Investments and Portfolios

Breaking news-more to follow. OpenAI sign a deal today with AVGO/Broadcomm for the design of a new Aspics accelerator. The size of the deal is circa 10GW which would be at least $150B. AVGO is up about $40

I follow Dylan Patel, founder of Semi Analysis. Very knowledgeable on the tech side, not the 'stock' side but for me, it compliments the knowledge and allows me to separate the hype from the reality . I know he met Chuck Liang recently to discuss their plans-he thinks something important is about to drop. Why Supermicro (SMCI) Gets the Spotlight in Dylan’s Tease—And Not Dell, Foxconn, or Wiwynn SM as a key supporter of his imminent “huge” framework on AI chips, inference, and infrastructure, due to drop by the evening of 9 October. SMCI is listed alongside hyperscalers/cloud players (CoreWeave, Nebius) and hardware/infrastructure specialists (Crusoe, HPE, Tensorwave), highlighting its pivotal role in the rack/server layer for optimised inference stacks. Notably absent are Dell, Foxconn, and Wiwynn (Wistron’s AI-focused ODM arm), despite their prominence in AI server markets. This isn’t arbitrary; it reflects SMCI’s unique position as the agile, high-density leader for the “neo-cloud” era,(think IREN) tailored to inference’s bursty(yes bursty-data that can burst from 1X to 100X in a nano second,) power-intensive demands. Here’s why SMCI gets the call-out, grounded in Patel’s reports, posts, and industry context:1. SMCI’s Edge: Speed, Customisation, and Hyperscaler Fit for InferenceRapid Prototyping and Deployment: SMCI excels at “just-in-time” manufacturing, delivering 100,000+ servers in weeks rather than months. This is critical for inference, where hyperscalers like CoreWeave (an SMCI client) demand swift iterations on hybrid NVIDIA/AMD setups to manage variable query loads. Patel’s September 2025 SemiAnalysis report on rack architecture (co-authored by him) praises SMCI’s modular designs for disaggregated PDUs and liquid cooling, enabling 250kW+ racks with a 30% better total cost of ownership (TCO) compared to rigid ODM builds. Foxconn and Wiwynn, as pure ODMs, prioritise volume for branded OEMs (e.g., Dell’s enterprise kits) but lag in bespoke hyperscaler customisation. Direct Hyperscaler Relationships: SMCI sells directly to neo-clouds (CoreWeave, xAI’s Colossus—partially SMCI-supplied) and AI labs (OpenAI’s AMD pivot), bypassing intermediaries. In his August 2025 “No Priors” podcast, Patel highlights SMCI’s vertical integration (from motherboard design to cooling), giving them a two-year lead on liquid-cooled hybrids, essential for inference’s 80%+ energy draw. Dell shines in enterprise/sovereign AI (per Patel’s May 2025 “How Dell Is Beating Supermicro” report), but their slower cycles, optimised for HPC stability, don’t match neo-cloud urgency. Inference-Specific Advantages: Backed by vLLM/SGLang (inference engines-sorry tech heavy!), the framework likely benchmarks rack-level metrics like tokens/second per watt. SMCI’s 8U/10U GPU trays (e.g., SYS-821GE with 8x Blackwell) blend NVIDIA prefill compute with AMD decode efficiency, reducing latency by 20-50%. Foxconn/Wiwynn handle high-volume NVIDIA HGX for cloud giants (e.g., Foxconn’s Oracle Stargate supply), but Patel’s critiques (e.g., 2023 posts on Foxconn’s public “reveals” being overhyped) underscore their ODM commoditisation—cheaper but less innovative for multi-vendor inference. Historical Context from Patel: SMCI as the “Crusher” in AI RacksPatel’s analyses consistently position SMCI as a leader in frontier AI infrastructure. A 12 September 2025 X post details his Supermicro factory tour with CEO Charles Liang, showcasing GB300/B300 and MI355X racks—directly tying to the tease’s NVIDIA/AMD backers. In contrast, his May 2024 report praises Dell for enterprise wins (e.g., Tesla, CoreWeave orders) but notes SMCI’s resurgence in neo-clouds via cheaper, denser cooling (e.g., April 2023 post: “I’m such an idiot for not going turbo long SuperMicro... they crush Dell and crew while being much cheaper”). Final thought- SMCI’s Unique PositionDylan’s call-out of SMCI reflects their role as the inference infrastructure “backbone” for collaborative, multi-vendor stacks, validated by factory tours, reports, and backers like HPE (also listed). Dell, Foxconn, and Wiwynn play critical roles in enterprise or ODM volume but lack SMCI’s hyperscaler agility and hybrid rack innovation for inference. If the full drop (expected ~evening 9 October BST) includes rack BOMs or benchmarks, SMCI’s prominence will likely grow.

