供应链分析
产业链结构、上下游与瓶颈环节研究 · 共 1315 条 · 滚到底自动加载更多
-
澄清自动驾驶技术等级与部署策略差异，避免重蹈Cruise覆辙。
@prudhviregula 当然，这是细微的语义差别。从技术层面看，它已达到第4级自动驾驶(L4)，例如 $NVDA、现代汽车以及分析师的观点。 但从其在德克萨斯州的启动来看，出于风险管理，它最初选择了监督式部署，以防止重蹈 Cruise 在加州的覆辙。 我与 Anthropic 进行了深入研究以再次确认。https://t.co/V87kdo0vX3
英文原文
@prudhviregula Sure, it's nuanced semantics. It's at level 4 tech wise, eg. $NVDA, Hyundai, and analysts. But from its Texas launch it chose supervised deployment for risk management at the start to prevent what happened with Cruise in CA. Did a deep research with Anthropic to double check. https://t.co/V87kdo0vX3
-
质疑市场对Uber与Nvidia自动驾驶估值逻辑的矛盾。
市场到底在抽什么？ 华尔街因自动驾驶汽车业务，将市值超1800亿美元的$UBER推高3.4%，今日市值增加超60亿美元。 然而，Uber所使用的自动驾驶汽车母公司$NBIS却下跌3%，市值现不足240亿美元。https://t.co/ahpz7B0mwx
英文原文
WHAT IS THE MARKET SMOKING??? Wall Street sent $UBER, a $180B+ company, up 3.4%, $6B+ today off of its self-driving cars. But $NBIS, the parent company of the self-driving cars Uber uses, got sent down -3% and is now worth less than $24B. https://t.co/ahpz7B0mwx
-
质疑AirWalletX架构能否防止中国政府强制获取美客户数据
人们并不关心工程师坐在哪里或数据中心建在哪里。 关键在于，中国政府是否能够像通过 TikTok（追踪记者）那样，强制访问或为 AirWalletX 植入后门以获取美国客户数据。 用一个简单的“是/否”问题来消除疑虑： 在您当前的技术架构下，是否不可能让任何身处中国/香港的人员或实体，即使在其政府依法强制要求下，也能从您的系统中获取美国客户/商业或交易数据？
英文原文
People aren't concerned where engineers sit or where data centers are. It's that the Chinese Government is able to forcibly access / backdoor US customer data like with Tiktok (tracking journalists), on AirWalletX. Simple Yes/No Question to shut down concerns: Under your current technical architecture, would it be impossible for any person or entity in China / Hong Kong Kong, even if legally compelled by their government, to obtain U.S. customer/business or transaction data from your systems?
-
补充AI供应链受益股并增加AMKR持仓
@B38B37 是的，内存也是。还需要补充联发科（I/O 芯片let）、$CLS、$JBL 和 $AMKR（TPU v7 模块封装的第二供应商）作为受益者。这正好说明了供应链的规模有多大。另外，正如 @Mexicancik1 提到的，我也增加了 $AMKR 的头寸。
英文原文
@B38B37 Yeah memory is. Also needed to add MediaTek( I/o chiplet), $CLS, $JBL, and $AMKR (second source for packaging the TPU v7 modules) up there as beneficiaries. Just goes to show how large the supply chain is. Also adding positions in $AMKR as @Mexicancik1 mentioned.
-
分析Google TPU v7供应链，建仓Lumentum以博弈TPU生态扩张。
对 $GOOGL TPU v7 Ironwood 供应商的分析。 以下是受 Google TPU 建设影响最大的公司列表。 + 我正在建仓的 TPU 相关股票。 [关键] 设计/IP： - 博通 [ $AVGO ]：共同设计并实现 Google 的 TPU ASIC（专用集成电路） [关键/高] 半导体晶圆代工： - #1 台湾半导体 [ $TSM ]：TPUv7 在 TSMC 3nm 工艺制造 - #2 三星电子：次要存储及晶圆代工合作伙伴 [关键/高] 存储： - #1 SK 海力士：为 TPUv7 Ironwood 提供 HBM3E - #2 三星电子：~TPUv7 特定报告强调 SK 海力士 + 三星。 [高] 光网络： - Lumentum [ $LITE ]：Google 广泛使用光电路交换 (OCS) - Coherent [ $COHR ]：OCS 参与者但较弱 [高] 电源管理 IC： - Monolithic [ $MPWR ]：这是一个投机性观点，即 Vicor 将被 $MPWR 取代，源自财报中提及 TPU [中] 热管理： - Vertiv [ $VRT ]：Vertiv 供应作为液冷系统核心的 CDU（冷却分配单元），将冷却液泵送至 TPU 芯片的冷板 - Modine [ $MOD ]：更投机性地认为他们提供大型冷水机组和空气处理单元 (AHU) [中] 互连： - TTM Technologies [ $TTMI ]，$ANET，Unimicron，Ibiden ______ Google TPU v7 “Ironwood” 的建设代表了一个旨在打破 $NVDA GPU 垄断的平行硅生态系统的构建。 实质性影响最集中在博通（作为硅架构师和商业载体）、存储综合体（SK 海力士/三星）以及光网络/电源领域（Lumentum/Vertiv），这是基于公开证据创建的，但很大程度上取决于采用率、供应商份额和竞争反应的实际表现。 从这项供应链研究中，我正在 $LITE 建立新头寸，以防 TPU 成为推理领域的主导 ASIC。 Lumentum 是 Google 致力于 OCS 的主要受益者，并构成了 TPU 吊舱中使用的 “Apollo” OCS 交换机的核心。 TPU v7 集群的爬坡直接转化为 Lumentum 光开关模块的出货量。由于 OCS 是 Google 超大规模方法独有的定制架构，Lumentum 在此处面临的 commoditization（商品化）压力小于标准收发器市场。 然而，如果 Anthropic、Meta、Apple 和其他公司购买 $GOOGL ASIC 导致 TPU v7 规模扩大，该供应链中的所有公司都将受益。
英文原文
Analysis of the $GOOGL TPU v7 Ironwood Suppliers. Here's the list of what comapnies are the most materially impacted by the Google's TPU buildout. + the TPU stock I'm taking a position on. [Critical] Design/IP: - Broadcom [ $AVGO ]: co-designs and implements Google’s TPU ASICs [Critical/High] Semiconductor Fab: - #1 Taiwan Semi [ $TSM ]: TPUv7 is fabbed at TSMC 3nm - #2 Samsung Electronics: Secondary memory & foundry partner [Critical/High] Memory: - #1 SK Hynix: HBM3E for TPUv7 Ironwood - #2 Samsung Electronics: ~TPUv7-specific reporting emphasizes SK hynix + Samsung. [High] Optical Networking: - Lumentum [ $LITE ]: Google uses extensively uses Optical Circuit Switching (OCS) - Coherent [ $COHR ]: OCS player but weaker [High] Power Management ICs: - Monolithic [ $MPWR ]: This is speculative that Vicor will be replaced by $MPWR, from earnings mentioning TPU [Medium] Thermal Management: - Vertiv [ $VRT ]: Vertiv supplies the CDUs that act as the heart of the liquid cooling system, pumping coolant to the cold plates on the TPU chips - Modine [ $MOD ]: More speculative that they provide provides the massive chillers and air handling units (AHUs) [Medium] Interconnects: - TTM Technologies [ $TTMI ], $ANET, Unimicron, Ibiden ______ The buildout of the Google TPU v7 "Ironwood" represents the construction of a parallel silicon ecosystem designed to break the monopoly of $NVDA GPU. The material impact is most concentrated in Broadcom (as the silicon architect and commercial vehicle), the Memory Complex (SK Hynix/Samsung), and the Optical/Power sectors (Lumentum/Vertiv) and was created from public evidence but is largely dependent on adoption, vendor shares, and competitive responses actually play out. From this supply chain research, I'm initiating a new position in $LITE, in the event the TPU becomes the dominant ASIC for inference. Lumentum is primary beneficiary of Google’s commitment to OCS and form the core of the "Apollo" OCS switches used in TPU pods. The ramping of TPU v7 clusters translates directly to unit volume for Lumentum’s optical switch modules. And because OCS is a bespoke architecture unique to Google’s hyperscale approach, Lumentum faces less commoditization pressure here than in the standard transceiver market. However all companies in this supply chain are set to benefit if the TPU v7 scales up from Anthropic, Meta, Apple, and others buying the $GOOGL ASIC.
