Reflections on an AI Road Trip
The pace of innovation around GenAI continues to accelerate and, as investors, we believe it is critical to hear unique insights from corporate leaders at the forefront of change. With that in mind, we recently travelled to the US West Coast to meet with executives from twenty public and private market technology companies across the market cap spectrum. Some of these businesses are building GenAI models or manufacturing the semiconductors required to run AI workloads, others are software companies hosting AI workloads or embedding GenAI into their products.
Here are our takeaways from the trip:
GenAI is Driving a Durable Tech Cycle, but is Still at an Early Stage
One of our biggest takeaways is how early we are in the development of frontier models and the investment needed for GenAI infrastructure. Frontier models are the most advanced and capable large language models (LLM) being developed currently. The companies leading the development of frontier models include Google, OpenAI, Anthropic, and Meta, with Gemini, ChatGPT, Claude, and Llama models, respectively. Each of these companies highlighted that their GenAI capabilities are advancing rapidly. However, they also emphasized that their models are still at an “undergraduate” level, rather than a “masters” or “PhD” level, indicating that we are still in the very early stages of frontier model development. GenAI has had a clear step change in capability roughly every 12-18 months and we believe this will continue. The key performance indicators for measuring frontier model performance center around intelligence and latency. Both have improved, not only due to advances in processing power, but also because the software used to train the models has become more efficient. NVIDIA believes that its graphic processing units (GPUs) will be one million times more efficient at GenAI processing over the next decade using the same kind of chip infrastructure, leading to further acceleration of frontier model development.1
Despite being at an early stage, frontier model development is moving extremely quickly. These models are integrating and embedding much of the functionality that initial AI start-ups had offered through application program interfacing on the earlier, less capable models. This has slowed due to rapid advancement in frontier models, which is an important factor leading us to believe that the largest potential investment opportunity is concentrated in the public versus private markets.
Investors Are Focused on the Return on Investment for Hyperscalers
A key focus as we met with these companies was determining expectations for return on investment (ROI). In order to create the most advanced and capable models possible, Microsoft, Google, Amazon, and Meta (the “hyperscalers”) are investing $188bn in 2024 to obtain the necessary hardware for the buildout of next-gen data centers capable of training GenAI, driving most of the S&P 500 capital expenditure for 2024.2 Any firm engaging in frontier model development needs vast resources as well as the willingness to spend on GenAI development. This leads us to believe that we will not see a wave of developers emerge with competitive frontier models.
Through numerous interactions with their management teams, we know firsthand how disciplined these companies are with their capital. A CFO at one of these enterprises was clear that they would obtain more GPUs if they could—despite the substantial cost. The escalating cost of GPUs is driven by the race among frontier model providers to build the most capable framework, and we don't expect that race to slow down any time soon. From an ROI perspective, while the hyperscalers are beginning to see incremental revenues from GenAI”, meaningful progress may be a year or two down the road, and this is what the market is grappling with.
Early use cases of GenAI include developers leveraging the technology for coding, ad placements, and enhancing productivity for sales functions. In addition, Microsoft may offer a premium version of its Office suite embedding a co-pilot feature; Google is integrating GenAI into its leading search engine and Gmail to enhance users’ experience; and Meta’s social media offerings could integrate GenAI to evolve from a content curating platform to a content generating platform. We believe these enhancements could lead the way to an acceleration in fundamentals and competitive advantages.
Semis Powering the Next Generation of GenAI Processing Are Likely to Broaden Beyond GPUs
While NVIDIA’s GPUs have enjoyed near total market share in the GenAI training market due to their unparalleled processing power and software ecosystem, we believe other semiconductor manufacturers are also poised to potentially benefit. Most notably, the hyperscalers have developed their own application-specific integrated circuits (ASICs) which, while not as powerful, are effective at running repeatable tasks in high volumes.
For example, ASICs are used in Google’s recommendation infrastructure for YouTube because the task is narrow and repeated at high volume. ASICs, which are purpose-built for specific workloads, can perform that task much more efficiently and at significantly lower cost than a powerful NVIDIA GPU. While the upfront cost to build ASIC infrastructure is high, the cost to run GenAI workloads on these chips will likely be lower than once the upfront investment is made. The hyperscalers are also uniquely positioned to spend on the software needed to support ASIC infrastructure for both training and inference. We are focused on identifying the potential winners in the semiconductor industry that may benefit from this transition.
In our view, semiconductor capital equipment companies also have a strong demand backdrop driven by AI investment. This is primarily due to the increased sophistication and demand for AI related semiconductors. We believe total Wafer Fab Equipment (WFE) spend could exceed $130bn in 2025, up nearly 40% from 2023 levels.3
Embedding Distilled Models On-Device Could Drive a Phone Replacement Cycle
The potential for GenAI to exist on the “edge” (accessing AI capabilities locally versus through cloud technology) is top of mind. This would be done with “distilled models,” which are derivatives of frontier models that have fewer parameters but are very capable for many consumer use cases. During our trip, we had the opportunity to observe live demos of GenAI models integrated into voice chat assistants, a significant improvement versus current offerings. We are focused on understanding the impact of GenAI capabilities existing on the edge and whether that could initiate a new replacement cycle for personal phones. This would represent one of the first and most tangible examples of GenAI having an impact on revenues and earnings for companies integrating GenAI technology into new device offerings.
