Two years ago, Goldman Sachs research head questioned “AI spending too much, earning too little,” now admits “got it wrong,” but says “the shovels are fully priced in, now bullish on the cloud.”

Two years ago, Goldman Sachs research head questioned “AI spending too much, earning too little,” now admits “got it wrong,” but says “the shovels are fully priced in, now bullish on the cloud.”

```

In June 2024, James Covello, Head of Research at Goldman Sachs, once raised a soul-searching question in a report: “Generative AI: Spending Too Much, Earning Too Little?”. Nearly two years later, in April 2026, James Covello admitted he had been “wrong,” personally “marking to market” the valuations.

On May 3, according to Chasewind Trading Desk, in April 2026, Goldman Sachs admitted in its latest lengthy research report that it had been wrong on the pace of consumer adoption and the resilience of hyperscaler cloud capital expenditure, but had more conviction than ever regarding poor ROI on the enterprise side and the unsustainable concentration of semiconductor profits.

According to the report, semiconductors remain the strongest financial beneficiary in the AI cycle, but this is accompanied by high expectations and highly concentrated profits. Cloud vendors are under pressure in the short term, but they own distribution, customers, data, computing platforms, and application ecosystems.

If AI eventually achieves success on the enterprise side, cloud providers will regain the right to be valued accordingly; if not, the first thing to be cut will also be their upstream capital expenditures.

Based on this logic, Goldman Sachs has given a clear relative value strategy: Go long hyperscale cloud providers, underweight semiconductors. Goldman believes that the market has already fully priced in the profits of “pick and shovel sellers,” whereas the cloud providers’ multiples have been severely compressed due to doubts about their ROI.

Whether enterprises ultimately achieve positive ROI, or cloud providers cut capex due to unprofitability (which would repair free cash flow and seriously hit semiconductor revenues), a long cloud/short semiconductor strategy will benefit in both scenarios.

Admitting Mistakes & Holding Firm: Surging Consumer Enthusiasm but Enterprises Still “Burning Money”

The report points out that consumer AI adoption rates broke historical records, but the massive capital expenditures of hyperscalers have not translated into real profits on the enterprise side, and FOMO (fear of missing out) has obscured deteriorating cash flow.

Goldman Sachs admits it was wrong on two core expectations:

Surprising speed of consumer adoption: Data show that generative AI reached about 53% penetration within three years of its launch, far beyond the early trajectories of personal computers and the Internet. Despite Americans spending over 5 hours a day on their phones and 95% of weekly active users using the free version of GPT, this still proves enormous consumer enthusiasm.

Cloud vendors increased spending even under share price underperformance pressure: Goldman had thought that if cloud vendors’ stock prices continued to underperform the market, they would cut AI capex. In reality, due to “FOMO” in the AI arms race, hyperscalers have massively increased spending, and their operating cash flows have been depleted.

Nevertheless, Goldman remains highly confident on the enterprise ROI side: despite enterprise AI adoption, making money remains extremely difficult.

An MIT lab report shows that enterprises invested $30–40 billion in generative AI, but shockingly, 95% of organizations had zero return.

An even harsher survey by Ernst & Young found 99% of respondent companies reporting financial losses from AI, with an average loss of $4.4 million per company (totaling about $4.3 billion).

Goldman says AI is, in fact, pushing up IT budgets (with global IT spending expected to rise from $5 trillion in 2024 to $6.15 trillion in 2026), rather than saving costs.

Profit Extraction Along the Value Chain: Semiconductor Profit Dominance Is Unsustainable

Goldman emphasizes that the current economic benefit from AI is only in favor of semiconductor firms, while cloud and model companies are maintaining capex through borrowing, and this degree of value concentration is unsustainable.

This is the core supply chain contradiction in the report. Historically, the prosperity of semiconductor companies depended on prosperous downstream customers. But in the current AI cycle, the semiconductor supply chain is generating record profits at the expense of everyone else upstream in the value chain.

Since the launch of ChatGPT 3.5, Nvidia’s net profits increased 25-fold, while the profit growth of Microsoft, Google, Amazon, and Meta was modest.Hyperscale cloud vendors have already burned through their operating cash flow and now must finance AI expansions with debt. Data show data center debt issuance doubled in 2025 to $182 billion.

Goldman bluntly states that this situation is both unprecedented and unsustainable. Either the leading enterprises must begin to extract economic value from AI (thus feeding value back to the supply chain), or they’ll ultimately have to cut spending on underlying chips.

