louise

[TMT Breakout @TMTBreakout]

The Great Gross-to-Net Panic of 2026: OpenAI’s ARR in two charts & glass half empty vs. half full

Glass half empty: People had thought $70B was net and the perceived dollar base just reset ~25–30% lower. OpenAI added ~$20B of ARR over August and September, not ~$30B. The first half was also weaker than people thought. The net run rate sat at ~$24–25B from February through June, and y/y growth slid to ~130% y/y by June, a bit decel from Dec 2025. That’s a lower starting point for YE26/YE27 net ARR targets & for valuation, and for the math on funding OAI’s compute commitments.

Glass half full: The growth rate hasn’t changed. $30B → $50B net is the same ~70% in two months as $40B → $70B gross, and the monthly additions picked up into September. The y/y line bottoms at +131% in June, then climbs to +150% in July, +182% in August and +229% (~3.3x) in September. Had July really been $40B, September’s $50B would have meant y/y growth flat at ~230%. On the net base, it’s actually sharp re-acceleration of ~100% y/y in 4 months. Against Anthropic nothing changes, because the ~$70B gross figure is still the like-for-like comparison.

[North@CreaoAI @anorth_chen]

https://t.co/tuhPesfxGp

[Tech Fund @techfund1]

We believe HBM-related strength should also be accretive to MU's margins as HBM commands a much higher premium compared to traditional DRAM (DBe HBM ASPs are ≥5x greater than conventional DRAM). HBM offers significant performance improvements vs. standard DRAM, which drives pricing substantially higher, more-than-offsetting the higher manufacturing costs associated with a more complex and silicon-intensive technology. The net result is a significantly stronger GM profile. As per our estimates, MU's HBM GMs currently stand in the ~60% range, while non-HBM GMs stand at ~35%. However, there is downside risk to these estimates depending on competitor product qualifications (increased competition diminishes individual vendor pricing power, resulting in GM headwinds).
Due to its 3D architecture and larger die size, HBM typically consumes 3-4x more silicon during the manufacturing process as compared to standard DRAM (i.e., HBM requires much more wafer capacity to produce the same amount of bits). This is conceptually understood as the "HBM trade ratio" and is particularly important because it grows larger with each new generation of HBM technology (e.g., as customers transition from HBM3E to HBM4, the trade ratio rises from 3:1 to ~4:1). This gives rise to two outcomes: (1) tighter supply for standard DRAM, which results in favorable pricing for products that account for more than ≥50% of MU's revenues (DBe), and (2) a higher floor for DRAM utilization, which results in improved profitability via higher fixed-cost absorption.

[SemiAnalysis @SemiAnalysis_]

PDOOM ALERT 🚨 : ~48% of e2e LLM latency is prefill, ~52% is decode. Prefill itself breaks into 2 ops:

🟠 Prefill extend (cache write) — ingests new context/files, writes fresh KV tokens
🟠 Cache read — reuses existing KV cache from prior turns

[fin @fi56622380]

回顾2025年半导体市场，真的是有太多太多精彩的故事，最大的主题就是:

AI需求驱动导致半导体基建的估值体系重构 + 产业链的价值分配重写

从2024年开始，半导体基建正在飞速吞噬整个IT产业利润，SP500里半导体净利润EPS在IT行业里占比，在两年时间从不到20%上升了到了40%，而且还在呈加速上升姿态

半导体整体前瞻利润率从2023年的25%已经升到了2025年11月的43%，已经明显超过了几个互联网巨头的平均利润率，这也印证了半导体利润率超过互联网会是新常态。整个IT产业的利润分配，流向半导体的比例越来越大。

要知道，就算是20~22年的半导体芯片荒，短缺如此严重，半导体的利润率和整个IT利润分配也没有显著增长

这就是故事的上半篇：AI需求驱动导致半导体基建的估值体系重构，不再是互联网时期的基建从属地位

------------------------
这个现象背后的逻辑是商业模式随着技术特性的变迁：

互联网时代，每次请求的网络和算力成本，边际成本极低，scaling的效果极好，分发的边际成本几乎为零

在AI时代，这个互联网时代分发边际成本几乎为零利于scalable的特性遭遇了根本性的重大挑战：且不说训练成本从此不是一次性开销而是年年增长，就客户的AI推理请求而言，由于inference scaling成为共识，加上垂直领域仍然需要更大规模的旗舰模型来保持竞争力，推理的成本不会随着硬件算力价格的通缩而同步降低

互联网企业从前的最大成本只有OPEX尤其是SDE人工成本，而现在，互联网公司历史上第一次像半导体厂foundry那样背上高折旧成本的资产负债表，商业模型恨不得要慢慢从“流量 × 转化率”部分转向“每 token 毛利”了

简单的说，互联网时代到AI时代的成本分布，在人力成本opex的基础上又加上了沉重的硬件/算力成本capex(财报里占比：MSFT 33%, Meta 38%)。

上个时代的互联网公司+CSP+SAAS是收租行业里的大赢家，而AI时代，算力(半导体/芯片折旧)成为了新的收租行业，整个IT行业的利润分布发生了剧烈的重新分配(EPS利润流向半导体从20%升到40%而且持续攀升中)，这就是半导体基建估值体系重构最重要的原因

