SemiAnalysis

‘Jensen’s nightmare’: Huawei, DeepSeek narrow AI chip ecosystem gap with Nvidia DeepSeek is likely to rely ‘primarily’ on Huawei’s ‘already mature and well developed’ technology stack as domestic semiconductor industry gains steam at…

Japan’s wave of corporate breaches is the latest sign that AI is changing how hacking works. Since Sept 25, 18 Japanese companies have disclosed breaches, including Tokyo Metro, Japan Post, Lawson, Nikkei, Yamato and Sagawa. Attackers got…

Three teams in one month have pushed LLM decode forward with megakernels: Inferact on Google TPUs, AMD with TileRT and FlyDSL, and now NVIDIA on GB300, now live on InferenceX. We're excited to see these gains and dig into the engineering…

Huawei Ascend 950PR Deep dive & µbenchmarks, DSv4.1 Flash Analysis, DeepSeek support for Ascend, AICPU, Tilelang Ascend https://t.co/3j3AvY17wC

WHAT 1,730 CRYSTAL PULLERS TELL US ABOUT MUSK’S 20 GW PLAN: Musk says SpaceX has a “highly proprietary” plan to add 20 GW of power by end-2027, beyond just gas turbines. Solar is a part of it and things are moving. China shipped 1,730…

BREAKING NEWS🚨🚨: TODAY META CONFIRMS THEY ARE RUNNING TPU INFERENCE IN PRODUCTION. What does this mean for the CUDA moat? For those who read our public newsletter, this is not a surprise, as on Nov 28, 2025, we publicly wrote that Meta…

After eight long years, TSMC is finally shrinking their 6T High Density SRAM bitcell for their A14 process! At IEDM, they will present the world's smallest bitcell, <0.017μm2 on their 2028 node. TSMC has been stuck at 0.021μm2 since N5 in…

TorchTPU shows Google finally embracing open-source and offer serious support. As our InferenceX shown recently, Google engineers and external companies have demonstrated great TCO. https://t.co/T1LAQwJoSB

Is 50 GW in a single year too aggressive? "Dear Brad called us out publicly and cut our forecast in half." So Jeremie (@JeremieEO) and Reyk (@robotknower) came on with Jordan (@JordanNanos) to show their work. They go site by site from…

ALERT🚨: Vera Rubin NVL72 is up to 7.3x better in perf per dollar than GB300 NVL72 for 1 million context length agentic inference workloads! THE MORE U BUY, THE MORE U EARN 👇️(1/3)🧵 https://t.co/wD4BDNpJ5m

ALERT🚨: SGLANG JUST MERGED MEGAKERNEL SUPPORT, GETTING CLOSE TO 1000 tok/s/user on the GB300 NVL72 AgentX Harness & increasing interactivity by 1.5x on MI355X too! Great work by the SGLang, RadixArk, NVIDIA, and AMD teams! Best of all,…

ALERT🚨: NVIDIA has been working overtime, nights, and weekends after being shocked 🤯 by last week’s TPUv7 FP8 submission showing better performance per dollar. Unfortunately for NVIDIA, its first GB300 FP4 counter-submission is still…

Datacenters 2027: The Year of Acceleration, +50GW in a Single Year 60GW Under Construction, SemiAnalysis Leading Indicators, 1.3T of Lease Contracts Signed YTD, The Shortage Could Move to Chips https://t.co/lTJ1Sbv5nl

SemiAnalysis: Datacenters SemiAnalysis sees 51.2GW of datacenter capacity being added in 2027. "Year-to-date, slightly under USD 1T of binding / non-cancellable datacenter leases have been signed – that’s key to supporting our aggressive…

BREAKING NEWS🚨: @Inferact & Google submit their FIRST preliminary TPUv7 Kimi K3 2.8T Model Submission, and it is already delivering strong perf per TCO against GB200 NVL72, B300, and MI355 in the ultra-high-interactivity regime due to…

TOKENMAXXING: @vllm_project and @inferact achieve 1.3 BILLION tokens per $ of TCO (hyperscaler volume pricing) via disaggregated serving on AgentX. This result is 4x better than the B300 aggregated configuration. https://t.co/jVevqcSNiG

Moratoriums aren't killing the US datacenter buildout. Our analyst @mayabarkin finds only ~2.3 GW of ~21 GW in moratorium areas (excluding the TCEQ pause in Texas) is likely to be delayed, as cited in The Economist. For more, read her…

SHOCKING: We have confirmed that multiple Neoclouds have been refused access to GTC Berlin due to them adding AMD GPUs, TPUs, and other chips to their cloud offerings Note that this doesn't apply to hyperscalers, as they have more buying…

Sparse attention fixes memory bandwidth, not memory capacity. “The whole idea of HiCache is trying to offload KV cache from the HBM to a lower tier, higher capacity but lower bandwidth memory, and only load the most relevant set of tokens…

