@SemiAnalysis_ SemiAnalysis

SemiAnalysis posts on X about ai, in the, core, taiwan the most. They currently have [------] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours.

Engagements: [------] #

• [--] Week [-------] -38%
• [--] Month [---------] -31%
• [--] Months [----------] +264%
• [--] Year [----------] +1,545%

Mentions: [--] #

• [--] Week [--] +16%
• [--] Month [---] +36%
• [--] Months [---] +185%
• [--] Year [---] +1,967%

Followers: [------] #

• [--] Week [------] +2.50%
• [--] Month [------] +9.50%
• [--] Months [------] +100%
• [--] Year [------] +376%

CreatorRank: [-------] #

Social Influence

Social category influence technology brands 25.56% stocks 21.05% countries 11.28% finance 6.77% currencies 3.76% automotive brands 3.76% social networks 3.01% cryptocurrencies 2.26% travel destinations 2.26% celebrities 1.5%

Social topic influence ai 17.29%, in the 6.77%, core 5.26%, taiwan #1020, inference #392, claude code #351, anthropic #697, china 3.76%, gas #1596, systems 3.01%

Top accounts mentioned or mentioned by @institlpgp @elonmusk @rwang07 @johncoogan @zaiorg @fabknowledge @jaysyoon @treasureh8nter @ogawa_tter @jiahanjimliu @william_r2rclub @r_and_invest @lang__leon @dylan522p @ai_hyperbull @peterwildeford @friedmandave @whitfill_parker @midnight_captl @tbu12345678

Top assets mentioned Microsoft Corp. (MSFT) TeraWulf Inc. (WULF) Taiwan Semiconductor (TSM) Alltoscan (ATS) Axion (AXN) Amphenol Corporation (APH) Cipher Mining Inc. Common Stock (CIFR) Hut [--] Mining Corp. (HUT) Astera Labs, Inc. (ALAB) Camtek Ltd (CAMT)

Top Social Posts

Top posts by engagements in the last [--] hours

"Track key global shipments with your ChipBook subscription: (2/2) https://semianalysis.com/chipbook/ https://semianalysis.com/chipbook/"
X Link 2026-02-12T14:01Z 60.7K followers, [----] engagements

"Jan has historically been 8.1% of full-year revenue. January Y/Y correctly predicts full-year direction 75% of the time (12/16 years). Currently Jan [----] is tracking above the full-year guidance pace of 30% in USD. Jan strength means TSMC is on track to post sequential growth in 1Q a rare outcome historically as Q1 is the seasonally weakest quarter. Even if growth moderates through the year (unlikely) the January strength gives significant cushion. January is almost never the peak month. Only [--] out of [--] years (2001 2015) had January as the highest month meaning the rest of [----] is very"
X Link 2026-02-10T22:00Z 60.7K followers, [----] engagements

"In an environment defined by constrained supply will AI servers go the way of Rolex watches Will Neoclouds have to butter up their Authorized Dealers to get the "privilege" of being on a months-long waitlist to "get the call" for hard to obtain networking equipment and AI servers (3/3) https://twitter.com/i/web/status/2022068239895826774 https://twitter.com/i/web/status/2022068239895826774"
X Link 2026-02-12T22:00Z 60.7K followers, [----] engagements

"BREAKING: @elonmusk 's xAI has bought an [--] additional 380MW of natural gas turbines from South Korea's Doosan Enerbility. The first two units are scheduled for delivery by the end of [----]. This will power an additional 600000+ GB200 NVL72 equivalent size cluster (or [------] + VR200 NVL144 equivalent size cluster). @rpoo @elonmusk @BrentM_SpaceX is singluar in terms of their abitily to build the absolute largest datacenter in the world nobody else run as fast as them. https://twitter.com/i/web/status/2008327656840831189 https://twitter.com/i/web/status/2008327656840831189"
X Link 2026-01-06T00:00Z 60.7K followers, 337.8K engagements

"We put out a Core Research article [--] weeks ago highlighting the logic behind the imminent ABF bottleneck while calling out AT&S $ATS and some of the other key beneficiaries. Those who listened then would be up +66% by now. LFG ABF substrate will have a multi-year shortage Any player involved in the supply chain will benefit LFG ABF substrate will have a multi-year shortage Any player involved in the supply chain will benefit"
X Link 2026-02-06T00:00Z 60.7K followers, 26.3K engagements

"Hi @cerebras can u have cooler models like Kimi K2.5 or DeepSeek ppl are willing to pay premium if fast agentic coding but looks like your stack is not currently capable of deploying an 1T model yet. Fast inference = 6x markup Don't be giving us any ideas 😼 Fast inference = 6x markup Don't be giving us any ideas 😼"
X Link 2026-02-08T20:48Z 60.7K followers, 74.4K engagements

"This is why dram costs so much"
X Link 2026-02-11T02:00Z 60.7K followers, 16.2K engagements

"Memory costs skyrocketing is leading to huge increases in server BoM for Neoclouds. They either have to raise hour GPU prices to cover the cost increase or not build. Memory cost increasing by three times the past quarter has led to risks throughout the value chain with OEM behavior becoming increasingly disorderly: quotes are being shortened repriced or withdrawn and future cost risk is being embedded into todays pricing. This has moved traditional enterprise procurement into more like AI server surge pricing with OEMs having to cover their risk. Even though BoM increases have been 10% we"
X Link 2026-02-12T22:00Z 60.7K followers, 75.8K engagements

"When server prices rise so sharply Neoclouds project economics change considerably. Proposed GPU deployments that once cleared IRR hurdle rates now are not attractive and will lead to Neoclouds walking away from deals starving the market of compute. Uncertainty around final system costs makes forward planning difficult. Furthermore lead times continue to lengthen not only on the server side but also for networking equipment as strong demand leads to component tightness. Procuring compute was already difficult given strong demand to rent GPUs and the AI Server Pricing Apocalypse makes things"
X Link 2026-02-12T22:00Z 60.7K followers, 25.3K engagements

"Choose your fighter 😈"
X Link 2026-02-13T02:00Z 60.7K followers, 15.1K engagements

"Cryptominers are very familiar with the datacenter world. Building out their own 100s of MW using immersion liquid cooling and managing supply chains of industrial equipment. Even better they can quickly tear down a cheap mining DC and replace with AI DC leveraging existing infra. (4/5) https://twitter.com/i/web/status/2022370240181510191 https://twitter.com/i/web/status/2022370240181510191"
X Link 2026-02-13T18:00Z 60.7K followers, [----] engagements

"CPUs are Back: The Datacenter CPU Landscape in [----] RL and Agent Usage Context Memory Storage DRAM Pricing Impacts CPU Interconnect Evolution AMD Venice Verano Florence Intel Diamond Rapids Coral Rapids Arm Phoenix + Venom Graviton [--] Axion. https://newsletter.semianalysis.com/p/cpus-are-back-the-datacenter-cpu https://newsletter.semianalysis.com/p/cpus-are-back-the-datacenter-cpu"
X Link 2026-02-10T02:21Z 60.7K followers, 92.4K engagements

"Amphenol (APH US) sold off post-earnings despite beating Q4 [--] estimates on a softer than expected Q1 [--] guide. The concerns: smaller beat magnitude than recent quarters margin guidance that missed expectations and organic revenue declining mid-single digits QoQ. The 14% Connectivity and Cable Solutions (CCS) revenue growth guide for [----] also looks sandbagged versus 30%+ growth in [----] especially with Corning seeing strong fiber demand. Record orders hit $8.4B (a 1.31x book-to-bill vs 0.99x last quarter) signaling customers are pulling forward and locking in next-gen AI program ramps that"
X Link 2026-02-11T18:02Z 60.7K followers, 18.5K engagements

