Today is October 1 and we have a public holiday in Hong Kong. Hope everyone has a great month ahead! More importantly, this month, Intel will release its earnings on 22. $INTC Beat and raise please BTW, I don't post the way I used to anymore Somehow marketing materiel, or I prefer to call it, a tool The things that banks give them away to attract trading flow and corporate finance business Somehow they will want to make your post be taken down. I don’t post as much as other people, and now I have to generate a chart in order to post, I think I will post even lesser I simply don’t want to get banned because As the saying goes The real treasure was the friends we made along the way Thank you everyone for following I ❤️ Intel and I ❤️ Intel gang and I ❤️ 𝕏

You are probably looking at the wrong exhibit, exhibit 12 is not really that important, 116% isn’t really that surprising because Exhibit 13, on the other hand 2024: 19% → 54% actual
2025: 22% → 73% actual
2026: 36% → 96% actual Three years. Same mistake. Same direction. Three consecutive years of underestimating the same trend by a lot. The biggest surprise is that people are still surprised by it. What's even more surprising is that an AI company missed it.

Morgan Stanley has a more detailed estimations on the cost of always on AI agents Light users mainly consume CPU resources. Heavy users increasingly become a GPU inference problem. Takeaway: If Muse becomes wildly successful, does Meta have enough CPU capacity to support it? Yes. Meta's 2026 CPU purchases alone could theoretically support ~1 billion daily active Muse users. They will build or rent about 600m vCPUs, mostly from AMD. Why don't they need 1 billion instances for 1 billion users? Each user's agent is only active about 3 hours out of 24 hours per day Therefore only 12.5% are actively consuming CPU at a given moment. So, 1 billion × 12.5% = 125 million concurrently active users However, usage isn't flat. So, about 300M instances needed Source: Morgan Stanley $META Platforms Inc | North America 23 Days of Musing...Top Investor Questions and Our Answers

Agents are expensive because they need a lot more inference But in fact, most of the cost is NOT inference. It's keeping an agent alive AI agents need A LOT OF CPUs and A LOT OF memory to get things done!!!

By 2028: HBM becomes one of the largest profit pools in memory. AI servers consume a growing percentage of industry DRAM capacity. Conventional DRAM gets crowded out because more wafers are diverted toward HBM production. $MU $SKHY

Memory supply expected to tighten into 2028E $MU $SKHY $SNDK

MU already released earnings?
Why is the price not moving? $MU

Bank of America Top picks into Q4 $NVDA : Enhanced buybacks + multiple GTC events keep the AI narrative alive. $INTC : Agentic AI could be a CPU story as much as a GPU story. Add foundry win anticipation and the setup gets interesting. $MU : Buybacks starting Dec. 9 signal confidence as memory fundamentals remain strong. $MRVL : Analyst Day on Oct. 6 + XPU ramp could be key catalysts. $LRCX : Positioned for potential share gains across both memory and logic spending cycles.

Do we have enough electricity and data center capacity to support the AI boom through 2030? Maybe. We don't fully have it today, but we can plausibly get there.

SpaceX is no longer just building Grok. It is a hyperscale, leasing compute to other companies. By the time all planned phases are online Colossus I + Colossus II could reach roughly 1.44 million GPUs and 2.71 GW of power capacity. $54.5B annualized run rate from only five customers $SPCX Source: Deutsche Bank Got compute?

Current cycle peak DRAM margin: 91% CY2028 normalized DRAM margin: high-70s% Both the normalized DRAM and NAND margins remain above every prior cycle peak. Source: Bernstein Global Memory: Cycle is still elevated, but beware of the long-term threat from China EXHIBIT 14 and 15

Most underappreciated AI theme? Agentic AI's impact on CPUs. As AI agents take on orchestration, memory management, and tool use, CPU demand should rise alongside GPUs. CPU/GPU ratios could move from roughly 1:8 toward 1:1. Top 5 stocks in JPMorgan CPU ecosystem basket: $ARM, $INTC, $AMD, $MRVL, and $TSM Source: JP Morgan Equity Strategy Tech de-rating became advanced in a number of areas

TSMC may spend over $100B in capex by 2028. That's not a company preparing for slower demand. It's a company preparing for a lot more chips, a lot more packaging, and a lot more AI. Just like Intel $TSM $INTC

Not cheap but definitely not expensive The median stock at 18x, while the aggregate index is 19x. That gap is actually quite small.

The problem was never 5% yield itself The problem was the vulnerable market This time is different Earning is good enough to support the valuation!