---------------
半导体高利润率的新常态趋势能持续多久？

目前的高溢价来自于前期不计成本的军备竞赛造成的半导体订单积压过多
但很显然，hyperscalers都不愿意当冤大头，都在试图自建ASIC降低成本，那么可以从2030年远期的算力分布来回看这个问题

长线来看，openai已经明牌了标准答案，10GW Nvidia，10GW ASIC，6GW AMD，其他hyperscaler划分比例有类似考虑

比如说，推理端希望ASIC >50%，GPU里再细分的话，AMD和NV(legacy)对半分。训练还是得NV占大头，60%+,剩下的自研ASIC和AMD对半分

2030年按60%推理，40%训练比例划分，算下来NV 38%， ASIC 39%， AMD 23%，跟openAI比例是几乎完全一致的，算是一个标准答案参考值

当然了，微软，Amazon，Google，Anthropic这几家里AMD的比例会比这个标准答案中枢/参考值明显低一些，xAI则是没有ASIC只有Nvidia+少量AMD

AMD的风险在于，当2030年再往后的更长期，CSP的in house ASIC越来越成熟(微软除外)，推理端ASIC占比可能越来越高，很难有incentive新买入大量GPU了，除非卖的足够便宜

最近风头正劲的TPU呢？Meta是不是要转向TPU？对Nvidia的利润率影响大吗？

实际上，Meta今年capex72B，明年capex110B，未来六年capex平均值可能达到160B附近，而Meta 6年10B的TPU订单算下来年均只有1.6B，而且购买的是TPU云服务，并不是裸TPU

也就是说，Meta这笔TPU订单只占到Meta未来6年capex的1%，并没有严肃的考虑大规模部署，可能只是作为和Nvidia讨价还价的手段而已

另外从Meta最近几个月的招聘广告来看，也并没有看到任何TPU engineer方面的招聘，不像
Anthropic那样从五月就招一堆TPU kernel engineer，十月才宣布大规模采购TPU做训练

所以说，不管原因是diversify供货商，还是给自研ASIC延迟做退路，还是因为AMD的MI350X延迟，Meta买TPU基本上只有一个考虑：增加买Nvidia GPU的议价权，但顶多只有推理份额里能讨价还价，实际效果很有限，对Nvidia利润率影响也很有限。

要知道，22年加密货币熊市矿难的时候，NVDA库存上升到了198天，利润率只是从65%回撤到了56%，算上PE/宏观双杀股价才从300变100，现在一直供不应求，利润率没道理能降下来

再加上TPU v8设计过于保守(没用HBM4)，Kyber rack的Rubin方案会比TPU v8的TCO更好，到头来最后还是得继续依赖Nvidia，很难议价。只要Nvidia继续保持这样的大踏步前进，竞争对手其实要跟上还是不容易的。

总之，一方面，全产业链瓶颈，比如cowos扩张都很谨慎，供不应求的状态还能持续多年。

另一方面，AI变现的利润曲线和硬件投入曲线存在“时间错配”，应用端的增长曲线会落后几年，只要这个应用端和基建端的增长曲线的时间错位依旧存在，半导体在IT行业的利润分配就会一直占优势。

从OpenAI的到2030年的投入曲线来看，这个时间错位至少要持续到2030年附近。也就是说半导体行业的超级扩张期带来的在IT产业利润划分的主导地位，目前看至少能持续到2030年

而半导体高利润率可能会维持的更长远一些，因为从互联网时代一次性基建属性变成了现在的收租基建属性

---------------------------------------------------

AI 不是只养活了 GPU，而是在用算力预算把“能把电变成 token 的每一环”都抬了一轮，从内存，存储，互联，光纤，电力，储能…..等等

上半篇讲完了“半导体吞噬IT利润”，那么下半篇讲的就是“AI算力价值溢出效应（Spillover Effect）重塑半导体内部格局”：GPU算力增长 -> 内存/存储/互联/CPU瓶颈 -> 溢出效应 -> 结构性机会

2025 年更有趣的故事，是巨大的行业红利在半导体内部怎么诞生结构性新机会，比如说，一个super cluster需要几个数据中心互联，光纤互联的长度需要上百万mile这个级别，这就是新机会

半导体产业链的结构性趋势带来的新机会，最典型的例子就是内存(DRAM/HBM)和存储(SSD)，HBM的需求增长太夸张，连带挤压DDR4/5产能，直接让以周期性为标志的内存行业甚至喊出了“周期不存在”了，Hynix因为在HBM上领先，甚至都开始憧憬起了几年后年利润1000亿美元，妥妥一个万亿市值的公司

这两个板块背后，是结构性趋势的转变：AI workload从训练逐渐往推理延申，推理比例越来越大。

而推理是一个非常纯粹的吃内存带宽速度(memory bound)的事情，可以说带宽速度=token/s。模型尺寸越来越大，以及上下文context length的增加，对内存的尺寸要求也相应增大，导致了内存的需求激增：推理即内存

下一代的的GPU/ASIC内存已经成了暴力美学，配备的内存size之巨大，是三年前无法想象的，回看22年H100的80GB简直像个玩具，这才几年就增长了十倍：

Nvidia Ultra Rubin - 1024GB HBM
Qualcomm AI200 - 768GB LPDDR
AMD MI400x - 432GB HBM

