Although the AI investment cycle is shifting from &#34;investment&#34; to &#34;returns,&#34; this does not mean that &#34;slowing down&#34; equates to &#34;peaking.&#34;
According to the Chasing Wind Trading Desk, Goldman Sachs pointed out in its latest AI report that despite the slowing growth rate, AI infrastructure investment will remain sustainable in the next 2-3 years and the market's excessive focus on &#34;slow returns&#34; may overlook the fact that cost benefits have already begun to be released, and stock prices have yet to reflect this structural change.
The report emphasizes that although the commercialization of AI is still in its early stages, the first phase of returns based on cost reduction has already emerged. It is estimated that by 2030, AI automation could save Fortune 500 companies approximately $935 billion in costs. Analysts believe that these early gains are sufficient to support the current level of AI infrastructure investment, even though the growth rate may slow.
Analysts also pointed out that hyperscale cloud service providers, as the main investors in AI infrastructure, base their investment decisions more on long-term revenue growth opportunities rather than short-term cost savings.
<h2>Controversy Over AI Investment Returns: Cost Savings vs. Commercialization</h2>
Since the surge of generative AI at the end of 2023, capital expenditures have exceeded $350 billion. Is the investment worthwhile? Are the returns sustainable? These questions are becoming the focus of market attention.
Goldman Sachs' research team divides AI value creation into three phases: the first phase achieves cost reduction through automation (currently ongoing), the second phase involves reinvestment and reconstruction, and the third phase realizes monetization through incremental revenue.
The report's analysis shows that the automation applications of AI in functions such as customer service, sales and marketing, and IT have begun to yield actual benefits. For example, 43% of call centers have adopted AI tools, resulting in an average operational cost reduction of 30%.
Specifically, JPMorgan plans to reduce 10% of its back-office staff over the next five years, AT&amp;T aims to reduce call center traffic by 30% using AI, and T-Mobile expects to cut customer service contact volume by 75% by 2027.
The current core controversy revolves around the sustainability of AI infrastructure investment. Goldman Sachs predicts that by 2030, global Fortune 500 companies could save up to $935 billion in costs, accounting for about 14% of their total costs, with a net present value return of approximately $780 billion on AI investments, still showing positive returns relative to the cumulative investment of $350 billion.
Analysts point out that hyperscale cloud service providers, as the main investors in AI infrastructure, base their investment decisions more on long-term revenue growth opportunities rather than short-term cost savings. This time difference between investment and returns complicates ROI calculations.
<h2>Short-term Infrastructure Spending is Secure, Inference Demand Becomes New Momentum</h2>
Another core concern in the market regarding AI stocks is whether infrastructure spending has peaked, especially with the inventory backlog and soft demand expectations for training chips.
Goldman Sachs believes that this concern is somewhat excessive, stating in the report:
In the next 2-3 years, especially by 2026, large technology companies (such as Microsoft, Amazon, Google, and Meta) will still have the financial capacity to maintain AI infrastructure investments without significantly compressing profit margins
Key suppliers such as TSMC, Broadcom, and SK Hynix continue to raise their revenue expectations related to AI. TSMC has been consistently increasing its near-term and long-term AI revenue targets for six consecutive quarters, while SK Hynix expects its HBM revenue to double by 2025. These data support the judgment of a fundamentally balanced supply and demand.
The demand for &#34;inference&#34; computing power from enterprise customers and governments (sovereign AI) will become a new driver of spending, especially as small and medium-sized enterprises rapidly expand their deployment of customized models or edge AI application scenarios.
Taking NVIDIA as an example, Goldman Sachs responds by analyzing the supply and demand balance of the GB200 NVL72 rack in depth.
Based on NVIDIA's public statements, large-scale cloud service providers are currently deploying about 1,000 GB200 racks per week and plan to further increase the deployment speed in the second quarter. Goldman Sachs' model shows that to avoid excessive inventory among ODM manufacturers, large-scale cloud service providers need to deploy 36,000 and 58,000 equivalent GB200 racks in 2025 and 2026, respectively.
The report points out that the advancement of NVIDIA's Blackwell architecture is in line with expectations, with this product accounting for 70% of data center revenue in the first quarter. The transition from Hopper to Blackwell is basically complete. This progress helps alleviate market concerns about inventory backlog.
However, analysts also remind that the visibility of capital expenditures for large-scale cloud service providers in 2027 remains limited, which is a key uncertainty affecting long-term supply and demand balance. In contrast, NVIDIA and Broadcom have better business visibility due to their longer product delivery cycles (about 12 months) and relatively concentrated customer base.
<h2>How are stock prices pricing AI expectations?</h2>
Taking NVIDIA as an example, Goldman Sachs points out that the market has partially factored in the strong demand expectations for its next-generation GPU (Blackwell), but the expansion of its customer base and the explosion of AI inference business are still undervalued.
For Broadcom, the rise in its stock price is more based on the clear guidance provided by the company—AI revenue is expected to grow by 60% year-on-year in FY25 and FY26. This expectation also leads Goldman Sachs to believe that the rise in Broadcom and NVIDIA's stock prices is not a &#34;bubble,&#34; but rather reflects a clearer path of mid-term fundamental improvement.
In contrast, for companies like AMD, ARM, and Marvell, Goldman Sachs maintains a neutral rating, citing that their AI-related businesses are still in the early stages with limited market share

GOOG

NVDA

GOOGL

MSFT

Goldman Sachs stated that although the commercialization of AI is still in its early stages, the initial phase of returns based on cost reductions has already emerged. It is predicted that by 2030, AI automation could save approximately $935 billion in costs for Fortune 500 companies. Analysts believe that this early benefit is sufficient to support the current level of investment in AI infrastructure, although the growth rate may slow down

- Goldman Sachs reports that while AI investment growth is slowing, it does not indicate a peak, as infrastructure investments are expected to remain sustainable over the next 2-3 years.  
- The report highlights that AI automation is already yielding cost savings, with Fortune 500 companies projected to save approximately $935 billion by 2030.  
- Concerns about the sustainability of AI infrastructure spending are deemed excessive, as major tech companies are financially capable of maintaining their investments without significantly reducing profit margins.

How long can AI infrastructure still be invested in? Goldman Sachs: 2-3 years is not a problem, the return window has just opened