Who will win the AI race? Goldman Sachs sees $7.6 trillion battle beyond Big Tech

Goldman Sachs’ Harnessing AI for the Real Economy report gives a different way of looking at the competition.

The report does not declare the United States, China, Europe or any individual company the winner of the AI race. Instead, it says that the next phase of competition will be decided by a much wider set of capabilities. Technology will matter. However, technology alone will not be enough.

The companies, sectors and countries that emerge strongest from the AI transition will also need capital, infrastructure, power, proprietary data and the ability to deploy AI inside the real economy. Goldman Sachs repeatedly returns to one central idea that the AI economy is not simply being built through better models. It is being built through an enormous reorganisation of industries, infrastructure and capital.

The winners, in other words, may not simply be those building the smartest AI. They may be the ones capable of turning AI into a functioning economic system.

AI race is becoming a race to build the real economy

Previous technology revolutions often followed a relatively clear sequence. Infrastructure was built first. Applications developed on top of that infrastructure. Capital structures then adapted to support the new economy.

AI is changing this pattern. According to the Goldman Sachs report, infrastructure, applications and capital structures are all evolving at the same time. Massive computing infrastructure is still being built even as AI is already beginning to reshape software, manufacturing, robotics, defense, energy and other industries.

This creates a much more complicated transition. Leaders are not simply competing to build better technology. They are competing to secure the resources required to support that technology at enormous scale.

Goldman Sachs puts the scale of the coming buildout into perspective by estimating around $7.6 trillion in capital spending between 2026 and 2031 across compute, data centres and power.

Hyperscalers alone are expected to invest trillions of dollars in AI through the end of the decade. Such numbers explain why the report sees capital strategy as one of the defining parts of the AI competition.

“To win in this era, leaders will need incisive capital strategy as much as the technology itself,” the report says. Building powerful AI is only one part of the challenge. Financing the computers, data centres, electricity systems and industrial infrastructure required to deploy it is another.

The biggest opportunity may lie outside software

Much of the public conversation around AI still focuses on chatbots, large language models and software applications. Goldman Sachs argues that the larger opportunity lies in the parts of the economy that AI has barely begun to transform.

Traditional Software-as-a-Service, or SaaS, represents less than 0.5% of global GDP, according to the report. The rest of the economy includes manufacturing, construction, energy, defense, logistics and other physical industries.

This is where the next stage of AI competition could happen.AI is moving beyond digital applications and into physical systems. Manufacturers are using technology to improve maintenance and operations. 

Automotive companies are using simulation and synthetic data. Utilities are looking at AI to forecast demand and optimise grid operations. Aerospace, life sciences and industrial companies are increasingly bringing AI into complex operational systems.The report calls this shift a move into the “real economy.”

Winning in this environment requires a different set of strengths from winning in consumer software. A company can build an impressive model without necessarily knowing how to integrate it into a factory. A powerful AI system may still struggle if it cannot operate safely inside an energy grid or an industrial facility. Technology also has to work with existing equipment, regulations and workflows.

This is why Goldman Sachs sees the next stage of AI as much more than a competition between models.

Value is moving beyond the model

One of the report’s most important findings is that the model layer itself may become increasingly commoditised. As more companies develop powerful AI models, simply having access to a model may become less of a competitive advantage.

Value could instead move toward the layers surrounding the technology. Goldman Sachs identifies the outcome layer, the orchestration and agent layer, and the data and context layer as increasingly important parts of the emerging AI economy.

The outcome layer focuses on completed work rather than software features or the number of seats sold to customers. The orchestration layer determines how work is routed and managed. The data and context layer becomes important because proprietary workflow history and institutional knowledge can provide differentiation. This changes the nature of competition.

The winner may not necessarily be the company with the largest model. It may be the company with access to valuable proprietary information, deep knowledge of a particular industry and the ability to turn AI into useful outcomes.

Goldman Sachs also points out that token consumption does not automatically translate into revenue.

The report’s discussion of China illustrates this point. Chinese models are serving inference at a fraction of the pricing of some US frontier models, creating what Goldman Sachs describes as a “China Cost Divide.” Chinese developers have also gained a much larger presence in the global model ecosystem.

A lower cost of inference can create a major advantage. However, the report does not conclude that this means China will automatically win the overall AI race. Value can migrate away from the model itself and toward data and outcomes. A cheaper model is therefore only one part of a much larger competitive equation.

Industrial AI could produce a different kind of winner

The Goldman Sachs report is particularly direct when discussing industrial AI. It identifies five factors that will determine industrial AI winners.