ALERT🚨: Last night, Google just submitted their FIRST TPUv7 Agentic benchmark, and it delivers 1.42x 🚀 better perf per dollar than GB300 NVL72 FP4, even on an external non-Google workload cost basis! TPUv7 is even 2.5x better than AMD…

We take chips apart all the way down to the transistor. Andrew Wagner on SemiAnalysis STEEL, our 18,000 square foot teardown lab in Hillsboro, Oregon: what we tear down, how far down we can see, and what our clients do with it.…

ASML Is Building a Hybrid Bonding Tool! https://t.co/WKf466XKw4

BREAKING NEWS🚨: Google has submitted its FIRST TPUv7 FP8 vLLM results to our Agentic Inference Benchmark and has demonstrated 1.67x better perf per dollar than even GB300 NVL72 FP4. 🚀 It is amazing to see Google go full steam ahead with…

ALERT🚨🚨: @vllm_project, @Inferact & NVIDIA TOOK CO-DESIGN TO THE EXTREME WITH A GB300 NVL72 INFERENCE OPTIMIZATION TO GET 🚀 7.2x 🚀 BETTER PERF PER TCO THAN LITERALLY EVERY OTHER AI CHIP, using an inference co-design optimization called…

TSMC’s wafer ASP gap with UMC, GlobalFoundries and Samsung keeps widening. TSMC is setting new blended ASP highs while most peers remain below their previous peaks. Per our Foundry Model, from 2Q22 to 2Q26, TSMC’s wafer revenue share…

Is GLM 5.3 worth sounding the alarms? 🚨 “As long as you have the budget, it will write faster kernels than humans.” Kimbo (@kimbochen) and Wenyao (@gao_wenyao) join Jordan (@JordanNanos) this week to unpack sparse attention, the memory…

ALERT🚨: NVIDIA SGLANG INFERENCE RUBIN IS 4.23x BETTER 🚀 PERF PER DOLLAR THAN EVEN GB300 NVL72. This is on the widely used production LLM engine called @sgl_project. (1/2)🧵 https://t.co/JNIMAeg2WW

As the need for training and inference infrastructure grows rapidly, there is demand for increased campus sizes, with the growth mandating campus links have 5 km reach or more. Full Coherent systems are predominantly used in long-haul…

As interposers and substrates approach their scaling limits, 2.5D packaging is moving toward 3.5D, with tighter D2D pitch, higher bandwidth density, and lower energy per bit to overcome reticle-driven constraints on lateral scaling. Higher…

ALERT🚨: AMD JUST LANDED SUPPORT FOR MEGAKERNEL DECODE IN UPSTREAM VLLM WELL BEFORE NVIDIA. This marks one of the first times that a major vLLM feature is only supported on ROCm and not on CUDA yet. MegaKernels is a technique that is…

HVLP (High Very Low Profile) is an electrodeposited copper foil engineered with an extraordinarily smooth bonding surface on the side that attaches to the resin or dielectric layer, departing from the naturally rough surface characteristic…

BSPD breaks a lot of the toolbox used for yield, fault isolation, inspection. BSPD adds new challenges for teardowns too, but we're getting pretty good at it. We present: Wildcat Lake, Intel 18A. Are you a passionate, high-achieving…

As industry attention turns to substrates, Japanese and Taiwanese ASP charts look like they're telling completely different stories. Japan pricing is increasing, but is not up nearly as sharply as the pricing data coming out of Taiwan,…

Copper is starting to hit its limits in advanced packaging. Pure Cu is ~400 W/m·K, while at Semicon Taiwan there are debut of Cu/diamond composites reaching ~800 W/m·K, helping spread heat faster and reduce local hotspots. They can also…

Every Robotaxi and Optimus bot needs memory to turn AI models into real-world action. As Tesla scales production, locking down that memory supply is becoming a critical pillar of its manufacturing strategy—and that’s precisely why Terafab…

The AI buildout needs more capital than backstops can cover. “There’s bank lenders, insurance lenders, they don’t want to take a lot of risk. You’ve got private equity, they want to take a lot of risk. You’ve got private credit, they’re in…

The "Lambda–Anthropic" deal is really a four-party trade: Nvidia–Lambda–Hut 8 Anthropic. Anthropic gets ~350 MW of Rubin (VR NVL72) capacity in 2028 at ~$5.80/GPU-hr. So how does the math work? (1/5)🧵 https://t.co/WlEirkhnuj

Can you run an AI datacenter campus on gas delivered by truck? Ellie Holbrook, who covers gas for the AI buildout at SemiAnalysis, told Bloomberg that it doesn't go far. (1/4)🧵 https://t.co/rPWUVdOTi7

Memory suppliers are finally acknowledging our HBM de-spec call. SK Hynix publicly said 16-hi is difficult, and the market is now prioritizing bandwidth over capacity. Hybrid bonding’s case in HBM was always stack height, and at 4-hi and…

Congrats to @AnjneyMidha for launching National Compute! Accessible & subsidized on-demand MI355X with availability for only $3/hr/GPU on Crusoe & Crusoe B300 for $4/hr/GPU. It is so cheap, we are worried that someone might start signing…

