Speaker A
Jim, we hear a lot about all these different types of models. We hear about open source, we hear about open weight. And if a lot of open source models are being used, how does that actually affect the value chain across the board? I
Goldman Sachs explores how new AI models impact enterprise adoption, profitability, and the AI investment landscape.
Key Takeaways
- Enterprise AI adoption is critical for monetizing AI investments beyond consumer use.
- Open source models enhance enterprise AI profitability and accessibility.
- The deployment and orchestration layer is essential for effective AI implementation in complex organizations.
- Economic benefits currently favor semiconductor companies but will shift as enterprises optimize AI use.
- Investors should track AI impact on company financials to identify long-term winners.
What the video covers
- Open source AI models benefit enterprise customers by enabling more profitable AI implementations.
- Consumer adoption of AI is rapid but mostly free, so monetization depends on enterprise adoption.
- Currently, enterprises are aggressively implementing AI but are not yet seeing net savings or profits.
- Economic value has mostly accrued to semiconductor companies, which is unsustainable without enterprise ROI.
- A key challenge is the deployment and orchestration layer that manages data and routes queries to appropriate AI models.
- Open source, open weight, and frontier models differ in accessibility, cost, and control over code and data.
- Open source models could shift economic value from semiconductor companies to enterprises and hyperscalers.
- The market is expected to become more disciplined and efficient as AI adoption matures.
- Investors should focus on companies showing clear AI benefits in their financials.
- The future of AI investment depends on profitable enterprise adoption and effective model optimization.
Chapters
- 00:00Introduction to AI Models and Value Chain Impact
- 01:11Expert Background and AI Adoption Overview
- 02:09Enterprise AI Implementation and Profitability Challenges
- 03:12Economic Value Distribution and Sustainability Concerns
- 04:02Deployment and Orchestration Layer Importance
- 05:05Explanation of Open Source, Open Weight, and Frontier Models
- 06:00Impact of Open Source Models on the Value Chain
- 07:01Market Discipline and Investment Implications
Full Transcript — Download SRT & Markdown
Speaker A
think open source really benefits the enterprise customer. I think open source models will make it more likely that enterprises can profitably implement AI in the organization. I think a lot of the companies that are going to be the long-term AI winners, most
Speaker A
people have never heard of yet. Hello, Jim, and thank you for sitting down with us again for an update on this very important issue of artificial intelligence. We've seen early on the most rapid adoption on the consumer side. On the other hand, we know consumers aren't paying
Speaker A
that much, but we haven't seen the pace of adoption on enterprise or corporate America.
Speaker A
But we haven't seen the pace of it. And there's a big question out there about when will we see this be effective? Obviously, you're one of the best people to address this question. A semiconductor analyst for about 16
Speaker A
years. The last nine of that journey, you were ranked the number one analyst by Institutional Investor. And now you head up all equity research analysts at Goldman. So we thought how great to actually hear from you. So we thought how great to
Speaker A
actually. How is the landscape evolving? Are things getting any better in terms of enterprise adoption in a profitable, effective way? Yeah, I really am happy you started there.
Speaker A
I really think that is the most important thing to talk about. I don't think enough attention is being paid by the market on that exact issue. It all starts with, does the end customer, does the end user make or save money implementing AI
Speaker A
or not? And as you said, on the consumer side, the adoption of the technology has been magnificent. And as you said, on the consumer side, the adoption of the technology, you know, faster than any other technology we've ever seen. But as you also
Speaker A
rightly said, 95% of consumers are using a free version of AI. So we're not going to monetize the significant infrastructure investment on the consumer side. It's going to have to come from profitable enterprise adoption. Now, enterprises are
Speaker A
very aggressively implementing AI in their organizations. And there's certainly a lot of use cases where it's been successful. But right now, enterprises collectively are not making or saving money on their AI implementations. And to me, everything flows from there.
Speaker A
ROI of the hyperscalers, the circular financing issue, all the societal issues around where data centers are looking. All of those things are byproducts of is the end customer making or saving money in their AI implementation. If you and I go
Speaker A
on Amazon, that's a cheaper option for us than going to the store. If you and I go on Amazon, that's a cheaper... It's certainly easier, but in most cases, it's cheaper as well. The advertisers are getting a better
Speaker A
ROI bang for their buck on advertising dollars, advertising on Google, than on print advertising.
