Summary of Key Points
In the past three days, two of the world's leading investment banks, Goldman Sachs and Morgan Stanley (MS), have offered completely opposite recommendations regarding the direction of investments in the AI industry chain: Goldman Sachs suggests shifting from the seven tech giants (Mag7) to semiconductor hardware, while Morgan Stanley advises reducing holdings in semiconductors and focusing on hyperscalers in the cloud computing sector. The disagreement between the two firms essentially stems from their differing views on the "tipping point" of the shift in AI's value—Morgan Stanley focuses on short-term (2-3 quarters) capital movements, whereas Goldman Sachs looks at the long-term (2-3 years) supply and demand dynamics of the industry. Additionally, third-party institutions such as a16z, Sequoia Capital, and Berkshire Hathaway have provided their perspectives from a longer-term standpoint. Ultimately, all parties have reached a consensus that AI investments are shifting from betting on the scarcity of computing power to focusing on the efficiency of converting that power into revenue.
Goldman Sachs vs. Morgan Stanley: One Buying Semiconductors, One Selling Them—What's the Argument?
Goldman Sachs: Semiconductors Are the “Shovel Sellers”; Now Is the Right Time to Buy
Goldman Sachs believes that the decline in the AI sector at the beginning of July was not due to a fundamental breakdown but rather an adjustment of positions. They argue that there is less capital in the market during the summer, institutions are repositioning their portfolios, and there are doubts about the efficiency of cloud companies' spending. The reasons for buying semiconductor hardware are straightforward:
- Cloud companies (such as Amazon Web Services and Google Cloud) are spending heavily on equipment for AI purposes, but their short-term profits will be eroded; in contrast, semiconductor companies, acting as "upstream shovel sellers" (similar to those who sell shovels during a gold rush), have stable earnings as orders are secured and performance can be measured.
- Three key sectors recommended: AI computing power chips (NVIDIA GPUs, Broadcom ASICs), high-bandwidth memory (HBM) from companies like Micron and SK Hynix (with both volume and price expected to increase by 2027), and semiconductor equipment and packaging services (as demand for HBM and new manufacturing processes is growing).
- Data support: AI server demand is expected to increase by 4.3 times between 2025 and 2030, directly driving the demand for these hardware components; Asian memory stocks (such as SK Hynix and Samsung) have already begun to rebound.
Morgan Stanley: Semiconductors Are Becoming Commodities
Morgan Stanley's core argument is that semiconductors are becoming commodities. They use an analogy: semiconductors, which previously rose in price like a parabola, will now follow the trends of commodities (such as oil and steel), with prices fluctuating based on supply and demand and losing their competitive advantages.
- Trigger point: Meta began selling excess AI computing power last week at discounted prices (H100/B200 chips at 40%-60% of the original price), indicating that cloud companies may slow down their spending.
- Logic chain: If Meta continues to sell off its excess computing power, the growth in capital expenditure for cloud companies will peak, affecting semiconductor demand. The current divergence in trends between semiconductors and cloud companies is unsustainable and will eventually return to normal.
- Note: They are not bearish on AI; rather, they believe that spending will slow down in the short term, marking the fourth phase of a cyclical adjustment.
What Do Other Institutions Think? Long-Term Trends and Compromising Paths
In addition to the two investment banks, other top institutions have offered more long-term perspectives:
a16z (Mark Anderson): Computing Power Will Become Excessive, but Demand Will Surge
He argues that anything in short supply will eventually become over-supplied, including AI chips. However, demand for AI will also surge, so the focus should not be on owning computing power but on using it to create value. Startups should develop their own models rather than merely leveraging existing frameworks.
Sequoia Capital/Dylan Patel: The Scale of AI Computing Is Even Bigger Than That of Oil
He emphasizes that the computational load for running AI models (such as ChatGPT) is enormous and predicts that the market for these services will exceed that of oil, accounting for several percentage points of global GDP.
- Data support: The cost of computing has decreased by 60 times in a year, but demand is growing even faster. For example, Anthropic generates significant profits from renting GPUs.
- Energy consumption: OpenAI and Anthropic will consume over 100 gigawatts of electricity by 2030 (equivalent to 100 large power plants), and the hardware demand will only increase by 2040.
Berkshire Hathaway (Abel): Investing $10 Billion in Google’s “End-to-End Stack”
Buffett's successor broke with the company's 61-year rule of avoiding investments in technology it didn't understand, investing $10 billion in Google to build an AI data center. This investment targets the entire end-to-end stack: from chips to cloud services to models. Google produces its own TPU chips (independent of NVIDIA), and its cloud business is growing rapidly, with the Gemini model attracting 900 million active users monthly. The investment has already generated a profit of $1.23 billion, indicating a bet on the efficiency of the entire chain.
The Essence of the Disagreement: Short-Term Sentiment vs. Long-Term Fundamentals
The conflict between Goldman Sachs and Morgan Stanley is not about who is right or wrong but about the time frame they are considering:
- Morgan Stanley focuses on the short term (2-3 quarters): GPU rental prices have dropped from $8 per hour to $2, and inventory levels have increased from 2 weeks to 8 weeks, suggesting that computing power is less scarce in the short term, and capital will shift towards cloud companies.
- Goldman Sachs looks at the long term (2-3 years): SK Hynix's HBM production capacity is fully booked by 2027, and AI server demand is expected to increase by 4.3 times, indicating strong fundamentals.
- These views are not contradictory; in the short term, semiconductors may experience a pullback, but in the long run, the industry will continue to grow.
Consensus: AI Investments Are Shifting from “Having Chips” to “Efficiency”
All institutions agree on a core trend: the value of AI is shifting, and the focus of competition is changing:
- In the early phase, having access to chips was key (e.g., companies that rented out computing power); in the future, efficiency will be crucial—only those who can convert computing power into revenue (such as Google's end-to-end stack or Anthropic) will gain a competitive advantage.
Berkshire Hathaway's investment in Google represents this shift, as it bets on the efficiency of the entire chain rather than just individual components.
Implications for Individual Investors
There's no need to argue about which institution is correct; the key is to consider your investment horizon:
- Short-term (few months): Follow Morgan Stanley and pay attention to cloud companies or short-term capital movements.
- Long-term (years): Consider Goldman Sachs' recommendations and focus on core semiconductor hardware sectors (HBM, AI chips).
More importantly, identify companies that can convert computing power into revenue—whether they are chip manufacturers or application developers. The goal is to find those that can generate more value from their investments.
In summary, the AI investment landscape has moved beyond the stage of competing for access to technology; now, the focus should be on finding companies that can efficiently utilize that technology to create profit.
(End of Article)
Note: The data mentioned for 2025/2026 is based on forecasted timelines and not actual years. Source: Compiled from Daily Angel and other reference materials.