#大佬观点

this is how finding GPUs goes: 1. someone says they have GPUs 2. you ask & turns out they have a lead on an OEM that will ship in maybe Q1 ‘27. 3. you ask & turns out they haven’t found colo space, haven’t ordered networking. 4. you ask & the OEM is SMC with optics in the hot aisle. back of your mind you also know SMC always rugs & ships late. 4. you ask & turns out they’re racked with hydra and they don’t even have final seller’s rights. 5. you say you’ll do $4.9/hr 10% prepay for RFS Jan ‘27. effectively betting there’s a 60% chance it goes live and you get ‘em 6. jan comes, not live, you get 50% of your prepay back and is what it is. everyone’s pissed. other scenario: 1. guy says he has GPUs RFS tmrw 2. you ask and get test node for 1hr, it has infiniband & storage nodes on network 3. he says $10/hr, sign in next two hrs 100% prepay or piss off. 3. startup knows they’ll make $20/hr so they grit their teeth & snap call sign plus wire. BMC without of band access is given same day, everyone’s happy! speculative market for yotta compute available maybe Q1 maybe Q3 ‘27 in yucatan is completely different than nodes with immediate RFS. anything under $7/hr probably isn’t real.

The "Rack Spread" is unfolding, token prices are falling while GPU prices climb Change since mid-July: - B300 GPU rentals up 106% - B200 up 66% - H100 up 17% - Average token prices down 58% Charts of the Week: https://www.a16z.news/p/charts-of-the-week-job-fud-fades

When agents become users, operators lose the assumption that every workload is trusted. Mohamed Awad explains why DPUs and hardware isolation move deeper into the AI control plane. Full paid interview: https://www.thediligencestack.com/p/executive-interview-arms-mohamed

Expert Takeaways: Nebius > Nebius operates an auction-driven cluster-access pricing model that drives up pricing levels by ~20% through competitive bidding. > Nebius requires 60–65% upfront contract payments, providing strong protection compared to debt-heavy peers like CoreWeave (facing up to 15% interest on new debt). > Short-term spot capacity fetches $40–50M ACV per megawatt (potentially up to $60M with auction uplift) vs. $20–25M for medium-term deals. > Tight supply for latest-generation GPUs through at least mid-2028 supports reserving 2027 capacity for high-margin spot workloads. > Meta and SpaceX control >1M GPUs and plan to deploy 25k–50k units monthly, pulling bare-metal contracts away from neoclouds. > Open-source models deliver comparable results at 70–80% lower cost, eroding OpenAI and Anthropic enterprise revenue. > Gartner projects >80% of independent neoclouds will fail by 2030 due to a K-shaped market favoring GB300/Vera-class hardware over older GPUs. > Nebius is insulated from cascading neocloud defaults because its core clients (Microsoft, Meta) are AA/AAA-rated, unlike BB-rated frontier labs. > High-end hardware (GB200/GB300) is shifting to inference workloads, which offer 4–8× higher margins than training. > Micron predicts HBM3/HBM4 memory shortages beyond 2028, while Microsoft’s 300,000 internal Maia ASICs displace demand for 300,000+ Nvidia GPUs. $NBIS $GOOGL $META $MSFT $AMZN

This is my conversation with Michael Kagan, the founder of Mellanox and current CTO of NVIDIA We speak about his life, the progression of networking over the span of his career, and how the work that NVIDIA is now doing to build "Planet scale" computing I hope you enjoy the conversation as much as I did 00:00 Highlights 01:16 Growing Up in Russia and Moving to Israel 03:42 Intel, the PC Revolution, and a Passion for Technology 06:14 Sponsor: Liquid 06:33 Lessons from Intel: Software and Networking 10:10 Founding Mellanox and the Vision for Cloud Computing 14:34 From Supercomputers to the Cloud 17:46 How GPUs Sparked the AI Revolution 20:52 Sponsor: Compute Exchange 21:14 NVLink and the Rise of Rack-Scale Computing 24:57 The Engineering Challenges Behind NVL72 27:39 Designing Hardware for Future AI Workloads 29:43 BlueField DPUs and the AI Factory 32:55 Making AI Agents Safer 35:37 What Mellanox Could Have Done Differently 37:03 NVLink, InfiniBand, and Spectrum-X 39:24 Connecting AI Factories at Planetary Scale 42:42 Keeping Massive AI Systems Running Through Failures

SAMSUNG: THE MEMORY SHORTAGE WILL WORSEN IN 2027 COMPARED WITH THIS YEAR.

This is my conversation with Michael Kagan, the founder of Mellanox and current CTO of NVIDIA We speak about his life, the progression of networking over the span of his career, and how the work that NVIDIA is now doing to build "Planet scale" computing I hope you enjoy the conversation as much as I did 00:00 Highlights 01:16 Growing Up in Russia and Moving to Israel 03:42 Intel, the PC Revolution, and a Passion for Technology 06:14 Sponsor: Liquid 06:33 Lessons from Intel: Software and Networking 10:10 Founding Mellanox and the Vision for Cloud Computing 14:34 From Supercomputers to the Cloud 17:46 How GPUs Sparked the AI Revolution 20:52 Sponsor: Compute Exchange 21:14 NVLink and the Rise of Rack-Scale Computing 24:57 The Engineering Challenges Behind NVL72 27:39 Designing Hardware for Future AI Workloads 29:43 BlueField DPUs and the AI Factory 32:55 Making AI Agents Safer 35:37 What Mellanox Could Have Done Differently 37:03 NVLink, InfiniBand, and Spectrum-X 39:24 Connecting AI Factories at Planetary Scale 42:42 Keeping Massive AI Systems Running Through Failures

