[email protected]

A tide of investment is pouring into generative artificial intelligence. Will it be worth it? “That’s the center of all debate right now,” says Sung Cho of Goldman Sachs Asset Management. Investment is flowing into everything from the silicon underpinning the training of artificial intelligence models to the power companies that supply electricity to acres of data centers. To see where the industry is headed, Cho and Brook Dane, portfolio managers on the Fundamental Equity team in Goldman Sachs Asset Management, met with executives from 20 leading technology companies driving AI innovation. Those conversations — with public and private firms, from semiconductor makers to software giants — indicate that some companies are already generating returns from AI, and some would buy even more AI hardware if they could get their hands on it. “Our confidence continues to increase that this technology cycle is real,” Dane says. “It’s going to be big, as they say.” But there are also risks. Cho and Dane say it’s possible that the large language models being built by a handful of companies will find they are competing in a winner-takes-all market. The use cases, or killer apps, that fully justify the intense investment are yet to emerge. They also point out that, in a year in which US stock indexes have set successive record highs, no rally in tech stocks ever goes in a straight line. “You get these waves of both investment-digestion and hype-reality,” Dane says. “And the two of them play out across a multi-year horizon.” Cho and Dane’s insights come as Goldman Sachs Research recently published an examination of the immense amount of investment in AI, featuring interviews with Daron Acemoglu, Institute Professor at MIT, and Jim Covello, Goldman Sachs’ Head of Global Equity Research. The report, titled “Gen AI: too much spend, too little benefit?” notes that mega tech firms, corporations, and utilities are set to spend around $1 trillion on capital expenditures in the coming years to support AI. We spoke with Goldman Sachs Asset Management’s Cho and Dane about the prospects for the industry’s return on its investments, whether this trend is primarily playing out in public or in private markets, and the companies that can hope to capitalize on the AI boom. How much investment are you seeing in these models? And do you think it’s realistic to see a return on investment that justifies that capital anytime soon? Sung Cho: That’s the center of all debate right now. Brook Dane: The biggest question in the marketplace right now is: Are we getting a return on the investment? I’m reasonably comfortable that we are seeing that return. And there are a couple of data points I’m looking at that give me comfort. First: We spent a lot of time on this trip talking with the CFO of a hyperscaler who had just come back from their strategic planning, where they’re doing their one-, three-, five-year forward looks. This person talked very openly, not with any kind of numbers whatsoever, about how they were doing the RoI calculations across the clusters where they were deploying GPUs and how they were finding it very accretive from a return standpoint. Now, this company is already running massive inferencing (using already-trained AI models to reason or make predictions) workloads across their infrastructure for recommendation engines. They’re seeing results in terms of increases in time spent on their platforms, as these models have predicted, with the next piece of content. So for them, the RoI calculation is probably the simplest to calculate, because you can deploy a cluster, you can do a more sophisticated algorithm that can then lead to more time spent, which can lead to more advertising surface, which can then drive revenue. The second thing, and this is from following the industry over the long term and having had lots of recent discussions with another hyperscaler around their capital spending plans: We know how disciplined they have always been historically, and how they’re seeing both incremental revenue pick up, and seeing the incremental returns that they get out of their capital spending. This CFO is emphatic that they have the money, and if they could get more GPUs to deploy they would. Having known this person for 20 years, and understanding how they approach capital budgets, how they spend their capital — this person wouldn’t be doing that if there wasn’t a genuine, real, tangible return that they can see in front of them. And they’re pretty emphatic. But it’s early, and the other downside is that for these frontier models you can’t fall off the front end of the wave. You can’t be the fourth frontier model that doesn’t spend the incremental $1 billion dollars to get your model to be better. So for those guys, there’s a bit of an arms race here, and there’s a little bit of a leap of faith embedded in that. Sung, what’s your view on the RoI question? Sung Cho: This is one of the most important questions. And that’s what’s going to dictate the direction of markets over the next six to 12 months at least, and whether tech continues to outperform or not. Obviously with any RoI question, you have to understand the scale of what’s been invested so far. If you look at NVIDIA’s revenues in calendar year 2022, they did $26 billion dollars in revenue. And they did $26 billion dollars in revenue this most recent quarter. So in basically two years, NVIDIA has quadrupled its revenues. If you compare what’s being spent on NVIDIA to total cloud capital expenditures, nearly 50% is going into NVIDIA chips. So the investment in AI has been massive. And if you think about RoI, the starting point is around the “I,” which has been very, very high. If you are a bull, the most important thing here, and Brook mentioned this, is that right now there’s a race to see who can build the best foundational model (general purpose models that can be applied