Speaker A
And that's what's missing right now, right? We're spending a lot of money. There's a lot of implementation. There's thousands and thousands of applications that are being used by enterprises, but it's not resulting yet in net savings, according to all of the survey work
Speaker A
that we've done. But it's not resulting yet in net savings, according to all of the... economic value is accrued to the semiconductor companies, and that's great for now for the semiconductor companies, but that's completely unsustainable unless the end customer, the end user,
Speaker A
the enterprises start to make or save money implementing AI. That's where I think everybody's research and focus should be at this point. Anthropic's done a really interesting study that shows what is the feasible market out there, what is the addressable market, and how
Speaker A
much actually enterprise has adopted. But even with that limited adoption and how much actually enterprise has adopted, relative to the potential, what they call the feasible market, they are blowing through their budgets for AI. Jim, one of the issues that you have raised is this layer that you are calling
Speaker A
deployment and orchestration layer. What is it exactly and why does it matter so much?
Speaker A
We have this situation today where, again, people talk about, you know, this technology will get unlocked when the new model comes out, the new, more capable model comes out.
Speaker A
Unlocked when the new model comes out, the new more capable model. I think the models are incredibly powerful and incredibly capable. I think a lot of the agents are very powerful and very capable. But right now, I think we're building agents on top
Speaker A
of data that isn't ready to be agented. And I think that the bigger your organization, the more complex that issue is. And so I think a lot of the things that need to be addressed in order to successfully, profitably implement AI in the
Speaker A
big enterprises are this data management layer and then this model optimization or are this data management layer. Orchestration layer. I think you're going to have a model optimization layer built into all enterprises that routes the high consequence queries to the big frontier models that are a little more expensive to use and
Speaker A
then it'll route the low consequence queries to the open source model. That model optimization or model router really is yet to be developed in most organizations. It's really important.
Speaker A
It's not talked about enough. There's a whole lot of really interesting companies, mostly private companies working on it. And I think that's going to be one of the big keys to unlocking the economic value of AI in the enterprise. We hear a
Speaker A
lot about all these different types of models. We hear about open source. We hear about open weight. And then at the higher level, frontier models. Can you just explain what these are? Sure. Absolutely. It's purest form. Open source means everybody has access to
Speaker A
the code. Everybody can share. Everybody can edit. And you can use it for free.
Speaker A
Open weight is a derivation of that, where you can use the output, but the source and the code are proprietary to the designer. So in the world of AI, you might be able to use an open weight model without paying for it, but
Speaker A
the data that that model was trained on is the property of the person who developed the model. A frontier model or a closed model is something that no one has the access code for, or no one and you have to pay to use,
Speaker A
and you certainly can't edit the model. And if a lot of and you have to pay to use, and you certainly can't edit the model. Open source models are being used and there's an optimization being done who actually loses in this how does
Speaker A
that actually affect the value chain? Yeah, across the board I think open source really benefits the enterprise customer. I think open source models will make it more likely that enterprises can profitably implement AI in the organization. So for investors I think
Speaker A
there's going to be the potential, you know, in a way that we really haven't seen yet to say, well, this is a company that's benefiting from AI because you'll be able to see it on the P&L, you know, if this vision of
Speaker A
the world plays out. Going to be the potential, you know, in a way that we really haven't seen. I think it's really good for the hyperscalers because it's more likely than you're going to be able to profitably fill up all this capacity that you're adding. And then
Speaker A
I think it's more of a challenge to the semiconductor layer that's benefited from the massive compute power that the frontier models require. If you can build models that don't require as much compute power, then I think that the customers can start to shift
Speaker A
some of the economic value from the semiconductor companies, which have made all the profits from now, further up in the chain, which again, for me, that's always where I start with this, that the end customer really has to be the one that's benefiting
Speaker A
from the implementation of th
Speaker A
hyperscalers keep on spending so much money? Our colleagues in G .I .R. estimate that between now and the end of 2030, there's another six trillion dollars.