"The Two Sides of Jensen Marketing vs Business Genius"
X Link 2026-02-12T02:00Z 60.7K followers, 15.7K engagements

"This week AMAT introduced Producer Viva Radical Treatment System. It is an AtomicLevel Polisher system that enables precision engineering of GAA Transistor Nanosheets. In a GateAllAround (GAA) architecture the core structure consists of horizontally stacked nanosheets. These ultrathin(only a few nanometers wide) silicon layers have surface conditions that directly determine chargecarrier transport efficiency. Even atomicscale roughness or contamination can significantly degrade device performance. (1/3) 🧵 https://twitter.com/i/web/status/2022309821186228224"
X Link 2026-02-13T14:00Z 60.7K followers, 17.5K engagements

"Applied Materials specifically developed to address these critical industry pain points: 🟠 Proprietary Ultra-Pure Radical Architecture The core of the Viva system utilizes a proprietary delivery architecture to generate ultra-pure radicals. By combining a remote plasma source with innovative hardware it effectively filters out high-energy charged ions that would otherwise damage fragile chip structures. 🟠 Damage-Free and Deep Treatment Environment The high concentration of neutral radicals creates an extremely gentle non-destructive environment. This allows for superior penetration enabling"
X Link 2026-02-13T14:00Z 60.7K followers, [----] engagements

"🟠 Dramatic Boost in Transistor Performance Through angstrom-level precision in surface modification the system creates an exceptionally pure and uniform channel surface. This significantly enhances electron mobility and switching speeds resulting in more efficient power-saving transistors capable of meeting next-generation AI demands. 🟠 Extended Application in Interconnect Technology Beyond channel engineering for logic chips Viva demonstrates powerful process scalability. When integrated with the Producer Pyra thermal anneal process its complementary radical treatment further reduces"
X Link 2026-02-13T14:00Z 60.7K followers, [----] engagements

"But how do you determine which ones are the best Which sites are capable of retrofitting How much capacity do they have in [----] Only our Datacenter Model subscribers get to find out. (5/5) https://semianalysis.com/datacenter-industry-model/ https://semianalysis.com/datacenter-industry-model/"
X Link 2026-02-13T18:00Z 60.7K followers, [----] engagements

"This is how OpenAI comes back and beats Anthropic"
X Link 2026-02-13T22:00Z 60.7K followers, 38.4K engagements

"DeepSeek Debates: Chinese Leadership On Cost True Training Cost Closed Model Margin Impacts H100 Pricing Soaring Subsidized Inference Pricing Export Controls MLA https://semianalysis.com/2025/01/31/deepseek-debates/ https://semianalysis.com/2025/01/31/deepseek-debates/"
X Link 2025-01-31T05:03Z 60.1K followers, 561.1K engagements

"Zareen is one of the go to places for many SF Bay Area AI researchers to get a quick bite. Most of the food is very good and was even on the Michelin guide in [----]. AI researchers not experienced with the Indian cuisine will commonly order their chicken tikka masala with garlic naan and mango lassi"
X Link 2025-10-26T15:00Z 60.1K followers, 243.1K engagements

"How to Kill [--] Monopolies with [--] Tool Substrate X-Ray Lithography a New American Foundry $10k Logic Wafers https://newsletter.semianalysis.com/p/how-to-kill-2-monopolies-with-1-tool https://newsletter.semianalysis.com/p/how-to-kill-2-monopolies-with-1-tool"
X Link 2025-10-29T21:05Z 60.1K followers, 406.6K engagements

"The high cost and power proportion taken up by DSPs has motivated the industry to find technologies that can disintermediate the DSP. The first wave of attack on the DSP was linear pluggable optics (LPO) which attempt to remove the DSPs altogether and have the SerDes from the switches directly drive the TX and RX optical elements in the transceiver. CPO takes the LPO concept to the next step by placing the optical engine on the same package as the compute or switch chip. A key benefit of CPO is that the DSP that was found in the transceiver is no longer required because the distance between"
X Link 2025-11-07T19:13Z 59.9K followers, 23.7K engagements

"The economics of AI has been a big question mark in many investors' minds - What does the value chain look like How do you model out the ROIC of AI What would the ROIC look like We built up an end-to-end economics stack to answer this question - how we go from a chips silicon cost through full system integration all the way down to the dollar cost per million inference tokens.(1/4)🧵"
X Link 2025-11-14T16:28Z 60K followers, 103K engagements

"Soitec is a French company known for manufacturing SOI wafers using its patented Smart Cut technology. Great technology companies continue to find new markets to apply their technology to and that's exactly what Soitec has done. Recently it discovered a new application for Smart Cut: its own share price. This year it has Smart Cut its own share price in half not once but twice Who said that Europe can't innovate https://twitter.com/i/web/status/1993351263598112981 https://twitter.com/i/web/status/1993351263598112981"
X Link 2025-11-25T16:09Z 59.6K followers, 155.8K engagements

"Massive IT load growth. A transforming electric grid. Five-year lead times for turbines. Why not build more of them Well GE and Siemens have seen this story before. (1/8)🧵"
X Link 2025-12-04T18:58Z 59.8K followers, 32.1K engagements

"Back in the '90s parts of the American electric grid were "deregulating." These reforms gave us commodity markets for electricity--aka ISOs and RTOs. INDEPENDENT POWER PRODUCERS (IPPs) often utilities from other states could build and run their own power plants and make money on these new electricity markets. Their generator of choice The COMBINED CYCLE GAS PLANT (CCGT) particularly the then-new F-CLASS. (2/8) https://twitter.com/i/web/status/1996655416579698736 https://twitter.com/i/web/status/1996655416579698736"
X Link 2025-12-04T18:58Z 59.8K followers, [----] engagements

"This is What Winning Looks Like From Space Satellite imagery just caught xAI doing something no press release ever could: painting MACROHARD across the roof of its Colossus [--] datacenter in Tennessee. A building-sized wink at Microsoft visible from orbit impossible to ignore. It reads like a joke until you look at whats underneath. ⚆ Colossus [--] isnt a concept site or a slide in a deck. Power and cooling infrastructure are expanding in lockstep with the campus pushing toward [---] MW of active capacity in the near term. Gas turbines are arriving. Cooling towers are multiplying. The build is loud"
X Link 2025-12-24T17:54Z 59.5K followers, 2M engagements

"IMPORTANT: Anthropic will directly purchase close to [-------] TPUv7 units and deploy them in facilities it controls. These [--] million chips will be purchased directly from Broadcom selling the systems to Anthropic. Under this structure for Anthropic owned TPUs TeraWulf (WULF) Hut8 Cipher Mining (CIFR) will deliver the data center infrastructure while Fluidstack will assume responsibility for on-site deployment including cabling burn-in acceptance testing and remote-hands services effectively outsourcing physical server operations for Anthropic."
X Link 2026-01-02T23:00Z 59.8K followers, 563.6K engagements

"AWS's Trainium3 NL72x2 rack has over 368+ Astera Labs scale up switches. Each compute tray will have four Trainium3 chips & eight [--] lane PCIe [---] PHY intra-tray scale up switches. In addition at the rack level there is [--] switch only trays with each switch tray having four [---] lane switch chips. Our public December [----] article analyzes this in depth. https://twitter.com/i/web/status/2008989969122267408 https://twitter.com/i/web/status/2008989969122267408"
X Link 2026-01-07T19:51Z 59.8K followers, 19.8K engagements