As companies explore new generative artificial intelligence technologies, immense spending on the infrastructure to make it possible is underway. At the same time, investors are debating the extent to which reality will keep pace with the extraordinary hopes for AI technology. In Europe, some of that money is being spent to boost digital enablers — companies with technology vital to build the chips powering the data centers that run the AI models. Double digit boosts to longer-term earnings power for certain digital enablers in Europe are possible, with beneficiaries in Germany, France, and especially the Netherlands, according to Goldman Sachs Research. We talked to Goldman Sachs Research analyst Alexander Duval about the impact of the AI buildout on the European technology ecosystem. How is AI capital spending affecting the European technology sector? The sheer scale of investment on the hardware side globally is clearly having an impact on a regional basis in Europe, affecting the digital enablers — the companies needed to make the AI expansion possible. We see technology giants set to spend more than a trillion dollars on AI-related capital expenditures in the coming years. And that’s starting right now. Near-term investment levels are pretty substantial. The largest US tech companies are spending around $200 billion already this year. To put that in context, that would be equivalent to roughly a quarter of the entire European continent’s capex for all sectors. So it’s truly a phenomenal amount of money. What are they spending it on? They need to beef up their AI data centers to do the parallel processing that’s needed. Those are US companies, though, and the semiconductor foundries themselves are often in Taiwan and Korea. Where do European companies fit in? Two specific angles. First, semiconductor equipment. If you want to make semiconductors that are this powerful, you need to buy lithography machines that can print the detailed circuitry onto the chip. There’s basically only one place you can go for these extreme ultraviolet (EUV) lithography machines: You can only get them from ASML in the Netherlands. Beyond that, there are things like atomic layer deposition, a very precise way of putting materials onto wafers to build up transistors, and another company in the Netherlands does that. For all of these players, we are talking about a pretty substantial boost. Another company in this country is a leader in hybrid bonding, which is needed to make AI chips more power- and heat-efficient. All this stuff from an infrastructure perspective is kicking off already. 1. Link; 2. Based on our US team’s forecasts for US Hyperscalers; 3. Nvidia, 4. BESI, 5. Link; 6. Reuters, 7. ASML Source: Compiled by Goldman Sachs Global Investment Research What other technologies or companies may get a boost? Another aspect is the amount of power that’s required to run the AI infrastructure. That’s a major issue. A German company produces semiconductors that regulate the flow of power into the data centers, and they have some quite innovative solutions to help with efficiency. Equally we see a need for extra sensors and microcontrollers at the edge of the network, which can be supplied by French technology. More generally, what are you watching for in terms of AI technology? The challenge, if we think more broadly about AI, is that we don’t exactly know how the contact-with-reality scenario will play out. We don’t know what issues it may bring up. There might need to be regulation; there probably needs to be guardrails put around it. That’s more about the application of the technology and how we program the software, as opposed to issues with the hardware per se. What potential use cases are you watching? You can already envisage how a lot of this could be used. If you can distill the sum total of the internet into a bunch of vectors, you can basically do that for anything. You can do this for music. You can do this for proteins, so you could use AI for protein engineering. You could do this for enzymes. You could ensure that certain crops will grow better. You could generate healthcare solutions. There definitely are things that, with a bit of imagination, could be pretty interesting and easy to access. Further work is clearly required, but clearly there are enterprises looking at ways of leveraging AI to control cost in multiple domains.  How will the best uses be sorted out? Investors are keen to see that there is a return. Some companies have a sensible plan as to what can be done with this technology and a clear idea about how they will get a return on investment. I think investors are happy with that. Some other companies have just said they’re spending money on AI, up to several billion dollars, and the stocks have been selling off.  We caution in the report that we need to ensure that these use cases are coming through. Your report cites the potential for a trillion dollars of investment. What is the timeframe? Over the next handful of years, I would say. The $200 billion this year is supposed to actually go up. It’s not one-off for this year. Obviously, we have to see how these use cases take off. Can we really engineer proteins? Can we really use AI to advance carbon capture? Can enterprises truly leverage it to save meaningful cost or drive substantial productivity gains? Several semiconductor players are forecasting extraordinary AI-related revenue growth and by implication extremely large capex plans. Clearly some of the most important technologists on the planet are of the view that this is something which has got a lot of momentum for the next few years. But in our report, we look out to 2030. If we don’t see the return on investment in that timeframe, that could suggest a more bearish scenario. How does that affect the outlook for European chip companies overall? We have constructed long-term scenarios for a slate of European companies. For a number of them, we can see

Cloud computing sales are expected to rise to $2 trillion by the end of the decade, according to Goldman Sachs Research. Generative artificial intelligence is forecast to account for about 10-15% of the spending. The total addressable market for cloud services is poised to expand at a 22% compound annual growth rate from 2024 to 2030, writes Kash Rangan, head of US software research in Goldman Sachs Research, in the team’s report. Generative AI could constitute $200 billion to $300 billion of cloud spending, as investment moves beyond mega technology companies and foundation model providers. Companies spending on digital transformation and cloud modernization will contribute to the surge in cloud computing sales, writes Rangan. Only about 30% of workloads have moved to the cloud, according to a recent survey by Goldman Sachs Research. The estimate for cloud revenue growth is also based on recent historical precedent — the market more than doubled between 2019 and 2023 to $496 billion, representing a 26% compound annual growth rate. At the same time, Rangan anticipates that spending for, and adoption of, generative AI will broaden out to more companies. As that happens, it will be a further catalyst for the cloud sector. AI investment poised to broaden beyond semiconductors Much of the recent technology spending, and subsequent rise in stock prices, has centered on infrastructure companies like semiconductor makers. The next phase is expected to generate opportunities in platform companies that allow the best use of that infrastructure while providing building blocks for next-generation applications, as well as software companies that create generative AI applications, Rangan writes. “We are starting off with infrastructure and that should lead to growth in platforms, which can help manage all the data and facilitate processing by applications,” he writes. Rangan forecasts