The first is physics-based architecture. Industrial systems operate in the physical world. Machines have to deal with material properties, thermodynamics, mechanical tolerances and other physical constraints. AI designed for industrial use therefore cannot operate in the same way as a system built purely for digital tasks.

The second factor is proprietary data. The report says some of the most valuable training data exists inside manufacturers, utilities and simulation libraries rather than on the public internet. Companies with deep access to operational data could therefore have an important advantage.

The third factor is edge deployment capability. Many industrial AI systems cannot depend entirely on distant cloud data centres. They may need to operate inside machines, embedded controllers and grid systems.

The fourth factor is certifiability. Industries such as aviation, energy and automotive cannot simply experiment with unreliable systems. Failures can have serious consequences. AI therefore needs to meet demanding standards of safety and reliability.

The fifth factor is workflow integration. Industrial companies cannot always replace decades of existing infrastructure. AI has to work with the systems, equipment and processes already in place.

Goldman Sachs says companies that combine all five factors can build durable competitive positions. Those that build only one or two may become commoditised.

This is perhaps the clearest answer the report provides to the question of who will win. The winners in industrial AI will be those capable of bringing several difficult capabilities together rather than excelling in only one area.

Controlling multiple layers could create outsized value

The report also examines the rise of physical AI. Physical AI includes intelligent systems operating in the real world, including autonomous vehicles, industrial robots, drones and humanoids.

Such systems require a much broader technology stack. The report describes a physical AI stack involving AI brains, simulation and digital twins, synthetic training data and purpose-built hardware.

Companies that can orchestrate multiple layers of this stack are positioning themselves to set standards and capture outsized value, according to Goldman Sachs.

This is significant because the future AI economy may become increasingly interconnected.A company working only on a model may control one part of the system. Another may build simulation tools. Another may manufacture hardware. Another may own the operational data.

Companies capable of bringing several of these layers together could have greater influence over how entire ecosystems develop. The growing importance of digital twins also reflects this trend.

Digital twins are virtual representations of physical assets, facilities and processes. They allow companies to simulate changes before making them in the real world.

Goldman Sachs describes the consolidation of simulation, design and industrial software as part of the creation of a new operational “control plane” for physical industries. The AI race could therefore increasingly involve control over the systems that connect the digital world with the physical one.

Robotics shows why technology alone is not enough

Humanoid robots have become one of the most visible symbols of the AI revolution. Goldman Sachs takes a more measured view of their commercial progress. Hardware is improving. Robots are demonstrating credible task performance in controlled environments. Capital has also flowed into the sector.

Workflow integration remains a major challenge. Companies have to determine how robots fit into existing operations. Workers have to be trained. Businesses need to achieve institutional-grade unit economics.

The report identifies capital, hardware and workflow integration as three important variables determining the viability of humanoid robotics. Capital, according to the report, has arrived ahead of revenue.

Broad commercial deployment is not expected until 2027–2029. Non-humanoid robotics is already operating at much greater scale in areas such as warehouses, logistics and industrial automation. These sectors have clearer unit economics and are therefore closer to existing financing frameworks.

The lesson is consistent with the rest of the report. A technology can be impressive without immediately becoming a successful business. Commercial deployment depends on economics, financing and integration. The AI race is therefore also a race to make technology economically deployable.

China has an important advantage, but not the final answer

China appears throughout the Goldman Sachs report as a major force in several parts of the AI economy. Chinese models have developed a significant cost advantage in inference. Chinese developers are also advancing embodied AI and robotics at lower costs.

The report shows a growing group of Chinese companies working across humanoid robotics and embodied AI. China also has a distinctive advantage in defense and industrial development through its military-civil fusion ecosystem, according to the report.

Goldman Sachs says this could create a significant competitive advantage because it does not face the same friction between venture capital and government procurement experienced in some Western systems.

Still, the report does not turn these advantages into a declaration of overall victory. The AI economy has too many layers for one advantage to decide the entire competition.

China’s cost advantage may matter greatly at the inference layer. Its manufacturing capabilities may matter in robotics. Its state-backed industrial system may matter in strategic sectors. Other parts of the AI economy will depend on different strengths.

Power may become the ultimate bottleneck

Every major AI system eventually encounters a physical constraint that is power. Goldman Sachs calls energy “the binding constraint.” AI infrastructure requires electricity. Data centres require electricity. Industrial AI and autonomous systems also increase demand for computing and power.