内存的另外一个潜在的爆发点在端侧，也就是手机/PC/汽车/机器人的端侧LLM，这两年主流的手机旗舰机已经从6GB升级到了8GB/12GB/16GB，提前为可能的端侧LLM生态做准备，毕竟手机算力下一代就能达到150TOPS量级，妥妥的桌面级，非常暴力

潜力上来说，端侧内存升级是比云端内存增量要更大的市场，毕竟端侧终端device的数量太惊人了，每年都是billion级别，一旦端侧LLM生态繁荣起来，内存用量翻倍轻而易举，针对端侧低功耗内存/存算一体的各种设计都会跟上

但端侧genAI的软件生态，似乎明显滞后，一直比我想象的进度要慢，可能是因为这方面还处于摸索期，并没有云端那么确定的ROI，厂商们在投入上都很谨慎，我在23~24年时候看好27年，可能还是太乐观了

互联网->移动互联网用了10~15年，端侧genAI/LLM可能也需要7~10年，可能得等云端ROI开发的差不多了，边际收益下降了，才能轮得到端侧genAI/LLM拿到开发资源，跑通端侧ROI。

--------------------------------------

另一个2025年半导体内部结构性转变的故事是NAND存储，特别是企业级eSSD硬盘

结构性趋势来源也是同一个，AI workload的推理需求越来越大。内存红利也外溢到了SSD存储，甚至HDD存储，因为内存不够用就用高速SSD作为多级缓存

主要逻辑是AI推理过程中内存溢出KV cache offloading到下一层SSD存储，以及向量数据库检索/indexing，都在增加SSD存储的需求

Micron财报说的精准又直白：“AI inference use cases such as KV cache tiering and vector database search and indexing, are driving demand for performance storage.”

至于为什么存储价格在第四季度才爆发，这需要区分一下合约价格和现货价格，合约价格涨幅会温和一些，就算是最紧缺的企业级eSSD合约Q4上涨大概25%。而当NAND产能在2025年被合约慢慢的吃光，现货的价格就造成了观感上强烈的冲击，一个月上涨50%以上。

另一个未经验证的逻辑是多模态的爆发，特别是AI图片和AI视频的需求爆发，也会加剧存储的短缺，我觉得这条线只能说未来可期，但目前的视频/图片精细程度，可能还不到当年GPT3的水平，要达到出圈效果还需要一些时日。

------------------------
那么下一步还有什么趋势转移带来的半导体结构性的机会呢

那么就要先看下一步AI推理端的需求趋势是什么，毫无疑问，agentic flow的比例会越来越大，2025并不是year of agent，而是一个decade of agent

从CPU视角去看agentic workload，routing和工具处理都在CPU上，如果把常用的agentic框架做profiling，比如SWE-Agent, LangChain, Toolformer，CPU最长可以占到90%的E2E端到端延迟，throughput瓶颈也更多的卡在CPU，甚至CPU能耗也超过了总能耗的40%

Agentic AI目前是一个CPU瓶颈更多的事情，在 agentic 框架里，CPU 是永远在忙的总指挥orchestrator, 很可能会成就CPU需求的新一波回暖

AMD 2025年Q2财报（8月5日），Lisa Su明确表述了这一现象：​"In particular, adoption of agentic AI is creating additional demand for general-purpose compute infrastructure, as customers quickly realize that each token generated by a GPU triggers multiple CPU-intensive tasks."
"agent AI的采用正在对通用计算基础架构产生额外的需求，因为客户很快就意识到GPU产生的每个令牌都会触发多个CPU密集型任务。"

​Q3 财报里Lisa又明牌了一次CPU TAM increasing due to Gen AI. "Many customers are now planning substantially larger CPU build outs over the coming quarters to support increased demands from AI, serving as a powerful new catalyst for our server business."

Nvidia也是把agent flow视为CPU需求，GB200/300 架构配置的CPU比例也比以往大的多，36颗 Grace CPU : 72颗 Blackwell GPU，直接达到了1：2的水平，AMD的路线则是用1~4个256核的EPYC去服务MI400系列72~128个GPU

以后的硬件架构，一定会往优化agent workload方向发展，比如agent task graph的调度和load balancing，CPU/GPU协同micro-batching

算力上的比较，说不定以后也会摆脱现在的纯GPU token rate比较，转向整个系统级全栈agentic benchmark比较.

--------------------------
半导体结构性转变带来的机会同时，下一步，可能也会带来一些意想不到的次生效应

云端AI数据中心需求爆发，造成内存和存储的暴涨，给消费电子的成本带来了很大压力，在2026年，这也许会演变成消费电子产业潜在的黑天鹅

PC厂商最近的股票大跌，也是这个原因。HP已经说了要减少内存配置，暗示要把PC重回8GB内存+256GB存储的时代了。

DRAM内存和存储再这么涨下去，可能会出现很离谱的情况：内存/存储现货价格比CPU和GPU还要更贵。尴尬的是，这可能直接延缓了消费电子期望的AI PC的进程，毕竟大内存是更有利AI PC的表现力的。

