While semiconductor companies are seeing record revenue and profits, the overall dynamic is “unprecedented and unsustainable,” Covello writes.
Where are the investment opportunities in AI?
The team points out that about two years ago they recommended that investors take a “picks and shovels” approach to the AI boom, investing in semiconductor and semiconductor equipment companies (semi cap). Since then, these stocks have outperformed the market, while the hyperscalers mostly have not.
From here, Goldman Sachs Research expects hyperscalers to outperform semiconductor companies and those that make equipment for semiconductor companies, Covello writes.
For one thing, investors have become fairly skeptical of the returns the hyperscalers are likely to deliver. If enterprises begin to show returns from AI spending, investors may be willing to pay higher multiples on these stocks again. Even if returns on enterprise AI spending continue to be challenging, the hyperscalers may decide to trim their capital expenditures. This may be the best scenario for this trade, the team writes, causing a relief rally for the hyperscalers on better cash flow prospects and a selloff for semiconductors.
The worst scenario for this trade would be the status quo persisting and the hyperscalers continuing their huge capital outlays despite challenges for successful enterprise AI adoption. Investors would lose money on the trade in this scenario, as all the value of AI spending would continue to accrue to the semiconductor companies.
How can enterprises better use AI?
The big question according to Goldman Sachs Research is how companies can create economic value from their spending on AI. The researchers suggest they need to ensure that they are building their agentic AI on data that is structured properly, and they need to deploy, or orchestrate, AI models in a cost-effective manner.
“Models will continue to improve at a fast pace, but, right now, model capability is not what’s holding back successful enterprise use cases,” Covello writes. “We believe data structure and orchestration are critical factors to unlock AI in the enterprise.” A new layer needs to be added between the enterprise and the model developers, he adds.
One goal for this “orchestration and deployment layer” would be to ensure that workflows are routed properly based on complexity and cost considerations. Low consequence, higher-volume workflows should be routed to simpler AI technology—open-source or lightweight models. Higher consequence tasks can go to the most advanced (and expensive) models, which will be reserved for environments where the cost of failure is high.
How should executives approach AI in their businesses?
The researchers provide an example of how this might play out at a hedge fund that wants to put AI tools in the hands of its employees. When someone queries the AI with what is essentially a glorified web search, that would be routed to a lower-end model. On the other hand, when building a new financial model or conducting complex valuation analysis across industries, the work might be routed to a more expensive, higher-powered model.
One approach that holds promise is to deploy small language models (SLMs), Covello writes. In contrast to the foundational large language models that attract attention, SLMs can be optimized for speed and lower cost, and they can be tailored to a specific workflow. Small models can be faster and increase accuracy even as they are much less expensive to train and use—showing the cost-effectiveness needed for successful enterprise AI deployment.
The new layer between the enterprise and the model providers would also address data issues. A basic example might be to ensure that a retailer’s data is aligned and accessible so that an AI tool can give helpful customer suggestions. That will only be possible if the databases for recommendation algorithms, customer behavior profiles, and current inventory are all properly organized and are not siloed.
“We believe organizations getting these building blocks in place will be key to unlocking AI economics in the enterprise,” Covello writes.
The researchers suggest that C-suite leaders may want to play the long game, given that some of what’s needed for enterprise AI success is not yet in place. “Slow down now so you can speed up later,” Covello writes. This won’t be easy, he admits, given the immense pressure executives face to show that they have a strong AI story—and the fear of missing out that seems to be driving shareholders and markets.
Which jobs and industries are likely to be disrupted by AI?
One place where markets have been watching for the impact of AI at the enterprise level is in large scale job replacement—and this has not yet occurred. Goldman Sachs Research has shown that, while AI is replacing some workers, these job losses are being partially offset by rising employment where AI augments human workers and boosts their productivity.