DRAMA ALERT🚨: In the YouTube comment section of the NVIDIA Chief Scientist Computer Museum panel, OpenAI Head of Jalapeno replies, saying that Jalapeno has already demonstrated on the SemiAnalysis InferenceX benchmark that Jalapeno is already running very fast (which, btw, beats NVIDIA chips like July Rubin) on non-OpenAI models, refuting the NVIDIA Chief Scientist’s misunderstanding about Jalapeno only being optimized/codesigned to run OpenAI-specific models. InferenceX is the established industry standard, endorsed globally, including by Meta, Oracle, Microsoft, Moonshot, NVIDIA, Weka, vLLM, SGLang, dg.

DeepSeek’s memory lights up for "Wright : Ace Attorney" We probed V4.1 Flash’s Engram gates to see which text patterns it uses. The results go well beyond names and facts. (1/6)🧵

SemiAnalysis ClusterMAX Lead @JordanNanos explains how Nebius picked up demand CoreWeave couldn’t handle and then executed well enough to reach the Platinum tier of its GPU cloud rankings: “CoreWeave has set the standard. I mean, they are the best in terms of technology.” “These guys deploy 10,000 GPUs a week at the peak when they’re building data centers. That’s incredible.” “But their balance sheet is full. There’s only so many people and resources you can marshal to build out another gigawatt when they’re going to 15 and beyond.” “We have customers coming to us and saying, ‘I saw you write such good things about CoreWeave. I just have $100 million I’m trying to give to them. They’re telling me they can’t accept it until May of next year because they’re backed up.’” “And therefore Nebius has been capitalizing.” “They’ve also executed very well. The technology that they’ve developed is strong.” “We think that managed clusters are approaching this feature-complete world where it’s hard to tell the difference in quality between Nebius and CoreWeave.” “Certainly our testing experience has been really solid on both of them.”

The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model 1,000+ facilities across 60+ operators mapped, built retail-first and flipped by AI, largest hyperscaler leases 1/5 national capacity, 100MW in 12 months, Eastern Data Western Compute

MONEY PRINTER ALERT🚨 NVIDIA vLLM B200 CAN GENERATE UP TO💰️$15 BILLION💰️OF ANNUAL PROFITS PER GIGAWATT serving the open DeepSeekv4.1 Flash model at the official interactivity & official selling prices. Using Engram DRAM offloading on NVIDIA results in a 50% increase in revenue per GigaWatt.

SemiAnalysis' @JordanNanos on the neocloud talent wars: "It's across the entire supply chain. We see this in electricians, data center technicians, and operators. Salaries for electricians in Louisiana and Abilene, Texas, are up 3x to 5x." "Crusoe just came in and paid these guys so much more money. But they're out of people. They need to train. No matter what job we're talking about, people just need to be trained." "The most successful neoclouds that we're seeing have this pipeline to take talent from other related areas and get them contributing to neocloud management and day-2 operations where a software engineer can start to work as an SRE on one of these teams."

PSA TO NEOCLOUDS COPING ABOUT CLUSTERMAX RATING: If you run a restaurant and you get bad Google reviews, don't attack the review. Instead listen and it might make your service better. For more food-related metaphors about neoclouds, check out:

Happy Thursday. On today's show: - @CompleteSkeptic (TypeSafe) - @JordanNanos (SemiAnalysis) - @leifthunder (Public) - @hoomanrenezhad (Solcoa) - @erika_alden_d (Pioneer Labs) - @Shalev_lif & @RomiLifshitz (Enclosure) See you on the stream.

ClusterMAX 3.0: The Industry Standard GPU Cloud Rating System Returns In gory detail: reliability, performance, support, pricing —and, of course, security— in our most thorough analysis of GPU cloud providers globally

ALERT ALERT ALERT 🚨 🚨 🚨 VLLM MAINTAINERS HAVE JUST SHOWN THAT TPUv7 CAN GET 700 tok/s/user,  56% BETTER PERFORMANCE THAN NVIDIA GB200 NVL72 THROUGH MEGAKERNEL OPTIMIZATION ON KIMI K3. As we said awhile ago, the TPU externalization of software is full steam ahead. This is ultra important to follow the progress of this.

MI355X IS UP TO 1.7X BETTER 💰️PERF PER DOLLAR 💰️THAN DGX B300. The AMD Mainland China UMBP team co-designed, in collaboration with Alibaba & the @sgl_project community, a new feature in SGLang that removes the duplicated KVCache contained between local L2 DRAM & distributed L3 DRAM, allowing for up to 2x more KVCache to be stored in DRAM. This feature is called UnifiedRadixCache external cache. But importantly, this marks the trend of AMD increasingly being first-class co-designed for new features in widely used top production engines like SGLang.

We tracked China's STAR Market semiconductor IPOs since Jan 2025, 13 in total with 60 days of trading, from IPO price → first-day close → day 60. All four chip designers are red from the first close, though none below issue. Packaging & test fared better: 3 of 4 green.

ALERT🚨: @alexandr_wang has spent more time on the app owned by the guy who wants to cage-fight Zuck than he has on his own Threads app. Maybe he should get Muse to screenshot his tweets & replies over to Threads.

Massive shoutout to TJ & Team at @EmbeddedLLM; they are the lead maintainers & CODEOWNERS of vLLM ROCm! Up until May 2026, they did not even have persistent access to an MI355X cluster, but after SemiAnalysis worked with & convinced AMD leadership of something that is quite obvious in hindsight, they finally got a persistent cluster to make MI355X great on AMD! Super proud of the @EmbeddedLLM Team, highly respected within the @vllm_project community.

ChipBook August data caught a +467% YoY jump in HBM exports from Korea to Malaysia, ~$3B in just two months. We attribute this primarily to Intel's advanced packaging capacity in Penang, with LinkedIn showing ~50 back-end job openings in Malaysia. Not an insignificant number. Exports to Taiwan slowed in return. Trade flows and job openings are starting to confirm customer shift. (2/3)