夸张的说，每个PC厂商和手机厂商的员工，甚至是消费电子厂商的员工，都应该买入存储和内存，作为职业风险对冲

明年年初开始，安卓阵营的内存以及存储成本要压不住了，三星，小米的手机售价都提高的话(美国市场现在已经提高不少了)，利好最大的就是苹果

苹果的内存产能，nand产能都是专属长约锁价特供的，顺带还把Kioxia给坑了好多不涨价产能，导致苹果的成本优势进一步扩大，苹果全球手机销量市占率增长可能会非常可观，接下来一阵子可能会是iphone辉煌的时光。

-----------------------

2025年半导体市场真的是太多精彩的故事了，Nvidia/AMD/TPU和各家hyperscaler的恩怨情仇引得各路下注的吃瓜群众心情跌宕起伏。
HBM/内存厂商吃到了memory-bound的红利，NAND厂商意外收获了KV cache的溢出效应，CPU在沉寂近十年后，可能会因agent orchestration再次回到增长叙事的中心

不再是Nvidia/AVGO几家算力厂商独大，而是AI workload算力价值溢出后的每一次演进，从训练到推理，从文本到多模态，从单模型调用到agentic flow，都在重写产业链的价值分配。

云端AI的繁荣正在挤压消费电子的生存空间——当PC厂商被迫讨论重回8GB时代，苹果却因供应链优势坐收渔利。这场算力军备竞赛的次生效应，可能在2026年以意想不到的方式重塑整个消费电子格局

半导体的故事不再是一条单线，而是一张持续自我重构的网。而 2025 年，大概只是合纵连横的第一回合

[Figure @Figure_robot]

Helix 02 is Figure’s most capable humanoid model yet

Today we posted a long form write-up on the site, read more about it here: https://t.co/y7qcqyvr8t

[Figure @Figure_robot]

Introducing Helix 02

It's our most powerful model to date - it's using the whole body to do dishes end-to-end and it's fully autonomous

[TheValueist @TheValueist]

$SNDK $NVDA REVISED NAND DC BOM: Under a 1 GW AI data center build costing $50B, the Bernstein Research framework for a $35B “1 GW” build can be rescaled to provide a reasonable first-order capex structure: approximately 39% of total capex is directed to GPUs ($19.5B), 13% to networking equipment such as switches and optics ($6.5B), 10% to power distribution including transformers, switchgear and UPS systems ($5.0B), 4% to thermal management and cooling ($2.0B), and 10% to land and buildings ($5.0B), leaving the remaining 24% or $12.0B for other IT hardware and site infrastructure. To tie these system-level percentages to a component-level bill of materials, J.P. Morgan’s AI server analysis of an 8×H100 system with a total server BOM of roughly $240K is used, in which the GPU complex accounts for 83% of the server BOM, CPUs 9%, DRAM DIMMs 2%, NAND-based SSD storage 3%, and other server elements (NICs, chassis, PSUs, backplane, labor, markup, etc.) 3%. Combining Bernstein’s estimate that GPU vendors capture 39% of total data center capex with J.P. Morgan’s finding that GPUs are 83% of server BOM implies that AI compute servers in aggregate (GPUs plus their surrounding server hardware) represent about 46.9% of the $50B, or $23.44B. Within this $23.44B server pool, the implied dollar allocations are: GPUs $19.50B, equal to 39.0% of total capex and 83.2% of server spend; CPUs $2.09B, equal to 4.18% of total capex and 8.9% of server spend; host DRAM $0.45B, equal to 0.90% of total capex and 1.9% of server spend; NAND-based SSDs $0.67B, equal to 1.33% of total capex and 2.8% of server spend; and other server components $0.74B, equal to 1.47% of total capex and 3.1% of server spend. The resulting top-down capex distribution for the full $50B build, consistent with these constraints, is therefore: AI servers $23.44B (46.9%), networking $6.50B (13.0%), power distribution $5.00B (10.0%), thermal management $2.00B (4.0%), land and buildings $5.00B (10.0%), and other infrastructure $8.06B (16.1%).

The next step is to isolate NAND-specific spending at the data center level rather than at the server level. The $0.67B attributed above to NAND-based SSDs represents the SSD hardware cost inside AI compute nodes only. AI and HPC operators typically also deploy separate storage tiers (scratch, training-data repositories, checkpoints, logs) whose cost is not fully captured by node-local SSDs; historical HPC procurement evidence suggests storage systems can account for roughly 10–15% of system contracts in traditional CPU-centric clusters, although this share has compressed materially in GPU-centric AI clusters. As a working assumption anchored in the observed shift toward GPU-heavy spending, a mid-case scenario is adopted in which stand-alone storage infrastructure (additional SSD shelves and storage servers beyond node-local drives) adds 50% incremental NAND spending relative to the NAND that resides inside AI nodes. Under this assumption, external NAND-backed storage contributes another $0.33B, equal to 0.67% of total capex, on top of the $0.67B of node-local SSDs. The resulting total NAND-based storage hardware outlay for the 1 GW, $50B data center is therefore approximately $1.00B, corresponding to 2.0% of total capex. A reasonable range, reflecting uncertainty about how aggressively a given 1 GW AI build provisions stand-alone NAND capacity, is roughly $0.83–1.33B of total NAND-based storage hardware, equal to 1.7–2.7% of the $50B budget.