Speaker A
I have to remind myself, we're not talking about billions, we're talking about trillions, which is pretty significant. So two questions. One is, why do they continue to focus on that capex and making the frontier models better and better, rather than the layer that
Speaker A
you're talking about? And can the market actually absorb all of that investment in terms of equity issuance and bond issuance? Yeah, I think that latter point is absolutely critical because I think the market is ultimately going to be the forcing mechanism
Speaker A
here. Up until relatively recently, every time one of these companies would announce more higher capex, the market would reward that company. Over the last quarter or so, you've really seen a We've really seen a significant shift in that, where the market is questioning
Speaker A
that a lot more. I think that's really healthy, because I think the market is really smart. I think the market is very efficient. It makes for a safer, better market. And I think it really is going to force a discipline on the companies,
Speaker A
where they have to be a lot more thoughtful in what they're spending, and the ROI on that spending, because the market is now demanding more line of sight, more visibility on an immediate ROI. We're in a different part of the AI investment cycle
Speaker A
today than I think we were at any time over the last couple of years where the than I think we were at any time over the last couple of years. stock market and investors are demanding a lot more visibility and a lot more
Speaker A
capital discipline for companies. And I think that's a significant positive. It seems like the market is coming your way in terms of your overall thesis that we've been talking about. Let's say we look at mid -May where the price of some of these
Speaker A
stocks all peaked to the close at the end of July. Approximately S &P 500 down two, but we have the hyperscalers down around. 13 low teens.
Speaker A
We have the semiconductors down in, let's say, the higher teens. And then we have large GPU manufacturers at a much higher number, more like nearly 20%.
Speaker A
Do you think the market has discounted this imbalance in the ecosystem well enough? And now you see the hyperscalers trading at parity to the S &P 500, which they've never traded at? Or is there more downside as our clients think? about their investment
Speaker A
profile? So I see two scenarios. One is enterprises start to economically profitably implement this technology. The cash flow starts to flow back to the hyperscaler companies. I think they would outperform the semiconductor companies in that scenario. The alternative is the enterprises
Speaker A
don't see the profitable returns. The market is now forcing more capital discipline on the a little discipline on the hyperscalers. They would need to slow down under that scenario.
Speaker A
We're not talking again about, you know, companies going from $200 billion a year to zero. We're talking about just digesting some of the capacity that they've added. I think that would be very bullish for the hyperscaler companies and obviously, you know, a lot
Speaker A
more problematic for the semiconductor companies who are getting all the revenue from those hyperscalers.
Speaker A
If we look at what's going on in terms of the financing, one of the issues that we raised in our 2026 outlook from the investment strategy group was this circular finance. Outlook from the investment strategy group was this circular financing and for a
Speaker A
while that was a big focus and people say whenever you've seen vendor financing in the past usually there are some big downdrafts that will eventually follow. So how do you think about the path of this circular financing especially when credit spreads are widening
Speaker A
in the bond issuance of the hyperscapers? It was very prescient for you to write about that in the year ahead outlook piece because that's becoming a bigger and bigger issue. I do think there's two different kinds of circular financing, if you will. If
Speaker A
the company is subsidizing the investment of one of their customers in order to sort of incentivize that customer to use their chips instead of somebody else's chips, I think that's okay. That's just another form of discounting. I don't have any issue with that
Speaker A
at all. That's very different from if I don't give this customer this money, they're not going to be able to make those investments. That's much more problematic. I don't think that's the sign of a very healthy supply chain. of a very healthy supply.
Speaker A
Again, I think it's a byproduct of the fact that all of the economic value has accrued to the semiconductor companies. And as I've said multiple times, if that doesn't change, this investment cycle is not going to continue. One of the recent
Speaker A
topics that's come up a lot is the use of Chinese open source and open weight models. When you think of Chinese versus non Chinese models. What's exactly going on?
Speaker A
Can you explain that whole non -Chinese models, what's exactly going on? issue that's getting so much headline? I think what's important for me in this is I think this has become a little bit of an issue of open source is China and frontier
Speaker A
is the US. I don't actually see that as being the long term dynamic here.
