ye

$AMD CPUs are superior to all $ARM based CPUs 🧵 Whether it is $NVDA $AMZN $MSFT $GOOGL or $ARM Not Financial Advice! DYOR! Educational Purpose! Analysts and research firms frequently overstate ARM-based server CPUs (AWS Graviton, Google Axion, Microsoft Cobalt, NVIDIA Grace/Vera, AmpereOne...) as superior or inevitable winners over AMD EPYC due to power-efficiency narratives, hyperscaler customization hype, and growth projections. In reality, flagship EPYC Venice (Zen 6) and EPYC 8005 series deliver superior real-world performance, ecosystem advantages, bandwidth, and total value across most workloads making broad ARM superiority claims misleading or outright wrong. I will link different threads relating to this, where you can understand the full picture from supply chain to TSMC capacity expansion. The reposted thread will have a very comprehensive comp on $AMD CPUs all others. 1. EPYC Venice vs $ARM designs Venice entered production ramp on TSMC 2nm in 2026 as the first and only high-performance computing product on that node. It targets leadership in AI factories, agentic inference orchestration, cloud, and HPC. EPYC Venice: Up to 256 Zen 6 cores / 512 threads (33%+ increase over Turin’s 192 cores). ~70% performance uplift (compute) and efficiency gains over 5th Gen Turin. Memory: Up to 16 channels DDR5, delivering 1.6 TB/s per-socket bandwidth (more than double prior gens). I/O: PCIe 6.0 support with significantly increased lanes (doubled CPU-to-GPU bandwidth), massive L3 cache (potentially up to ~1 GB in high-end configs via dense chiplets). New SP7 socket, power envelopes scaling high (hundreds of watts to 1.4 kW peaks in dense setups) for maximum throughput Why Venice Outclasses ARM: Raw Throughput & Features: ARM designs ( Grace with 144 Arm cores, or Vera successors) rely on high core counts but lag in per-core performance, SMT (most lack it), AVX-512-like vector capabilities, and massive I/O/bandwidth. Venice’s Zen 6 architecture + 2nm process + extreme memory bandwidth excels in bandwidth-hungry workloads like Agentic AI orchestration, databases, HPC, and large-model inference where CPUs pair with GPUs at shifting 3-5:1 ratios. Benchmark, Current-gen EPYC (Turin) already beats Grace Superchip by 2–3.7x in key workloads (SPEC, Java, databases) with better efficiency in balanced tests. Venice’s 70% leap widens this gap dramatically. Full x86 compatibility runs trillions of lines of legacy/enterprise code natively. ARM requires porting, dual maintenance, and suffers fragmentation fatal for hybrid cloud/on-prem/enterprise deployments. Hyperscaler ARM wins only in their locked-in, greenfield scale-out niches (web serving, simple microservices). Analysts projecting ARM to 40-50%+ server share are just wrong and often ignore these realities, focusing on power walls and custom silicon savings while downplaying software costs and performance shortfalls. 2. EPYC 8005 Series (Zen 5, Edge/Telco/Storage Optimized) vs. ARM The 8005 is a single-socket SP6 design for power-constrained, dense deployments not a direct flagship rival, yet it still outperforms many ARM chips in targeted use cases. EPYC 8005: ~Up to 84 Zen 5 cores / 168 threads. ~TDP: 70–225W (e.g., 8635P at 225W). ~Memory: 6-channel DDR5-6400, up to 3 TB. ~I/O: 96 PCIe Gen5 lanes. ~Strong perf/watt gains over prior gen. Why 8005 woms pver ARM-based $NVDA Grace: ~Perf/Watt: 24,408 ssj_ops/watt vs. Grace’s 13,218 (~85% better). ~Effective Performance: SMT advantage, higher per-core throughput, AVX-512, and x86 compatibility deliver better real output in vRAN, storage (Ceph), edge AI, and telco despite lower headline core count. ~Power Envelope: 225W vs. Grace’s 500W where it matters for dense/edge racks. Analysts makes many claims that ARM wins in ultra-low-power hyperscale web, but EPYC 8005 provides balanced, deployable superiority without porting headaches. 3. Average Selling Prices (ASPs) and Pricing Dynamics AMD EPYC Venice/Turin maintains significantly higher ASPs than most ARM offerings, reflecting premium positioning on performance, features, and broad compatibility while still delivering superior perf/$ in independent tests. However, EPYC 8005 list pricing starts at $529 (8-core 8025P) up to $5,799 for the flagship 84-core 8635P (1K unit). This positions it much more competitively for edge/telco/storage while commanding premiums over lower-end ARM for equivalent capabilities. Custom hyperscaler ARM (Graviton, Axion, Cobalt) claims to benefit from internal vertical integration with lower effective costs and cloud instance pricing often 20-40% below comparable x86, but looking at Token bills, the claim does not match. Merchant ARM can be similar-priced per core but lacks AMD’s per-core performance and ecosystem. NVIDIA Grace systems command high system-level pricing (tens of thousands in superchips), but per-CPU economics favor hyperscalers’ captive designs. Analysts projecting ARM revenue share gains often cite lower ASPs as a volume driver, yet this understates AMD’s higher realized prices and better TCO in non-optimized workloads So the question is, why Analysts often try to shape this false narrative? Power efficiency and "custom silicon" stories dominate because data centers face energy constraints. Hyperscalers claim internal savings (20-40% in optimized cases), but these don’t translate universally. Independent tests show x86 (especially AMD) often wins on perf/$, TCO, and throughput when factoring runtime and ecosystem. ARM server share grew because of claims, not actual benchmarks and remains ~15-25%; x86 (AMD gaining to ~40%+ server revenue share) retains dominance due to compatibility. Projections assume seamless ARM adoption that ignores software. Conclusion: Hyperscalers heavily market their custom ARM CPUs (plus ASICs like Trainium/TPU) for 20-50%+ lower TCO and token costs in inference, citing vertical integration, power efficiency, and scale. AWS, for example, promotes Graviton + Trainium for superior price-performance, with Anthropic committing billions to these for Claude workloads. But the disconnect with Enterprise Bills: Enterprises report exploding AI spend (often 85%+ of budgets on inference) despite per-token price drops. Claude Enterprise starts ~$20/seat/month + full API token billing (no included usage cushion for large orgs), with real per-user costs often $60–$250+/month driven by agentic/multi-turn usage. Total bills for mid/large deployments easily hit to millions annually to tens of millions depending on the size of the enterprises/projects. Hyperscalers capture efficiency gains as higher margins (Anthropic inference margins reportedly rose to ~70%). Token prices fall (40-70% in some cases), but consumption explodes with agentic workflows, outpacing unit cost reductions. Providers subsidize entry but monetize at scale via usage metering, features, and lock-in; AKA Enterprises never get this saving. ARM/custom silicon shines in hyperscaler-optimized, scale-out inference but underperforms in general orchestration, hybrid environments, or high-bandwidth tasks where EPYC Venice/8005 excels. Enterprises on Anthropic (often via AWS or Google) pay premium API rates that embed provider margins rather than raw silicon TCO savings. AMD shows Venice/Verano racks driving token costs to $0.0001–$0.0002/M (or lower), with rack-scale advantages in agentic AI (2.37x+ throughput vs. ARM baselines today, higher for Venice). On-prem/hybrid EPYC deployments avoid cloud markups and porting overhead for better end-to-end TCO long term. Flagship EPYC Venice represents architectural and process leadership that ARM designs cannot broadly compete with. Even EPYC 8005 outperforms in many segments at competitive value. Hyperscaler TCO claims sound compelling but fail to translate to controlled enterprise costs (as seen in Anthropic bills), where usage growth and margins dominate. Analysts over-index on hype while underweighting compatibility, bandwidth, premium ASP positioning, and real TCO. For most enterprises and agentic AI, AMD EPYC is the superior, practical choice. Workload-specific testing is key, the blanket "ARM is better" narrative does not hold. Not Financial Advice! DYOR! Educational Purpose! Clip source: Q1 2026 Earning call.

