Artificial intelligence is attracting enormous investment, reshaping markets, and changing how businesses think about the future. But beneath the excitement lies a more difficult question: which parts of the AI ecosystem will ultimately justify the capital being invested in them?

Daniel J. Arbess, Founder & CEO at Xerion Investments, from the United States, brings an unusually broad perspective to this question. His background spans public securities, macro-driven investing, healthcare, neurobiology, and neuroimmunology. He has followed the AI investment landscape closely since investing in Nvidia in 2015 and has also built a multi-agent AI version of the fundamental investment process he once carried out with a full team of analysts. His perspective combines years of market experience with hands-on engagement with the technology itself.

His central argument is neither an uncritical endorsement of AI nor a rejection of the sector’s potential. Instead, it is a question of economics: where will demand grow, which parts of the technology stack can sustain their value, and how will the enormous investment in computing infrastructure generate returns?

The Real Test of AI Is Whether the Investment Pays Off

Major technological shifts can attract capital long before their ultimate business models become clear. Investors may recognize that a technology is transformative while still disagreeing about which companies will capture its economic benefits. This distinction matters because the success of an industry does not guarantee that every company within it will deliver attractive returns.

AI infrastructure has drawn investment on an extraordinary scale, helping drive economic activity and supporting the expansion of computing capacity. Yet the size of that investment creates its own challenge: the industry must generate enough sustained demand to justify the resources committed to it. The important question is not simply whether AI will become more capable, but whether its commercial applications can produce returns proportionate to the cost of building and operating it.

Arbess frames the issue around the relationship between capital spending and future demand, “My concern from the beginning has been to try to visualize how those capital investments will be returned.”

This question is becoming increasingly important as investors evaluate AI companies. Technical progress may establish what is possible, but durable business value depends on customers adopting the technology, paying for its use, and generating enough economic benefit to support the costs involved.

The Next Growth Engine May Come From AI Entering the Physical World

Much of the early AI boom has centered on digital tasks: generating text, analyzing information, writing software, and automating knowledge work. These applications are significant, but the potential market becomes much broader when AI begins to influence activities in the physical world.

Autonomous defense systems, transportation, and robotics could create demand for computing beyond conventional data analysis. These applications require systems to interpret changing environments, process information, and support decisions or actions in real time. If adoption expands, they could become important drivers of additional computing demand.

The significance is economic as well as technological. New physical-world applications could widen the range of industries purchasing AI capabilities, creating additional markets for computing infrastructure and specialized hardware. However, their ultimate impact will depend on whether the systems can be deployed reliably, safely, and at commercially viable costs. Arbess identifies this transition as a potential turning point, “I believe that the turning point will be when real-world uses of AI in the physical world start to take off.”

He points to autonomous defense and transportation as examples already moving in that direction, while identifying humanoid robotics as a possible future source of substantial demand. His argument suggests that the next stage of AI growth may be determined less by the novelty of digital demonstrations and more by the technology’s ability to operate in the world around us.

Computing Demand Must Expand Beyond Conventional Data Analytics

The growth of AI depends not only on the number of models being developed, but also on how extensively those models are used. If computing demand remains concentrated in a limited set of applications, the industry may face a different economic reality from one in which AI becomes embedded across many sectors and activities.

This is why the distinction between existing workloads and new categories of demand matters. Conventional data analytics can be valuable, but it may not, by itself, support every expectation built into the current scale of infrastructure investment. Broader adoption across physical systems, industrial processes, and autonomous technologies could create a larger base of computing-intensive activity.

For investors, this introduces a practical way to assess the market: look for evidence that AI is opening new sources of demand rather than merely shifting existing tasks onto more expensive technology. The question becomes whether additional computing capacity will be used productively enough to sustain the economics of the infrastructure behind it.

Arbess explains the scale of demand he believes may be necessary, “Demand for computing will need to extend beyond AI applications in basic data analytics.”

His point is that the industry needs more than incremental expansion of familiar applications. A broader range of real-world uses may be necessary to support the scale of investment already underway.