The Way Forward: Rise of Small Language Models (SLM) and the Data Orchestration Layer

The report states, Large models are not the cure-all for enterprises; small task-specific models (SLMs) and the emerging “data structure and orchestration layer” will be key to unlocking enterprise AI economic returns.

For enterprises to achieve ROI, the value chain must move downstream. Goldman points out that the capability of models is not the hindrance to success; the real obstacles are data structure and workflow orchestration:

  • SLMs outperform LLMs: Compared with broad large language models (LLMs) requiring millions of dollars to train and significant GPU support, smaller parameter SLMs perform better on the enterprise side.
  • For instance, Datadog demonstrated at an analyst conference that SLMs, trained with internal data and domain expertise, deliver higher accuracy at lower cost. A supply chain firm reduced response latency by 47% and cut costs by 50% after switching to dedicated SLMs.
  • Necessity of the orchestration layer: Enterprises cannot use expensive large models to solve low-value, commoditized problems.
  • For example, if a hedge fund analyst simply queries S&P 500 performance, the request should be routed to a cheaper model; only complex valuation modeling should use advanced LLMs. Lack of an orchestration layer leads to wasteful use of costly tokens — the main culprit of enterprise IT budget overruns.

Profit Pool Disruption and Employment Impact: A Realistic Assessment

Goldman believes that the "augmentation" effect of AI on employment offsets its "substitution" effect, and the real disruption to profit pools may arise in autonomous driving, software, and advertising.

Contrary to media hype of “AI taking jobs,” Goldman’s macro team analysis shows the substitution and augmentation effects basically offset each other. The baseline forecast suggests that at most, the unemployment rate peak rises 0.6 percentage points during the transformation period.

So where are the profit pools capable of supporting AI’s enormous expenditure? Goldman points out several potential directions:

  1. Transportation: By 2035, the global robotaxi (autonomous rideshare) market could reach $415 billion (cumulative gross profit ~$440 billion); global heavy truck market could be about $560 billion.
  2. Software: AI will not eat software — total addressable market (TAM) is actually expanding (from SaaS to agent transformation). Software giants with deep domain experience have a stronger moat than “pure AI native” applications.
  3. Advertising: AI is profoundly changing advertising through automated content generation ($114B market) and digital channel shift ($170B market); Google’s PMax and Meta’s Advantage+ already show high adoption.

Management Should Resist FOMO, Investors Should Go Long Cloud Providers

Goldman says company management should resist FOMO and shift from “building” to “buying” AI solutions; investors should go long cloud providers and underweight semiconductors.

The report concludes with a historical metaphor from the Internet era: The pioneers got the arrows and the settlers got the land. Uber was built on the bankrupt assets of first movers, but this was a full 20 years after the dot-com bubble.

Goldman’s advice to the C-suite: Slowing down is speeding up. Don’t be swept up by FOMO into building AI applications lacking a solid data foundation.

“In the Internet era, pioneers got the arrows, settlers got the land.” Enterprises should clarify their position, and with the high cost of self-building AI becoming clear, “buy,” rather than “build,” is becoming mainstream.

There is increasing evidence that the hidden costs of self-built AI systems far exceed expectations; enterprises are reassessing the cost-effectiveness of leveraging mature platforms.

In terms of investment strategy, Goldman strongly recommends focusing on relative value trading between cloud vendors and semiconductors:

Scenario 1 (Win-win but clouds lead): Enterprises begin to show positive ROI, market concerns about cloud capex are dispelled, and cloud valuation multiples recover sharply; semiconductors, already highly anticipated, lag behind.Scenario 2 (Core profit scenario): Enterprise ROI remains weak, cloud vendors decide to cut AI capex. Cloud stocks rebound as cash flow prospects improve, while semiconductor stocks suffer heavy losses due to earnings disappointments.Scenario 3 (Risk scenario): Current situation continues. Cloud vendors ignore poor ROI and keep ramping up spending, while semiconductors keep extracting profit from the whole value chain, causing losses for this relative value trade. However, Goldman believes the probability of the first two, more rational scenarios is rising significantly.

 

~~~~~~~~~~~~~~~~~~~~~~~~

The above excellent content comes from Chasewind Trading Desk.

For a more detailed interpretation, including real-time analysis and first-line research, please join the [Chasewind Trading Desk▪Annual Membership]

Risk Warnings and DisclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice and does not consider individual user’s specific investment objectives, financial status, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article are appropriate for their own circumstances. Investments based on this article are at your own risk. ```