-
EIP-7918对ETH销毁量影响微乎其微，未改变核心销毁机制。
EIP-7918 设定了与执行成本挂钩的价格下限，但并不会导致销毁率(burn rate)出现实质性跃升。 1. 以 Base 交易为例，L1 数据成本为 0.000000000390881 ETH。若 ETH 价格设为 4000 美元，该成本仅约 0.0000015 美元。应用 EIP-7918 后，在 Base 等 OP Stack 二层网络上的成本增加约 39%，但即使增加后也不超过 0.0000021 美元。微乎其微。 2. “假设以太坊基础费用(base fee)保持在 1 Gwei，且全天每块包含 6 个 Blob，若应用 EIP-7918，因 Blob 费用每日销毁的 ETH 如下：换言之，每日销毁量约为 0.32 ETH。考虑到当以太坊基础费用为 1 Gwei 时，仅基础费用每日就销毁约 100 ETH，这并不算多。”来源：All About Fusaka: Seungmin Jeon 这并未保留机械式的销毁结构。
英文原文
EIP-7918 sets a price floor tied to execution costs but does not translate to a material jump in burn rate 1. In the case of a Base Transaction, 0.000000000390881 ETH was used as the L1 data cost. If we set the ETH price at $4,000, this cost is only about $0.0000015. With EIP-7918 applied, this cost increases by roughly 39% on OP Stack rollups like Base, but even after the increase it does not exceed $0.0000021 Negligible. 2. "Assuming Ethereum’s base fee remains 1 Gwei, and with 6 blobs per block throughout the day, if EIP-7918 is applied, the daily ETH burned due to blob fees would be as follows: In other words, the daily burn would be approximately 0.32 ETH. Considering that around 100 ETH is burned daily from the base fee alone when Ethereum’s base fee is 1 Gwei, that’s not a large amount." Source: All About Fusaka: Seungmin Jeon This does not preserve the mechanical burn structure.
-
L2扩容优化抵消销毁，底价机制下ETH销毁量依然微小。
情况可能比我读到的更糟。你假设我们看到的是L2（Layer 2）创纪录的增长。我则基于适度+轻微增长的视角。借用 @Seungmin Jeon 的数学模型： - 所以并不显著。L2在大多数情况下几乎不在blobs（数据块）上花钱。底价机制防止了费用螺旋降至1 wei。有人分析了EIP-7918：当L1基础费为1 gwei时，底价每天增加约0.32 ETH的blob销毁量。底价是L1基础费的1/16，相对于1 wei有意义，但绝对值依然极小。 - Base的L1数据成本增加39%，从每笔交易$0.0000015升至$0.0000021。PeerDAS增加了8倍的blob容量。 供应扩张而押注需求并非保证的销毁飞轮，L2总是可以优化绕过DA（数据可用性）销毁。更多容量 -> L2围绕其优化 -> 费用维持在底价 -> 销毁保持最低。
英文原文
So it might be even worse from what I've read. You're going under the assumption we're seeing record levels of L2 growth. I'm going under the lens we see modest + slight growth. Just using someone else's math @ Seungmin Jeon - So not meaningfully. l2s currently spend almost no money on blobs in most situations. The floor prevents the spiral to 1 wei. Someone else did an analysis on EIP-7918: the floor adds ~0.32 ETH/day in blob burns when L1 base fee is 1 gwei. The floor is 1/16th of L1 base fee, meaningful relative to 1 wei, but still tiny in absolute terms. - A 39% increase on Base's L1 data cost takes it from $0.0000015 to $0.0000021 per tx. PeerDAS adds 8x blob capacity. Supply expanding while betting on demand isn't guaranteed burn flywheel, L2s can always optimize around DA burn. more capacity ->L2s optimize around it -> fees stay at floor- > burn stays minimal.
-
L2分流导致ETH主网销毁量与交易量脱钩
起初确实如此。现在已大不如前。 附两张图对比交易量与销毁量（细节虽多，但核心观点依然成立）。 简单来说，大量网络活动转移到了二层网络(Layer 2)，即以太坊主网(Ethereum Mainnet)之外。 对于数十亿的交易量，他们只需发布数据块(Blobs)，这基本上只是显示交易发生的压缩数据。 因此，所有这些交易产生的销毁量微乎其微（故图表如此显示）。
英文原文
Originally yes. Not much anymore Attached two photos volume vs. burn (there’s a lot of nuances but point still stands). In simpler terms, lot of network activity went to layer 2, which is off the ethereum mainnet. And then for billions in volume, they can just post blobs, which is basically compressed data showing that transactions happened. The resulting burn for all those transactions is minimal (hence the chart)
-
分析NBIS、IREN和CRWV商业模式优劣及新云厂商时间窗口。
好问题。麦肯锡曾就此话题发文（我觉得写得极差，因为他们以 $CRWV 为主要锚点）。 但其中部分观点成立，并对 $IREN 等公司发出警示。他们的观点： - 当前的裸金属租赁商业模式薄弱且脆弱 - 避免过度依赖少数大客户 - 开辟可防御的利基市场（如主权计算、专用工作负载） - 通过收购整合或成为超大规模云服务商 这些确实正确，但未能捕捉到一些细微差别。 关于 $NBIS： - 极度多元化（这构成了利用率的强大护城河，对利润率计算至关重要） - 全栈式（可防御的利基市场） - 通过收购整合（旨在成为超大规模云服务商，拥有4家同步增长的子公司） 这就是我说它具有最高非对称上行潜力的原因。 关于 $IREN： - 当前的裸金属租赁业务目前是护城河。文章指出长期来看它很脆弱，这是正确的。因此 $IREN 正通过与 $MSFT 合作开展 GPU 基础设施即服务(IaaS) 向上攀登全栈阶梯，并可能尝试构建上层软件层（尽管这很难） - 我们将拭目以待，这需要极高的执行力。 关于 $CRWV - 老实说，我不知道他们如何摆脱债务陷阱 - 他们试图用 $NVDA 作为后盾，但这充其量也很脆弱（例如 OpenAI 拥有 1 万亿美元以上的资本支出，试图争取政府 + 科技七巨头提供资金担保） 新云厂商是一场与时间的赛跑，我同意文章的观点（这就是我说高确信度持有2年，而非5年以上的原因）。 他们拥有从科技七巨头(Mag7)弱势中获取收入的绝佳窗口期 -> 将收入转化 -> 建立长期差异化和护城河。 我不知道最终结果如何，但我们将拭目以待。
英文原文