Within that $1.00B of total SSD and storage-hardware spend, only part is the raw NAND silicon that SanDisk and its peers sell. Industry BOM analyses and SSD cost calculators consistently show that NAND flash is by far the largest material cost in an SSD, with NAND typically comprising on the order of 60–80% of the SSD BOM depending on capacity and design, while controllers, DRAM caches, packaging, firmware development and manufacturing overhead make up the remainder. Taking a 70% mid-point for raw NAND as a share of SSD cost implies that the $1.00B of total NAND-based storage hardware decomposes into roughly $0.70B of raw NAND chips (1.4% of total data center capex) and about $0.30B of controller, packaging, system integration and vendor margin (0.6% of total capex). Applying the range of total SSD spend discussed above yields a plausible band for raw NAND cost of approximately $0.50–0.93B, equal to 1.0–1.9% of the $50B build, with the central estimate at $0.70B or 1.4%. For comparison, Bernstein’s vendor-share analysis for 1 GW AI data centers indicates that memory suppliers as a whole (DRAM plus NAND) capture about 3–4% of total capex, or roughly $1.5–2.0B per $50B data center, which is broadly consistent with a memory split of approximately $1.0–1.3B to DRAM/HBM and $0.5–0.9B to NAND under the component-level breakdown above. Global memory revenue mixes showing DRAM at roughly $97B and NAND at $68B in 2024 (about 60% DRAM and 40% NAND of the combined memory market) further support the view that a $50B AI data center allocates low-single-digit percentages of capex to NAND.

On these numbers, and keeping the $50B total build constant, the impact of NAND price moves on overall project economics is small. Using the central estimate that total NAND-based storage hardware represents 2.0% of capex ($1.00B), a 50% increase in SSD pricing would raise aggregate capex by 1.0% or $0.50B, and a 100% increase would raise aggregate capex by 2.0% or $1.00B. If the focus is specifically on raw NAND chip pricing, using the 1.4% of capex ($0.70B) estimate, a 50% increase in NAND prices translates into only a 0.75% increase in total data center capex (an additional $0.38B), and a 100% increase in NAND prices lifts total capex by about 1.5% (an additional $0.70B). These sensitivities sit against a backdrop in which GPU, networking and power-infrastructure line items individually consume 10–40% of the budget and themselves exhibit significant price and lead-time volatility. Current market commentary from flash controller and SSD vendors indicates that NAND prices have already more than doubled over the past 6 months and that 2026 NAND output is effectively sold out, with further 10–30% or greater price increases expected into 2027 as suppliers restrict capacity additions. Taken together, the sharpened cost breakdown shows that, in a 1 GW, $50B AI data center, total NAND-based storage hardware likely represents roughly $0.8–1.3B or 1.7–2.7% of capex, with raw NAND content itself at approximately $0.5–0.9B or 1.0–1.9% of capex, while the largest capital line items remain GPUs (around $19.5B or 39%), the broader AI server complex ($23.4B or 46.9%), networking ($6.5B or 13%), power distribution ($5.0B or 10%), thermal management ($2.0B or 4%) and land and buildings ($5.0B or 10%).

[郭明錤｜Ming-Chi Kuo @mingchikuo]

Nvidia Vera Rubin/VR200 NVL72重點更新，整合我的最新產業調查與Nvidia CEO黃仁勳的2026 CES演講

1. Nvidia將AI伺服器VR200 NVL144 (die-based) 改名為VR200 NVL72 (package-based)。我的調查指出，VR200 NVL72將分為Max Q與Max P兩種power profile。
➢ 黃仁勳提到了NVL72，而非先前的NVL144
➢ Max Q與Max P硬體設計相同
➢ Max Q GPU/機櫃功耗 (TGP/TDP)：約1.8/190 (kW)
➢ Max P GPU/機櫃功耗 (TGP/TDP)：約2.3/230 (kW)
➢ 兩者均顯著高於GB300 NVL72的1.4/140 (kW)
➢ 因應資料中心供電規格，不同power profile提升安裝彈性。此改變也意味著Nvidia開始注意到物理世界對建置AI伺服器的限制，並將其反映在產品規格上

2. VR200 NVL72 Max Q與Max P的GPU散熱設計升級。
➢ 兩者均採用微通道冷板 (micro-channel cold plate；MCCP) 搭配鍍金散熱蓋 (gold-plated lid)
➢ 市場高度期待GPU TGP達2.3kW後開始採用微通道蓋板 (micro-channel lid；MCL)，但MCL最快需至2H27才會量產

3. 受益於GPU、HBM與NVLink Switch規格升級，VR200 NVL72的AI訓練/推理算力約是GB300 NVL72的3.5/5倍，機櫃的供電也因此大幅提升。
➢ VR200 NVL72的power shelf升級到3*3U 110kW (6×18.3kW PSU) (vs. GB300 NVL72最普遍的8*1U 33kW (6×5.5kW PSU))
➢ VR200 NVL72 power shelf採3+1備援設計
➢ VR200 NVL72的電力進線 (power whip) 提升到100A (vs. GB300 NVL72的60A)，這對資料中心供電要求更高 (如母線 (busway) 與插接箱 (tap-off box) 等設計)。

