Semiconductor portfolio adjustment strategy & Some thoughts on the prosperity of AI chips

Reduced half of the position in $AMD(AMD.US) and increased the position in $NVIDIA(NVDA.US)

1. Increased position in $Taiwan Semiconductor(TSM.US)
2. Significantly increased position in $Meta Platforms(META.US) during the big drop
3. Slightly increased position in $Tesla(TSLA.US) ahead of earnings
4. Funding for the adjustment came partly from reducing AMD and partly from reducing XBI

Reasons for buying META:

First, Meta's big drop was clearly spooked by the boss's firm commitment to AI investment, reminiscent of its stubborn all-in on VR years ago, which led to massive spending without business growth. But upon closer thought, AI is different from VR. VR was a lone wolf, while AI has industry-wide consensus. META's social and ad businesses have clearly benefited from AI (here, AI doesn’t entirely equal LLM/genAI), and its brand image has also gained from AI (llama3 LLM). This part has already delivered ROI, with strong cash flow, and its product and business moat remains unchallenged (even the TikTok situation favors it). Investing in AI is no issue.

In fact, when it comes to cutting-edge AI/genAI/LLM in monetization scenarios, the number of large-scale companies that can or are likely to form a positive feedback loop can be counted on one hand:

For B2B scenarios:

1. Microsoft, as clearly stated in its earnings; Azure AI isn’t monetized yet, but Copilot/Office scenarios are already impressive.
2. CRM, Einstein Copilot

For B2C scenarios:

1. META
2. Google (Google Cloud isn’t monetized yet); AI improves product experience and ad efficiency.
3. TESLA, with FSD v12 and future iterations. This part is more uncertain but represents a massive, entirely incremental B2C AI monetization scenario.

Besides the above fundamentals, two timing-related reasons for buying META:

1. A 15% single-day drop at open, deemed irrational based on fundamentals.
2. Microsoft’s post-market earnings showed firm AI investment expansion and solid returns, which should quickly calm market concerns.

Reasons for buying NVDA:

Two reasons:

1. TESLA, META, GOOGLE, and MSFT are all aggressively investing in AI. Data center plans are long-term, with annualized spending. Why worry about NVDA in the mid-term? Let’s hit $1000 again.
2. AI’s technical trajectory ensures scaling law will last at least 1-2 years. GPT-5 will be the next frontier, sparking another arms race.

Technical evolution (quoting a group chat):

A simple explanation of what’s happening with cutting-edge LLMs like GPT:

The core challenge now is embedding larger search spaces and evaluation strategies to improve logic and reasoning when answering complex queries.

GPT-4’s answers are still "unthinking"—grammatically fluent but not always aligned with human expectations, based solely on supervised training data.

Users often guide it to "think step by step" or follow frameworks. For translation, the best current method involves three LLM roles:
Role 1: Translate
Role 2: Critique Role 1’s output
Role 1: Revise based on feedback
Role 3: Final polish

This process could be integrated into model training/inference, universally improving performance (e.g., for math problems).

Expectations based on these practices:
1) Text data may peak (at most doubling), but video data has huge potential. While video models don’t boost reasoning, they expand capabilities (e.g., video understanding). Data scaling will continue driving compute demand.
2) Model complexity (search spaces, evaluation steps) will rise, requiring more advanced chips.
3) Rising costs mean fewer buyers for training chips, but each buys more. Inference demand will grow.
4) Complexity may erase the perceived lower barrier for inference chips, keeping NVDA ahead of AMD.
5) Power constraints in the U.S. make energy efficiency critical. NVDA optimizes entire cards like mini data centers, sustaining its lead.

Reasons for reducing AMD:

1. Confidence in NVDA’s lead
2. Short-term unlikelihood of NVDA’s inference-chip moat weakening vs. AMD

Reasons for buying TSMC:

1. AI chip demand is solid, long-term, and booming
2. TSMC’s absolute lead in high-performance chips and advanced packaging
3. Underappreciated AI upside—it hasn’t rallied much yet
4. Unconcerned about weak iPhone chip demand; bullish on iPhone + AI and Apple’s AI server plans (3nm in 2025)

Reasons for buying TESLA:

As mentioned, it’s the only large-scale, purely incremental B2C AI play I see. I’ll watch at high prices but buy the dip for FSD’s "ChatGPT moment."

That’s all—thanks for reading!