"From Tokens to Burgers: A Water Footprint Face-Off https://newsletter.semianalysis.com/p/from-tokens-to-burgers-a-water-footprint https://newsletter.semianalysis.com/p/from-tokens-to-burgers-a-water-footprint"
X Link 2026-01-15T19:06Z 59.5K followers, 40.1K engagements

"In [----] we flagged Lam Research expanding in Malaysia. Today most of its high-volume production is there not the U.S. That trend has only accelerated. (2/6) https://newsletter.semianalysis.com/p/lam-research-is-outsourcing-more https://newsletter.semianalysis.com/p/lam-research-is-outsourcing-more"
X Link 2026-01-22T17:02Z 59.5K followers, 11.9K engagements

"Applied Materials is following. Its Singapore [----] plan aims to double manufacturing capacity in the country over this decade. (3/6)"
X Link 2026-01-22T17:02Z 59.7K followers, 11.1K engagements

"IMPORTANT: NVIDIA just announced the ability for their new BlueField-4 DPU to control [--] CX-9 backend NICs allowing for unified management of both frontend north-south network & backend high speed east-west networks. The only issue with this is that DPUs are mainly seen as an cost center and BlueField-4 is fucking expensive especially considering that each BlueField-4 has 128GB of LPDDR5X. Remember the whole point of an DPU is to save CPU cores but when the DPU is priced at 75% margins (i.e. 4x cost of goods) the cost advantages of BlueField quickly fade away. This is why many of the largest"
X Link 2026-01-30T22:00Z 59.6K followers, 30.2K engagements

"A couple days ago Nvidia has accepted AWS's patch for NVIDIA's inference KV Cache transfer library to allow for interop over EFA between NVIDIA GPUs & Trainium Chips. What this means is AWS can do inference prefill on small scale up machines like NVIDIA HGX while doing inference decode on large scale up world size racks like Trainium3 NL32x2/NL72x2 or GB200 NVL36x2. Prefill does not require large scale up domain sizes. https://twitter.com/i/web/status/2018072496117854627 https://twitter.com/i/web/status/2018072496117854627"
X Link 2026-02-01T21:22Z 59.4K followers, 42.6K engagements

"Australia's datacenter market is blowing up Major hyperscalers making massive moves and leased capacity is set to triple over the coming years. Our datacenter model subscribers found out weeks ago so now twitter gets to learn. (1/5)🧵"
X Link 2026-02-02T18:00Z 59.4K followers, 12.4K engagements

"Chat is the same person that put out the reassuring NVIDIA PR telling everyone to keep calm also in charge of Oracle's twitter account The NVIDIA-OpenAI deal has zero impact on our financial relationship with OpenAI. We remain highly confident in OpenAIs ability to raise funds and meet its commitments. The NVIDIA-OpenAI deal has zero impact on our financial relationship with OpenAI. We remain highly confident in OpenAIs ability to raise funds and meet its commitments"
X Link 2026-02-02T21:07Z 59.6K followers, 44.8K engagements

"REUTERS: "OpenAI is unsatisfied with some Nvidia chips and looking for alternatives sources say""
X Link 2026-02-02T21:56Z 59.7K followers, 73.2K engagements

"Shoutout to @elonmusk's Colossus [--] being more environmentally friendly than @innoutburger CRAZY😭 (Video Credits: @pleometric)"
X Link 2026-02-03T02:00Z 60.4K followers, 35.1K engagements

"China inspection equipment imports reached record levels in mid [----] but have since begun to normalize. The slowdown in imports is more pronounced in imports from Singapore and the USA (KLA & ONTO) than imports from Israel (Nova & Camtek). (1/2)🧵"
X Link 2026-02-03T18:01Z 59.5K followers, [----] engagements

"Jensen framed a benchmark: about $50B of capex for a 1GW datacenter. That number belongs in the same mental shelf as iconic public works. The Golden Gate Bridge cost roughly $35M in the 1930s and opened in [----]. It still carries daily life nearly a century later. Both types of projects are bets on flow. One moves people and goods across water. One moves computation across networks. Both require power systems supply chains maintenance cultures and rules that decide who benefits and who pays. A useful lens is simple: cost use life and spillover. The sticker price is visible. The real story is"
X Link 2026-02-03T22:00Z 59.6K followers, 13.6K engagements

"The default UI visuals of Kimi K2.5 is much better than Claude Code Claude Code gives off the infamous purple hue by default that looks ugly while Kimi K2.5 default generates an much more cleaner UI"
X Link 2026-02-04T02:00Z 59.5K followers, 10.5K engagements

"China experienced a massive surge in ion implanter demand through [--------] fueled by aggressive silicon carbide (SiC) capacity expansion and a broader push for domestic semiconductor self-sufficiency. However following the SiC market correction and a cooling off of overall fab expenditures imports from the U.S. have stabilized at historically normalized levels. These volumes have likely reached a floor with further significant reductions now improbable. (1/2)🧵 https://twitter.com/i/web/status/2019048568024129560 https://twitter.com/i/web/status/2019048568024129560"
X Link 2026-02-04T14:01Z 59.7K followers, [----] engagements

"The third year of Jensens Annual AI supply chain dinner in Taiwan. The participants are executives and leaders of Nvidias supply chain partners and the list has grown even bigger this year. The media dub the dinner as the Trillion Dollar Dinner teasing the importance of these companies as a part of the trillion dollars AI server industry as well as the aggregate market cap of these companies. The companies attended are the following: Semiconductor manufacturing packaging testing: 🟠 TSMC 🟠SPIL 🟠KYEC ODM/OEM AI server assembly: 🟠Foxconn 🟠Quanta 🟠Wistron/Wiwynn 🟠Pegatron 🟠Inventec"
X Link 2026-02-04T18:01Z 59.9K followers, 25K engagements

"If you go to the website rentahuman you can have an AI agent rent a person to do tasks in the real world for it. However if you wanted to instead rent a robot to do those tasks for you then you should check out our post on Robotics Levels of Autonomy below. (1/2)🧵"
X Link 2026-02-05T00:00Z 59.9K followers, [----] engagements

"Due to the Agent Swarm innovation Kimi K2.5 Agent feels so much better Instead of trying to increase the interactivity via more memory bandwidth the 10x cracked Chinese engineers take a different approach by dividing an task into any parallel subtasks such that they can provide an much faster experience to the end user without requiring any additional TB/s. Constraints breeds innovation https://twitter.com/i/web/status/2019471105682002313 https://twitter.com/i/web/status/2019471105682002313"
X Link 2026-02-05T18:00Z 60.1K followers, 26.9K engagements

"Taiwan server production was stable at historically high levels for much of [----] but most recently - ASPs on those units skyrocketed +280% YoY and +140% sequentially as new server SKUs continue to ramp. (1/2)🧵"
X Link 2026-02-06T14:01Z 59.9K followers, 10.2K engagements

"Tracking components in the semiconductor supply chain is important for two reasons: 1) It provides early upstream leading indicators of downstream performance. 2) It offers direct insight into key enabling suppliers and often highlights rich potential investment opportunities. CCL exports is a great example for this; as a core component in every PCB or substrate supporting an AI accelerator CCL datasets provide an early read-across for the GPU vendors as well as meaningful information about material suppliers themselves (Nan Ya Plastics EMC etc.) many of whom have outperformed $NVDA over the"
X Link 2026-02-10T14:00Z 60K followers, [----] engagements

