NVDA · NVIDIA CorporationNVIDIA Corporation
🔬 NVIDIA is shifting from selling chips to selling complete AI factory solutions, with Vera Rubin now in full production and Agent AI becoming mainstream, driving massive demand for AI infrastructure
NVIDIA is shifting from selling chips to selling complete AI factory solutions, with Vera Rubin now in full production and Agent AI becoming mainstream, driving massive demand for AI infrastructure.
Evidence & details
- +SemiAnalysis reported NVIDIA halves CPU SOCAMM DRAM to 96GB modules in Rubin NVL72, cutting total DRAM from 55TB to ~28TB per rack.
- +The reduction is due to supply bottleneck of 16-layer high-density LPDDR5X SOCAMM from Hynix/Samsung.
- +Rack BOM cost drops from $7.6M to $6.8M, saving ~10%.
- ⚡Faster Rubin deployment and higher volume production capacity.
- ⚡Improved TCO for hyperscalers like MSFT, encouraging larger orders.
- ⚡Potential increased demand for SSD and optical interconnect as workload shifts from CPU DRAM.
- −Reduced CPU DRAM may limit performance for very long context inference, multi-agent parallelism, and large KV cache, requiring more cross-node communication or SSD offload.
Chart
Post Timeline · 7 posts
英伟达最新的GPU少用Dram是怎么回事？昨晚SemiAnalysis这个报道说英伟达主动将下一代AI超级计算机平台——Rubin NVL72机架的CPU侧SOCAMM DRAM容量砍半。消息一出，整个存储板块大跌。详细聊聊： 1、这到底在说什么？ NVIDIA 2026年主力AI机架平台，一柜装72颗Rubin GPU + 36颗Vera CPU，通过NVLink全互联，形成“一块巨型GPU”。GPU侧用HBM4（每柜约20.7TB），CPU侧用SOCAMM——一种可插拔的LPDDR5X系统内存模块，便于维护和故障隔离。 原计划：每颗Vera CPU…
Thesis: NVIDIA actively reduces CPU-side SOCAMM DRAM in Rubin NVL72 to mitigate LPDDR5X supply constraints, lowering cost and accelerating deployment, but may impact memory-intensive workloads.
SemiAnalysis: NVIDIA is reducing the SOCAMM capacity for Rubin. via @TMTBreakout https://t.co/xe6QLgMKMN
Thesis: NVIDIA is reducing SOCAMM capacity for Rubin, potentially indicating lower demand or strategic shift.
Highly bullish implications today for Shunsin (6451). Which is Foxconn’s SiPH and CPO packaging/test arm. -> UDN reported upward revision in volume and Foxconn began shipping cpo switch racks to $NVDA -> CPO products are being produced in Foxconn's Vietnam plant… Guess whic…
Thesis: Shunsin is Foxconn's SiPH and CPO packaging/test arm, now benefiting from CPO volume ramp as Foxconn ships CPO switch racks to NVDA and expands Vietnam facility, indicating start of massive growth.
Kevin Warsh is the next Federal Reserve Chair. Markets may confuse him as a "Hawk". His actual stance in 2026 is nuanced. Here's his policies and how they affect the markets: 1. AI/Semis ( $NVDA, $MU): Extremely Bullish 2. Metals (Silver, Gold): Extreme Bearish 3. Crypto ( …
Thesis: Kevin Warsh as Fed Chair is expected to be bullish for AI/semis (NVDA, MU), crypto (BTC, CRCL), banking (JPM, BOA), small-caps (RUT), and cautious for European stocks (VGK, EZU), based on his pro-AI, …
Thinking about buying $CRWV on the $NVDA deal? Here's the way to think about it: A casino gives you more money, so you could take on more debt with collateral to gamble. Then all the extra money goes back to the Casino (aka. $NVDA). That doesn't make Coreweave a good long. …
Thesis: CoreWeave's deal with Nvidia is analogous to a casino giving more money to gamble, leading to increased debt that ultimately benefits Nvidia. CoreWeave is not a good long due to high debt and interest…
@EarningsMindset Yeah, the topic was about how the TPU v7 Ironwood is the first competitive alternative to $NVDA and performs better for specific inference workloads. The current $NVDA backlog is a security guarantee for medium term. It's just a question of whether TPUs continue…
Thesis: TPU v7 Ironwood is a competitive alternative to NVDA GPUs for inference, but next-gen NVDA likely maintains lead; bear case if TPUs outperform, demand shifts, and AI compute demand lessens.
@LoganMacGregor9 Good question about risk, I enjoy talking about the downsides too. 1. Hyperscalers completely their own datacenter buildout with GPUs (TPUs with $GOOGL, Tranium with $AWS), energy, locations in 5+ years. This is the analogy of Qualcomm with Apple. Apple just u…
Thesis: Lists four main risks for NBIS and other AI infrastructure companies: hyperscaler vertical integration, GPU depreciation, credit tightening/dilution, and overbuildout due to software efficiency breakt…
Market liquidation cascades from silver crash and hawkish Fed nomination fears, but overall fundamentals remain bullish with expected rate cuts and AI growth. Recommend repositioning to FCF-generating names and being cautious on speculative high-beta names.
source ↗Hyperscaler AI capex is increasing dramatically, benefiting AI supply chain companies, especially ASICs, memory, and key enablers like TSM and NVDA. Hidden beneficiaries like AAOI, POET, MRVL, LITE also benefit.
source ↗Tesla Optimus bottlenecks identified: Real World Intelligence (NVDA), The Hand (possibly VPG), and Scale Manufacturing (Chinese suppliers).
source ↗Hyperscalers going into debt for AI capex benefits chip suppliers like NVDA, TSM, AVGO as they transfer balance sheets to them.
source ↗The 1 million job revision signals a permanent structural shift due to AI/automation, not a slowing economy. Corporate profits will explode as companies replace human labor with AI and robotics. Investors should buy AI supply chain equities to hedge against unemployment and benefit from margin expansion.
source ↗Trump administration proposal for AI chip export permits is unlikely to pass; current selloff is overreaction.
source ↗US Commerce Department denies return to AI diffusion rule, countering negative reporting that caused sell-off in semis, supporting NVDA and AMD.
source ↗Portfolio designed to profit from a doomsday scenario involving US ground troops in Iran and Iranian retaliation, with specific positions betting on financial crisis, oil spike, tech crash, Lucid collapse, volatility, and Nvidia decline.
source ↗Nvidia at $4.1T market cap faces downside risk from liquidity crunch and energy crisis exacerbated by potential Iran escalation, with Israel maintaining conflict stance.
source ↗Nvidia prefers SIVE independent, but potential share acquisition by MRVL or acquisition by AVGO to choke competitors' CPO roadmap.
source ↗NVIDIA is aggressively advancing its CPO (co-packaged optics) architectural roadmap and securing major industry players into its standards, positioning itself to lead in photonics.
source ↗Speculation that NVDA buying a stake in Win Semi would be bullish due to Win Semi's history of attracting strategic investments like from AVGO.
source ↗LITE sold out into 2029 with CPO mass production confirmed for H2 2026; NVDA upstream supply chain for CPO/photonics benefits.
source ↗CPU supply shortage driving price hikes and capex increases, benefiting CPU makers and foundry.
source ↗Nvidia's VR200 product delayed by two months due to cooling challenges, indicating potential near-term headwinds for the company.
source ↗Nvidia's upcoming AI PC (N1/N1X) based on Arm SoC with Blackwell GPU will drive agentic AI era, benefiting Nvidia, Arm, and Microsoft through ecosystem dominance and new revenue streams.
source ↗Supply chain surveys predict 10 million NVIDIA N1X device shipments in 2 years, but the author believes this is too optimistic as AI PC demand is fake and offers no value to consumers.
source ↗If software companies' tools are used in agentic AI, they will race to secure compute capacity, benefiting Nvidia.
source ↗NVIDIA is shifting from selling chips to selling complete AI factory solutions, with Vera Rubin now in full production and Agent AI becoming mainstream, driving massive demand for AI infrastructure.
source ↗