The AI Value Chain Will Not Reward Every Layer Equally

The AI ecosystem spans several distinct layers: the land and power needed to support infrastructure, semiconductors and interconnects, integrated computing data centers, AI models, and the entrepreneurs and enterprises building applications. Each layer plays a different role, faces different competitive pressures, and captures value in different ways. Understanding these distinctions is essential to identifying where investment returns are most likely to accrue.

A company supplying critical components may benefit from demand across many competing AI systems. A company developing a particular model, by contrast, may face pressure from rivals offering similar capabilities, including models that are available with fewer restrictions on access or modification. As capabilities spread, differentiation can become harder to maintain.

This does not mean every layer outside hardware is unattractive. It means investors need to distinguish between technological importance and the ability to capture lasting economic value. A product can be central to the industry and still face intense pricing pressure, high operating costs, or weak barriers to competition.

Arbess describes his evolving view of the AI stack, “I remain comfortable with investments in compute and critical hardware components, but believe the pressure will increasingly fall on frontier models.”

This is a distinction between the overall growth of AI and the relative position of individual businesses. As the ecosystem matures, the winners may be those that occupy defensible positions in the value chain rather than simply those most closely associated with the latest breakthrough.

When Models Become More Accessible, Differentiation Becomes Harder

The competitive landscape for AI models is changing as more developers gain access to increasingly capable systems. Open-weight models and alternatives from different regions can put pressure on companies that previously benefited from having a relatively exclusive technological advantage.

For many businesses, the decision will not be about choosing the most advanced model available, but the model that delivers sufficient performance, protects customer data, and makes economic sense for the task.

Greater accessibility can be positive for adoption: businesses may have more options, developers can experiment more freely, and applications can be built without depending entirely on a small number of providers. For model companies, however, this can make it harder to maintain pricing power if customers can obtain comparable capabilities elsewhere.

The business implications extend beyond model performance. Companies must consider distribution, integration, reliability, customer relationships, and the cost of serving users. A technically impressive model does not automatically become a defensible business if competitors can replicate its core functionality or deliver similar results at lower cost. Arbess highlights this competitive pressure directly, “I think that the large language models are being commoditized by open-weight models from China and elsewhere.”

The broader lesson is that technological progress can create value for the market while reducing the advantage of individual providers. As capabilities become more widely available, lasting differentiation may increasingly depend on how effectively a company turns AI into a product customers need and are willing to pay for.

A Positive View of AI Does Not Require Buying Every AI Valuation

One of the most important distinctions in technology investing is the difference between believing in a sector’s future and believing that a particular asset is worth its current price. A market can be built around a genuine technological transformation while still containing companies whose valuations leave little room for uncertainty.

High expectations can create pressure when growth, margins, competition, or future demand fail to match what investors have priced in. This is particularly relevant in capital-intensive industries, where companies may need to spend heavily before their business models produce sufficient cash flow.

A disciplined approach therefore requires evaluating individual opportunities on their own economics. Investors must consider what assumptions are already reflected in a valuation, what could weaken those assumptions, and whether the business has a credible route to returns.

Arbess makes clear that enthusiasm for AI as a broad investment theme does not translate into indiscriminate enthusiasm for every company associated with it, “I wouldn't touch an OpenAI IPO at, you know, a trillion and a half dollars with a ten-foot pole.” He follows that strong valuation judgment with a broader distinction. “I'm basically positive on the broad AI trade.”

Together, these comments capture a more selective investment philosophy. AI may continue to expand across the economy, but the outcome for investors will depend on entry price, competitive position, and the economics of each business - not on the strength of the overall narrative alone.

Conclusion: Follow the Economics, Not Just the Excitement

AI will remain a powerful force in technology and investment, but its future depends on turning infrastructure into productive applications, expanding compute demand, and building businesses that can sustain a competitive edge. The strongest opportunities may lie where customer demand, defensible technology, and sustainable economics converge, while physical-world AI and competition among model providers reshape the market.

Arbess’s perspective highlights the need to look beyond technological excitement. For investors, the key questions are which AI applications will deliver real value, which businesses can capture it, and whether the returns justify the investment. Distinguishing lasting commercial opportunities from hype will be critical to navigating AI’s next phase.