"We ran a small benchmark with WideSearch on Claude Teams: [--] unique English tasks [--] trials each 30-min timeout GPT-5.2 judge. We compare it to a Opus [---] baseline. Teams were more expensive and slower on average: Total $93 for baseline vs Total $131 for teams. We got 46/60 baseline vs 47/60 teams completed runs due to our 30-min cutoff (WideSearch tasks can run for 3+ hours with 100+ sources) (2/9)"
X Link 2026-02-10T18:00Z 59.9K followers, [----] engagements

"Teams also scored lower than the Claude Opus [---] baseline in our setup. Not just slower and more expensive. From their technical report Kimi2.5 scored 72.7% on item F1 and reaches 79.0% with agent swarm. This lines up with the idea that K2.5s gains come from training + RL not just orchestration prompting. Whereas from our experiments Claude Opus [---] received a score of 64.8% and Agent swarm scored lower 53.8%. (Note: Kimi vs Claude numbers arent directly comparable across different task sets timeouts and trial counts. We also did not change CLAUDE_CODE_SUBAGENT_MODEL and used the default of"
X Link 2026-02-10T18:00Z 60K followers, [----] engagements

"Sequential tool-calling scales linearly with steps. K2.5 explicitly trains on wide-search + deep-branch tasks without telling the model to parallelize so parallel decomposition is learnt not copied or prompted. Trainable orchestrator: plans decomposes schedules Frozen subagents: execute subtasks concurrently Orchestrator aggregates results into the final answer Mechanically the swarm is just two extra tools: create_subagent(name system_prompt) assign_task(agent prompt) A trainable orchestrator plans/decomposes/schedules while frozen subagents execute concurrently. (4/9)"
X Link 2026-02-10T18:00Z 59.9K followers, [----] engagements

"The reward shaping targets two classic multi-agent failure modes: r_parallel: prevents serial collapse (one agent does everything) r_finish: prevents spawn spam / pointless parallelism r_perf: the actual task outcome reward There is also a new token-level clipping mechanism designed to mitigate off-policy divergence which occurs when the updated policy drifts too far away from the policy that generated the data. This is more crucial over long sequences which are typical for agentic workloads. (5/9) https://twitter.com/i/web/status/2021283061632008293"
X Link 2026-02-10T18:00Z 59.9K followers, [----] engagements

"The swarm also functions as proactive context management or context sharding: Parallelize without contaminating the global context. Each subagent keeps an independent working memory/local context and the orchestrator only receives task-relevant outputs not full traces. The technical report explicitly points this out as better than reactive truncation (e.g. discard-all summary) (7/9)"
X Link 2026-02-10T18:00Z 59.9K followers, [----] engagements

"With Kimi [---] showing the possible effectiveness of agent swarms we might stop treating multi-agent as a prompt pattern and start treating it like a planner + distributed runtime problem: optimize the critical path not total steps; shard context; schedule tool I/O like jobs. (9/9)"
X Link 2026-02-10T18:00Z 60.1K followers, [----] engagements

"RT @johncoogan: AGI is here but the NFL's software system still takes [--] days to upload group photos"
X Link 2026-02-10T19:46Z 59.9K followers, [--] engagements

"PCBs are the core building block of any integrated circuit or electronic device. It doesnt matter if its an Nvidia GPU or the Apple A18 they all get mounted on a PCB. Tracking the state of the PCB industry is simple yet effective way to understand where we are in the cycle. (1/3) 🧵"
X Link 2026-01-28T14:01Z 60.7K followers, 13.5K engagements

"In the metrology and inspection market In the metrology and inspection (M&I) market KLA stands unrivaled at the top far ahead of all competitors driven by its comprehensive product portfolio and longstanding acquisition strategy.However in recent years we have seen Onto Innovation and Hitachi HighTech rapidly closing the gap as they strengthen their portfolios and gain traction in advanced nodes. (4/4)"
X Link 2026-01-28T22:00Z 60.7K followers, [----] engagements

"xAIs Colossus [--] First Gigawatt Datacenter In The World Unique RL Methodology On Site Turbines + Mississippi Expansion with Solaris Energy Can xAI afford it Middle East Funding Tesla Talent Exodus API revenue Consumer Growth RL Environment https://semianalysis.com/2025/09/16/xais-colossus-2-first-gigawatt-datacenter/ https://semianalysis.com/2025/09/16/xais-colossus-2-first-gigawatt-datacenter/"
X Link 2025-09-16T18:30Z 60.7K followers, 872.7K engagements

"Even Toto the worlds premier toilet manufacturer has been benefitting from the memory cycle No not because procurement teams are cr**ping themselves over their next round of their DRAM contract negotiations. but because Toto is a leading supplier of electrostatic chucks (ESCs) used to hold wafers during critical fab steps like etch and deposition. (1/2) 🧵 https://twitter.com/i/web/status/2017296971690541488 https://twitter.com/i/web/status/2017296971690541488"
X Link 2026-01-30T18:00Z 60.7K followers, 12.5K engagements

"IMPORTANT CLAUDE TIP: it is important that contains essentials such as prompting claude code to use Master branch by default instead of Main branch http://CLAUDE.md http://CLAUDE.md"
X Link 2026-02-09T22:00Z 60.7K followers, 20.6K engagements

"Core Scientific is the quintessential cryptominer that pivoted into AI datacenters. Back in [----] the company announced it movement and since then the stock has 5xd and aims to serve 1GW of workloads soon (2/5)"
X Link 2026-02-13T18:00Z 60.7K followers, [----] engagements

"RL Environments and RL for Science Data Foundries + Multi-Agent Architectures Worker Automation RL as a Service Anthropic's next big bet GDPval + Utility Evals Computer Use Agents LLMs in Biology Lab Procurement Patterns Platform Politics + Access https://newsletter.semianalysis.com/p/rl-environments-and-rl-for-science https://newsletter.semianalysis.com/p/rl-environments-and-rl-for-science"
X Link 2026-01-06T18:50Z 60.7K followers, 148.2K engagements

"Claude Code is the Inflection Point What It Is How We Use It Industry Repercussions Microsoft's Dilemma Why Anthropic Is Winning. https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point"
X Link 2026-02-05T19:05Z 60.7K followers, 916.5K engagements

"Agent swarms are moving from prompting tricks to math. Kimi K2.5 trains an orchestrator that spawns + schedules specialist sub-agents in parallel and reporting [----] lower wall-clock on WideSearch plus higher scores. Anthropic also recently released agent teams where multiple Claude Code instances work together. It is still experimental but has been used to write Claude's C compiler. (1/9) 🧵 https://twitter.com/i/web/status/2021283054019330194 https://twitter.com/i/web/status/2021283054019330194"
X Link 2026-02-10T18:00Z 60.7K followers, 33.2K engagements

"Swarms convert a single serial GPU inference into a branching distributed workload: many concurrent subagent jobs each with its own context and tool calls plus a fan-out/fan-in control loop. That and agentic load in general shifts the bottleneck from just GPU decode toward scheduler overhead tail latency and I/O so the inference stack starts to look like a workflow engine running at datacenter scale. In that world CPUs start to matter more. They drive orchestration tool execution retrieval pipelines sandboxing and caching while GPUs still do the core model compute. Find out more about the"
X Link 2026-02-10T18:00Z 60.7K followers, [----] engagements