Hi great question. So there was an article by Mckinsey on this topic (which I think is terribly written since they use $CRWV as the main anchor). But some points holds true, and gives warnings to $IREN and others. Their claims: - current bare-metal rental business model is weak and fragile - avoids overreliance on a few giant customers - carve defensible niches (sovereign compute, specialized workloads) - consolidate through acquisitions or be a hyperscaler Are definitely correct, but fail to capture some nuances. So for $NBIS: - Extremely diversified (so this is more as a powerful moat for utilization, which is huge for margin calculations) - Full-stack (defensible niche) - consolidate through acquisitions (it's aiming to become a hyperscaler, has 4 subsidiary companies growing alongside it) That's kind of why I've said it has the highest asymmetrical upside of the bunch. For $IREN: - current bare-metal rental business is a moat as of today. The article is correct in stating long term it's fragile. That's why $IREN is moving up the full-stack ladder doing GPU iaas with $MSFT, and will likely try and build software layers on top (though it's hard) - We will see what comes out of this, it's high execution. For $CRWV - idk how they're going to get out of the debt trap tbh - they're using $NVDA to backstop it, but it's shaky at best (eg. openai with $1t+ in capex trying to get gov + mag7 to backstop funding) Neoclouds are a race against time, I agree with the article (which is why I said 2 year high conviction hold, not 5 years + ). They have this brilliant window of opportunity of weakness from mag7 -> funnel revenue down -> build long term differentiation and moats. I don't know what will happen, but we'll see
-
AI云股被算法归篮联动，CRWV财务堪忧，NBIS/IREN长期价值或超CRWV。
是的，完全同意。我认为算法/市场目前将 $NBIS、$IREN 归入 $CRWV、$NVDA 的篮子中。因此，任何关于 Coreweave 的负面报道都会对其他公司产生负面影响。我们终于看到 $WULF、$CIFR 等被归入数据中心(colocation)篮子（相对不受 GPU 贬值论点影响），并表现优异。但坦率地说，$CRWV 是一个财务噩梦，只要它被视为新云(neocloud)行业领导者，就会影响其他公司。至于 $NVDA，$GOOGL 的 TPU 论据是我目前看到的针对 $NVDA GPU 云业务最强的看空理由，但这些公司已经从 $META、$MSFT 锁定了 5 年的超大规模云服务商(hyperscaler)合约。我认为市场最终会正确定价，我相信 $NBIS 和 $IREN 有一天会比 $CRWV 更有价值，但其中只有一家拥有自动驾驶 Robotaxi lol。
英文原文
Yep absolutely. I think algos/market put $NBIS, $IREN in the $CRWV, $NVDA basket right now. So any negative hit piece about Coreweave does negatively affects the others. We've finally seen $WULF, $CIFR and others get put into the colo basket (which are relatively unaffected to GPU depreciation arguments), and outperform. But $CRWV is a financial nightmare to put it bluntly, so it does affect the others as long as it's treated as the neocloud sector leader. As for $NVDA, TPU arguments from $GOOGL is the strongest bear case I've seen to date though on $NVDA GPU clouds, but these companies already have have 5 year hyperscaler deals locked in from $META, $MSFT. imo markets will price things in correctly in due time, I do think both $NBIS and $IREN will be worth more than $CRWV one day. but only one of those has self-driving robotaxis lol
-
澄清NBIS成本估算逻辑及与微软合作路线图
与 $MSFT 的对比有些偏差，但 H100 的标准化处理虽然带有推测性，应该更稳健。 我使用了整个投资组合的混合估算值，例如芬兰设施（Mantsala），那里的数据中心租赁费实际上为 $0（仅包含折旧摊销+运营支出）。当他们从 DataOne 租赁美国设施（新泽西州 Vineland）时，我对其进行了平均处理。 正如其他人指出的那样，仅就 $MSFT 的交易而言，租赁成本可能要高得多，接近每兆瓦 $180万-$220万的市场价格。 $NBIS / $MSFT 的合作不仅限于 H200。GB200/B300 也在路线图之中。我当时正在对比 H200、H200 和其他 GPU，最后顺手加上了这个。可惜帖子发晚了没法编辑。 感谢大家的提问。
英文原文
The $MSFT comparison is off but the h100 normalization, while speculative, should be more robust. I used a blended estimate of entire portfolio eg. Finland facility (Mantsala), where colo rent is effectively $0 (just D&A + OpEx). When they lease the US facility (Vineland, NJ) from DataOne and I averaged it. For the $MSFT deal only, the lease cost is likely much higher closer to market rates of $1.8M-$2.2M per MW as someone else pointed out. $NBIS / $MSFT is not limited to H200. GB200/B300 is on the roadmap too. I was doing comparisons with H200, H200, and other GPUs and threw that in at the end. Too late to edit the post though. Appreciate the questions.
-
对比NBIS与IREN的微软交易，指出NBIS收入溢价及IREN利润率被高估。
是的！关于收入溢价的观点确实很有帮助。但我认为细微差别体现在 $MSFT 的交易中（这显示了 $NBIS 和 $IREN 每兆瓦的利润率差异）。 Nebius 的微软交易使其每兆瓦年的收入比 IREN 的微软交易高出约 19-20%。许多毛利率数据因资产负债表会计处理而被夸大，因此我发此帖以标准化利润率。 $IREN 的实际杠杆内部收益率可能更接近 20%，鉴于其与戴尔的数十亿美元支出，使用这一指标可能优于 85% 的项目息税折旧摊销前利润（EBITDA）数据。 $MSFT 基于合理推测（结合靠近 Azure 服务器的地理位置和软件优势）更看重 $NBIS 的完整人工智能云平台。如果 $IREN 在顶层软件和基础设施即服务（IaaS）层面补齐短板，其未来利润率和合同有望缩小这一差距。
英文原文
Yep! Definitely some helpful points about the revenue premium. But I think the nuance did show up in the $MSFT deal (which shows the margin difference between $NBIS and $IREN per MW). Nebius’s MSFT deal gives it ~19–20% higher revenue per MW-year than IREN’s MSFT deal. A lot of the gross margin figures are inflated by balance sheet accounting, hence why I made this post to normalize margins. $IREN 's realized levered irr is probably closer to 20%, it's probably better to use that over the 85% project EBITDA figures since they're spending billions with Dell. $MSFT values $NBIS full AI cloud platform more from an educated guess (mix of location closer to azure servers and software). If $IREN closed the software on top level and iaas level, its future margins/contracts could close that gap.