infrastructure as a service (IaaS) will account for $580 billion of the cloud market by 2030, or 29%. Platform as a service (PaaS) is expected to make up $600 billion in that same time period, or roughly 30%, while software as a service (SaaS) is expected to contribute $780 billion for 41% of the market. https://www.goldmansachs.com/infographics/v2/flourish/forecast-split/index.html?auto=1 “The infrastructure layer is poised to be the initial beneficiary, as we are already observing with the AI revenue ramp from the hyperscalers,” he writes. “This should be followed by the platform and application layers, respectively. An inherent tethering exists between PaaS and SaaS —  where PaaS solutions are needed to support the emergence of a killer application but the value in the platform layer can’t compound until more compelling applications emerge.” The next phase of generative AI Five of the biggest US technology companies are forecast to spend $215 billion on generative AI this year (up from $125 billion in 2022). But sky-high capital expenditures for generative AI are expected to gradually decline. Lowering costs isn’t entirely straightforward, as Rangan points out that some aspects of model training are relatively fixed. But the team still expects companies to eventually produce models that extract more efficiency from the hardware, for model training to give way to using the models (known as inferencing), and for smaller and specialized models to emerge. The software sector, which has had three straight years of decelerating growth, is set to potentially re-accelerate. The uptick will be driven by declining interest rates (lowering the hurdle rate for some IT projects), more certainty about economic policies after the US election in November (which has delayed some spending decisions), and key software conferences in the fall that will provide insight on generative AI products. IT budgets for generative AI are resilient Even as the investor outlook for generative AI gyrates from excitement about its prospects to skepticism about its viability, there are signs that investment in the technology is resilient. Goldman Sachs Research’s survey of IT buyers shows respondents expect 9% of their budgets to be potentially allocated to generative AI in three years, up from 7% indicated in an earlier survey in January.  https://www.goldmansachs.com/infographics/v2/flourish/future-it-budget/index.html?auto=1 The history of cloud computing can be instructive for understanding the development of generative AI. Rangan points out that it took time for the killer applications that underpin cloud computing to reach mainstream status. The return on investment for generative AI is hard to quantify, just as it was in the early stages of cloud computing. “Going through transition points like this can be very unnerving,” he writes. There are other parallels with cloud computing, Rangan writes. It took time for cloud-based applications to become more robust than their on-premises counterparts, but now they have more functionality. Likewise, once training has produced AI models of sufficient maturity, that will pave the way for more sophisticated applications.  That said, if generative AI doesn’t materialize as a disruptive force, software company valuations could rise, as there will be less competition for IT budgets, and there’s less risk of displacement, Rangan writes. But that’s not his base case. “There is a much greater probability that the generative AI opportunity is indeed real and that software applications and platform companies are able to re-invent and therefore re-accelerate growth — especially as interest rates start to come down,” he writes. “With or without generative AI, software platforms and applications companies are in a very good position to deliver attractive returns for investors over the next several years.”

Superfast chips are in high demand — not just by the artificial intelligence industry, but also by companies that work in computer graphics, robotics, autonomous vehicles, or drug discovery. “It’s fun to see all these amazing applications being created,” says Jensen Huang, the CEO of Nvidia. Speaking to David Solomon, the CEO of Goldman Sachs, at the Communacopia + Technology conference in San Francisco, Huang explained how computer graphics, for example, rely heavily on AI infrastructure. “We compute one pixel, and we infer the other 32,” he says in an edition of Goldman Sachs Talks. “Computing one pixel takes a lot of energy. Inferring the other 32 takes very little energy, and you can do it very fast. And the image quality is incredible.” Given this speed and flexibility, this infrastructure more than pays for itself, Huang says, responding to a question from Solomon about returns on investment for customers. By spending on such equipment, “the computing cost goes up a little bit — maybe it doubles,” Huang says. “But you reduce the computing time by a factor of about 20. You get 10x savings.” How Huang sees the data center market Chips that accelerate computing are everywhere, but there is no such thing as a universal accelerator, Huang says. Instead, every time a chip company enters a new market, it must learn new algorithms. They differ according to purpose; the algorithm for image processing would be different from the algorithm to model fluid dynamics. “Usually, some 5-10% of the code represents 99.999% of the run time,” Huang says. “So if you take that 5% of the code and offloaded it onto an accelerator, then technically you should be able to speed up the application a hundred times.” The promise of this kind of accelerated computing has led to keen investor interest in the data center market, Huang says. He thinks this infrastructure can yet be improved. For one thing, the average data center is “super-inefficient, because it’s filled with air, and air is a lousy conductor of electricity.” Making data centers denser — eliminating the air, in other words — will make them cheaper and more energy efficient. Another revolution lies in how data centers now understand not just how to process data but the meaning of the data itself, and how to translate one form of data to another, Huang says: “English to images, images to English, English to proteins, proteins to chemicals.” The chip supply chain needs to be resilient The ecosystem of manufacturers and suppliers to the chip industry is sprawling and complex, and particularly concentrated in Asia. As a result, Nvidia tries to design diversity and redundancy into every aspect of its supply chain. Companies need to have “enough intellectual property” to be able to shift their manufacturing from one “fab” — or chip-making facility — to another if they have to, Huang says. “Maybe the process technology won’t be as great, or you won’t get the same level of performance or cost, but you will still be able to provide the supply.”