SK HYNIX: EXPECTS DEMAND TO OUTPACE SUPPLY THROUGH 2030

Daishin Securities CSPs and AI labs: CSPs and AI labs had requested the development of lower-stack HBM products, such as 4-high and 6-high configurations, for use in low-end inference products. Those plans now appear to have been shelved, with 8-high once again emerging as the minimum stack height. NVIDIA: Rubin’s specifications have changed amid base die supply constraints and requests from customers prioritizing bandwidth over cost efficiency. The original plan called for a mix of 8-high and 12-high HBM, but the 8-high option appears to have been dropped. For Rubin Ultra, scheduled for release in 2027, multiple architectural variants are currently undergoing tape-out. Initial versions are expected to use the highest-specification 8-high HBM4, while subsequent versions are expected to adopt either 8-high or 12-high HBM4E. AMD: AMD is expected to use a mix of 8-high and 12-high HBM4 in its 2027 products. One notable development is the upcoming MI450, which is expected to increase the number of HBM stacks within the CoWoS package to 12, delivering 288GB of HBM4. The increase from eight to 12 HBM stacks per CoWoS package is a positive development for HBM content over the medium to long term. OpenAI and Anthropic: OpenAI is expected to adopt 12-high HBM4, while Anthropic is expected to use 8-high HBM4.

拉远视角看，GPU 的租金仍然高于标准的 6 年折旧时间表 租金价格与时间表的差距： - H200: +143% - B200: +111% - RTX PRO 6000: +73% - H100: +54% - A100: +9% 大部分差距是在 4 月之后拉开的，当时 token 需求开始起飞 本周图表：https://www.a16z.news/p/charts-of-the-week-attention-is-all

安迪·贾西（Andy Jassy）担任 CEO 五年来，AWS 规模几乎扩大至三倍，亚马逊营收增长约 80%，并计划今年投入 2200 亿美元以赢得 AI 竞赛。亚马逊的 AI 业务在三年内达到 250 亿美元的年化收入，增速是早期 AWS 的数百倍。 贾西做客 a16z 的 Ben Horowitz 和 Erik Torenberg 的节目，畅谈这一历程背后的决策：从疫情后重建亚马逊的履约网络，到 AWS 的起源故事。 他们还探讨了：数据中心为何成为政治斗争焦点、亚马逊的自研芯片、对 Anthropic 和 OpenAI 的投资、推理成本为何需要降低 10 倍、AI 最被低估的风险、为何每家公司都必须从零开始重塑产品，以及贾西给当今 18 岁年轻人的建议。 0:47 2021 年以来亚马逊的变化 1:51 疫情如何重塑零售业 3:30 AI 为何为 AWS 增长注入超级动力 5:20 领导力原则冲突时怎么办 9:36 6000 万美元 vs. 250 亿美元：早期 AWS 与早期 AI 13:20 AI 资本支出的现金循环如何运作 15:44 美国为何必须建设数据中心 20:55 从实验室到企业的 AI 哑铃模式 23:12 需求增长 vs. AWS 老客户 28:11 Annapurna、Graviton 与 Trainium 33:55 押注 Anthropic 和 OpenAI 37:15 为何 AI 意味着更小的团队 42:13 AI 被低估的风险：安全 46:29 亚马逊为何重构 Alexa 的大脑 53:07 贾西如何跟孩子谈 AI 56:36 押注 AI、机器人与太空 YouTube: https://youtu.be/-YMMRQ1qscI @ajassy @Amazon @bhorowitz @eriktorenberg

I spent two days touring Intel’s most advanced fab and talking to its technical leadership to understand what the chips of the future will look like. A lot of what's coming next – gate-all-around transistors, backside power, CFET, hybrid bonding – is hard to follow without actually seeing it, so we made all the awesome animated explainers that I wish I had while I was prepping. I also asked Intel’s leadership how the semiconductor industry will cope with the enormous shock in demand from AI. AI compute has grown 3x a year for the last decade, but according to Intel's CTO, Intel could only double leading-edge wafer production with today’s global stock of chipmaking tools. What kinds of tricks can foundries pull to extend this crossover point where demand surges way ahead of supply? This was a totally new format for us - hope you enjoy! 0:00:00 – Intro 0:00:47 – How Intel brings up a new process node 0:03:22 – What Intel learned from 10nm 0:06:28 – From planar transistors to gate-all-around 0:13:13 – Why High-NA EUV pays off for Intel but not TSMC 0:19:20 – Could chips have evolved completely differently? 0:23:33 – Inside Intel's "Hot Topics" meeting 0:29:30 – Replacing copper wires with ruthenium 0:32:41 – Why stacked transistors take a decade to ship 0:39:45 – Hybrid bonding and the chip-to-chip bottleneck 0:48:00 – There aren't enough tools to 10x chip production 0:53:14 – Can Intel catch up to TSMC? 1:01:02 – Terafab, and rethinking the fab from scratch

We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent. It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

https://x.com/i/article/2112588009790943232

RIP long term contracts. I've been saying this forever now and even Mr. @elonmusk is waking up to it.