-
AI算力需求指数级增长抵消GPU迭代贬值，NVDA客户优质，非泡沫崩盘。
答案很微妙。 主要看两个因素： 1. GPU 变得更节能。 2. 大语言模型(LLM) 在容量/能效上更高效。 在 LLM 方面，我们看到像 DeepSeek 这类模型在处理不需要高精度的任务（如回答烹饪食谱或知识库查询）时极其高效。 然而……随着计算力的增加，准确率（尤其是复杂研究问题）也在提升。Elon 和 Magnificent Seven 意识到了这一点，所以他们正在扫货市场上的所有 GPU 以创造超级智能。这也是为什么 Anthropic 和 Google 正在建设耗资 400 多亿美元的数据中心，用于运行需要更多算力进行批判性思维（如 Genesis 任务）的更高级 Opus 和 Gemini 模型。 在 GPU 方面，每一代新 GPU（例如 H100 -> B200）在能效和每美元性能上都有显著提升。例如，Blackwell B200 是 Hopper H100 的 30 倍。 如果基于这个假设，那么到 2027/2028 年，市场上将出现大量过时的低效 H100 和 B200，导致二手 GPU 市场崩盘。 但是：这是假设我们没有看到对新 AI 能力的指数级需求（我们很可能会看到，且正在发生）。正因为这种指数级需求，今天旧模型（如 7 年前的 TPU 和 2020 年的 GPU）仍被用于低优先级的推理任务。 $NVDA 的订单已积压数年，人们正在购买 $AMD 的 GPU 和 $GOOGL 的 TPU 来构建任何新增产能。 至于思科类比，思科的客户是互联网泡沫时期无盈利能力的公司。$NVDA 的客户是 $META、$AMZN、$GOOGL、$MSFT，这些是世界上最盈利的公司。所以最坏的情况我们可能看到回调，而不是互联网泡沫式的崩盘。
英文原文
Answer is nuanced. So two factors: 1. GPUs get more power efficient. 2. LLMs get more capacity/power efficient. For the LLMs case, we're seeing that on deepseek type models be extremely efficient on stuff that don't require much accuracy. Basic stuff like responding to questions about cooking recipes, or knowledge-base stuff. However... accuracy increases, especially with complex research questions, scaled with compute. And people like Elon + mag7 realize this, which is why they're just buying up all the GPUs on the market to create superintelligence. And why antrhopic/google is building $40B+ datacenters for more advanced opus and gemini models that require more compute for critical thinking (eg. Genesis Mission) For the GPUs case, every new generation of GPU (e.g., H100 -> B200) offers dramatic improvements in power efficiency and performance per dollar. eg. Blackwell B200 is 30x than the Hopper H100. If we go off that assumption, then there would be a massive useless supply of less-efficient H100s and B200s in 2027/2028 creating a used GPU market crash. HOWEVER: This is if we don't see an exponential demand for new AI capabilities (which we likely will, and what we're seeing now). Because of this exponential demand, TODAY, older models are still used (eg. TPUs from 7 years ago and GPUs from 2020), for lower inference task in lower priority inference tasks. $NVDA is backlogged for years and people are buying GPUs from $AMD /TPUs from $GOOGL to build out any new capacity. As for Cisco analogy, Cisco's customers were .com bubble companies with no profitability. $NVDA's customers are $META, $AMZN, $GOOGL, $MSFT the most profitable companies in the world. So worst case scenario we might see a correction, not a .com bubble crash.
-
指出基于旧版TPU的分析已过时，强调TPU v7性能飞跃使旧成本分析失效。
你发布的 @ArtificialAnlys 分析对于评估当前 $GOOGL TPU 与 $NVDA GPU 的局势已经过时/错误。他们基于 TPU v6e 进行了分析。 我的帖子是关于更新型号（TPU v7 Ironwood 与 $NVDA Blackwell B200）在架构上的对等性。 TPU v7 Ironwood 的性能飞跃如此巨大（计算能力提升高达 10 倍），以至于他们对 v6e 的成本分析已不再相关。
英文原文
The @ArtificialAnlys analysis you posted is outdated/ wrong for analyzing the current $GOOGL TPU vs. $NVDA GPU situation. They did theirs on TPU v6e. My post was on architectural parity on newer models (TPU v7 Ironwood vs $NVDA Blackwell B200). The TPU v7 Ironwood performance leap is so substantial (up to 10x compute jump) that their v6e cost analysis is irrelevant.
-
Scale被Meta收购后，AI训练平台用户被迫转向NBIS等替代品
@RKLBMan 还有那个同比增长150%的AI训练/标注平台呢 lol 这是在Scale刚以290亿美元估值被$META收购之后——现在被迫转向$NBIS等其他平台的使用者（超大规模云服务商）
英文原文
@RKLBMan And the AI training/labeling platform growing 150% y/y lol This is after Scale just got bought by $META at a $29B valuation -> hyperscalers that used it now are forced into other platforms like $NBIS
-
分析Robotaxi竞争格局：Waymo领先，Uber联手Avride和WeRide应对巨头威胁。
欢迎反驳关于 $TSLA 是 FSD SAE 2级自动驾驶的观点。特斯拉被归类为2级，是因为其 Robotaxi（自动驾驶出租车）始终有人远程监控。埃隆正在追求一个更宏大的计划，即采用摄像头+非地理围栏解决方案，所以我做的并非简单的同类比较。以下是三点看法：1. 如果他能制造出可重复使用的大型太空火箭，我毫不怀疑他最终能在地球上实现他的计划。2. Waymo 明显领先于其他所有竞争者。3. 我的观点是，来自 $GOOGL、$TSLA、$AMZN 等万亿美元市值公司的个位数 Robotaxi 玩家正在与 $UBER 竞争。Uber 视此为威胁，因此正与 $NBIS 的 Avride（美国）、WeRide（中东、中国）合作，以在明年扩大规模。
英文原文
Feel free to dispute the point that $TSLA is FSD SAE level 2 automation. Tesla is classified as level 2, because in their robotaxis, you have people remotely behind the wheel at all times. Elon is pursuing a more ambitious plan with cameras + non-geofenced solution so it's not an apples to apples comparison I'm making. That being three things: 1. I have no doubt Elon can achieve what he's planning in due time back in planet earth if he's able to create large reusable rockets that go into space. 2. Waymo is clearly ahead of all other players. 3. Point I'm making is there's single digit robotaxi players that all come from trillion dollar company exposure in $GOOGL, $TSLA, $AMZN that compete vs $UBER. Uber sees this as a threat so they're working with $NBIS Avride (United States), WeRide (Middle East, China) to scale it up over the next year.
-
对比巨头与Avride在自动驾驶出租车领域的竞争格局及Uber布局。
🎯 对我观点的评论一针见血。人们将自动驾驶出租车(Robotaxis)视为一个万亿美元以上的市场。 $GOOGL Waymo 正在快速扩张并拥有先发优势。$TSLA 押注低成本+摄像头以实现大规模完全自动驾驶(FSD)出租车，但尚未完全到位。 人们通过购买万亿美元市值的 $TSLA 或 $GOOGL 来获取这种自动驾驶出租车敞口。 然而，有一家市值仅 210 亿美元的小公司 Avride，它是少数几家 Level 4 开发者之一（自 Yandex 以来已开发 7 年以上）。 Uber 认为 Tesla + Waymo 构成巨大威胁，因此通过与 $NBIS Avride 合作进行扩张（因此有 3.75 亿美元投资+合作），我们可以看到这将在未来一年与其核心业务一起极其迅速地扩大规模。
英文原文
🎯 Spot on commentary of my point. People see robotaxis as a trillion + dollar market. $GOOGL Waymo is expanding rapidly and has first mover advantage. $TSLA is banking on low cost + cameras to achieve FSD robotaxis at scale but aren't quite there yet. And people buy $TSLA or $GOOGL for that self-driving car robotaxi exposure through trillion dollar companies. Yet, sitting in some small $21B marketcap company is Avride, one of the only Level 4 developers out there (that's been in development for 7+ years since Yandex). Uber sees an immense threat from Tesla + Waymo so they're expanding with $NBIS Avride (hence the $375m investment + partnership) and we can see this scale up extremely rapidly over the next year alongside their core business.
-
看好传统云商，CRWV或成新云代表，板块成员间接受益。
鉴于数据的重要性，概率上更倾向于像 $ORCL 这样的联邦云(FedRamp)提供商以及像 $AMZN、$GOOGL 和 $MSFT 这样的传统超大规模云服务商(Hyperscalers)。 $CRWV 可能是最有可能成为新云(Neocloud)代表的公司，但他们仍处于应用阶段。$NBIS 的可能性不大。 话虽如此，新云板块的其他成员也是间接受益者。
英文原文
Prob just fedramp providers like $ORCL and traditional hyperscalers like $AMZN, $GOOGL, and $MSFT given how critical the data is. $CRWV is probably the most likely Neocloud but they’re still in application phase. $NBIS is unlikely. That being said, other Neocloud sector members are indirect beneficiaries.