With the world’s biggest companies racing to build the most sophisticated artificial intelligence, questions about how far and how fast the technology can be scaled are coming into focus, according to Goldman Sachs Research.  The advent of generative AI has created a surge of excitement about the future of the technology. Unlike other types of AI, gen AI can create its own outputs in natural language. Because it’s “multimodal,\” it can also generate responses in other formats, including text, numbers, videos, and sound. Large tech companies are citing initial use cases, and some enterprise players see scope for major productivity gains in certain areas such as code writing, which could make the most valuable employees even more productive. However, the exact size of future economic benefits is the subject of debate. The cost of building gen AI at scale is extremely high, with big tech companies investing hundreds of billions of dollars — although the cost per query has come down considerably since the technology first launched. At the inaugural Goldman Sachs European Virtual AI and Semis Symposium, 20 speakers — from CEOs and technologists to macro economists — came together to assess the prospects for AI. In particular, they discussed key topics including use cases, the total addressable market, challenges for further development and adoption of the technology, and implications for European hardware and semiconductors. We talked to Alexander Duval, head of Europe Tech Hardware & Semiconductors in Goldman Sachs Research, about the main findings of the symposium. What role are the large tech companies known as “hyperscalers” playing in the development of AI? So far, US tech giants have been at the vanguard of generative AI use. They\’ve been developing large language models, which they can use both for their existing business and also potentially in creating new business tools. The symposium heard how the technology is generating a quarter of one hyperscaler’s code and saving meaningful engineering time for another. Broader use cases in the real economy include its use to predict the structure of proteins, and even to “de-age” an actor\’s appearance in a movie. Those are striking examples. But it\’s worth bearing in mind that hyperscalers have been spending hundreds of billions on this. Together, they have spent around $200 billion on AI this year, and that will probably increase to $250 billion next year. Developing large language models can cost tens or hundreds of millions of dollars. And that\’s why at this symposium, we really wanted to look at whether it\’s feasible or desirable that the technology could scale to address many more use cases. These hyperscalers have a lot of free cash flow, and we are starting to see examples of use cases, but a number of industry observers believe that at some point we need to see a return on investment for a broader array of use cases and verticals. Have any key use cases emerged for artificial intelligence in the broader economy? Because generative AI is multimodal, it could theoretically apply to multiple fields: customer support, coding, medical analysis, marketing and many others. Given that there is a very significant level of investment in AI, the aggregate benefit of such use cases will need to be demonstrated in order to justify a solid return on investment. That being said, some participants at the symposium said that it might not be imperative for AI to scale in one particular area — in other words, a single key use case may not be necessary — as long as the economic benefits from all the different use cases are sufficient in aggregate. You could see efficiency gains across the board. Some speakers pointed out there are a number of examples of very large successful tech businesses where you could argue that there wasn\’t a key use case at first. Take the example of ride hailing apps. There was already a perfectly good solution: Walking to the end of the street and hailing a taxi physically. But by leveraging software and network effects, you could create very large economic benefits, as well as benefits to consumers. Is there still room for smaller technology companies to compete? Some speakers at the symposium had interesting insights on small language models. At first, technology players were focused on building large language models — and those are still important. But there is also a trend of developing smaller and more efficient models. Small language models are easier to fine tune, they may have lower energy consumption, and they can be customized to meet an enterprise\’s specific requirements in a given domain (such as legal, medicine, or finance). They’re also generally less expensive, because they\’re smaller and use less power. Large language models will remain important, and tech behemoths have the resources, free cash flow, and balance sheets to drive the development of those. But speakers pointed out that there will be other, perhaps smaller players in the ecosystem who can innovate and develop small language models that will sit on top of those larger models. Some speakers thought this presented an opportunity for small companies to drive innovation at the top of the stack and highlighted the large number of companies being founded daily to do so. Could the high cost of generative AI hold back development? Training LLMs requires very high levels of capital investment. You need to build a data center, you need all the semiconductors — that includes both GPUs and memory chips — and you need hardware, power, and utilities. Speakers mentioned that the cost per query in some domains is multiple times higher than for a regular search algorithm.  That said, there has been steady progress on reducing costs. The cost of a generative AI query at some large tech companies has come down significantly since the launch of the technology, and a new gen AI company has said that revenue generated by the latest generation of LLMs exceeded the cost of training prior models. While some speakers stated there could be a risk that spending on AI could reduce if significant returns