"TSMC's Jan [----] set a new all-time monthly record: $12.7B (+42.6% Y/Y +19.1% M/M). January has been positive every single year since [----] for [--] straight years. The current streak of Jan M/M growth (2020-2026) is the longest in TSMC's history. Previous record was [--] years (2011-2016). And Jan 2026's +19% is the strongest January M/M since 2013's +28%. Jan perfectly maps to Q1 guidance. Jan is typically 35.9% of Q1 revenue implying $35.5B revenue in 1Q26 right in the $34.6-$35.8B guidance range. (1/2)🧵 https://twitter.com/i/web/status/2021343445512487377"
X Link 2026-02-10T22:00Z 60.7K followers, 14.1K engagements

"As Korean Memory makers continue to ship both mainstream and high-end memory to Taiwan and China we are also monitoring an uptick in exports to Malaysia a likely read across for $INTCs Advanced Packaging capacity ramp in that locale. (1/2)🧵"
X Link 2026-02-11T14:01Z 60.7K followers, 28.7K engagements

"Keep an eye on the latest Intel developments with your ChipBook subscription: (2/2) https://semianalysis.com/chipbook/ https://semianalysis.com/chipbook/"
X Link 2026-02-11T14:01Z 60.7K followers, [----] engagements

"We like APH's AI content growth in FY26. The VR200 NVL144 will see gen-on-gen increase in compute tray copper content versus GB300 and NVIDIA's shift to a cableless compute tray design is actually a net positive to APH 's Paladin board-to-board connectors replace vulnerable cables and capture PCIe content. Dollar content per compute tray in VR NVL144 will be much higher than that of current-gen systems. Amphenol is also winning Paladin design slots at AWS for Trainium systems. (2/3) https://twitter.com/i/web/status/2021646016445809052 https://twitter.com/i/web/status/2021646016445809052"
X Link 2026-02-11T18:02Z 60.7K followers, [----] engagements

"Paladin connectors are Amphenol's high-speed board-to-board solution that enables the cableless architecture in next-gen AI racks. Current designs use DensiLink OverPass and PCIe cables inside compute trays but these are vulnerable to assembly damage and create failure points. Paladin eliminates these cables by routing signals through the PCB with board-to-board connections. (3/3) https://twitter.com/i/web/status/2021646017536340176 https://twitter.com/i/web/status/2021646017536340176"
X Link 2026-02-11T18:02Z 60.7K followers, [----] engagements

"Two AWS sites are quietly shaping the companys next wave of growth while the rest play catch-up. Heres whats happening: Project Rainier in Indiana and Mississippi is doing most of the heavy lifting. Both campuses are on track driving the bulk of AWSs near-term capacity growth. Mississippi is already preparing the next phase: [--] additional buildings (660MW) are in the works though vertical construction hasnt started yet. (1/2) 🧵 https://twitter.com/i/web/status/2021823798517920084 https://twitter.com/i/web/status/2021823798517920084"
X Link 2026-02-12T05:48Z 60.7K followers, 18.5K engagements

"Outside these two no other site is moving fast enough for a GW-scale delivery in late [----] or [----]. AWS has broken ground on some GW-scale sites in Georgia North Carolina and Pennsylvania but progress there is slower. Big picture: AWSs plan to double power capacity by [----] is back-end loaded. Indiana Mississippi and international expansion will define whether the next phase of growth hits on time and it looks like it could be a major acceleration in [----]. (2/2) https://twitter.com/i/web/status/2021823799964946887 https://twitter.com/i/web/status/2021823799964946887"
X Link 2026-02-12T05:48Z 60.7K followers, 13.2K engagements

"After a slowdown in the first half of [----] we are once again seeing material shipments of IC chips from Taiwan to Mexico as $TSM seeks to leverage trade agreements such as USMCA to reduce tariffs and make trade easier. (1/2)🧵"
X Link 2026-02-12T14:01Z 60.7K followers, [----] engagements

"The November ChipBook focused on key names in the increasingly important Test space and called out [--] publicly traded companies including WinWay ($6515.TW). On December 23rd we went further in Core-Research and explained why WinWay is uniquely poised to benefit from increasingly large AI packages. Since that date WinWay is trading up +75%. (1/3) 🧵 https://twitter.com/i/web/status/2022008212564615648 https://twitter.com/i/web/status/2022008212564615648"
X Link 2026-02-12T18:01Z 60.7K followers, [----] engagements

"The other two names listed on the same Chipbook slide are up +38% and +127% since the Chipbook Released (2/3)"
X Link 2026-02-12T18:01Z 60.7K followers, [----] engagements

"In China an online recruiting platform trains a 3B param LLMs that outperforms Qwen 🚀 Announcing Nanbeige4.1-3B our latest open-source 3B model mastering reasoning preference alignment & agent capabilities Try it now: https://t.co/FrR6RL9DjW https://t.co/Hwd6M4PCBP https://t.co/sKLEjUMUjA 🚀 Announcing Nanbeige4.1-3B our latest open-source 3B model mastering reasoning preference alignment & agent capabilities Try it now: https://t.co/FrR6RL9DjW https://t.co/Hwd6M4PCBP https://t.co/sKLEjUMUjA"
X Link 2026-02-13T16:53Z 60.7K followers, 132.4K engagements

"Many are putting up impressive numbers as well. TeraWulf with 360MW+ officially announced with Google. IREN with Microsoft for 200MW. Others are hitting huge numbers as well. Why are they so successful (3/5)"
X Link 2026-02-13T18:00Z 60.7K followers, [----] engagements

"Congrats on the release @Zai_org Support for DeepSeek Attention (DSA) is still maturing but we're already seeing great performance with vLLM High sustained input token throughput across all input sequence lengths shows the power of DSA Great work @vllm_project 🔥Congrats to @Zai_org on launching GLM-5 744B parameters (40B active) trained on 28.5T tokens integrating DeepSeek Sparse Attention to keep deployment cost manageable while preserving long-context capacity. vLLM has day-0 support for GLM-5-FP8 with: 📖 DeepSeek Sparse https://t.co/YbqhtHiq1i 🔥Congrats to @Zai_org on launching GLM-5"
X Link 2026-02-14T20:15Z 60.7K followers, 21.9K engagements

"Memory Mania: How a Once-in-Four-Decades Shortage Is Fueling a Memory Boom Prices are doubling again Supercycle is bigger and could last longer than you think. https://newsletter.semianalysis.com/p/memory-mania-how-a-once-in-four-decades https://newsletter.semianalysis.com/p/memory-mania-how-a-once-in-four-decades"
X Link 2026-02-06T17:48Z 60.7K followers, 99.4K engagements

"Day [--] support for GLM-5 744B frontier MoE and recipes on CUDA SGLang Disappointed that @AnushElangovan hasnt landed GLM5 Day [--] ROCm SGLang support and recipes. 🎉 The mysterious Pony Alpha is finally revealed congrats to @Zai_org on releasing GLM-5 SGLang is ready to support on day-0. 🛠 744B params (40B active) model built for complex systems engineering & long-horizon agentic tasks 📚 28.5T tokens pretraining for a stronger https://t.co/VIokMU7aqV 🎉 The mysterious Pony Alpha is finally revealed congrats to @Zai_org on releasing GLM-5 SGLang is ready to support on day-0. 🛠 744B params (40B"
X Link 2026-02-11T20:51Z 60.7K followers, 13.2K engagements

