[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

@SemiAnalysis_ "ASML: why an engineering company will need to become an economics company 🧵"
@SemiAnalysis_ on X 2025-07-22 16:42:45 UTC 27.1K followers, 213.6K engagements

"TSMC says the tools are too expensive even first-mover Intel is shying away. And hyper-NA the successor to high-NA is even more in doubt"
@SemiAnalysis_ on X 2025-07-22 16:42:49 UTC 27.1K followers, 5887 engagements

"For most of its history ASML won by solving exceptional technical challenges. They threw 10000 engineers at a problem until it was solved. Even if it took a decade"
@SemiAnalysis_ on X 2025-07-22 16:42:45 UTC 27.1K followers, 6242 engagements

"Happy 4th July Speed is the Moat & @AnushElangovan & his team Keeps Running Faster & Faster. Still lots of areas where ROCm has gaps but many are already closing"
@SemiAnalysis_ on X 2025-07-04 20:56:59 UTC 27.1K followers, 63K engagements

"Well all that's left now is prayer. Repost to join my prayer circle for Intel 🕯 🕯 🕯 🕯 🕯 Intel 14A 🕯 🕯 Customer 🕯 🕯 🕯"
@SemiAnalysis_ on X 2025-07-24 21:09:14 UTC 27.1K followers, 37.7K engagements

"Moonshot the maker of Kimi K2 just released a technical paper diving into the model's development Kimi K2 is the best non-reasoning model: let's dive into a critical part of why the model is so good🧵"
@SemiAnalysis_ on X 2025-07-21 22:20:14 UTC 27.1K followers, 100.9K engagements

"Great to see AMD supporting events from all the relevant AMD neoclouds such as the DigitalOcean AI Summit today Furthermore AMD is fostering all relevant neoclouds by offering them large finance incentives & GTM partnerships as we explained in our AMD Advancing AI Event article"
@SemiAnalysis_ on X 2025-07-23 02:20:46 UTC 27.1K followers, 11.3K engagements

"OpenAI Stargate is applying the same strategy as @elonmusks Colossus cluster and is currently installing XX gas turbines. Each GE LM2500XPRESS gas turbines generates 34MW When all XX will operational it be enough to generate 1000MW enough to 500k GB200 NVL72 chips in the first Stargate site"
@SemiAnalysis_ on X 2025-07-24 05:58:48 UTC 27.1K followers, 113.3K engagements

"In addition to X platform exclusive content we will also be posting platform exclusive content & insights on Xiaohongshu about GPUs & the AI industry GPU AI Link below to our XHS page👇"
@SemiAnalysis_ on X 2025-07-17 05:20:33 UTC 27.1K followers, 11K engagements

"They solved ArF immersion in the 2000s when many thought it would be impossible to put water between a lens and a rapidly accelerating wafer"
@SemiAnalysis_ on X 2025-07-22 16:42:46 UTC 27.1K followers, 5904 engagements

"Intel the home of Moore's Law for the first time in history is evaluating if it will continue at the leading edge. From its 10-Q "However if we are unable to secure a significant external customer and meet important customer milestones for Intel 14A we face the prospect that it will not be economical to develop and manufacture Intel 14A and successor leading-edge nodes on a go-forward basis. In such event we may pause or discontinue our pursuit of Intel 14A and successor nodes and various of our manufacturing expansion projects." Just like that we could be could be talking about a TSMC"
@SemiAnalysis_ on X 2025-07-24 20:50:25 UTC 27.1K followers, 127.6K engagements

"But the challenge for the future is totally opposite: the technology is there (high-NA is remarkably good already) but the economics might not be"
@SemiAnalysis_ on X 2025-07-22 16:42:48 UTC 27.1K followers, 5256 engagements

"Ever since the release of Cerebras inferencing in HotChips2024 Cerebras has been handing Groq massive Ls. DeepSeek R1 Llama3 70B: Cerebras: 2256 tok/s/user Groq: XXX tok/s/user"
@SemiAnalysis_ on X 2025-07-24 16:13:45 UTC 27.1K followers, 26.1K engagements

"Intel 18A Details & Cost Future of DRAM 4F2 vs 3D Backside Power Adoption (or Not) Chinas FlipFET Digital Twins from Atoms to Fabs and More VLSI 2025 Roundup"
@SemiAnalysis_ on X 2025-07-21 18:52:06 UTC 27.1K followers, 36.8K engagements