More than two years after ChatGPT’s debut, Goldman Sachs Chief Information Officer Marco Argenti says the potential for generative artificial intelligence is coming into focus. The development of increasingly powerful large language models (LLMs), supercharged by advances in robotics will, in Argenti’s view, begin to bring sweeping changes to everything from employment to regulation of the technology in 2025. Argenti, the former vice president of technology of Amazon Web Services, makes five predictions about how AI could evolve and interact with businesses and society in the near future: The new hybrid workforce: AI systems are becoming more like humans. So why not employ them like humans? The capabilities of AI models to plan and execute complex, long-running tasks on humans’ behalf will begin to mature. This will create the conditions for companies to eventually “employ” and train AI workers to be part of hybrid teams of humans and AIs working together. The question becomes: How best to get humans and AI to work together. Companies will reskill human managers to oversee a hybrid workforce. The role of human resources will evolve into a department for human and machine resources. The first AI “layoffs” could eventually emerge, in which AI models will be replaced by better AI tools or humans if they perform poorly compared to their peers. The emergence of expert AI: The AI version of PhDs will arrive. Companies will integrate AI with their proprietary data, either with retrieval-augmented generation (RAG) — an architecture that can connect LLMs to external, specialized datasets — or via a process known as fine-tuning, which involves enhanced training of an LLM with a smaller, specialized dataset. As a result, expert AI systems, or large expert models, will gradually emerge with advanced capabilities and industry-specific knowledge — for example, specialized models for medicine, robotics, finance, or material sciences. Robotic breakthroughs powered by AI: So far, AI models have been trained by reading essentially all the books in the world. What if they\’re trained on the world itself? Children learn to walk before they learn to read. In the same way, the intersection of LLMs and robotics will increasingly bring AI into, and enable it to experience, the physical world, which will help enable reasoning capabilities for AI. At the same time, these models will transform commodity hardware into specialized components capable of performing far beyond their default capabilities. Advanced cameras using cheap sensors, studio-quality microphones using low-cost transducers, and off-the shelf mechanical joints capable of performing complex manipulation tasks will drive down costs for combining advanced AI with robotics and will speed up innovation. Regulation goes from global to local: As the world awaits regulatory clarity, responsible AI principles will take center stage for CEOs and boards.   In addition to (and somewhat separate from) national, state, or sectoral regulations, companies across sectors will continue to see the benefit of implementing proper controls, such as responsible AI principles (i.e., a form of self-regulation). Responsible AI will become an even bigger priority for CEOs and boards of major companies. Large model consolidation: The Formula One experience arrives for AI. Given the cost and complexity of engine development in the Formula One motorsport competition, there are many cars but only a few engine makers. Likewise, the investment required to train and maintain large frontier models (those that are the largest and most advanced) for AI will eventually result in only a handful of providers. Consolidation will mirror what has taken place in cloud infrastructure, databases, and operating systems, where the total number of companies developing large AI engines will be countable on one hand. Startups that are now “model-centric” will shift towards building solutions that are model-agnostic, focusing instead on other aspects of AI such as compliance, safety, data integration, orchestration, automation, and user experience.

Nuclear power will be a key part of a suite of new energy infrastructure built to meet surging data-center power demand driven by artificial intelligence. But nuclear can’t meet all of the increased data-center power needs. Natural gas, renewables, and battery technology will also have a role to play, according to Goldman Sachs Research. Several big tech companies looking for low carbon, round-the-clock energy signed contracts for new nuclear capacity in the last year, and there could be more such deals ahead. Those efforts come as electricity usage by data centers is expected to more than double by 2030, according to reports led by Brian Singer, Jim Schneider, and Carly Davenport.  In total, the team forecasts 85-90 gigawatts (GW) of new nuclear capacity would be needed to meet all of the data center power demand growth expected by 2030 (relative to 2023). But well less than 10% will be available globally by 2030. As power needs ramp up, the efficiency gains of data center infrastructure are beginning to slow, according to Davenport, a US utilities research analyst in Goldman Sachs Research. “Growth from AI, broader data demand, and a deceleration of power efficiency gains is leading to a power surge from data centers,” she writes. How much is AI power consumption expected to rise? Power demand from data centers is on track to grow more than 160% by 2030, compared to 2023 levels, Goldman Sachs Research projects. A scenario in which 60% of that increased demand was met by thermal sources such as natural gas would lead to an expected emissions increase of 215-220 million tons globally, equivalent to 0.6% of the world’s energy emissions. While renewables have the potential to meet most of the increased power needs from data centers at some times of day, they don\’t produce power consistently enough to be the only energy source for data centers, explains Schneider, a digital infrastructure analyst in Goldman Sachs Research. “Our conversations with renewable developers indicate that wind and solar could serve roughly 80% of a data center\’s power demand if paired with storage, but some sort of baseload generation is needed to meet the 24/7 demand,” Schneider writes. He adds that nuclear is the preferred option for baseload power, but the difficulty of building new nuclear plants means that natural gas and renewables are more realistic short-term solutions. Nuclear energy has almost zero carbon dioxide emissions — although it does create nuclear waste that needs to be managed carefully. But the scarcity of specialized labor, the challenges of obtaining permits, and the difficulty of sourcing sufficient uranium all pose a challenge to the development of new nuclear power plants. By the 2030s, though, new nuclear energy facilities and developments in AI could start to bring down the overall carbon footprint of AI data centers. In the meantime, companies trying to supply energy for new data centers are likely to focus on a mix of power sources, writes Singer, global head of GS SUSTAIN in Goldman Sachs Research. “Our outlook on power demand growth warrants an ‘and’ approach, not an ‘or’ approach, as we see ample opportunities for generation growth across sources,” he writes. How much will nuclear power increase? Recent contracts for nuclear energy facilities along with signs of countries’ greater appetite for nuclear power suggest a significant increase of investment in the next five years, and a corresponding rise in power supply in the 2030s. The proliferation of AI data centers has boosted investor confidence in future growth in electricity demand at the same time as big tech companies are looking for low-carbon reliable