IREN continues to come up in our conversations with neocloud customers, and usually for the wrong reasons. In Prince George and Mackenzie in Northern BC, Canada, we’ve heard the stories: multi-day power outages, network upgrades, storage…

Intel shipped the first commercial backside power delivery and its first GAA transistors in Panther Lake. SemiAnalysis's teardown puts 18A compute logic density on par with TSMC N3E, and not ahead of N3P, N2 or Samsung SF2 — while the…

Every transistor needs power. Traditionally, it comes down from the chip’s frontside: package contact → wide upper wires → finer wires/vias → local rails → transistors. The wires are largely copper, built layer by layer in insulating…

Optical backplanes are gaining traction and so is the ecosystem surrounding it. Senko showcased the expanded beam backplane connector at CIOE and we estimate that there are ~1500 strands or more of fiber per connector. Optical backplanes…

Experiments have shown that offloading Engram to DRAM results in up to 50% better performance on H200, B200, B300, and even GB300 NVL72. Our technical team has collaborated with 10x engineers from @EmadBarsoumPi & @AnushElangovan to…

We took the silicon out of the silicon. Quick turn from SemiAnalysis STEEL Teardown Lab: iPhone 18 Pro Max, A20 silicon on TSMC N2. more to follow... https://t.co/RURlD49MP1

Many Google TPU customers are already asking Google for SemiAnalysis AgentX performance results, as customers are saying that it is a representative workload of their agentic inference traffic. We are excited to announce that we are…

Recommended reading for the weekend: 1. @damnang2 — Is the 2028 Memory Downcycle Real? 2. @AnthropicAI — Scenarios for our Economic Future 3. @BrookingsInst — Workforce policy for the age of AI 4. @Microsoft — What we’ve learned from…

You’re bearish on AI infrastructure while the US is about to 4x data center capacity in just two years even with all the local opposition (Save this). SemiAnalysis expects the United States to deliver 38 GW of data center IT capacity in…

SemiAnalysis: Our Datacenter Industry Model forecasts +38GW of Datacenter IT capacity delivered in the US in 2027, more than double that of 2026. Today, we estimate that roughly 2.3 GW of planned capacity is genuinely delayed because of…

HBM per GPU is shrinking, not growing. Rubin Ultra uses 8-high stacks: 192GB versus 288GB on standard Rubin and B300. A year ago the industry expected 16-high. Per ChipPub and SemiAnalysis, 4-high is next: inference is bandwidth-bound, and…

SemiAnalysis: we estimate ~$11T of cumulative capex from CY24 through CY29. $GOOGL $AMZN $META $MSFT $ORCL

Wow. Incredible...Citrini has been bought by SemiAnalysis 🤯 https://t.co/cAqYWklbJ5

Wow. “Citrini Research Founder Van Geelen Sells Firm To Semianalysis” https://t.co/8D8dPTbSEj

You should be buying as much NVIDIA as possible (Save this). SemiAnalysis projects that NVIDIA could hold roughly $1.4 trillion in cash and investments by fiscal 2031, giving it more capital than almost any customer, competitor or bank…

SemiAnalysis projects Nvidia's cash balance to grow to $1.4T and off-balance-sheet liabilities to grow to ~$2.5 trillion by FY31. $NVDA https://t.co/45TsGwZzZ1 https://t.co/jq8rfFcHw3

SemiAnalysis on memory for robots: "Although total DRAM wafer capacity continues to grow, most of the incremental capacity is being absorbed by HBM for AI accelerators. That leaves the commodity and LPDDR supply that robot brains depend on…

Nvidia Chief Scientist Bill Dally: "What everyone is showing, acutally did see some slides on this at Hot Chips, is the SemiAnalysis InferenceX Benchmark. Cause it hits exactly the data that people want" Nvidia Chief Scientist stated this…

Our SemiAnalysis Industrials Model (https://t.co/YBYz1XWpf1) maps datacenter capacity additions to construction and equipment demand and tracks the full universe of suppliers positioned to capture that spending. Looking back at Q2 results…

ALERT🚨🚨: On apples-to-apples InferenceX Comparisons, TPUv7 Ironwood achieves 50% better perf per dollar than Blackwell Ultra on the new TorchTPU external inference stack! TorchTPU brings native PyTorch to TPUs, & along with Google…

SemiAnalysis: Ironwood costs approximately $0.181 per million total tokens, compared with $0.222 for B200 and $0.276 for B300. That is approximately 19% lower cost than B200 and 34% lower than B300, while delivering the same per-user…

Error 401 (Bad Request)!!1 401. That’s an error. The server cannot process the request because it is malformed. It should not be retried. That’s all we know. Back Skip navigation Search Search with your voice Sign in Ep. 027 - OpenAI…

Manage Consent To provide the best experience, we use cookies to store and/or access device information. FunctionalFunctional Always active The technical storage or access is strictly necessary for the legitimate purpose of enabling the…