4. VR200 NVL72的散熱設計更依賴液冷方案。
➢ Compute & NVSwitch tray均採用無風扇設計 (fanless)
➢ 機櫃冷卻液流量 (TCS flow) 幾乎增加100% (vs. GB300 NVL72)，有利CDU、分歧管、水冷版與QD規格與/或數量升級
➢ 機櫃風量 (Air flow) 需求則降低約80% (以CFM計) (vs. GB300 NVL72)。

5. VR200 NVL72的Compute tray首度採用midplane，落實無線設計 (cableless)。
➢ Midplane關鍵規格包括：44層 (22+22)、M9 CCL (EM896K3)、尺寸約420*60 (mm)。
➢ 黃仁勳提到，受益於此設計，Compute tray的組裝時間可由2小時顯著降低到5分鐘 (vs. GB300 NVL72)。

6. 2026年Rubin CoWoS投片量約30-35萬片，預計在1Q26初開始小量試產並於2Q26末量產。VR200 NVL72的機櫃組裝量產預計在3Q26底，考慮到良率爬坡，預計2026/2H26機櫃出貨量約5,000-7,000櫃。

7. 當機櫃功耗達約200kW，54V配電會開始遭遇顯著的空間占用 (銅排/線材) 與轉換效率瓶頸，故VR200 NVL72可視為Oberon機櫃的最後一代方案；後續面對AI算力提升帶來的挑戰，將採用支援800V HVDC的次世代Kyber機櫃設計因應。

[汉松 @Yonah_x]

这两天听完 Manus 首席科学家季逸超（Peak）的最新一期播客，整理一下技术相关的、我觉得有意思的观点。

他有个暴论：200k 以上的上下文其实不重要。乍一听反直觉，但他的逻辑我觉得说得通。他的核心观点是：比起无限扩展上下文，更重要的是让模型具备“压缩意识”。

人类的工作记忆也不强，但我们知道可以把信息整理成文档（比如放进 Notion），需要时再检索。模型也应该学会类似的策略：识别哪些信息可以放到外部存储（文件系统/数据库），哪些历史可以压缩成更紧凑的表达继续保留在上下文里。

这里的关键不只是“压缩”，而是让模型理解：信息是“被压缩”而不是“消失”。换句话说，压缩后的摘要要能在需要时被展开、被引用、被追溯。这种能力需要专门训练，但目前多数 Chatbot 模型并没有针对它进行系统训练。

把这点放到产品形态上，就会引出另一个矛盾：Chat 模型与 Agent 场景的优化目标往往是冲突的。Chat 模型倾向于尽快给出一个看起来完整的答案；而 Agent 任务更像 ReAct 流程：反复调用工具、观察、再决定下一步行动。这也是为什么 ChatGPT 现在有 GPT5.2、GPT5.2-Chat、Codex-5.2 等 N 个版本：不同场景需要的能力不一样。

更反直觉的是：一些在榜单上推理表现更强的模型，放到 Agent 场景（多轮工具调用、长轨迹执行）里反而更容易“掉链子”。为什么？因为长上下文场景下，O 系列这类模型的指令遵循能力下降，幻觉和幻觉工具调用概率反而上升。也就是说如果你直接把为竞赛编程或数学设计的推理模型平移到 Agent 场景，效果反而会变差。

Peak 给的解法是“交错式思考”：获得观察结果后，进行短暂的中间推理，再决定下一步行动。而不是用户问一个短问题，模型哗一下在脑内想几千字。这个解法已经是业界共识了，最近的 minimax 模型用的就是这种范式。

[TMT Breakout @TMTBreakout]

The Non-Bubble that disappointed both Bulls and Bears -- how Sam's Splurge changed everything

The worst kept secret among Tech market participants — just something AI bulls don’t admit out loud: they want a price-action bubble every bit as much as the bears do.

Both want to see that steep, “blow-off” ascent that characterizes parabolic tops. Why? The AI bulls are all fully loaded for a vertical melt-up and AI bears want the aftermath so they can yell “I told you so.”

It’s obvious to AI bulls (us included) that we’re not in an AI bubble. The non-argument is simple: valuations are reasonable (NVDA near 20x), the equity risk premium is almost 300bps above where it troughed in the tech bubble, operating margins are rich, and we’re still very early in the demand/build out of the AI supercycle.

But bulls will typically follow this argument up by saying “‘it’s more like ‘97/’98.”

Implicit in that statement is that they’re hoping the inevitable outcome is a ‘97-’99 style ramp, with all hoping it would occur as soon as possible. Why? The simplest answer is usually the right one: everyone likes bigger bonuses as soon as possible.

Stated succinctly: the “AI bubble” ascent was the paradigm that both bulls and bears were operating under for most of this year, or longer.

Bad news for the AI bulls and bears: the past few weeks has brought an end to that paradigm and led us to an unexpected turning point in the dynamics of the AI trade/narrative. On the 3 year anniversary of ChatGPT’s release, no less.

And we have Sam’s $1.4T 30GW splurge to thank for it.

Sam’s Splurge (we’ll call it “SS”) opened up AI “pandora’s box,” shifting the AI narrative in unexpected ways.

First, the overarching discussion has shifted to a greater focus on OAI’s ability to monetize and what that means across the Tech ecosystem, from ad platforms to software/services companies to GPUs to infra hosting like ORCL. Despite the behemoth it already is, the market began to appreciate it was taking implicit bets on what is still a 3 year old start up industry/company.