Speaker A
Open source has been around forever in different flavors. Linux was an open source operating system. Red Hat on top of Linux was an open source software company. Meta's model is an open. -weight model. This isn't a China versus US thing, an open -source
Speaker A
versus frontier. I think ultimately you're going to have US open -source and open -weight as well. I don't think enterprises will ever successfully implement AI only using frontier models. I think they're going to have to use frontier models because the
Speaker A
high -consequence queries are going to demand the most powerful models, but then there will be open -source and open -weight models that can take care of a lot of the rest of the queries that are a little bit lower consequence to allow the
Speaker A
profits back into the enterprise in this equation. We don't know for a fact whether these open -source models in China were built on U .S. frontier models, but a lot of people make comments about that. The technology world will find ways to
Speaker A
replicate technology capability in cheaper forms, and it's not to me about distillation or stealing code or anything like that. I think that's just how the technology industry works and smaller, faster, cheaper is what powered technology innovation since the beginning of time. I
Speaker A
think you're going to wind up with really powerful models that are much cheaper because that's what technology does. And you get into the issue of, you know, what's how long of an advantage do you have by having that expensive frontier model versus, you
Speaker A
know, how quickly the fast followers can come with cheaper technology. I don't think that's specific to AI or frontier models. I think that's just how the technology industry. that's you know specific to AI or frontier models versus open source models at all one
Speaker A
of the big Versus open source model. concerns out there is this concept of jailbreak that these AI models go out of their sandbox and they can create harm we recently read about open AI testing one of their latest models the AI goes
Speaker A
out of its sandbox into the internet and hacks into a hugging face now some of their reaction has been oh rogue AI but others have said it's just human hubris. So how should people think about that? My big takeaway from that
Speaker A
is when people want to understand why enterprise adoption has been so much slower than maybe people thought three years ago or why it is difficult to, you know, profitably implement AI, this is why. You have to be so careful. You have to have
Speaker A
so many guardrails. You have to put things in a sandbox and even then sometimes it's dangerous. If you're a consumer, you can use this and feel If you're a consumer, you can use this. comfortable that it's not going to come steal a lot
Speaker A
of your personal data. If you're an enterprise, you can't do that. In some of our conversations with our clients, we quote Jeffrey Hinton, the so -called godfather of AI and Nobel laureate now. And initially he used to come out with very
Speaker A
strong statements. And one of them was that we have achieved immortality, just not for us. And He's just not for us, and it's all the AI models that will be immortal. More recently, he said, nobody should do any forecasts. It's just city to
Speaker A
forecast anything about AI. We just don't know how it's going to evolve. Yet, we're in the forecasting business. So how do you think about your responsibilities in guiding clients versus someone like Jeffrey Hinton saying, one shouldn't forecast anything because we don't know how
Speaker A
it's going to evolve? Excellent question. I think we have relevant history to lean on.
Speaker A
I take my responsibility in all this is to be very fundamentally driven as opposed to maybe I take my responsibility in all this is to be very... thematically driven.
Speaker A
And it's why I'm so focused on, you know, who's profiting in the supply chain and the fact that the semiconductors are the ones profiting right now. And that's not sustainable unless everybody else in the supply chain starts to make money. So I'm using
Speaker A
the current state of play to drive a lot of the forecasting as opposed to just sort of blind belief on things. I grew up covering tech stocks as the internet bubble was bursting, right? And so as a result, covering tech stocks as the
Speaker A
internet bubble was bursting. Right. And so you know, I'm going to try to draw as many lessons from that as I can. And, you know, again, I think there's going to be some great long -term winners here, but my lesson and experience from
Speaker A
the internet, it was a different group of companies that were the ultimate winners and it wasn't, you know, all the companies. Look, Google, one of the greatest companies in the world, came out of that era. There was a whole lot of companies that
Speaker A
Google was competing with in that era that nobody ever heard of again, right? And I think it's going to be era that nobody ever heard of again, right? And I think. exactly the same this time. Part of our job as forecasters is to
Speaker A
determine which companies are going to be able to rise above and which companies are going to be the winners. Again, I come back to this data management, orchestration, model optimization layers, a bunch of companies that most people have never heard of yet. I
Speaker A
think that's a key technology bottleneck that needs to be solved in order to unlock the economic value. And so that's how I try to do my forecasting in that regard. Great. Thank you. so much, Jim. It's been very helpful. Really appreciate your time.
Speaker A
Thank you so much. It's been a pleasure.
Topics:AI investmententerprise AIopen source AIfrontier modelsAI profitabilityGoldman SachssemiconductorsAI deploymentmodel orchestrationAI adoption