energy. This is leading to the de-mothballing of recently retired nuclear generators, as well as consideration for new larger-scale reactors. In the US alone, big tech companies have signed new contracts for more than 10 GW of possible new nuclear capacity in the last year, and Goldman Sachs Research sees potential for three plants to be brought online by 2030. Meanwhile, governments are also broadly more supportive of nuclear power. Switzerland is reconsidering the use of nuclear generators for its electricity supply, while nuclear power enjoys bipartisan support in the US, and the Australian opposition party has put forward plans to introduce nuclear reactors. Participants at the COP28 conference in late 2023, an annual summit convened by the UN to address climate change, agreed to triple global nuclear capacity by 2050. Building the ‘green’ data center Green energy sources are also receiving considerable investment from AI providers. The team forecasts that 40% of the new capacity built to support increased power demand from data centers will be renewables. The supply cost of renewable energy sources is cheaper than generating electricity from natural gas, before taking into account transmission considerations and filling gaps when the sun isn’t shining and the wind isn’t blowing. Analysis by Goldman Sachs Research shows that, at face value, the average cost of energy for onshore wind hosted on the site of a data center is $25 per megawatt hour in the US, while solar energy costs $26/MWh, and combined cycle natural gas (the most fuel-efficient type of gas-fired power plant) costs $37/MWh before accounting for the cost of carbon capture. In practice, though, utility-scale solar plants only run around 6 hours per day on average, while wind plants run for an average of 9 hours per day. There is also day-to-day volatility in the capacity of these sources, depending on the radiance of the sun and the strength of the wind. Transmission costs are also a consideration for data center operators. Because renewable energy sources often take up a much greater land footprint than natural gas or nuclear, they are more likely to be located away from big cities, where much of the energy that they generate is used. As a result, the energy they generate may have to travel further before it is used. On the other hand, thermal plants — such as those powered by nuclear reactors or combined cycle natural gas — can run throughout the day, without hourly intermittency challenges.

The launch of three new Chinese generative artificial intelligence models sent shockwaves through the tech sector. The share price of a basket of AI-related stocks in the US dropped 10% in the first two days of this week. The development highlights the sheer amount of capital that’s been invested in AI by US technology giants so far, as well as the investment that will be needed to scale the technology going forward, according to Goldman Sachs Research. It remains to be seen exactly how some of the new models were trained, with emerging questions on sources of data. But there are also signs that the developments in China could lower the cost of running chatbot apps, which can be used for everything from coding software to writing sonnets, and make them more widely available. “What’s clear to us is that lowering the cost of AI models will drive much higher adoption, as it would make the models much cheaper to use in future,” says Ronald Keung, the head of Goldman Sachs Research’s Asia internet team. “Some of these Chinese models have driven the industry to focus not just on raising the performance, but also on lowering the cost.” We spoke with Keung about his take on the new generative AI models in China, the cost savings they may provide, and breakthroughs they may enable. As costs decline and AI models become smarter, Keung says, we may be a small step closer to reaching artificial general intelligence — an AI that displays excellence across all human fields of knowledge. What is important about the recent AI developments in China? Three Chinese AI models were launched last week, as well as two multi-modal text-to-image models this week. And while most of the attention has been on DeepSeek’s new model, the other models are at around the same level in terms of performance and cost per token (a token is a small unit of text). The cost of inferencing (the stage that comes after training, when an AI model works with content that it has never seen before) has fallen by more than 95% in China over the past year. We expect this much lower inferencing cost to drive a proliferation of generative AI applications. Some of the models launched over the past week are focused on deep thinking modes or reasoning. That means that the chat bot goes through each of its steps when you ask a question, telling you what it’s thinking before it arrives at an answer. That takes around 5-20 seconds for every question. The process makes sense when you look at how human beings interact — if you ask me a question, and then I give you an immediate answer in milliseconds, then the chance is that I might not have thought it through. These models think before they speak. The performances of these models seem to have improved a lot as a result. It\’s mostly because they assess their own answers before giving a final output. Are these developments likely to change the way that capital is invested in AI? Chinese players have been focused on driving the lowest cost, and also maybe trying to use minimal chips in doing the same tasks. I think over the last week, there\’s also been more focus on whether edge computing is becoming more popular, which could allow smaller AI models to run on your phone or computer without connecting to mega data centers. I think these are all questions that investors have on how the landscape will evolve. What is clear to us is that lowering the cost of AI models will drive much higher adoption, as it would make the models much cheaper to use in future. Both our research teams in China and our US teams expect this year to be the year of AI agents and applications. The good news is that some of these Chinese models have pushed the industry to focus not just on raising the performance, but also on lowering the cost. That should drive higher and higher adoption of artificial intelligence. How much cheaper are the AI models in China relative to the incumbent AI providers in the US? When it comes to how much the companies charge per use of the model, which is measured on a per-token basis, the charges are significantly lower. As of last weekend, a Chinese AI model’s pricing was 14 cents per million input tokens. That’s only a single-digit percentage of the amount that an equivalent reasoning model from a large US technology company charges. It’s clear that prices are starting to come down as a result. Already, we’ve seen some US big tech companies adjust their pricing, including making some of