One of the technical challenges with distilled models is how to handle queries that are too sophisticated to be completed on the edge. Distilled models need to be smart enough to send any sophisticated queries to a data center with more processing power. We are still working through what that might look like for the companies potentially sending or receiving sophisticated queries.
The First Quarter for Software was Cyclical, not Secular
During solid revenues during the first quarter of 2024, we saw software companies maintain—rather than raise —their full-year guidance. The market took this as an indication that GenAI could potentially disrupt these companies or that their customers were spending less on software to compensate for higher spending on GenAI integration. As a result, we believe that the slight softness in software companies’ first quarter results was likely cyclical, rather than an indication of a secular decline.
As opposed to being at risk of disruption by frontier models, we believe that many of the biggest software companies are more valuable in the context of frontier model development, as they have vast amounts of proprietary customer data. Indeed, there are many potentially exciting GenAI integration initiatives in the works among the largest software companies. We may also see partnerships develop between enterprise software companies that have valuable datasets and frontier model providers, creating clear advantages for the providers that can train their frontier models on the highest quality data.
Our Conviction in Cybersecurity Remains High
Cybersecurity continues to be a top priority for both governments and corporations, and we maintain that demand for the most innovative solutions will remain robust. Geopolitical tensions and political uncertainty remain elevated. With more than 50 countries—home to half the world’s population—holding national elections in 2024, we are seeing an increase in advanced cyber threats and, as a result, growing investment to protect against these threats. Effective cybersecurity is also critical to GenAI, given the huge amounts of data involved in training frontier models. As corporations and governments try to keep pace with the evolving threat landscape and secure their proprietary data, next generation cybersecurity providers have opportunities to help clients protect their most valuable assets.
In addition to an increase in the volume of cyber threats, technological advancements have led to an increase in the sophistication of attacks. GenAI not only creates new threat vectors, but also democratizes complex and damaging techniques. The increase in volume and sophistication of cyberattacks has led to an increase in spending on cybersecurity solutions, underscored by the critical need for organizations to employ the most robust and comprehensive cybersecurity solutions.
Sovereign AI Development is Accelerating and China’s Capabilities Have Advanced Faster than Expected
There is a global race to build and deploy GenAI, not just at the corporate level, but also among sovereigns. Governments around the world are increasingly recognizing the transformative potential of the technology—with applications across national security and governance, in addition to significant potential gains in labor productivity—driving an increase in government-backed funding and strategic investments. This is a clear tailwind to the GenAI buildout cycle on a global scale and gives us additional confidence in the durability of this tech cycle. We believe localized AI infrastructure (or AI factories) will be critical to countries looking to develop AI capabilities, but which do not want to rely on US-based data centers and the need for multilingual GenAI models (large language models) in different geographies.
China has been able to develop its domestic AI semiconductor sector much more rapidly than the market had expected, despite the US government imposing tight restrictions on the export of leading-edge semiconductors to China, used to develop GenAI. The speed at which China has scaled its capabilities has taken many by surprise and leads us to believe that China will continue to develop its own technology ecosystem, which will exist in parallel with, but operate independently from, the one being developed by the West.
The Road Ahead
We believe GenAI is one of the most profound technological innovations of our lifetime. From an investment perspective, we also believe it may drive a durable and significant tech cycle. In our view, we are still at the very early stages of this cycle and, as we look forward, we believe the opportunity set is likely to broaden out, creating new winners and losers.
Given the significant investment required to develop AI capabilities and the need for vast quantities of high-quality data, we continue to believe that the key beneficiaries of GenAI transformation will be found in the public market, rather than the private equity market. While the so-called “Magnificent 7” 4 companies have been some of the most immediate beneficiaries in terms of share price appreciation, we believe that other companies further down the market-cap spectrum may also emerge as GenAI beneficiaries. The space is dynamic and changing rapidly. As GenAI technology advances and new frontrunners emerge, investing in the right companies will be critical to long-term investment success.
Goldman Sachs Asset Management’s Fundamental Equity Technology team conducts in-depth analysis on industry trends and companies across the market cap spectrum, and meets regularly to discuss market dynamics and portfolio holdings. The team also collaborates with the broader Fundamental Equity team, consisting of 100+ research analysts globally.
1 NVIDIA. As of June 6, 2024.
2 MSCI, Wind, Bloomberg, FactSet, Goldman Sachs Global Investment Research. As of May 24, 2024.
There is no guarantee that objectives will be met. The economic and market forecasts presented herein are for informational purposes as of the date of this presentation. There can be no assurance that the forecasts will be achieved. Please see additional disclosures at the end of this presentation.
3 Goldman Sachs Asset Management, Bloomberg, Visible Alpha. As of June 30, 2024.
4 Refers to Microsoft, Apple, Google, Amazon, NVIDIA, Meta, and Tesla.