I did say $MU looked like the next $NVDA. Now we're at a $1.23T MC. Started talking more about Samsung Electronics/Sk Hynix back in 2025. Put more concentration into the memory theme like $SNDK and others, Jan of this year. And I'm glad my prediction with Micron + memory is playing out well! Hope people had fun with $EWY longs too, those are up a lot.

Gonna be pain this week.

GUESS WHAT ANON? After today’s new news with Ayar joining Nvidia NVLink fusion. $SIVE is now the laser source for likely: The entire Nvidia’s NVLink CPO listed supply chain ecosystem partners. From Marvell Celestial, Lightmatter, and now Ayar today (the three listed in NVLink CPO). This is why I call $SIVE a structural photonics laser chokepoint over CPO and now Nvidia ecosystem supply chains. -> Celestial was likely a direct customer to Sivers, not through Poet. (2023 investor presentation mapping), then bought by Marvell. -> Lightmatter was also listed there as a customer in 2023 investor presentation deck mapping. And… Guess what else? Then they all happen to use GlobalFoundries. Which Sivers is now the GFS silicon photonics foundry-level reference laser (also new news yesterday). Supply chain mapping all starting to make sense now anon? Sivers is also likely now the primary laser source for Ayar after they removed Macom/Lumentum their laser supply chain section (now just gfs/sivers), as a cherry on top. Algorithms completely miss this type of image based mapping. After this announcement, I personally think current valuations are very undervalued: Given Sivers now holds one of the most important structural laser chokepoint over Nvidia CPO NVLink ecosystem supply chains.

最新的情况是自从Anson Advisors退场之后，另外俩家的仓位变化有了一些更新： 1. Two sigma 他们虽然持仓来到了 2.38% 比上次公布的 2.36%高，但其实他们是已经减仓了的，为什么呢？ 根据数据显示 他们在27日到28日先持续加仓，5月28日财报前夕小幅减仓，但整体仓位仍处高位。 所以他们的数据变化从2.10%（5/22）→ 2.22% → 2.36% → 2.42% → 2.38% 从盘末情况，他们是减仓了一点，但还是有2.38% 所以他们的亏损来到了 2. Voleon 保持和上次一样的水平 虽然 $sive的价格因为财报有明显的回落但是俩家机构的亏损仍然有着近6千万美金 我们也需要注意Anson 平仓（~$800万）和 Qube 平仓（~$3000万）带来的涨幅相近，说明边际买盘在萎缩，散户筹码越来越集中，卖方力量在减弱，这种结构在逼空行情末期非常典型，但所有的前提是多方齐心协力，而不是争相的卖出

Bought some $SIVE here at 75 SEK/share. Its a completely overvalued company but my bet is that US gamblers will skyrocket the share price today due to the news about Global Foundries partnership/collaboration. Its up 24% right now, could as well continue up 35-40% today. This stock has zero connection to fundamentals, its a play on the human psychology.