their paid models free. So I think there will be a continuing race on efficiencies. Your colleagues from the US tech team have said that the market reaction is being driven by questions around the amount of capital expenditure needed to scale AI, the return on investment on the money already spent, and the pace of investment going forward. What is your take on those questions as it relates to Chinese firms? I think that applies a lot more to US companies. The major Chinese internet giants increased capital expenditure by 61% last year — but that was from a low base. The listed Chinese companies have not been spending as much on capex over the last two years, in aggregate. Instead, they’ve been very focused on shareholder returns. Spending has only just started to pick up for these companies in 2024. However, in absolute terms, Chinese Internet companies have been spending just a fraction of what their global counterparts have been spending, so there are fewer questions on the return-on-investment that they can expect from high spending on AI. Investor focus on AI has been concentrated on the US. Do you think this will mean more investor interest in Chinese companies in the future? It\’s a bit early to tell. Overall, the China Internet base is

The explosion in interest in generative artificial intelligence has resulted in an arms race to develop the technology, which will require many high-density data centers as well as much more electricity to power them. Goldman Sachs Research forecasts global power demand from data centers will increase 50% by 2027 and by as much as 165% by the end of the decade (compared with 2023), writes James Schneider, a senior equity research analyst covering US telecom, digital infrastructure, and IT services, in the team’s report.   Recent Chinese developments, and particularly the AI model known as DeepSeek, have raised concern about the returns on current and projected AI investment. Still, several questions remain about DeepSeek’s training, infrastructure, and ability to scale. “In the long run, if we see efficiency driving lower capex levels (from either hyperscalers or new investment plans from new players), this would mitigate the risk of long-term market oversupply we see in 2027 and beyond – which we think is an important consideration that could drive more durability and less cyclicality in the data center market,” says Schneider. On the demand side for data centers, large “hyperscale” cloud providers and other corporations are building large language models (LLMs) capable of natural language processing and understanding. These models must be trained on vast amounts of information, using power-intensive processors. On the supply side, hyperscale cloud companies, data center operators, and asset managers are deploying large amounts of capital to build new high-capacity data centers. Taken together, the balance of data center supply and demand is forecast by Goldman Sachs Research to tighten in the coming years. The occupancy rate for this infrastructure is projected to increase from around 85% in 2023 to a potential peak of more than 95% in late 2026. That will likely be followed by a moderation starting in 2027, as more data centers come online and AI-driven demand growth slows. How much power will data centers require for AI? At present, Goldman Sachs Research estimates the power usage by the global data center market to be around 55 gigawatts (GW). This is comprised of cloud computing workloads (54%), traditional workloads for typical business functions such as email or storage (32%), and AI (14%). By modeling future demand for each of these workload types, our analysts project power demand will reach 84 GW by 2027, with AI growing to 27% of the overall market, cloud dropping to 50%, and traditional workloads falling to 23%. This baseline scenario could, however, be affected by a deceleration in usage by AI — for example, if the transition to AI-driven work and AI monetization doesn’t develop as quickly as anticipated. In such muted scenarios, demand could diverge from the baseline estimate by 9-13 GW. The global landscape of data center supply The current global market capacity of data centers is approximately 59 GW. Roughly 60% of this capacity is provided by hyperscale cloud providers and third-party wholesale data center operators (these providers usually have a small number of very large enterprise customers). The remaining belongs to more traditional corporate and telecom-owned data centers. The AI-dedicated data center is an emerging class of infrastructure. Although very few exist so far, they’re designed for the unique properties of AI workloads — high absolute power requirements, higher power density racks, and the additional hardware (such as liquid cooling) that comes with it. They’re usually owned by hyperscalers or wholesale operators. Regionally, Asia Pacific and North America have the most data center power and square footage online today — most notably in regions such as Northern Virginia, Beijing, Shanghai, and the San Francisco Bay Area. These places have high compute and data traffic as well as robust corporate campus demand. How data center supply is rising around the world Goldman Sachs Research estimates that there will be around 122 GW of data center capacity online by the end of 2030. The mix of this capacity is expected to skew even further towards hyperscalers and wholesale operators (70% versus 60% today). Although Asia Pacific has added the most supply over the past ten years by a wide margin, North America has the most scheduled capacity coming online over the next five years Given the higher processing workloads demanded by AI, the density of power use in data centers is likely to grow as well, from 162 kilowatts (kW) per square foot to 176 kW per square foot in 2027. (These figures exclude power overheads such as cooling or other functions related to data center infrastructure.) “Data center supply — specifically the rate at which incremental supply is built — has been constrained over the past 18 months,” Schneider says. These constraints have arisen from the inability of utilities to expand transmission capacity because of permitting delays, supply chain bottlenecks, and infrastructure that is both costly and time-intensive to upgrade. Power demand from data centers will require additional utility investment As data centers contribute to a growing need for power, the electric grid will require significant investment. Goldman Sachs Research estimates that about $720 billion of grid spending through 2030 may be needed. “These transmission projects can take several years to permit, and then several more to build, creating another potential bottleneck for data center growth if the regions are not proactive about this given the lead time,” Schneider says. In Europe too, a data center-led surge in power demand is under way, after 15 years of decline in the power sector. Having surveyed utilities across the continent, Goldman Sachs Research found that the number of connection requests received by power distribution operators (a leading indicator of future demand) has risen exponentially over the past couple of years, mostly driven by data centers. \”We estimate a potential 10-15% boost to Europe’s power demand, over the coming 10-15 years,” Alberto Gandolfi, head of the pan-European utilities team for Goldman Sachs Research, writes in a separate report. Goldman Sachs Research estimates a data center pipeline for Europe amounting to about 170 GW, equivalent to about one-third of the region\’s power consumption. “It is unclear