I thought it would be that, but it was worth asking, just in case. Many Thanks

Very impressive if it is real, but I'm skeptical. Seems I was justified - click the three dots for the description. Still, it's a mighty impressive render. [image: 1758571262139-a7c0ef52-d4b8-453d-8270-81742c9881f3-image.png]

No fees other than the 0.35% annual management fee, applied monthly

Thanks for taking the time to reply Adam << thumbs up thingy>>

Thanks Adam

From JP Morgan: AMZN has grown its Advertising Services revenue from ~$13B in 2019 to ~$56B in 2024, implying a +39% CAGR. The success of Amazon’s Advertising Services business highlights key advantages AMZN & retailers have over GOOGL & META. According to WSJ, AMZN has deployed more than one million robots in facilities, which is the most it has ever had and near the count of human workers at the facilities.

Colossus-2 expansion to 1 million GPUs gets closer....... xAI is raising $25 billion, comprising $5 billion in debt through Morgan Stanley (a term loan B, fixed-rate term loan, and senior secured notes) to fund the Colossus supercomputer with commitments due by June 17, 2025. Additionally, xAI is seeking $20 billion in equity at a valuation between $120 billion and $200 billion to support its growth.

Interesting update which is a big positive for companies like Oracle. 'You can have our tech if we manage it' -seems to be a link to Stagate/open AI for countries. Quoted from a news wire below New AI Diffusion Rule is COMING SOON and New Version Will Let Allies Buy US Chips With Conditions The Commerce Department is drafting a replacement for its recently repealed AI diffusion rule to ensure the new controls don’t impede U.S. exports to allies, Commerce Secretary Howard Lutnick said June 4. “Our view is we are going to allow our allies to buy AI chips provided they’re run by an approved American data center operator and the cloud that touches that data center is an approved American operator, so we control it while it’s over there,” Lutnick testified before the Senate Appropriations Subcommittee on Commerce, Justice and Science. The rule that was rescinded last month (see 2505130018) was “very confusing,” Lutnick said. “For example, the prime minister of Poland hunted me down and said, ‘What did I ever do to you that you have me as a Tier 3’ country, the most stringent of three tiers? ‘I’m part of Europe -- what are you doing?’ It was illogical. It was hastily rushed through at the very end of the Biden administration.” He expects Commerce will release the new AI diffusion rule soon. “I can’t say with more specificity but pretty soon,” he said.

The previously posted tables containing predicted min Blackwell (packaged) chips has proven very accurate in fact we also said TSMC would likely find additional capacity(and apparently they have). And from that we forecast the revenue growth, knowing what they sell these chips for. The start of Blackwell was clunky due to the transition from Hopper because Hopper was packaged on the 'S' line and Blackwell on the 'L' line and there were some heating issues which is no surprise because it was all very new. We predicted about $10 billion Q o Q growth vs $4-5B during Hoppers reign. We are now at this point today. Blackwell is bug free and at ramping as fast as CoWoS will allow. This is the old schedule by quarter in 000's: [image: 1748798151117-screenshot-2025-06-01-at-18.14.59.png] . TSMC have indicated circa 500k chips per month and growing from June. They are all sold(5-10X). Losing China completely, and it wont be but let's just ignore it all together. All that has happened is we take the hit and next quarter they achieve $47B and the next $57, $67, $77 and so on. The only caveat is that when Rubin arrives at year end, Dec/Jan they will add further revenue to each quarter simply due to the ASP of Rubin being higher-so id expect $12-$13B q o q increases. Next year automotive will be meaningful as all car makers equip their cars with ADAS and robotics/omniverse will start adding revenues-the q o q rhythm will grow again. But this is exactly how I see their revenue grow over time. A very long time and yes I expect 100B+ per quarter some time at the end of next year or Q1 the following. And to anyone who thinks they will have transition issues again, well, Rubin and Rubin Ultra are all packaged on the same line so the transition will be seamless . It is not until 2027 that packaging will move to a new process called Sow-X which is when the real party starts. We are looking at racks 40X more powerful than today's. Racks containing 500 chips and consuming 1 megawatt each and could cost $20M or more. This is the roadmap from 2027 to 29. The experts have been predicting a plateau in revenue for over a year now-remember Cathy Wood and the 'Dean of Valuation'. They were all wrong and continue to be so. Cisco-look at Cisco. A very scientific analogy. Exciting times ahead imo.