-
梳理新云领域中小玩家定位，类比AI版AWS。
@AustranSkolSwft $WULF 作为像 $CIFR 一样的数据中心托管(Colo)服务商，属于新云(Neocloud)领域。$IREN 凭借带有 GPU 的全栈 IAAS 服务也归入此类。 但唯一的全服务纯新云玩家是 $CRWV 和 $NBIS。 不过它们都属于那个新类别：充当 AI 版 AWS 的小型参与者。
英文原文
@AustranSkolSwft $WULF is in the neocloud sector as a colo player like $CIFR. $IREN falls there too under full-stack IAAS with GPUs. Only full-service pure Neocloud would be $CRWV and $NBIS. But they're all under that new category of small players acting as AWS for AI.
-
警惕OpenAI泡沫，看好Mag7资本支出受益的AI基础设施股。
我对“AI泡沫”的主要担忧是OpenAI及其1万亿美元资本支出(capex)的承诺。这显然是一个泡沫（以及大语言模型LLM的私人估值）。其他大多数方面则不然。 任何直接依赖他们的公司，如$ORCL、$CRWV，鉴于AI模型在技术上已超越GPT，可能会陷入困境。所以简单的做法就是远离它们！ 就我个人而言，ChatGPT 5.1的表现糟糕透顶，我实际上取消了订阅，转而使用Gemini/Claude。Claude Opus 4.5在编码任务上优于Codex。Gemini在图像生成上优于ChatGPT。此类例子不胜枚举。 无论如何，AI将长期存在，任何与Mag7相关的（$GOOGL -> $CIFR, $WULF），（$MSFT -> $IREN, $NBIS），以及连接性如$ALAB (AWS), $CRDO (mag7)都极具前景，因为它们是Mag7自由现金流(fcf)增加资本支出的直接受益者。
英文原文
The main fear I have in the "AI Bubble" is OpenAI and their $1T capex promises. That is a clear bubble (and private valuations of LLMs). Most other things, no. Any company directly reliant to them $ORCL, $CRWV might be in trouble given how AI models leapfrogged GPT. So the simple thing to do is stay away! Personally speaking, ChatGPT5.1 is horrendous and I actually cancelled my subscription to go with Gemini/Claude. Claude Opus 4.5 outperforms Codex in coding tasks. Gemini outperforms ChatGPT in image generation. Can go on and on. Regardless, AI is here to stay, and anything Mag7 related ( $GOOGL -> $CIFR, $WULF ), ( $MSFT -> $IREN, $NBIS), connectivity like $ALAB (AWS), $CRDO (mag7) is extremely promising since they're the direct beneficiaries of increasing capex from mag7 fcf
-
列举AI领域资本支出加速增长的10大证据，强调AI赛道持续高景气度与投资机遇
对于AI领域的任何人来说，很难不看好。 资本支出正在加速增长，而且是以指数级的速度。 直接流向以下几个方面： 新云服务商：$CIFR、$NBIS、$WULF、$IREN， 连接性：$ALAB、$CRDO、$CLS， 能源：$VST、$FLNC、$TE、$EOS， 半导体/晶圆厂：$NVDA、$AMD、$GOOGL、$TSM， 存储：$SNDK、$MU和$STX。 仅在过去几周，我们就看到： 1. AI曼哈顿计划——美国政府正给予顶级模型访问专有实验室数据的权限以加速研究。 2. $GOOGL在德克萨斯州投资400亿美元建设数据中心。 3. Anthropic投资500亿美元建设边缘计算基础设施以支持其Opus 4.5+模型。 4. $TSM公布创纪录的远期收入数据（AI支出）。 5. $NVDA确认创纪录的远期收入数据（AI支出，锁定2年产量）。 6. $META将2025年数据中心/AI资本支出提升至400-450亿美元，用于llama5-6。 7. 今年三次降息以加速增长并降低融资成本。 8. Dominion Energy警告AI数据中心带来大规模电力负荷激增。 9. $AVGO表示AI网络订单达到前所未有的规模。 10. 阿联酋和主权国家推进AI发展。 我们没有看到任何放缓。只有创纪录的增长。 事实上，随着Claude Opus 4.5、Gemini 3的最新模型进展，以及美国政府的新承诺，感觉我们才刚刚看到人工智能新前沿的冰山一角。 （该推文引用了 @aleabitoreddit 的推文，引用内容仅供理解语境）: Nebius [$NBIS]是当前被低估最多的成长型公司。 它有潜力以210亿美元的市值成为下一个$GOOGL。原因很简单： 它的投资组合公司令人惊叹。 这个概念最令人难以置信的例子是$FTX公司。以下是故事： 当我们观察$META如何增长成为万亿美元公司时，不仅仅是Facebook。他们的投资组合公司Instagram、Whatsapp和其他应用使Meta主导了社交媒体领域。 $FTX在数字资产和前沿技术领域做着类似的事情。 四年前，即2021年，$FTX向一个大资产篮子投资了58亿美元。其中很大一部分投入了这三家核心公司： 1. Anthropic，持股13.56%，估值25亿美元。 2. Robinhood [$HOOD]，持股7.6%，估值85.4亿美元。 3. Solana [$SOL]，4100万+代币。 快进到今天，那将是： · Anthropic最新一轮估值3500亿美元。那部分股份价值约474亿美元。 · Robinhood现在市值超过1000亿美元。那部分股份价值约76亿美元。 · Solana现在每个代币价值131.5美元，使那部分股份远超57亿美元。 仅这三家公司就在4年内产生了超过550亿美元的价值，这甚至还不包括FTX的数百亿美元加上其他数十项投资，以及Chime、Layerzero、Aptos、Hidden Road（被$COIN收购）和加密货币的持股。 他们的投资组合公司比他们的核心业务更持久（想象一下，如果核心业务像$GOOGL搜索和YouTube一样持续增长，那将价值多少）。 $NBIS现在有着与$FTX在加密领域、$META在社交媒体领域相同的布局，但在人工智能领域拥有合法且飞速增长的核心业务。 Nebius拥有： 1. Clickhouse，28%持股，估值约70亿美元（2025年上半年为63亿美元）。 2. Avride，83%持股，估值约60亿美元（优步融资后）。 3. Toloka AI，约65%持股，估值约6.4亿美元。 4. TripleTen，100%持股，估值约3亿美元。 · Clickhouse为Anthropic、$META、$TSLA、$NET和许多财富500强公司提供支持。 · Avride是一家自动驾驶出租车机器人公司，从Yandex分拆出来，$UBER在3.75亿美元融资轮中投资以与Waymo竞争。 · Toloka是一个AI标注平台，亚马逊、微软、Anthropic和Shopify都在使用。 19.6亿美元+49.6亿美元+4.16亿美元+3亿美元=76亿美元的投资组合公司估值，这些公司的增长速度超过大多数公开成长型公司。 但如果我们看看他们以每年700%+的速度增长至70-90亿美元ARR的核心业务，拥有48亿美元现金，为$META、$MSFT、Cursor、政府和更多客户提供支持…… 这可能是它以低于90美元的最后一个月，因为今天MSCI纳入将为其带来从数亿美元到低数十亿美元的额外资金流入。如果我们看看$IREN或$CIFR等热门选择，没有任何其他数据中心成长型公司有这种类型的投资组合。 $NBIS估值仅210亿美元，市场正在忽视这个机会。
英文原文
It’s hard for anyone in the AI space not to be bullish. Capex is ramping up. Exponentially. And flowing directly down to: Neoclouds: $CIFR, $NBIS, $WULF, $IREN, Connectivity: $ALAB, $CRDO, $CLS. Energy: $VST, $FLNC, $TE, $EOSe Semi/foundries: $NVDA, $AMD, $GOOGL, $TSM Memory: $SNDK, $MU, and $STX In the past few weeks alone, we got: 1. Manhattan Project for AI - US government is giving top models access to propriety labs data for accelerating research 2. $GOOGL spending $40 on DC buildout in Texas 3. Anthropic spending $50 on EC buildout to support their Opus 4.5+ models 4. $TSM confining record forward revenue numbers (AI spend) 5. $NVDA confirming record forward revenue numbers (AI spend, 2Y production locked in) 6. $META upping 2025 DC/AI capex spend to $40-$45B for llama5-6 7. 3x rate cut this year to accelerate growth and make funding cheaper. 8. Dominion Energy warning of massive AI power load surge from AI datacenters 9. $AVGO signaling AI networking orders at unprecedented scale 10. UAE and sovereign countries pushing into AI We’re not seeing any slowdown. Only record growth. In fact, with the recent model developments from Claude Opus 4.5, Gemini 3, and now new commitment from the US government, it feels like we're just seeing the tip of the new frontier for Artificial Intelligence.
