In June 2024, Goldman Sachs Research Head James Covello raised a critical question in a report: "Generative AI: Spending Too Much, Earning Too Little?" Nearly two years later, in April 2026, Covello acknowledged his misjudgment and conducted a market value adjustment.
According to trading desk sources on May 3, Goldman Sachs admitted in its latest extensive research report that it had been wrong about the speed of consumer adoption and the resilience of capital expenditures by hyperscale cloud providers. However, the firm expressed stronger conviction than ever regarding persistently low return on investment at the enterprise level and the unsustainable concentration of profits within the semiconductor sector.
The report stated that semiconductors remain the segment with the strongest financial realization within the AI cycle, but they carry the burden of high expectations and extreme profit concentration. While hyperscalers face short-term pressure, they hold key advantages in distribution, customer bases, data, computing platforms, and application ecosystems.
If AI ultimately proves successful at the enterprise level, cloud providers will regain control over their valuation narratives. If it fails, they will be the first to cut the capital expenditures paid upstream.
Based on this logic, Goldman Sachs proposed a clear relative value trading strategy: go long on hyperscale cloud providers and underweight semiconductors. The firm believes the market has already fully priced in the profits of the "picks and shovels" suppliers, while cloud providers' valuation multiples have been severely compressed due to ROI doubts.
This strategy would benefit regardless of the enterprise outcome. If enterprises achieve positive ROI, cloud providers' valuations would recover. If cloud providers cut unprofitable capex, leading to free cash flow improvement but hurting semiconductor revenue, the long cloud/short semi trade would also profit.
**Admitting Error and Holding Firm: Consumer Enthusiasm Exceeds Expectations, but Enterprise AI is Still "Burning Cash"** The report noted that consumer AI adoption rates have broken historical records, yet the massive capital expenditures by hyperscalers have not translated into actual profits from enterprise use, with FOMO masking deteriorating cash flows. Goldman Sachs admitted misjudging two core expectations:
* **Surprising Consumer Adoption Speed:** Data shows generative AI reached approximately 53% penetration within three years of launch, far exceeding the early trajectories of personal computers and the internet. This demonstrates immense consumer enthusiasm, even though Americans spend over 5 hours daily on phones and 95% of weekly active GPT users are on the free version. * **Hyperscalers Ramped Spending Despite Stock Underperformance:** Goldman Sachs initially believed cloud providers would cut AI capex if their stocks continued underperforming the market. Instead, driven by FOMO in the AI arms race, hyperscalers significantly increased spending, depleting their operating cash flow.
Despite this, Goldman Sachs maintains high conviction on poor enterprise ROI: making money with AI remains extremely difficult for businesses.
An MIT lab report indicates enterprises have invested $30-40 billion in generative AI, yet a staggering 95% of organizations report zero return. An EY survey was even starker: 99% of respondent companies reported AI-related financial losses, with an average conservative loss of $4.4 million per company (totaling approximately $4.3 billion).
Goldman Sachs stated that AI is actually increasing IT budgets—projected to grow from $5 trillion in 2024 to $6.15 trillion in 2026—rather than saving costs.
**Supply Chain Drain: Unsustainable Semiconductor Profit Dominance** Goldman Sachs emphasized that AI's economic benefits currently favor only semiconductor companies, while cloud providers and model companies are sustaining capex through debt—a concentration of value deemed unsustainable. This is the core supply chain contradiction identified. Historically, semiconductor booms were predicated on their downstream customers' prosperity. In the current AI cycle, however, the semiconductor supply chain is generating record profits by sacrificing the interests of others upstream.
Since the release of ChatGPT 3.5, Nvidia's net profit has grown 25-fold, while profit growth for Microsoft, Google, Amazon, and Meta has been relatively modest.
Hyperscalers have exhausted their operating cash flow and now rely on debt to fund AI expansion. Data shows data center debt issuance doubled in 2025, reaching $182 billion.