"Subscribe to ChipBook and Core-Research here: (3/3) https://semianalysis.com/core-research/ https://semianalysis.com/chipbook/ https://semianalysis.com/core-research/ https://semianalysis.com/chipbook/"
X Link 2026-02-12T18:01Z 60.7K followers, [----] engagements

"Time-to-market is dominating the datacenter race. It justifies new business opportunities like powered land and the onsite gas boom. But another segment is taking more share with massive plans of scaling in mind: cryptominers. (1/5) 🧵"
X Link 2026-02-13T18:00Z 60.7K followers, 15K engagements

"Nvidia AI Demand Unlocking Billions in New Growth"
X Link 2026-02-14T02:00Z 60.7K followers, 13.2K engagements

"We have been cookin with Seedance-2 🤤"
X Link 2026-02-15T02:00Z 60.7K followers, 19K engagements

"Everything makes sense now TSMCs earlier acquisition of a massive land parcel in Arizona provides enough space for an additional [--] to [--] fabs. Meanwhile a landmark tariff agreement between Taiwan and the U.S. has been finalized. Under this deal Taiwan will commit a total of $500Bcomprising $250B in direct investment and $250B in government-backed credit guaranteesin exchange for a tariff cap of no more than 15%. (1/3)🧵 https://twitter.com/i/web/status/2014443175717343341 https://twitter.com/i/web/status/2014443175717343341"
X Link 2026-01-22T21:00Z 57.8K followers, 14K engagements

"Furthermore the Section [---] semiconductor framework establishes two pivotal conditions. First Taiwanese companies expanding U.S. semiconductor capacity are permitted to import up to [---] times their planned output exempt from Section [---] duties during the construction phase. Second upon completion of these production projects companies remain eligible to import [---] times their actual U.S. capacity duty-free (2/3) https://twitter.com/i/web/status/2014443177168511100 https://twitter.com/i/web/status/2014443177168511100"
X Link 2026-01-22T21:00Z 57.9K followers, [----] engagements

"In other words if Taiwan keeps growing its semiconductor output it must proportionally expand its U.S. footprint as well.And thats exactly why TSMC has been moving so quickly in Arizona. (3/3)"
X Link 2026-01-22T21:00Z 57.8K followers, [----] engagements

"Following $TSMs $52$56 billion CapEx guide for [----] the ChipBook Taiwan WFE tracker serves as a valuable tool for monitoring equipment arrivals and installations as leading indicators of capacity expansion and node ramps. (1/2) 🧵"
X Link 2026-01-23T02:00Z 57.8K followers, [----] engagements

"IMPORTANT: At CES [----] this week NVIDIA marketing has renamed VR200 NVL144 back to VR200 NVL72. Gone is NVL's Jensenflation. VR200 only ever had [--] logical GPU within an single scale up domain so the name never made sense & led to customer frustration. Glad to Jensen has stopped with this naming scheme. https://twitter.com/i/web/status/2014699994557620423 https://twitter.com/i/web/status/2014699994557620423"
X Link 2026-01-23T14:01Z 57.7K followers, 19.6K engagements

"NVIDIA is an natural treasure & Jensen says he has access to the @POTUS anytime 🚀"
X Link 2026-01-23T22:00Z 59.2K followers, 52.7K engagements

"Can we say something without everyone getting mad @tbpn @johncoogan @jordihays"
X Link 2026-01-24T02:00Z 57.7K followers, 42.4K engagements

"Emissions control is an underrated aspect of onsite gas generation. Modern air permitting typically requires strict emissions controls which for a continuously-running gas plant requires exhaust limit of 2-5 parts per million (ppm) of nitrogen oxide (NOx). NOx emissions contribute to smog and acid rain and they're a common byproduct of combustion turbines and engines. (1/4) 🧵 https://twitter.com/i/web/status/2015182873892942139 https://twitter.com/i/web/status/2015182873892942139"
X Link 2026-01-24T22:00Z 58K followers, 16.9K engagements

"A modern GE Vernova LM2500 can do this easily with a Dry Low NOx (DLN) system in the combustion stage and a Selective Catalytic Reduction (SCR). The DLN system works by finely mixing fuel and air to deliver a lean (low fuel high air) burn. This reduces combustion temperature: a hotter burn generates extra NOx. This system alone gets you down to [--] ppm of NOx. From there the SCR system combines catalysts and ammonia to cut emissions output down to the 2-5 ppm required for air permits. (2/4) https://twitter.com/i/web/status/2015182875432223115 https://twitter.com/i/web/status/2015182875432223115"
X Link 2026-01-24T22:00Z 57.9K followers, [----] engagements

"The modern Mitsubishi Power FT-8 an aeroderivative turbine that competes with the LM2500 doesn't have DLN. Instead it relies on older water injection technology to reduce NOx emissions in the combustion stage. This technology can't use normal water though--it needs demineralized water filtered down to [--] microns This requires not only water input but also water treatment beyond normal potable water. (3/4) https://twitter.com/i/web/status/2015182876581511375 https://twitter.com/i/web/status/2015182876581511375"
X Link 2026-01-24T22:00Z 57.7K followers, [----] engagements

"What's strange is that older FT-8 systems had this technology Mitsubishi Power bought the FT-8 aeroderivative design from Pratt & Whitney in [----] but the P&W version DID have a Dry Low NOx option. What gives Without this technology package Mitsubishi Power will struggle in the aeroderivative market because without expensive water treatment infrastructure its aero package can't meet the air permitting necessary for prime power--even if smaller engines and other gas turbines can. (4/4) https://twitter.com/i/web/status/2015182878020178025 https://twitter.com/i/web/status/2015182878020178025"
X Link 2026-01-24T22:00Z 57.7K followers, [----] engagements

"China front-end WFE imports have leveled off after a significant spike in 2023/4 but still remain high relative to historic averages. In particular lithography imports are on the rise again. (1/3) 🧵"
X Link 2026-01-26T14:00Z 57.7K followers, [----] engagements

"Scale-across is proving to be one of the brightest spots of growth in the broader networking space in [----]. The longer distances over which scale-out networks are extended in order to connect datacenters tens or even hundreds of kilometers apart necessitates the use of deep buffer switches and Aristas Big Boy Modular Chassis Switch is the default option across all hyperscalers. Unlike its Tomahawk-based switches where Arista faces competition from other WhiteBox and branded switching vendors the Big Boy chassis enjoys a far wider moat and Arista currently dominates the chassis switch market."
X Link 2026-01-26T18:01Z 58.2K followers, 11.9K engagements

"Arista came to dominate the modular chassis switch market as Cisco and Junipers go to market was oriented around their internal silicon while Arista worked with Broadcom. Aristas go to market was also cloud centric while the others built around telecoms and their high degree of co-engineering with Broadcoms Jericho and Ramon fabric switches led to a product that dominated the market. (2/2) https://twitter.com/i/web/status/2015847466684309744 https://twitter.com/i/web/status/2015847466684309744"
X Link 2026-01-26T18:01Z 58.2K followers, [----] engagements

"IMPORTANT: Blackwell Ultra solves an attention operation performance issue found in Blackwell. As @tri_dao presented at the SemiAnalysis Hackathon one of the biggest bottlenecks in the core attention operation is not GEMMs but the softmax (exponential) part. In Hopper & Blackwell the exponential operation's throughput was too slow to effectively overlap with the [--] GEMMs in attention. Blackwell Ultra solves this by having 2x higher exp throughput than Blackwell Note that other chips have this performance mismatch issue too such as Trainium3 having to increase their exponential function"
X Link 2026-01-26T22:00Z 57.7K followers, 36K engagements