The emergence of new artificial intelligence models in China could drive faster development and adoption of the technology in the country than previously projected, according to Goldman Sachs Research. Should AI gain traction, it could boost productivity and GDP growth in the world’s second-largest economy. China’s AI advancement has accelerated since our researchers examined, in 2023, the potential impact of generative AI on the country’s economic growth. The number of foundation models, which represent cutting-edge AI research, climbed to 20 by the end of that year, surpassing the combined total of the EU and UK. The release of DeepSeek’s model, which may have been developed at lower cost than other leading models, suggests a faster adoption rate and greater economic upside for China than previously anticipated, Goldman Sachs Research economists Hui Shan, Joseph Briggs, and Xinquan Chen write in the team’s report. Faster adoption of generative AI in China could translate into lower labor costs and higher productivity as more tasks are automated.  Goldman Sachs Research now estimates that generative AI will start raising potential growth in China by 2026 and provide a 0.2-0.3 percentage point boost to China’s GDP by 2030, up from 0.1 percentage point previously. The team adjusted its estimate for the potential uplift to GDP but didn’t change its forecast for real GDP growth “given the uncertainty associated with the future path of AI development and adoption.” How will AI affect China’s economy? Our researchers’ initial forecasts assumed China would see a similar AI boost as other advanced emerging market economies, implying only a 10-20% rate of adoption by 2030 before peaking in the mid- to late-2030s. However the AI development in China over the past 12-18 months and the recent DeepSeek breakthrough suggest the timeline for adoption could be faster. Goldman Sachs Research now expects AI adoption rates in China to exceed 30% by 2030, peak in the early 2030s, and achieve full adoption within the next 15 years. If that occurs, AI adoption in China would track more closely with the experience of developed economies than those in emerging markets.   While AI adoption in China may happen faster than expected, our economists think the impact may be front-loaded, and the overall effect could be slightly smaller than previously predicted. Compared to the US, the Chinese labor market is less prone to AI automation, due to its higher share of physical jobs. For example, agriculture, manufacturing, and construction trades account for about half of all jobs in China, well above their 19% share of total employment in the US. Meanwhile, sectors that are more exposed to AI-driven task automation — such as finance and insurance, and professional and technical services — constitute less than 3% of jobs in China, compared with 14% in the US. After incorporating more detailed industry-level employment data in China, Goldman Sachs Research nudged down its forecasts for AI’s likely uplift to overall GDP over the next 10 years from 9% to 8%. That’s well below the expected 15% lift in the US. “Although the total effect after full AI adoption is marginally lower than our previous estimate, the impact over the next few years is likely to be more positive due to the faster adoption timeline,” the team writes. How much is China investing in AI? Chinese companies’ ability to improve performance while reducing costs and computing power requirements will continue to boost capital expenditures and investment, according to Goldman Sachs Research analysts. The race to develop AI agents and applications is resulting in higher investment throughout the tech ecosystem in China — including semiconductors, data centers, cloud services, software, and telecommunication companies. Our equity analysts expect China’s four biggest AI model developers to increase their total capex by 38% in 2025. AI-related spending by tech companies is expected to increase sharply over the next few years as they build up AI infrastructure, platforms, and applications. And as AI applications are developed and mature, the non-tech sector is likely to increase spending to incorporate end-use AI services into regular production later in this decade. Our economists expect total AI-related spending will account for close to 1% of China’s annual GDP in the coming years.  How will AI affect China’s labor market? The main economic impact from generative AI is expected to come from task automation that raises productivity. Our economists point out that implementation of these new technologies will need to be managed carefully, particularly given the current weak state of China’s labor market and persistent deflationary pressures. Some 15% of young people are unemployed, and in recent years more than 10 million students have graduated from colleges annually. “Amid a severe housing downturn, increased efforts to rein in local government implicit debt, and regulatory tightening in the financial industry, job losses have been reported in the real estate sector, among civil servants, and in the financial sector,” the team writes. “In this environment, job destruction by AI adoption, while raising labor productivity, could exacerbate deflation, erode confidence, and further weaken the economy.” In the near term, the race for AI development could boost tech sector employment modestly, while the lack of mature AI applications implies minimal job displacement in the non-tech sector. Then, a few years down the road, the number of displaced workers may jump in both tech and non-tech sectors as AI becomes ready for mass adoption. The experience of other productivity-enhancing technology cycles suggests it may take several years for displaced workers to eventually find jobs in other sectors. But eventually, employment may grow again in the last stage of the labor market adaptation to AI adoption. That said, AI could help China reckon with a population that’s expected to age more quickly than other major economies. The country’s working-age population (15-64-year-olds) may contract by 25% over the next 25 years. “AI and robotics could provide an answer to the aging society,” according to Goldman Sachs Research.  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,