-
分析超大规模云厂商数据中心模式差异及潜在合作逻辑
我不会像某些特定的 $BMNR 或 $IREN 投资者那样盲目吹捧，认为街上的绿灯就意味着对公司利好。所以我倾向于认为这只是巧合。因为看起来 $AMZN、$GOOGL 倾向于偏好数据中心托管(Colo)模式（因为他们可以插入自己的 TPU、未来的 Trainium 芯片），而 $MSFT、$META（以及 99.9% 使用 $NVDA 且没有现成定制芯片的 AI 公司）则偏好 $NBIS、$IREN 类型的模式。所以 $AMZN 宣布在印第安纳州为其 AWS 数据中心部门投资 150 亿美元。但关键要注意的是，他们此前在那里已经花费了 313 亿美元，所以这并不算全新投入。你可能看到了 $NBIS 正在印第安纳波利斯附近建设 1000 多英亩的绿地数据中心。这只是一个有利于建设的区域，可能并非针对合作伙伴关系或像 Anthropic（最近承诺 400 亿美元资本支出）那样的另一笔超大规模云厂商交易。说实话，我现在可能更相信 Anthropic 的资本支出承诺而不是 OpenAI。话虽如此，也许如果 $AMZN 耗尽容量并达到与 $NVDA 的最大采购订单上限（因为他们正在插入自己的 $CIFR 芯片），他们可能会转而使用 Nebius。
英文原文
I'm not going to be a blind shill like some specific $BMNR or $IREN investors that can say a green light on the street means bullish for the company. So I would lean coincidental. Since it looks like $AMZN, $GOOGL tends to favor colo models (since they can plug in their own TPUs, Trainium chips in the future) while $MSFT, $META (and 99.9% of AI companies that use $NVDA and don't have custom chips readily built out) prefer $NBIS, $IREN type models. So $AMZN announced they're investing $15B in Indiana for their AWS DC segment. But key thing to note is they've already spent $31.3B there before so it's not exactly new. You probably saw how $NBIS is doing a 1000+ acre greenfield DC near Indianapolis. It's just a favorable area for buildout, probably not directed at a partnership and another hyperscaler deal like Antrhopic (who committed $40B in capex spend recently). I'd probably trust Antrophic more than OpenAI right now with capex spend lol. That being said, maybe if $AMZN runs out of capacity and hits max purchase order with $NVDA (since they're plugging in their own chips with $CIFR), they would use Nebius instead.
-
分析11月市场下跌原因：政府停摆、Binance清算、套利交易解除等多因素叠加
谢谢！我想在发帖前先了解一下情况，因为我最初以为是跟降息预期有关、政府停摆导致私人系统流动性被抽走，或者可能是某种算法因素。 政府停摆确实起了很大作用，加上25个基点的降息，但它们不应该导致股票/加密货币下跌30%以上（且重新开放后几乎没有反弹），因为美联储并没有收紧政策，我们刚经历了2次降息，而且很多公司的基本面正在改善。 Binance清算只是加密货币最初的个殊催化剂，但美联储降息而日本加息导致套利交易在流动性重新加载中解除，同时叠加其他并发因素，这可能就是我们现在在这里的原因。
英文原文
Thanks! I wanted to get a grasp of the situation before posting, since originally i was thinking it was it had something to do with rate cut odds, aftermath of gov shutdown draining some private system liquidity, or maybe something algorithmic initially. Gov shutdown does play a large role + .25 bps but they shouldn’t affect stocks/crypto 30%+ (and little recovery after reopening) since the fed isn’t tightening, we just got 2 rate cuts, and fundamentals of a lot of companies are improving. Binance liquidation was just the initial idiosyncratic catalysts fr crypto but fed cutting while Japan hikes causes carry trade unwind from the reload amid the other concurrent factors is prob why we’re here now.
-
分析套利交易平仓规模及宏观因素对高贝塔资产的影响
我对2024年套利交易平仓(carry trade unwind)的尸检报告进行了更深入的分析，其中快速资金（对冲基金、自营交易台）的短期资本约为2000亿至3000多亿美元（约占5000亿总额的40-60%左右）。然而，过去一年内这些敞口可能已重新建立，因此类似但规模较小的平仓可能再次发生。日本加息+美联储降息导致已重新建立的套利交易出现平仓。当你提到加密货币溢出到高贝塔(high beta)股票时，如果没有宏观逆风，币安清算导致的比特币跌破9万美元很可能已被买回至10万美元（尤其是考虑到三次降息），所以情况并非如此。
英文原文
So had a deeper breakdown on the 2024 autopsy report from carry trade unwind but figure was ~$200B – $300B+ish in fast money (hedge funds, prop desks) short term capital (which is ~40-60ish of total of $500B) However it's likely been reloaded in terms of exposure in the past year so a similar but less amount might be unwinding again. Japan hiking rates + Fed cutting leads to that carry trade unwind in terms of what got reloaded, When you mention crypto spilling into high beta stocks, if there weren't any macro headwinds, the Bitcoin drop sub $90k from binance liqudations probably would have been bought up to $100k already (esp with 3x rate cut) so that's probably not the case.