"In modern AI pipelines storage choices can significantly impact performance and efficiency. Two primary approaches are direct-attached storage (DAS) which is local to a server and network-attached storage (NAS) which is accessed over a network by multiple systems. Direct-attached storage (DAS) tightly couples flash to the GPU server minimizing latency and avoiding network hops. This makes it attractive for boot scratch and narrowly latency-sensitive paths but capacity is constrained by server form factor and cost scales poorly as clusters grow. Network-attached storage (NAS) disaggregates"
X Link 2026-01-27T02:00Z 57.9K followers, 18K engagements

"AVGO stock has underperformed over the past few months driven largely by the market narrative that a growing share of TPU volumes will shift toward the MediaTek designed SKU. This perception has contributed meaningfully to Broadcoms recent de rating with MediaTek materially outperforming AVGO over the same period. As we highlighted in our TPU deep dive last December Googles decision to partner with MediaTek is primarily about reducing the margin it pays to its silicon design partners. Broadcom has historically charged Google for the full System in Package allowing it to layer on a substantial"
X Link 2026-01-27T18:00Z 59.2K followers, 29K engagements

"For TPU v8 this strategy is explicitly visible through a dual track approach with one SKU co designed with Broadcom and another with MediaTek. TPU v8AX codenamed Sunfish features two compute dies one IO chiplet and eight stacks of HBM3E while TPU v8X codenamed Zebrafish uses a simpler configuration with a single compute die one IO die and six stacks of HBM3E. While TPU v8X initially faced tapeout delays related to issues on the compute die those challenges have since been resolved and the chip is now expected to ramp aggressively starting in the fourth quarter of [----] and into [----]. These"
X Link 2026-01-27T18:00Z 59K followers, 12K engagements

"Looking ahead TPU v9 will continue to follow a dual track strategy with both Broadcom and MediaTek each retaining a SKU. While the current market narrative paints an increasingly negative picture for Broadcom the reality is considerably more nuanced. For our detailed quarterly unit and ASP forecasts for TPU and our full view on how this plays out across v8 and v9 refer to our Accelerator Model. (3/3) https://semianalysis.com/accelerator-hbm-model/ https://semianalysis.com/accelerator-hbm-model/"
X Link 2026-01-27T18:00Z 58.9K followers, 18.5K engagements

"Even Tech Twitter AI Cultists are not hyping up Claude Code enough Humans are linear thinkers but AI progresses exponentially. Think about the quality of AI coding agents [--] months ago [--] year ago and where they are now and then draw an curve of best fit and that is where they are be in [--] months when more AI coding agent models come out. There is no middle ground those that don't fully embrace AI will perish from lack of abundance. Opus [---] is infinitely times better than previous gen models. Think about how good Opus [--] will be for coding"
X Link 2026-01-27T22:05Z 57.9K followers, 13.4K engagements

"IMPORTANT: NVIDIA announced that their compute tray assembly time fell by 36x from [--] hours to [--] minutes due to the new VR200 cableless & hose less design. NVIDIA follows the footsteps of AWS Trainium2/3 cableless design that leds to faster manufacturing time & full automation as robotics still have a difficult time plugging cables into the correct ports. Note that due to the strength of NVIDIA's world class serdes team they are able to have cableless nvswitch trays too without any retimers leading to faster assembly for switch trays too. In contrast due to AMD's lack of rack scale experience"
X Link 2026-01-28T02:00Z 58.9K followers, 61.5K engagements

"However not all PCBs are created equal. Within each cycle some PCB suppliers grow faster (think AI accelerators) while others will grow slower if at all (think auto). The ChipBook datasets are helpful in providing a birds-eye view of the totality of the semi industry but can dive deep to identify supply chain winners and losers. (2/3)"
X Link 2026-01-28T14:01Z 58.9K followers, 20.9K engagements

"Want to learn more Sign up here: (3/3) https://semianalysis.com/chipbook/ https://semianalysis.com/chipbook/"
X Link 2026-01-28T14:01Z 58.3K followers, [----] engagements

"Tokens are the an unit of intelligence the more tokens you use the more intelligence u asborb"
X Link 2026-01-28T18:00Z 58K followers, [----] engagements

"When people think about semiconductor manufacturing many picture highly automated robotic arms and overhead transport systems. However the true unsung heroes behind highvolume chip production are metrology and inspection. If you cannot see it you cannot manufacture it; and if you cannot measure it you cannot achieve yield. (1/4) 🧵 https://twitter.com/i/web/status/2016632524609315164 https://twitter.com/i/web/status/2016632524609315164"
X Link 2026-01-28T22:00Z 58.9K followers, 30.6K engagements

"Why have metrology and inspection become so critical in advanced nodes [--]. Shrinking margin for error:In advanced process nodes the alignment requirements between layers are extraordinarily tight. Even a slight deviation in metrology can cause interconnects to either break or short directly impacting device reliability and overall yield. [--]. Increasing structural complexityAs transistor architectures transition to FinFET or even GAA traditional optical inspection tools are no longer sufficient. Metrology must evolve to effectively see through these highly complex gateallaround structures and"
X Link 2026-01-28T22:00Z 58K followers, [----] engagements

"Over the past three months there has been a market spike in Logic IC production in Japan. Likely coming from JASM (Japan Advanced Semiconductor Manufacturing) this is an important dataset to monitor for $TSM production in Japan. (1/2) 🧵"
X Link 2026-01-29T14:01Z 59.2K followers, [----] engagements

"Want to learn more Sign up here: (2/2) https://semianalysis.com/chipbook/ https://semianalysis.com/chipbook/"
X Link 2026-01-29T14:01Z 58.4K followers, [----] engagements

"Can we please get a fact check @fabknowledge @AnthropicAI"
X Link 2026-01-29T18:00Z 59K followers, 10.1K engagements

"Ever wonder how an Nvidia GB200 NVL72 gets from the factory to the data center You cant just throw a state-of-the-art [----] lb rack (compared to [---] lb for CPU-based server) on a standard FedEx truck. Moving AI infrastructure is more like transporting a heart for transplant than shipping electronics. (1/9) 🧵 https://twitter.com/i/web/status/2016994844246241458 https://twitter.com/i/web/status/2016994844246241458"
X Link 2026-01-29T22:00Z 59K followers, 21.9K engagements

"A fully loaded AI rack can weigh up to [----] lbs. Thats the weight of a Ford Explorer concentrated into a 2x4 foot footprint. Standard pallets would crush instantly. These AI racks require custom-engineered reinforced bases with shock-absorbent foam just to keep them from tearing through the floor. (2/9)"
X Link 2026-01-29T22:00Z 57.7K followers, [----] engagements

"Unlike Ford Explorers that are made for ruggedness and the great outdoors - AI servers are packed with thousands of costly and sensitive components (think accelerator HBM stacks liquid cooling loops etc.) that are hyper-sensitive to any sudden movement. A single hard "thud" can dislodge a single component rendering the rack inoperable and making maintenance necessary. To protect delicate components from damage during transit these racks sit on integrated heavy-duty shock absorbers and specialized shock-mount bases designed to neutralize impact. This setup is paired with custom foam layers for"
X Link 2026-01-29T22:00Z 57.8K followers, [----] engagements

"You wont see these servers in a standard semi-trailer. Shipping AI hardware requires Air-Ride Suspension trucks and not the traditional leaf spring system. Standard cars and trucks transmit every road crack directly to the passengers and cargo. Air-ride trailers "float" on cushions of air ensuring the delicate hardware doesn't vibrate loose during a 500-mile haul. (Time to copyright Air-Jensen) (5/9)"
X Link 2026-01-29T22:00Z 57.7K followers, [----] engagements