Second, SS and connected deals brought more focus to the interconnectedness of the whole ecosystem, OAI’s outsized role in it, circular financing, and “too big too fail” discussions.

The interconnectedness of the AI ecosystem📷
Third, SS and his “Give us a few months and it’ll all make sense … We are not as crazy as it seems. There is a plan.” opened up discussions about government’s role in AI. While government intervention would help accelerate the AI buildout, it also opened a doorway of investor doubt. Reader CIO At CG expressed these opposite outcomes well in TMTB Chat:

“It’s bullish if/when it happens. But until it happens it creates doubt if it is gravy (more upside) or if it’s needed to execute the 1trn+ commitments. Any doubts on ability to execute the 1.4trn is just bearish sentiment vis-à-vis today. So Friar opened a door that was closed. And by opening it, it opened both the left and right side of the distribution. It also makes people realize that they are too big to fail: if they fail to execute they will bring the entire ecosystem multiple down. And hard.”

Fourth, the sheer scale of the SS $1.4T plan, which is nearly the size of the whole private credit market, nudged both public and private lenders to reprice AI-linked risk, most notably seen in the rise of Oracle and Coreweaves’ CDS spreads. At the same time, off-balance-sheet structures—e.g., Meta’s $27B Hyperion SPV with Blue Owl— didn’t help by concentrating risk with private creditors and muddying system-wide leverage mapping.

The ironic thing is, if SS would have been half the size, things would have continued to grind along, investors would have enjoyed the ‘27 and ‘28 visibility, maybe even building the energy for a large vertical ascent in price action. Instead, it had the opposite effect: pouring too much gasoline on the fire and drowning out the energy for a big move up.

Fifth — by locking in commitments eight years out, SS dragged the long-horizon AI debate into the present. Over the last few weeks I’ve heard an increasing amount of bulls give voice to risks they’ve been able to normally wave away over the first 3 years of the AI trade, in an unusual sign of humility. Some of the key existential questions that now feel more present in the discussion:

How does the grid support the post-’28/’29 buildouts and what about water, land use, and local pushback? We’ve already heard of local governments slowing DC buildouts, and this week the WSJ wrote how Bernie and others are dialing up scrutiny of Data CentersIf inference moves to phones/PCs/cars, how does that rebalance hyperscaler capex, useful life assumptions, and who captures value? What’s the risk of stranded assets if models plateau or workloads shift to cheaper/edge solutions?The AI catch-22 no bulls want to talk about: If enterprise agents and automation work as advertised, what’s the path for unemployment and wages? If white-collar unemployment rises, what happens to ad spend and consumer wallets — remembering that GOOGLE and META are cyclically exposed ad businesses at their core? How does the seep into their top line and capex trajectory? If AI models don’t deliver, do we get a capex hangover and productivity disappointment?All of this sits against a U.S. backdrop that’s still skeptical of AI — worried about job loss and asking for a slower, safer rollout — which can swing sentiment and policy quickly. Will the current administration still be as supportive of the AI rollout if sentiment and unemployment shift in a more negative direction?

These are issues that will be a lot more prominent in the next 3 years of the AI trade than they were in the first 3 years of the AI trade.

This all began to seep into the price action of AI stocks several weeks ago: ORCL giving back all of its “monster RPO” move and more, very speculative sectors like Nuclear/Quantum rolling over, and the AI ecosystem progressively rallying less and less on each Open AI deal that was announced.

It all culminated in the last two weeks. We can give thanks to some hawkish fed speak and Sam’s now infamous BG2 pod appearance for providing the spark needed to ignite the fire spreading. In a period of time where nothing has changed fundamentally in respects to the AI trade, the market began more heavily digesting the overarching effect of SS: more unknowns and more uncertainty in the minds of investors. After all, the market isn’t just a mechanism for discounting fundamentals and perceived risk, but also the current emotional state of participants. With belief shifting from inevitable euphoria (read: vertical ascent price action) to verification, SS has had the opposite effect of what Altman likely intended: more multiple compression and less belief in out year estimates.

With greater uncertainty, it’s no wonder certain pockets of the market have underperformed: names with perceived questionable business models / debt issues (ORCL, CRWV, NBIS, Miners, etc.), names with perceived AI top of funnel / structural issues (DUOL, MNDY), names with rising opex as the market is less confident in how long heightened spend will be here to stay (META).

It’s also no surprise that as the market digests these new developments, the profitability factor has outperformed while names with good narratives and fast growth but little in the way of valuation support have underperformed: NET, PLTR, SHOP, TSLA, U. This is also why memory has been so strong: EPS revisions are currently happening —> there’s nothing uncertain about opening up your favorite DRAM/NAND spot price checker, seeing how much DRAM/NAND has risen overnight, and plugging it into your model. These names are arguably more attractive in the current environment than they were before.

The market is currently doing what it always does after a narrative/paradigm shock: digest, recalibrate, reassign risk premia. NVDA EPS and Gemini 3 are the next events on the docket to absorb. We’re running low gross while we let the market do its thing, letting the overarching narrative/price action stabilize and become clearer.