Goldman Sachs stated this is both unprecedented and unsustainable. Either top-tier enterprises must start deriving economic value from AI (thereby supporting the supply chain), or they will eventually be forced to reduce spending on underlying chips.
**The Path Forward: Rise of Small Language Models and Data Orchestration** The report suggested that large models are not the solution for enterprises; instead, small models for specific tasks and emerging "data fabric and orchestration layers" are key to unlocking AI's economic value. For enterprises to achieve ROI, the value chain must shift downward. The real barriers to enterprise success are not model capabilities but data structure and workflow orchestration:
* **SLMs Outperform LLMs:** Small Language Models, with fewer parameters, perform better for enterprises compared to general-purpose Large Language Models that require millions in training costs and powerful GPUs. For example, Datadog demonstrated at an analyst event that an SLM trained on internal data, leveraging domain expertise, delivered higher accuracy at lower cost. A supply chain company reportedly reduced response latency by 47% and cut costs by 50% after switching to a specialized SLM. * **Necessity of Orchestration Layers:** Enterprises should not use expensive large models to solve low-value, commoditized problems. For instance, a hedge fund analyst querying S&P 500 performance should route the request to a cheaper model; building a complex valuation model warrants using an advanced LLM. The lack of orchestration leads to wasteful use of expensive tokens, a primary cause of IT budget overruns.
**Profit Pool Disruption and Employment Impact Realities** Goldman Sachs believes AI's "augmentation" effect on jobs offsets its "replacement" effect, with true profit pool disruption likely in autonomous transport, software, and advertising. Contrary to media narratives about job losses, analysis by Goldman's macro team indicates AI's substitution and augmentation effects are roughly balanced. Under a baseline forecast, the peak unemployment rate increase during the transition would be only 0.6 percentage points.
So where will the profits to support massive AI spending come from? Goldman Sachs highlighted several potential areas:
* **Transportation:** The global autonomous Robotaxi market could reach ~$415 billion by 2035 (cumulative gross profit ~$440B); the heavy-duty truck market could reach ~$560 billion globally. * **Software:** AI will not consume software; the total addressable market is actually expanding (shifting from SaaS to Agents). Software giants with industry expertise have stronger moats than "pure AI-native" apps. * **Advertising:** AI is transforming advertising through automated content generation ($114B market) and shifts in digital channels ($170B market), with high adoption rates already seen in Google's PMax and Meta's Advantage+.
**Management Should Reject FOMO; Investors Should Favor Cloud Providers** Goldman Sachs advised management to reject FOMO and shift towards "buying" instead of "building" AI solutions. Investors were advised to favor cloud providers over semiconductors. The report concluded 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 pioneers, but this occurred a full 20 years after the dot-com bubble.
Goldman Sachs counseled corporate C-suites: moving slowly can lead to faster progress. Avoid being swept up by FOMO to build AI applications lacking a data foundation.
"During the internet era, the pioneers got the arrows, the settlers got the land." Companies should understand their position, and as the high costs of building AI become apparent, "buying" is becoming mainstream over "building." Increasing evidence shows the hidden costs of in-house AI systems far exceed expectations, leading companies to reassess the value of leveraging mature platforms.
On investment strategy, Goldman Sachs strongly recommended focusing on the relative value trade between cloud providers and semiconductors:
* **Scenario 1 (Win-Win, Cloud Leads):** Enterprises show positive ROI, doubts over cloud capex fade, cloud valuation multiples expand significantly; semiconductors, already highly valued, see smaller gains. * **Scenario 2 (Core Profit Scenario):** Enterprise ROI stays low, cloud providers cut AI capex. Cloud stocks rally on improved cash flow prospects ("relief rally"), while semiconductor stocks fall sharply on damaged revenue expectations. * **Scenario 3 (Risk Scenario):** The status quo persists. Cloud providers ignore poor ROI and continue aggressive spending, semiconductors continue draining the supply chain, causing the relative value trade to lose. However, Goldman Sachs believes the probability of the first two, more rational, scenarios is rising significantly.
Comments