-
四大流动性冲击引发高杠杆资产暴跌，但AI巨头基本面未变。
市场刚刚经历了近代史上最严重的去杠杆冲击。 高贝塔资产正在崩溃： • $NBIS, $IREN, AI 股票从高点下跌约 40%+ • $MSTR, $BTC 在过去一个月暴跌 30-40%+ • 加密货币市值蒸发超过 1.2 万亿美元 这不正常，这是针对市场上增长最快且杠杆最高的板块的强制平仓。 以下是同时发生的四个流动性流失因素： 1. 加密货币前兆冲击（2025 年 10 月 – Binance 故障） • Binance 的定价错误导致 USDe 跌至 0.65 美元，在 24 小时内触发了超过 190 亿美元的强制清算 • 使用 USDe/wBETH/BNSOL 的高杠杆头寸（25x–50x）引发了全行业连锁保证金清算的连锁反应 • 该事件使加密货币结构脆弱，波动性现在威胁到流入 $MSTR/国债股票超过 70 亿美元的 MSCI 资金，存在 BTC/NAV 错配和强制出售比特币的风险 2. 美联储政策不确定性（降息鞭打效应） • 在美联储发出矛盾信号后两天内，市场对 12 月降息预期从 97% → 35% → 70%+ • 这种波动起到了隐性紧缩的作用，迫使杠杆基金和算法（在 $NVDA 财报后看到的情况）提前去杠杆 • 政策模糊性提高了全球风险，将美国的不确定性传导至全球融资市场的广泛抛售 3. AI 信贷压力（投机性债务破裂） • AI 建设需要 3.5 万亿美元的外部融资，促使公司大量进入债务市场 • Google + XAI 在 Similarweb 上的使用量提升引发了人们对 OpenAI 和循环融资的担忧，以及 $1T+ 的资本支出流向 $ORCL, $CRWV, $AMD 等公司，而这些公司并没有必要的资本。 • Applied Digital [ $APLD ] 23.5 亿美元垃圾债券（评级 B+）暴露了交易对手和集中度风险，由于需要更多债务来资助建设，其股票及相关数据中心股票下跌。 • 资本市场开始区分拥有真实现金流的公司和依赖投机性债务及 OpenAI 合同的公司，惩罚了如 $ORCL 和 $CRWV 等标的。 4. 日元套利交易平仓（催化剂） • 日本央行加息至 0.5% 及正常化缩小了美日利差，挤压了 80 万亿日元（约 5000 亿美元）的杠杆头寸 • 日本国债收益率上升引发回流资金，给美国收益率带来压力并抽干全球流动性 • 强制平仓导致投资者出售美国科技股和加密货币以偿还日元贷款，加剧了避险情绪的连锁反应 底线： 所有四个冲击都从市场的同一角落抽干了流动性——高贝塔、高杠杆资产，导致成长型科技和加密货币剧烈去杠杆，并增加了进一步强制抛售的风险。 过度的杠杆、脆弱的资产负债表和集中度风险决定了哪些资产跌幅最大。 美联储开启新一轮降息周期是一个“拐点”，投资者需要判断这种宽松是预防性措施还是对更严重衰退的反应。然而有一件事是确定的： 宏观冲击重置了估值，但并没有改变从 $NBIS 到 $META 的企业基本面。
英文原文
Markets just suffered their worst deleveraging shock in recent history. High-beta assets are collapsing: • $NBIS, $IREN, AI stocks are down ~40%+ from peaks • $MSTR, $BTC plunged 30-40%+ in the past month • Crypto erased over $1.2 trillion in value This wasn’t normal, it was a forced unwind across the markets with the most growth and leverage. Here's the four liquidity drains hitting at once: 1. Crypto precursor shock (Oct 2025 – Binance failure) • A pricing error on Binance sent USDe to $0.65, triggering $19B+ in forced liquidations in 24 hours • Highly leveraged positions (25x–50x) using USDe/wBETH/BNSOL cascaded into a chain reaction of cascading margin liquidations across the industry • The event left crypto structurally fragile, and volatility now threatens $7B+ MSCI inflows into $MSTR/treasury stocks, risking BTC/NAV mispricing and forced Bitcoin sales 2. Fed policy uncertainty (rate cut whiplash) • Markets priced a December cut at 97% → 35% → 70%+ in two days after conflicting Fed signals • This volatility acted as a stealth tightening, forcing leveraged funds and algorithms (seen post $NVDA earnings) to deleverage pre-emptively • Policy ambiguity raised global risk, transmitting U.S. uncertainty into broad selling across global funding markets 3. AI credit stress (speculative debt cracks) • AI build-out requires $3.5T in external financing, pushing companies heavily into debt markets • Google + XAI raise in usage per similarweb raised concerns about OpenAI and circular financing, alongside how $1T+ in capex spend going into $ORCL, $CRWV, $AMD, and others without having the ncessary capital. • Applied Digital [ $APLD ] $2.35B junk bond (rated B+) exposed counterparty and concentration risk, sending its stock and related data center stocks down due to the need of more debt to fund buildout. • Capital markets began differentiating between firms with real cash flow and those reliant on speculative debt and OpenAI contracts, punishing names such as $ORCL and $CRWV. 4. Yen carry trade unwind (the catalyst) • BoJ rate hikes to 0.5% and normalization narrowed the U.S.–Japan rate gap, squeezing ¥80T (~$500B) in leveraged positions • Rising JGB yields triggered repatriation flows, pressuring U.S. yields and draining global liquidity • Forced unwinds led investors to sell U.S. tech and crypto to repay yen loans, amplifying the risk-off cascade Bottom line: All four shocks drained liquidity from the same corner of the market, high-beta, leveraged assets, driving a violent unwind in growth tech and crypto and raising the risk of further forced selling. Excessive leverage, fragile balance sheets, and concentration risk determined which assets crashed the hardest. The beginning of a new rate-cutting cycle by the Fed is an "inflection point" as investors need to determine whether the easing is a preventative measure or a reaction to a more serious downturn. However one thing is for certain: The macro shock reset valuations, but it didn’t change the fundamentals of businesses from $NBIS to $META.
-
基于财报与宏观因素，发布Neocloud板块个股梯队排名及评估框架。
Neocloud（新云）板块梯队排名。 Q3财报后+市场板块回调： [S] 级：$NBIS [A] 级：$CIFR, $WULF, $IREN [B] 级：$GLXY, $CORZ [C] 级：$APLD, $CLSK [D] 级：$WLAC, $DGDX, $WYFI [F] 级：$CRWV*, $SLNH [U] 级：$GRRR, $MARA, $DGDX, $CLSK, $BITF, $HIVE, $RIOT ** 从左到右依次排名。 *：短期价格下跌创造了诱人的入场点（例如 $CRWV），但中期（12个月）吸引力仍然有限。 **：不确定，决定不评级（但在梯队图中列为F级）。一些纯矿企尚未完全转型为HPC（高性能计算）+ 合同不确定性太大。 很多人询问财报后的更新，这些只是基于以下加权评估的个人想法： a. 合同可见性和收入确定性 b. 对宏观收紧和信贷条件的韧性 c. 资产负债表强度和利润率概况 d. HPC建设风险（产能执行和编排） e. 收入增长轨迹（1-2年视野） f. 当前市值相对于收入增长和基础设施价值的比率 整个Neocloud板块都很有吸引力，但截至Q3，某些公司的非对称回报显然高于其他公司。
英文原文
The Neocloud Sector Tierlist. Post Q3 earnings + Market Sector Drop: [S] Tier: $NBIS [A] Tier: $CIFR, $WULF, $IREN [B] Tier: $GLXY, $CORZ [C] Tier: $APLD, $CLSK [D] Tier: $WLAC, $DGDX, $WYFI [F] Tier, $CRWV*, $SLNH [U] Tier: $GRRR, $MARA, $DGDX, $CLSK, $BITF, $HIVE, $RIOT ** Ranked in order from left to right. * : Short-term pricing drops created compelling entries (e.g., $CRWV), but med-term (12m) attractiveness remains limited. **: Uncertain, decided not to rate them (but put them in tierlist photo as F). Some pure miners haven't fully pivot into HPC yet + too much contract uncertainty. Lot of people asked about updates post earnings, these are just personal thoughts based on weighted assessments of: a. Contract visibility and revenue certainty b. Resilience to macro tightening and credit conditions c. Balance sheet strength and margin profile d. HPC buildout risk (capacity execution and orchestration) e. Revenue growth trajectory (1–2 year horizon) f. Current market capitalization relative to revenue ramp and infrastructure value The whole Neocloud sector is compelling but some have clearly higher asymmetrical return over others as of Q3.
-
TSM盈利是AI核心指标，超大规模云厂商建设不受NVDA财报影响。
我的观点是，$TSM 的未来盈利（forward earnings）是迄今为止人工智能交易（AI trade）最大的指标，因为它涵盖了从 $GOOGL TPU 生产到 $AMZN，以及 $NVDA、$AMD 等所有超大规模云服务商（hyperscalers），且我们已看到其盈利和利润率大幅超出预期。 即使 $NVDA 的未来营收被大幅下调（我对此表示怀疑），也不会使 $MSFT 到 $IREN、$AMZN 到 $CIFR 等公司已签订的多年度算力（compute）积压订单失效。 我们已看到 Anthropic 与 $MSFT Azure 的算力协议（这将惠及 $NBIS、$IREN 和 $CRWV 等），$GOOGL 近日建设超 400 亿美元的数据中心（此前已与 $WULF、$CIFR/FluidStack 签署托管设施（colo）协议），$NVDA 的财报不会改变超大规模云服务商的建设步伐。 但它确实对整个 AI 交易产生重大影响，并直接影响 $NVDA GPU 板块（例如重度依赖 $NVDA 的 AI 云厂商）。
英文原文
My opinion is that $TSM forward earnings was the biggest indicator of the AI trade so far since they span from all hyperscalers such as $GOOGL TPU prod to $AMZN, as well as $NVDA, $AMD, and we've already seen them blow away earnings + margins. Even if $NVDA forward revenue is quoted heavily downward (which I doubt), it won't invalidate multi-year contracted compute backlog from $MSFT to companies like $IREN or $AMZN to $CIFR made already. We're already seeing Anthropic x $MSFT compute deals with Azure today (which flows down to $NBIS, $IREN and $CRWV others), $GOOGL build out a $40B+ datacenter the other day (they've made colo deals with $WULF, $CIFR / fluidstack previously), and $NVDA earnings won't the change the hyperscaler buildout. But it does have a large impact on the overall AI trade as well + directly affect $NVDA GPU parts of the sector (eg. $NVDA heavy AI clouds).
-
澄清全栈基础设施定义，辨析IREN与NBIS等技术层级差异。
我觉得你可能把自己搞糊涂了，我只是想澄清一下术语。你在第一条评论中说 $ORCL 不是全栈(full stack)，因为他们做了数据中心租赁(DC leasing)，而 $IREN 是全栈。 现在你从 $ORCL 转移话题，开始谈论 $IREN。 我们只是在争论“全栈”服务提供的语义定义。如果你想论证 $IREN 不是全栈，而是做垂直整合的基础设施，那部分也是成立的。 我提到的 $IREN + 软件编排，是指调度、监控分配、故障恢复、集群配置(cluster prov)、容器运行时(container runtime)等，这些是为 $MSFT 提供算力所需的 GPU 编排。 集群编排 + 配置层 + 购买 GPU 而不仅仅是托管机房(colo)，我认为这就是你提到的全栈基础设施。 如果你想再深入一步，$NBIS $CRWV 做的是模型 API、推理运行时（平台/运行时层），比如 $NBIS 的代币工厂(token factory)，这是更深层次的全栈服务，所以 $IREN 确实不算全栈。 这只是在争论全栈之上叠加更多软件的语义问题。如果你接受这个论点，那么对于 $IREN 来说确实如此，且因公司而异。
英文原文
I think you might be confusing yourself and I'm just trying to clarify the terms. You were saying $ORCL is not full stack because they did DC leasing and $IREN is full stack in your first comment. Now you're shifting from $ORCL and talking about $IREN. And we're just going into semantic definition on full-stack offerings and if you wanted to argue $IREN is not full stack but does vertically integrated infrastructure, then that's partially valid too. With $IREN + software orchestration I was referring to scheduling, monitoring allocation, fault recoveriy, cluster prov, container rutime, etc for GPU orchestation needed to deliver compute to $MSFT. Cluster orchestration +provisioning layer + buying GPUs rather than just colo I'd argue is full-stack infrastructure as you mentioned. If you wanted to go one-step deeper that $NBIS $CRWV does it's model APis, inference runetimes, (platform/runtime layer) like $NBIS token factory which is a deeper full-stack offering, so $IREN is not full-stack sure. This is just arguing semantics of more software on top of full-stack and if you go with that argument, sure with $IREN and it differs company-to-company.
-
澄清AI算力租赁中软件编排对利润率及自由现金流的关键作用。
将全栈定义为GPU之上的软件层是正确的，因此$ORCL符合这一标准。你之前混淆了垂直整合和全栈服务，这是两个截然不同的概念。 这是一个非常细微的差别，在讨论$NBIS和$CRWV时，你混淆了“纯GPU租赁”与包含编排能力的收入/利润率。 让我再次澄清： 对于$NBIS和$CRWV，人们说其收入来自纯GPU租赁是一种过度简化。以$NBIS为例，它包含一个优化GPU利用率和整体效率的管理软件栈，这就是为什么他们相比$IREN在与$MSFT的交易中获得了更好的价格/兆瓦(MW)优势。这不仅仅是任何人都能做的纯GPU租赁。 再次强调，$CRWV表示软件是其护城河，正是软件将$NBIS和$CRWV的利润率与$ORCL等市场其余部分区分开来（后者毛利率极低，例如14%）。 你可以通过购买GPU并出租来获得极高的营收数字，但如果不能转化为自由现金流(FCF)和利润，那就毫无意义。而这正是通过增加内部运营支出(opex)进行软件编排发挥作用的地方。
英文原文
That's a correct with full-stack as software as a layer on top of GPUs, so $ORCL would fit that mark. Earlier you were conflating vertically integrated and full-stack offering, which are two distinct terms. This is a very nuanced, with $NBIS, $CRWV you're confusing you're confusing "straight GPU leases" and revenue/margins + orchestration. Let me clarify again: For $NBIS, $CRWV, it's a oversimplification when people say it comes from straight GPU leases. When you use $NBIS for example, it includes a managed sfotware stack that optimizes GPU utilization and overall efficiency, which is why they got a better deal/MW compared to $IREN with $MSFT. It's not just straight GPU leases that anyone can do. Again $CRWV said software was their moat, it's what differentiates margins between $NBIS | $CRWV, and the rest of the market like $ORCL (which had incredibly low gross margins, eg. 14%). You can have the extremely high revenue numbers just by buying GPUs and leasing them out, but if it doesn't convert to FCF, profit, it doesn't mean anything. And that's where increasing internal opex with software orchestration comes in.
-
澄清AI云全栈与裸金属区别，指出$IREN因执行风险派发股息。
我觉得你有点混淆了这些术语，让我澄清一下。 全栈服务(Full stack offering) = AI云 $NBIS $IREN $CRWV 裸金属(Bare Metal) = $CIFR, $WULF 类型的服务（AWS将自家GPU接入 $CIFR） $IREN 在决定为 $MSFT 交易购买数十亿美元GPU时，转向提供全栈AI云服务。这导致了因执行风险而产生的特别股息(DG)。 $IREN 在财报电话会上明确表示，他们之所以做全栈（包括GPU采购+软件编排），是因为“更高的收入潜力和更强的回报”。 关于你提到的软件编排，它主要用于改善内部运营支出(opex)和利润率，但在面向客户的差异化方面略有优势。 与电力相比，GPU利用率和编排对最终利润率的影响巨大，我在之前的帖子中做过定量分析。 对于为 $NBIS 等支付裸金属服务的客户，他们购买的不仅是原始硬件，而是经过优化的软硬件堆栈。
英文原文
So think you're confusing the terms a bit, let me clarify. Full stack offering = AI Cloud $NBIS $IREN $CRWV Bare Metal = $CIFR, $WULF type offerings (AWS plugs in their own GPUs into $CIFR) $IREN is pivoted to doing a full stack offering AI cloud when they decided to buy billions in GPU for the $MSFT deal. This is what led to the DG due to execution risk. $IREN on their earnings call said it themselves that they did full-stack (including GPU purchases + software orchestration) due to "higher revenue potential and stronger returns" For software orchestration that you mentioned, it's mainly for improving internal opex and margins, but is a slight advantage for customer facing differentiators. GPU utilization and orchestration plays an incredible difference in terms of final margins, when comparing to power and I did a quantitative breakdown in an earlier post. For customers that pay for bare metal for $NBIS and others, they are paying for an optimized hardware and software stack, not just raw hardware.