@dylan522p at Semianalysis joked this week that time is now divided in BC (Before ChatGPT) and AD (After Da Launch of ChatGPT). We think the AI trade will eventually be divided between BSS (Before Sam’s Splurge) and ASS (After Sam’s Splurge). BSS and ASS.

Wait - that doesn’t have a nice a ring to it, so let’s say it differently. We think the straight-line giddy phase of the AI trade will give way to something healthier: a phase where fundamentals and idiosyncrasies matter even more. Tech will always be a narrative and boom and bust heavy investing sector (that’s part of the fun), but in a landscape where sentiment is more balanced, stock-picking will become more relevant. That’s a good thing.

SS popped the non-bubble. But the AI trade isn’t broken: it’s simply entering a more mature, scrutinized phase.

[Figure @Figure_robot]

Designed for Helix, the home, and the world at scale, Figure 03 establishes the foundation for true general-purpose robotics. Read our detailed write-up:

https://t.co/tQgWizBuvg

[Meituan LongCat @Meituan_LongCat]

🚀 LongCat-Flash-Thinking: Smarter reasoning, leaner costs!
🏆 Performance: SOTA open-source models
on Logic/Math/Coding/Agent tasks
📊 Efficiency: 64.5% fewer tokens to hit top-tier accuracy on AIME25 with native tool use, agent-friendly
⚙️ Infrastructure: Async RL achieves a 3x speedup over Sync frameworks

🔗Model: https://t.co/PqZGT2HNq1
💻 Try Now: https://t.co/IccK18HvhO

[Tesla @Tesla]

https://t.co/ZA7KCwX7N9

[TrendForce @trendforce]

📢 Jensen Huang brings good news from Beijing — #NVIDIA gets U.S. green light to restart #H20 sales in China. #Samsung, the top HBM3 supplier, is set to cash in.💡More: https://t.co/ykYiQH4zED 🔗

[louise @lotussucong]

Novaspace reports Starlink outpaces launches: SpaceX enters new era of profitability  https://t.co/opPuYybAft

[Win Smart, CFA @WinfieldSmart]

Gold 🟡 as a store of value
visualization

[歸藏(guizang.ai) @op7418]

Andrej Karpathy 新文章，主要介绍了为什么大语言模型颠覆了传统的技术扩散路径

未来已经到来，而且分布惊人地均衡。顶尖 AI 在一夜之间免费出现在每个人口袋里的设备上。

历史上，变革性技术通常遵循"自上而下"的扩散模式：最初由政府或军方开发，然后被企业采用，最终才到达普通个人。

LLMs呈现出截然不同的扩散模式 - 它们为普通人带来了不成比例的巨大益处，而对企业和政府的影响则相对滞后且有限。

最关键的是，使用门槛极低 - 模型费用低廉（甚至免费），速度快，任何人都可以通过网址或本地设备随时访问。

LLMs之所以对个人如此有价值，主要因为它们提供了一种非常特殊的能力组合 - 在极广泛的领域中提供准专家级知识/表现，尽管这种能力相对浅薄且容易出错。

个人通常最多只精通一个领域，因此LLMs的广泛准专业性让他们能够做以前做不到的事情。

对组织的影响较为有限的原因主要是：

- 组织的独特优势与LLMs的能力特征不匹配
- 组织处理的问题复杂度更高，需要更多协调
- 组织固有的惯性，包括文化、历史先例、官僚主义等

"利益分布"的未来格局特别值得关注，这主要取决于性能与资本支出的动态范围。

如果未来金钱能够买到显著更好的AI，情况将发生变化 - 大型组织将集中其巨大资源购买更多智能，精英人群也可能再次从社会其他人群中分离出来。

回顾各种科幻作品，几乎没有人预测到AI革命会以这种方式发展 - AI不是作为政府秘密项目出现，而是基本上在一夜之间免费出现在每个人口袋里的设备上。

这颠覆了William Gibson著名的"未来已经到来，只是分布不均"的名言。出人意料的是，未来已经到来，而且分布惊人地均衡。这是真正意义上的"人民的力量"。

[louise @lotussucong]

https://t.co/hcKHUvBUzN

[Manny Li @MannyNewsletter]

(1/4)

來自中國的 DeepSeek 事件昨晚擊沉了 NVIDIA 15% 股價。以下是一早起來速讀完一些雜食閱讀後，我目前還放在心中的 memo。
​
1. DeepSeek 令人印象深刻之處是它的訓練跟推論成本都極低，而高成本一直是過去三年來人們講到 AI 時最顧慮的，因為成本（以及連帶的電力需求）是發展的瓶頸，且突破瓶頸的方式似乎只有一個：跟 NVIDIA 買東西，因為 Jensen 說「買越多才會越便宜」（更大的叢集）。
​
2. 因此，DeepSeek 在價格上帶來的破壞，對所有其他產品與服務業者而言都是超級利多。例如，原本以為一年要花 100 億美元才能提供的服務，可能突然變成只要一半、甚至十分之一。
​
3. 承上，我一直認為一定規模的產品與服務業者會是這波 AI 革命的最大贏家。技術變化很快，但人性以及基於人性的僵化組織則變化很慢。掌握通路與銷售關係是關鍵。
​
4. 然而，DeepSeek 帶來的價格破壞也讓更多人有了彎道超車的機會，不論是新創，或是想要做 AI 升級的既有公司。