"Upon arrival at the data center - The "white-glove" premium-care crew takes over the moment the shipment hits the loading dock initiating a high-stakes maneuvering process for the 'last-mile'. Using precision rigger lifts (like the autonomous one in the video below - credit to @JordanNanos) they navigate the multi-thousand pound units through tight aisles with millimeter accuracy as a single wrong move or a slight tilt at this stage could turn a multi-million dollar asset into an incredibly heavy paperweight. (6/9) https://twitter.com/i/web/status/2016994851670183971"
X Link 2026-01-29T22:00Z 58.2K followers, [----] engagements

"However none of this applies if youre Elon Musk. In the @WalterIsaacson biography he depicts the Christmas eve heist that Musk and his cousins carried out where they cut off the Twitter servers in the Sacramento data center with pocket knives and just loaded the servers into rented U-Hauls to be redeployed in Portland. And yes it worked (granted these were CPU-based servers but still). Heres an excerpt for your entertainment: Musk turned to his security guard and asked to borrow his pocket knife. Using it he was able to lift one of the air vents in the floor which allowed him to pry open the"
X Link 2026-01-29T22:00Z 58.4K followers, [----] engagements

"To summarize as server costs rise alongside the complexities of silicon networking gear and cooling systems the price of a transportation miscalculation - whether through scrap costs or deployment delays - can be staggering. This in turn is transforming one of the worlds most 'boring' industries shipping into a high-stakes science. (8/9) https://twitter.com/i/web/status/2016994855386230908 https://twitter.com/i/web/status/2016994855386230908"
X Link 2026-01-29T22:00Z 58.9K followers, [----] engagements

"And as the doomsday narrative of AI-driven job displacement accelerates it is important to recognize that AI is actually fueling an insatiable demand for industrial roles. Consider the HVAC technicians construction workers electricians and - as this post illustrates - shipping logistics teams who have seen their job complexity and wages rise as a direct result of AI and the AI data center buildout. (9/9)"
X Link 2026-01-29T22:00Z 58.9K followers, [----] engagements

"O(100) tok/s/user interactivity is NOT fast enough. groq/cerebras at O(1000) tok/s/user is not fast enough Claude Code preferably should run at O(10000) tok/s/user interactivity. The AI enlightenment is near and Claude Code is only running at O(100) tok/s/user"
X Link 2026-01-30T02:00Z 59.1K followers, 13.9K engagements

"With investors running into key memory names like SKHynix Samsung and Micron it is vital to track the progress of HBM shipments to watch for early indications of production slowdowns or continued expansions. (1/3) 🧵"
X Link 2026-01-30T14:01Z 59K followers, 38.5K engagements

"Toto has leveraged decades of advanced ceramics expertise to engineer ESCs that maintain tight thermal uniformity minimize particle generation withstand harsh plasmas and process gases and remain ultra-flat across thousands of process cycles. Even small deviations in chuck flatness or resistivity can translate directly into CDU drift and yield loss. This makes ceramic ESC quality is critical and is why the Totos semiconductor business is also its highest margin (2/2) https://twitter.com/i/web/status/2017296973175435374 https://twitter.com/i/web/status/2017296973175435374"
X Link 2026-01-30T18:00Z 59K followers, [----] engagements

"As AI and HPC systems continue to increase rack-level power density per-rack power demand has grown from the traditional [----] kW to 100+ kW. Under conventional AC and 48V DC power architectures this increase results in extremely high distribution currents causing significant IR losses cable and busbar size constraints connector heating and reduced overall efficiency due to multiple power-conversion stages. 3:24 AMAn 800V HVDC architecture addresses these challenges by converting AC power to high-voltage DC at the data-center level and distributing DC power to rack-level units with final"
X Link 2026-01-31T02:00Z 59K followers, 15K engagements

"Backend networking architecture is instrumental to networking ownership cost analysis. How does networking architecture matter for networking ownership costs In general we observe the following trends: 🟠Hyperscalers typically get preferential terms on networking equipment compared to neoclouds; 🟠Infiniband-based networks are more expensive than Ethernet-based networks for the same networking type and cluster size; 🟠3-Layer networks would be more expensive than 2-Layer networks on a per rack server (1 rack server = [--] GPUs in the case of GB200 and GB300) basis; and 🟠All else constant G300"
X Link 2026-01-31T22:00Z 59K followers, 12.3K engagements

"Before we dive into several examples of how architecture influences networking costs we define the following terminology used to describe networking clusters: 🟠Network layers: Number of switch layers required to connect all GPUs within the same cluster typically [--] or [--] layers 🟠Rails: Number of pathways you can split a server tray which is equal to the GPUs per server tray or [--] in the case of GB200 or GB300 deployments 🟠Planes: Number of pathways you can split a NIC 🟠Attach Rate: GPU per Networking Component (or the reverse) (2/6) https://twitter.com/i/web/status/2017719643012559008"
X Link 2026-01-31T22:00Z 58K followers, [----] engagements

"A 36864-GPU Infiniband cluster consisting of [--] rails and [--] layers can support [----] GPUs per QM9700 Switch (25.6T per Switch). For the same cluster size and type assuming that we swap all GB200 chips for GB300 chips the number GPUs that can be supported per QM9700 Switch theoretically halves from GB200 to GB300 because the bandwidth of GB300 is double. (3/6) https://twitter.com/i/web/status/2017719644287590834 https://twitter.com/i/web/status/2017719644287590834"
X Link 2026-01-31T22:00Z 57.9K followers, [----] engagements

"AWS is planning over 1GW and most of it for Anthropic in Melbourne. This is just one part of Anthropic's international strategy to resolve data locality issues and they've even deployed Trainium in Australia Wouldn't you like to know where. 😉 (pic unrelated) (2/5)"
X Link 2026-02-02T18:00Z 58K followers, [----] engagements

"Cambricons rally is not solely sentiment-driven but evidence is still early: chip revenue has risen with Siyuan [---] and there are signs of demand for Siyuan [---]. Some engineers report an easier software experience than Huawei Ascend and certain Chinese hyperscalers that compete with Huawei may prefer a neutral supplier. That said Cambricon appears to have less ability to secure SMIC N+2 wafers and long-run HBM supply which could constrain scale. ByteDance is also Cambricons largest customer accounting for 80% of Cambricon's revenue highlighting concentration risk. 2/3"
X Link 2025-09-30T18:24Z 60.7K followers, 12.4K engagements

"How mofos sound pitching their custom inference engine to VCs:"
X Link 2026-01-22T13:00Z 60.7K followers, 25.3K engagements

"As Clive [--] 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"
X Link 2025-07-24T21:17Z 60.7K followers, 44.5K engagements

"Co-Packaged Optics (CPO) Book Scaling with Light for the Next Wave of Interconnect Scale-out and Scale-up CPO CPO TCO and Power Budgets DSP Transceivers vs LPO vs NPO vs CPO TSMC COUPE MZM vs MRM vs EAM Modulator Deep Dive CPO Focused companies and CPO Supply Chain https://newsletter.semianalysis.com/p/co-packaged-optics-cpo-book-scaling https://newsletter.semianalysis.com/p/co-packaged-optics-cpo-book-scaling"
X Link 2026-01-01T22:03Z 60.7K followers, 51.6K engagements

Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing