Corporate uptake of artificial intelligence (AI) started out gradual, but enterprises are beginning to rapidly deploy the technology in a development that will affect the entire AI ecosystem.

The catch is that the AI industry’s ability to scale is constrained by how much computing power is available, according to Goldman Sachs Asset Management.

This was a key takeaway for Brook Dane and Sung Cho, the co-heads of the US Mid and Large Cap and Technology Equity businesses at Goldman Sachs Asset Management, during their annual fact-finding trip to Silicon Valley in June.

Visiting with leaders of top semiconductor makers, internet platforms, developers of large language models (LLMs), and startups, Dane and Cho took the pulse of an industry that is spending unprecedented sums on infrastructure.

“We have never seen an up cycle like this,” Cho says. “And as an investor you are always trying to figure out the right water level of demand.”

Dane says the technology is “transformational.” “But we also recognize what we can actually analyze and know, and what we can only guess at.”

We spoke with Dane and Cho about “compute constraints,” AI’s impact on online advertising, and why valuations are still attractive in this latest tech boom.

You’ve just returned from your fact-finding trip to the Valley. What was your top takeaway?

Brook Dane: The most important thing we are seeing is the traction around deployment of AI in the enterprise space. The companies doing this at scale are using a tremendous amount more compute than what the typical company is using.

According to one company we met with, the top 5% of companies are consuming three times the number of tokens that the median company is using, and this gap continues to widen. (Tokens are the units of text LLMs read, process, and generate, and have become a key metric for AI growth.)

It feels like we are at the beginning of a tipping point where we are seeing use cases broaden beyond coding. We think enterprise adoption of AI is going to be a durable trend over the next several years.

Sung Cho: This is putting a lot of pressure on compute.

Pressure on compute?

Sung Cho: Yes. If you think about the adoption of AI, it started with chatbots three years ago and consumers jumped in at a rapid pace. Enterprises were slow to adopt AI. Now with the rollout of models designed for business workflows and other activities, you are seeing AI adopted to do far more. These models are providing companies with agentic features. Token use is accelerating. That has put pressure on computing infrastructure.

Brook Dane: We are in a compute-constrained environment that is real and durable. And this demand extends to ASICs, which are computer chips designed for a single purpose; memory chips, which store data; and to the suppliers of both of those supply chains. This buildout is accelerating and it’s forcing supply chain constraints across the entire ecosystem.

Does this mean the investable AI market is broadening?

Sung Cho: It is. Part of the reason why the enterprise AI chain is so meaningful is because it’s ushering in more demand for inference, which is when you use a trained AI model to perform tasks or produce results from fresh data. And the computing infrastructure that supports inference can be very different from the one that underlies training.

So as you transition from a pure AI-training-only world to an inference-computing world, you’re going to put pressure on different parts of the compute stack that didn’t have pressure before.

How will that affect the buildout for AI?

Brook Dane: In one example, we’re looking at how the demand is impacting fiber optics and the substrates that support fiber optics.

Sung Cho: Yes, one of the areas we continue to like is the optics space. As processing speeds increase, the bottleneck isn’t necessarily semiconductors but how fast processors can speak to other processors. As you build more data centers, you have to connect them with large amounts of fiber optics. Additionally, inside the data center, connections are primarily based on copper. As transmission speeds get faster, fiber optics will have to replace copper as well. This is why we like the optics space. It’s going to accelerate for the next five-plus years.

Switching to AI and the internet, are LLMs a challenge to online advertising?

Brook Dane: A year ago, the fear was that LLMs were going to cannibalize online search and depress the ability to advertise against these streams. What we have seen is the exact opposite. There is so much depth to AI queries that businesses can target consumers more effectively. You have more context around what a user is looking for and thinking about so the ad placement becomes more valuable.

Sung Cho: I would add that AI is accelerating the time we spend on these platforms by individualizing and creating more content and helping these programs understand what we like and don’t like. This creates more inventory or opportunities for online advertising, and that inventory gets monetized.

What is the state of play with AI agents in the enterprise space?

Brook Dane: While it’s still early, the thinking and models around what we call agentic commerce have evolved. Big online companies recognize there is always going to be a place for a human in a transaction with a customer. But agents will reach a point where you can enjoy whatever experience or execute whatever task you like faster and more efficiently.

What does that mean for investors?

Sung Cho: In terms of the winners in agentic commerce, they are going to be the platforms that understand what products or services the agents are looking for. In the previous world, online merchants just had to be at the top of the keyword search list, and that was how you optimized your marketing. Now it’s asking, what does an agent want?

Brook Dane: The big internet platforms, as well as retailers, are already using agentic platforms. They all have a slightly different take, but they are all leaning into this as a use case for AI.

This is all happening so fast, yes?

Sung Cho: It’s crazy.

Brook Dane: And that agentic workflow is coming through on the enterprise side as well. It’s not just about shopping and the disruption you might see in ecommerce. These agents are creating mini-applications and perform other workflow-related tasks.

In June, tech stocks, especially those connected to AI, recorded another selloff. Investors seem increasingly wary about an industry that is spending so much.

Sung Cho: There is always skepticism when things move really fast. What you can focus on are the near-term indications, so let’s return to compute constraints. Our ability to meet this inflection in demand might be less than what the market wants. But that actually makes me more bullish because if you can’t meet demand over one year, then you have to meet it over three to five years. The spending becomes more durable and sustainable.

Brook Dane: To Sung’s point, we spend a lot of time thinking about valuations. When you look at the implications of these supply constraints, the valuations are actually very attractive right now.

The fundamental trend driving this wave of technological innovation is really solid. It won’t be a straight line, but smart investors will take advantage of pullbacks in the market to build positions where they have conviction. That’s what we try to do every day.

This article is being provided for educational purposes only. The information contained in this article does not constitute a recommendation from any Goldman Sachs entity to the recipient, and Goldman Sachs is not providing any financial, economic, legal, investment, accounting, or tax advice through this article or to its recipient. Neither Goldman Sachs nor any of its affiliates makes any representation or warranty, express or implied, as to the accuracy or completeness of the statements or any information contained in this article and any liability therefore (including in respect of direct, indirect, or consequential loss or damage) is expressly disclaimed.