The stock doesn’t fit out criteria and moving 100% in a day doesn’t make it attractive. Odds are it will fade over the coming weeks. It’s very small, no track record, loses money and will be extremely volatile. And to add something you have to make a call on what to sell. That’s not to say it can’t go higher.

Morning, Air time is driven by: Weight or our position vs the market position. We were contrarians with SMCI (still are). SMCI drives a lot of opinion! The significance of the company vs the sector Companies driving change/innovation (driving news) My personal interests-and what I think others find interesting-esoteric sectors Cobens has 20k clients and we managed 3X more assets than IM before the rescue I appreciate that we talk about Tech a lot and a couple of holdings in particular. The fact is Nvidia drives/is a barometer of the wider sector(and the market) and what is good/bad for Nvidia impacts other big names.

Hi D, Netflix is set to drop its Q1 2025 earnings after the market shuts on 17 April. Analysts expect earnings per share (EPS) will hit $5.74, up 8.7% from $5.28 last year, with revenue expected at $10.54 billion, a 12.5% jump year-on-year. That’s pretty close to Netflix’s own forecast of $10.46 billion in revenue and an operating margin of 28.2%. They’ve got a history of smashing EPS forecasts, beating expectations in all four of the last reported quarters, like Q4 2024 when they pulled off $4.27 against $4.20 expected. That said, it's hard to know but their numbers should be solid. Everyone’s eyeing their ad-supported memberships, which made up 55% of sign-ups in ad markets last quarter, and their content line-up, with several content hits. They won’t disclose subscriber numbers anymore, focusing on revenue and margins instead.

Here is a complete list of ALL notable Robotics players. See the common thread? [image: 1744622289650-screenshot-2025-04-14-at-10.15.51.png]

SK hynix raises its 2025 CAPEX from $15 billion to $19 billion. SK Hynix has decided to increase its planned CAPEX this year by 30% due to the surge in demand for high-bandwidth memory (HBM), which was already high. The chipmaker initially planned to spend 22 trillion won this year in expanding its facilities but this has been upped to 29 trillion won, sources said. The decision was recently finalised and SK Hynix has also sent memos to suppliers to deliver the equipment to the M15X fab in Chungju by October, two months faster than initially planned. These moves are all responses from its customers requesting SK Hynix deliver more HBM chips faster. The company’s main customer is Nvidia, which is using HBM with its AI accelerators. The GPU giant is requesting SK Hynix deliver HBM faster than planned. SK Hynix will also be supplying HBMs to Broadcom starting this year. We are reading the room, folks. Demand is off the hook and I would speculate, supply is also improving. Whilst the market is distracted by and shunning everything due to tariffs speculation, these businesses are clearly articulating a shift.

Here is the compute comparison. One rack = the entire DC. The DC cost 3b usd to build and 150m annually to operate excluding the land/lease The single rack costs 30m and 12m to operate annually. Could live in a broom closet [image: 1742381904418-screen-shot-2025-03-19-at-10.52.08.png]

We will be making two changes to the portfolios next week(2 new stocks)-not sure of exact timing. Treat this as aheads up of an impending rebalance across Lifestyle/Tech