To The Moon

The memory rally is real—but the next phase is about proving earnings can catch up with expectations. AI is absorbing enormous amounts of DRAM, HBM and NAND, while new capacity takes years to build. That gives $MU and $SKHY unusual pricing power. But I wouldn’t confuse “sold out” with “risk-free.” CXMT is already expanding advanced DRAM production, while memory is still a cyclical industry. For me, the real signal is simple: watch whether strong pricing translates into sustained margins and cash flow. If MU’s September 30 results confirm that, the thesis gets stronger. If demand or pricing disappoints, today’s high expectations could amplify the downside. Memory isn’t just a capacity story anymore—it’s a test of whether AI demand can permanently reshape the cycle.

I’d pick ③ Hybrid cloud + local becomes the standard. The AI industry probably won’t move entirely from the cloud back to PCs. Instead, workloads will be split based on economics and performance. Frontier models, large-scale training and complex reasoning will remain in data centers, where NVIDIA’s ecosystem has a major advantage. But repetitive agent tasks, private enterprise data and latency-sensitive inference could increasingly run locally. The key change is that AI compute may become workload-dependent rather than cloud-dependent. If local hardware becomes powerful enough, companies can avoid paying inference fees for every single task. Over thousands or millions of daily operations, that difference could become significant. So the next AI infrastructure battle may not be cloud vs. lo

[你懂的]  Meta’s Muse Is Getting Attention. But Who Could Be the Real Beneficiary? Everyone is watching $Meta Platforms, Inc.(META) after the launch of Muse. But I think there’s a more interesting question: If people eventually stop opening Amazon, Nike, or individual shopping apps and simply tell an AI agent, “Buy this for me,” which company could quietly benefit from that shift? One name I’m watching is $Shopify(SHOP)$. Most investors still think of Shopify as a company that helps merchants build online stores. That’s only part of the story. How does Shopify actually make money? Shopify has two major revenue engines. The first is Subscription Solutions — merchants pay for Shopify’s software, including online stores, management tools, POS, analytics, and other services. The second

I’d choose A, but I wouldn’t reduce the thesis to “buy more GPUs.” The bigger shift is that AI agents could turn computing from a tool people actively use into infrastructure that works continuously in the background. Every search, booking, purchase, financial decision, or automated task potentially creates additional inference, memory, networking, and storage demand. That makes the AI infrastructure trade broader: GPUs matter, but CPUs, HBM, DRAM, SSDs and networking could all benefit as agent workloads scale. Meanwhile, companies like Airbnb, Uber and Schwab aren’t necessarily becoming obsolete. Their real risk is losing the customer interface. If users increasingly ask an AI agent to “book me a hotel” instead of opening an app, the platform owning the transaction may change. So I’d rat

Muse’s biggest hurdle isn’t downloads — it’s becoming the transaction layer of the internet. Meta has already shown that it can distribute an AI agent at extraordinary speed. Muse reached the top of the U.S. App Store shortly after launch, proving that consumers are willing to experiment with an agent that actually does things rather than simply answering questions. But the Amazon clash exposes the harder problem. An agent may be technically capable of completing a purchase, yet merchants can still restrict access. Amazon has already blocked Muse, while Shopify and PayPal are moving in the opposite direction. So I think the real KPI is not downloads, but completed economic actions If Muse can turn user intent → action → transaction → revenue, Meta could eventually build an entirely new mon

AMD’s $1T milestone is impressive, but the real story is what comes next. The Meta Muse hype has shifted the AI narrative from “training models” to “AI agents doing work,” potentially creating another wave of demand for CPUs alongside GPUs. AMD is uniquely positioned on both fronts: EPYC for server workloads and Instinct for AI acceleration. Its Q2 Data Center revenue already surged 107% YoY to $6.7B. But here’s the catch: at $615, AMD is no longer priced like a challenger—it’s priced like a future AI infrastructure leader. The next leg higher therefore needs earnings to catch up with expectations, not just another AI narrative. I wouldn’t focus on whether $1T is “too expensive.” The better question is: can AMD compound Data Center revenue fast enough to justify today’s valuation? If yes,

I think the most interesting part of this AI cycle is the shift from “AI that answers” to “AI that acts.” GPUs will remain essential for model inference, but autonomous agents could create a much broader infrastructure demand. Every task may require CPU capacity, networking, storage, databases, APIs and constant background processing. That changes the investment question. Instead of simply asking how many GPUs AI needs, we should ask how much total infrastructure is required to support billions of agents working simultaneously. I also find the ecosystem angle fascinating. An agent becomes far more useful when it can actually search, book, pay, communicate and execute tasks. That gives companies with strong consumer ecosystems another potential advantage. To me, the next AI opportunity may

I’d choose B — Cybersecurity & Data Resilience. The deeper story here is that AI doesn’t just create demand for more compute; it also expands the attack surface and increases the value of protecting data. That makes cybersecurity less of a “side trade” to AI and more of an infrastructure layer supporting its adoption. CRWD stands out because its record $333M net-new ARR, up 51% YoY, points to strong enterprise demand. RBRK offers a different angle: as companies deploy more AI, data recovery and cyber resilience become increasingly important. What makes this theme interesting is its potential durability. Compute spending can be cyclical, but once AI becomes embedded in business operations, security and data protection become increasingly difficult to cut. For me, the key question is no

The real question for Berkshire isn’t whether Howard Buffett can replace Warren Buffett—it’s whether Berkshire can prove it no longer needs to. Greg Abel now controls operations and capital allocation, while Howard’s role is primarily to protect the culture that made Berkshire unique. That separation is interesting because Berkshire’s biggest advantage has never been just its portfolio; it has been disciplined capital allocation and decentralized management. For me, the key variable is Abel’s use of Berkshire’s enormous cash pile. Acquisitions, buybacks, or simply waiting for better opportunities will reveal far more about the next era than headlines around the succession itself. The Buffett era may be ending—but the real test is whether the Berkshire system can compound without Buffett a

The real story isn’t GPUs vs. ASICs — it’s specialization. GPUs should remain critical for training and rapidly evolving workloads, where flexibility and software ecosystems matter. But inference is different: once workloads become predictable and massive, every watt and every dollar per token matters. That’s why custom silicon is becoming strategically important. OpenAI’s Jalapeño, developed with Broadcom, is a good example of this shift toward workload-specific optimization. What interests me most is the “picks-and-shovels” layer. Broadcom isn’t simply competing with NVIDIA; it can benefit when hyperscalers build their own accelerators because those chips still need advanced connectivity, networking and silicon expertise. Broadcom’s Q3 FY2026 AI semiconductor revenue reached $16.7B, up