The problem is becoming more difficult because electricity grids were not designed for the concentrated and geographically uneven demand created by massive AI infrastructure.

The report says grid interconnection queues in major data-centre markets can stretch from eight to 12 years. This creates a serious challenge for the AI economy. A company may have access to capital and computing hardware but still struggle if it cannot secure enough electricity.

The response is already changing infrastructure development. Goldman Sachs says behind-the-meter and islanded power generation is becoming a credible development pathway for data centres. Hyperscalers are also becoming increasingly involved in power development, including nuclear energy.

Water and skilled labour are additional constraints. The report estimates that the US power and grid value chain will need more than 500,000 additional workers by 2030. Capital alone cannot solve this problem if skilled workers cannot be trained quickly enough.

The future winners of the AI race may therefore depend partly on something far more traditional than software: the ability to build and operate energy infrastructure.

The financing race could decide which sectors survive

The report’s strongest statement about winners and losers may come from its discussion of financing. “Which sectors thrive and which stall turns less on the tech frontier than on who closes the financing gap,” Goldman Sachs says.

The technology, according to the report, is arriving faster than the capital architecture required to fund its deployment. This creates a divide between different parts of the AI economy.

Digital infrastructure can use relatively familiar financing instruments, including investment-grade bonds, private credit, infrastructure debt and other established structures.

Industrial AI software can access public markets and selective mergers and acquisitions. Physical AI has less mature financing structures and continues to rely heavily on venture capital and growth equity.

This means different technologies may advance at different speeds not because one is technically superior, but because one can be financed more effectively.

Goldman Sachs sees AI as requiring many forms of capital simultaneously. Private credit, insurance capital, sovereign wealth funds, infrastructure finance, public debt, venture capital and government-backed investment are all becoming part of the AI economy.

GPUs themselves have created a new financing challenge because they depreciate quickly and face technological obsolescence. Robotics is experimenting with Robot-as-a-Service models that can shift spending from capital expenditure toward operating expenditure.

Insurance companies and sovereign wealth funds are emerging as important pools of long-duration capital. The ability to connect these different forms of financing could become a major competitive advantage.

Governments and sovereign wealth are entering the race

The AI competition is also becoming increasingly geopolitical. Goldman Sachs describes the growing role of “instrumental capital,” where governments use financing to achieve strategic objectives.

Capital is increasingly connected with supply-chain security, domestic technology development, critical minerals, defense capability and compute sovereignty. Sovereign wealth funds are becoming important players because they can invest over decades without facing the same redemption pressures as shorter-term investors.

The report notes that announced AI and digital infrastructure capital commitments from UAE and Saudi vehicles exceed $300 billion in stated targets. Sovereigns are also approaching AI through two tracks.

One involves investing globally in data centres, semiconductors and high-performance computing. The other involves building domestic AI hubs, supporting national champions and securing technology-transfer opportunities.

The AI race is therefore no longer simply being fought between private companies. Governments, sovereign funds and national industrial strategies are becoming part of the competitive landscape. Geopolitics will increasingly influence which investments can be made, which partnerships are permitted and which supply chains remain secure.

The real winners have not emerged yet

One of the most important conclusions of the Goldman Sachs report is that the AI economy is still under construction. Most of the capital that will shape the industry has not yet been deployed. Much of the infrastructure has not yet been built. Many of the mergers and acquisitions that will determine the competitive landscape have not yet happened.

The next three to five years, according to the report, will be particularly important. This period could determine the architecture of the AI economy for a generation.

Goldman Sachs does not offer a simple answer to who will win the AI race. No single country is declared the winner. No individual technology company is presented as the inevitable champion.

Technology matters, but so does capital strategy

Powerful models matter, but proprietary data and workflow integration matter too. Cheap inference can create an advantage, but value may move toward outcomes and institutional knowledge.

Robotics can demonstrate remarkable technological progress, yet commercial success still depends on financing and unit economics. Massive computing infrastructure can be built, yet the entire system can run into the physical limits of electricity, water and skilled labour.

The next phase of the AI race is therefore moving beyond the question of who can build the best model. It is becoming a competition over who can combine technology with capital, infrastructure, power, data and physical deployment.

The institutions capable of sourcing capital, pricing risk, structuring deals and operating across sectors could play a defining role in shaping the outcome.

In the Goldman Sachs view, the AI revolution is no longer only about creating intelligence inside computers. The much bigger challenge is bringing that intelligence into the real economy. The winners will emerge from those capable of building the systems required to make that transformation work.