"ASML itself admits the economics of hyper-NA are challenging. Can they solve a cost problem by throwing 10000 engineers at it Its not a given"
@SemiAnalysis_ on X 2025-07-22 16:42:49 UTC 27.1K followers, 8397 engagements

"Most GB200 Clusters in the world are just 400G NIC per GPU except for OpenAI's Stargate giving OpenAI 2x stronger scale out ultra large scale training network performance As you can clearly see in the image there are X 400GbE OSFP cages per compute tray which means it is 800GbE per GPU (2 400GbE Cages per CX-8 NIC) For most other clusters they will only have 800G per GPU in GB300 but OpenAI specially negotiated a special GB200 SKU with 800GbE (8x112G) per GPU with Nvidia. Furthermore with 800GbE ConnectX-8 the GPU is able to talk directly to the NIC over PCIe Gen6 verus in 400GbE ConnectX-7"
@SemiAnalysis_ on X 2025-07-22 21:21:29 UTC 27.1K followers, 54K engagements

"Although "baking transformers in silicon" may sound cool it is mostly just a marketing slogan. XX% of Transformers FLOPS are just GEMMs and 256x256/128x128 systolic arrays in TPU/Trainium are already optimized for GEMMs. Even modern GPGPU (with Tensor cores) are optimized well for GEMMs. Even if you "bake transformers in silicon" aka just create an gaint systolic array most of your die area will still be taken up with SRAM cells and you will still face the memory wall since your HBM memory bandwidth will be the same as GPGPU/TPUs/Trainium"
@SemiAnalysis_ on X 2025-07-24 02:21:55 UTC 27.1K followers, 41K engagements

"Link to the full piece on meta where we break down what went wrong with llama4 bemoth meta's new Datacenter strategy and what changes they r making to prevent another llama4 bemoth disaster 2/2"
@SemiAnalysis_ on X 2025-07-14 17:35:28 UTC 27.1K followers, 11.3K engagements

"As Clive X said it is not possible for a chip released in same timeframe to have 10x perf per watt over TPU/Trainium/GPGPU using the physics in our universe despite claims from companies that are trying to bake transformers in silicon. If you looked at a profile before & done benchmarking23 most of your transformer wall time is spent in GEMMs rmsnorm residual embedding layer (and even attention at 10k ctx len) does not take up much time thus baking these layers into silicon won't have much affect on end to end performance. The attention layer constantly changes in frontier models (MHA MQA GQA"
@SemiAnalysis_ on X 2025-07-24 21:17:33 UTC 27.1K followers, 21.5K engagements

"The modern factory is an AI token factory. Raw Silicon electricity and water comes into a Datacenter and what comes out is intelligence (in the form of tokens). Our AI Token Factory Economic Stacks tracks the gross margins profits and efficiency at each layer of the stack from the neoclouds to inference API providers to the application layer (agents code editors etc): - Neoclouds convert power and silicon into rentable SLURM/Kubernetes/ssh GPU clusters - Inference API providers turn GPU clusters into tokens - Application layer turn tokens into productivity"
@SemiAnalysis_ on X 2025-07-07 16:38:09 UTC 27.1K followers, 30.1K engagements

"A few weeks ago Wells Fargo called SemiAnalysis "that investors in fact treat as a bible" So have you done your bible readings lately anon"
@SemiAnalysis_ on X 2025-07-21 12:47:10 UTC 27.1K followers, 12.7K engagements

"In both cases the economics were there it was just a matter of making the technology high-volume ready. ASML built a world-class engineering organization to make it work"
@SemiAnalysis_ on X 2025-07-22 16:42:47 UTC 27.1K followers, 5216 engagements

"EUV was an even greater challenge with louder naysayers but after 10+ years in development ASMLs bet paid off with a monopoly in one of humanitys greatest technologies"
@SemiAnalysis_ on X 2025-07-22 16:42:46 UTC 27.1K followers, 5523 engagements

"Night and Day mindsets towards competition: Google Deepmind vs OpenAI"
@SemiAnalysis_ on X 2025-07-24 21:29:02 UTC 27.1K followers, 47.4K engagements

"P.S. ASMLs (retired) leaders deserve credit for making decade-long bets that were far from a given. Its rare in the age of quarterly earnings. Care to bet against this guy"
@SemiAnalysis_ on X 2025-07-22 16:42:50 UTC 27.1K followers, 8047 engagements

"Zucc confirms that they are building multiple GW clusters. 1000MW Prometheus will come online in 2026 and Hyperion will be built in phases and be able to scale to 5000MW+. For context the largest H100/H200 clusters operational today is only 150-200MW. Link to full article below 👇 1/2"
@SemiAnalysis_ on X 2025-07-14 17:35:28 UTC 27.1K followers, 164.6K engagements

"Don't worry - H100/H200s won't be worthless anytime soon. It will still be a valuable baseline that impressive AI accelerator upstarts like mAtX EtChEdAi oPeNAI TiTan cHip MeTa MtIAv3 MsFt MaIA etc can use in their performance comparisons"
@SemiAnalysis_ on X 2025-07-23 15:07:59 UTC 27.1K followers, 21.8K engagements

"Moonshot also does RL for hard-to-verify domains like instruction following and faithfulness. Interestingly for Instruction Following Moonshot utilizes a hybrid approach. Hard checks to ensure the model's output comply with instructions but also an LLM judge to ensure that instructions were followed"
@SemiAnalysis_ on X 2025-07-21 22:20:15 UTC 27.1K followers, 4889 engagements

"CoreWeave CTO breaks his neck with ClusterMAX GPU Cloud Award"
@SemiAnalysis_ on X 2025-07-15 22:12:13 UTC 27.1K followers, 42K engagements

"DeepSeek Debrief: XXX Days Later Traffic and User Zombification GPU Rich Western Neoclouds Token Economics (Tokenomics) Sets the Competitive Landscape"
@SemiAnalysis_ on X 2025-07-03 20:17:34 UTC 27.1K followers, 175.5K engagements

"The US has signed two landmark agreements with the United Arab Emirates and Kingdom of Saudi Arabia (KSA) that that will noticeably shift the balance of power. The deals have economic geopolitical and national security implications and taken together we believe these bilateral deals reinforce Americas claim to AI leadership. They also raise real concerns but none that the US cant address. These deals do not weaken compute availability for the US rather they strengthen them. We believe that these agreements are transformative and set to take AI infrastructure accessible to American AI to the"
@SemiAnalysis_ on X 2025-07-16 02:35:00 UTC 27.1K followers, 18.9K engagements

"The OpenAI open source model is going to be really really good🍓"
@SemiAnalysis_ on X 2025-07-15 22:36:29 UTC 27.1K followers, 270.2K engagements

"One of the major technology innovations that has enabled LLM training and inference and a fundamental component of NVIDIAs datacenter GPUs was actually partially invented by AMD. Click below to read more. 1/5"
@SemiAnalysis_ on X 2025-07-02 17:29:58 UTC 26.9K followers, 49.9K engagements

"Theres a lot of fear mongering surrounding high purity quartz (HPQ) and Spruce Pine NC following the devastating flooding from Hurricane Helene The area contains the purest form of natural quartz but the significance of supply disruptions from the mines is exaggerated 1/8"
@SemiAnalysis_ on X 2024-09-30 21:47:07 UTC 27K followers, 427K engagements

"ClusterMAX billboard is in NYC Times Square today. SemiAnalysis' ClusterMAX is the premier GPU cloud rating system covering 90%+ of the market by rental GPUs volume. We test and assess across X evaluation categories from security to networking performance to pricing model to health checks etc. Full Article linked below 👇"
@SemiAnalysis_ on X 2025-07-08 21:06:23 UTC 27.1K followers, 53.9K engagements

"First - inventory levels for raw wafer companies are currently low but even then theres X months of DIO at Globalwafers and Siltronic and X months of DIO at SUMCO. Existing inventory is a buffer. Can mining activities restart within X months Probably. 2/8"
@SemiAnalysis_ on X 2024-09-30 21:47:09 UTC 26.9K followers, 30.8K engagements

"AI researchers when they discovered that torch.compile doesn't scale well to real multi-node production training workloads and is a giant footgun"
@SemiAnalysis_ on X 2025-07-21 19:29:59 UTC 27.1K followers, 37.7K engagements

"For other verifiable domains like Math and STEM Moonshot focuses on achieving diverse coverage ie including a breadth of topics as well as calibrated difficulty. Ensuring the right difficulty means the problem is not too hard or too easy for a model. If either of these happen the model may not be able to learn anything useful. This is a hard balance to strike"
@SemiAnalysis_ on X 2025-07-21 22:20:15 UTC 27.1K followers, 4983 engagements