条件。
- Falsifiable central evidence核心证据可证伪The KV cache comparison uses a public calculator and named models, letting any reader independently reproduce the 5.48GB vs 60-89GB numbers rather than trusting the author's word.KV 缓存对比使用了公开计算器和具名模型，任何读者都能自行复现 5.48GB 与 60-89GB 的数字，而不必单纯采信作者一面之词。
- Coherent technical narrative技术叙事连贯自洽Rather than listing DeepSeek innovations as disconnected trivia, the piece links each (MoE, GRPO, MLA/DSA, Engram, mHC) to a shared constraint — compute/memory scarcity — giving the list explanatory power.文章没有把 DeepSeek 的创新罗列成互不相关的琐碎信息，而是把每一项（MoE、GRPO、MLA/DSA、Engram、mHC）都关联到同一个约束——算力/内存稀缺——使整份清单具有解释力。
- Connects software to hardware ecosystem打通软件与硬件生态The move from 'small KV cache' to 'this benefits YMTC and CXMT' is a genuinely useful systems-level insight, tying model architecture choices to concrete industrial beneficiaries.从“KV 缓存变小”推导到“这利好长江存储和长鑫存储”，是一个真正有价值的系统级洞察，把模型架构选择与具体的产业受益者联系了起来。
- Names concrete sources引用来源具体Arxiv paper links and a SemiAnalysis reference give readers a path to verify claims further, rather than leaving the analysis as unsupported assertion.文中给出了 Arxiv 论文链接和 SemiAnalysis 的引用，为读者提供了进一步核实论断的路径，而非停留在无支撑的断言层面。
- Speculative endgame终局论断带有推测性The '$10 trillion ecosystem, $1 trillion valuation' framing is the author's inference, not a disclosed DeepSeek strategy — it should be read as an interpretive thesis, not confirmed fact.“10 万亿生态、1 万亿市值”的框架是作者本人的推断，而非 DeepSeek 公开披露的战略，应被当作一种解读性论断，而非已证实的事实。
- Cherry-picked comparisons对比对象存在选择性GLM5 and Qwen3 are chosen comparators; the piece doesn't address whether other labs (US or Chinese) have comparable or superior KV-compression techniques not covered here.GLM5 和 Qwen3 是作者选定的对比对象；文章没有讨论其他实验室（无论中美）是否已有同等或更优的 KV 压缩技术。
- Assumes execution, not just capability假设了执行力，而不仅是能力Having the technical means to shift load to NAND/LPDDR doesn't guarantee China's memory industry scales output, yield, or cost fast enough to realize the described trillion-dollar ecosystem.具备把负载转移到 NAND/LPDDR 的技术手段，并不等于中国存储产业就能在产能、良率和成本上快速跟上，从而真正实现文中描述的万亿级生态。
- No engagement with counter-risks未讨论潜在风险与反面情形The piece omits risks like export controls tightening around NAND/LPDDR too, potential quality/reliability gaps in offloaded KV cache at scale, or competitors closing the compression gap faster than assumed.文章没有讨论一些潜在风险，比如出口管制也可能收紧到 NAND/LPDDR 领域、大规模转存 KV 缓存可能存在的质量/可靠性问题，或者竞争对手比预期更快追平压缩技术差距。
Worth reading for anyone tracking Chinese AI hardware or long-context agent economics — the KV cache math is concrete and checkable. But treat the trillion-dollar valuation and 'grand strategy' framing as one enthusiastic analyst's extrapolation, not confirmed DeepSeek policy, and weigh it against the execution risk in scaling domestic NAND/LPDDR production.
对关注中国 AI 硬件或长上下文智能体经济性的读者值得一读——KV 缓存的数学计算具体且可核实。但“万亿美元估值”与“大战略”的框架，应被视为一位热情分析者的个人推断，而非 DeepSeek 官方确认的政策，读时还需权衡本土 NAND/LPDDR 产能扩张所面临的执行风险。
Excerpt原文节选
This is a short excerpt, not the full piece — the complete essay belongs to its original author; please read it in full at the link above.
以下仅为节选，并非全文——完整文章版权归原作者所有，请点击上方链接阅读全文。
The English text on this side is an AI translation provided for convenience; the authoritative version is the source in the other language.
DeepSeek's 10 Trillion Dollar Grand Strategy
By GDP (@bookwormengr) Title: DeepSeek's 10 trillion USD grand strategy
Have you ever wondered how exactly DeepSeek plans to make money — and make it big?
They haven't launched a competitive coding subscription plan like Zhipu (GLM), Moonshot, or MiniMax. They don't have multimodal, voice, or video models. To this day, they don't even have an evaluation framework (a "Harness" — a benchmarking tool used to test and evaluate model performance), though I've recently heard they're starting to hire for it. And DeepSeek has long been committed to open source, happily sharing its "secret recipes" with the world. Is this madness? Or are they simply burning money? Are the investors lining up to hand them $10 billion just throwing their money away?
No — in my view, it's exactly the opposite!!!
Here I want to talk about what I've observed in what they've done…
[…the source continues — read the rest at the link above]
[……原文更长，完整内容请点击上方链接阅读]
A detailed breakdown of these underlying technical innovations will be published as a long-form piece this weekend — if you're interested, follow my Substack column: https://polymath707.substack.com/ ...
DeepSeek 的 10 万亿美元宏伟战略
作者：GDP (@bookwormengr) 标题： DeepSeek's 10 trillion USD grand strategy
你有没有想过，DeepSeek 到底打算怎么赚钱，而且是赚大钱？
他们没有像智谱（GLM）、月之暗面（MoonShot）和 MiniMax 那样推出有竞争力的编程订阅计划。他们没有多模态、语音或视频模型。时至今日，他们甚至连一个评测框架（Harness， 用于测试和评估模型性能的基准测试工具 ）都没有（虽然最近听说他们开始招人做了）。而且，DeepSeek 还长期致力于开源，乐此不疲地分享自己的“独家秘方”。这难道是疯了吗？还是纯粹在烧钱？那些正准备给他们投资 100 亿美元的投资人们，难道是在把钱往水里扔吗？
不，在我看来，恰恰相反！！！
在这里，我想聊聊我对他们至今所作所为的观察，以及他们似乎正在践行的战略。DeepSeek 创始人梁文锋的眼光显然盯着一个大得多的终极奖杯——他们不仅自己能冲击 1 万亿美元的市值，还能顺便帮中国催生出一个高达 10 万亿美元的产业巨兽！
重新审视 DeepSeek 的“英雄之旅”
DeepSeek 总是逆风而行，他们不屑于去卷那种“比别人好一点点”的微调模型，也不急着去卖当下的应用（比如各种编程套餐）。我在 2025 年 1 月 27 日发过一条疯传的推文，谈到了我所看到的景象，而现在的剧情正变得越来越精彩。
当大家都在死磕稠密模型（Dense Models， 所有参数都参与计算的传统大模型结构 ）时，DeepSeek 却迎难而上，选择了极难训练的混合专家模型（MoE, Mixture of Experts）。
他们从“第一性原理”（First Principles）出发，发明了全新的 GRPO 算法，取代了在强化学习（RL, Reinforcement Learning）中虽然占据统治地位、但实现成本极高的 PPO 算法。
他们摸索出了基于验证奖励的强化学习（RLVR, Reinforcement Learning from Verified Rewards），并将其作…
[…the source continues — read the rest at the link above]
[……原文更长，完整内容请点击上方链接阅读]
关于这些底层技术创新的详细拆解长文将在本周末发布，感兴趣的朋友欢迎关注我的 Substack 专栏： https://polymath707.substack.com/ ...