TradingKey - While revenues of artificial intelligence companies continue to grow rapidly, investments surrounding their data centers, GPUs, power, and chip supply chains have already entered the "trillion-dollar era".
The question then becomes straightforward: Can such massive infrastructure investments ultimately translate into sufficient revenue, profits, and economic growth?
In an episode of the Big Technology Podcast, SemiAnalysis founder Dylan Patel and analyst Jordan Nanos discussed the scale, returns, financing, neoclouds, security risks, and the competitive landscape between Anthropic and Google regarding AI infrastructure. While their overall assessment leans optimistic, they also acknowledged that the AI industry is approaching multiple constraints spanning capital, energy, construction capacity, and societal allocation.
AI Infrastructure Could Reach $2 Trillion
Patel believes that the market generally underestimates the true scale of AI infrastructure buildout. According to his estimates, relevant U.S. capital expenditures next year could approach $2 trillion, accounting for 5% to 6% of U.S. GDP. This figure includes not only data centers and GPUs, but also investments in power, cooling systems, chips, construction equipment, and the entire supply chain.
There is another key difference between AI infrastructure and traditional railroad and highway construction: computing power built in the U.S. does not serve the U.S. alone. Some AI services used by users in Europe and Asia actually rely on U.S. data centers for training and inference. Therefore, U.S. capital expenditures largely support global AI demand, and their economic impact may exceed that of past domestic infrastructure cycles in a single country.
This means that AI infrastructure is no longer just an isolated tech sector, but is gradually becoming a macroeconomic variable that impacts GDP, capital markets, and industrial policy.
AI Investment Growth Will Slow, But Demand Has Yet to Peak
AI capital expenditure cannot sustain high-speed growth indefinitely. As the investment base expands, it becomes difficult for companies to continue doubling growth at a larger scale. Meanwhile, power, land, construction capacity, chip supply, and financing capabilities are all becoming increasingly evident constraints.
Jordan noted that for both major cloud providers and emerging cloud vendors, adding gigawatt-scale power facilities beyond original plans has become significantly more difficult. Even if returns on investment are strong, companies must first raise massive capital before proceeding with construction.
However, slowing growth does not equate to a collapse in demand. The core issue currently facing the industry remains insufficient supply rather than excess computing power capacity. Chip capacity, data center construction speed, and power grid access capabilities are all still struggling to fully meet customer demand.
The true risk trigger point would be a significant drop in enterprise demand for intelligence and knowledge work. As long as demand continues to exceed supply, it is unlikely that the industry will immediately enter a downturn simply due to a slowdown in capital expenditure growth.
AI Revenue Lags Investment, Returns May Come From the Entire Economy
Another characteristic of the AI industry is that revenue typically lags capital expenditure. Data centers and GPUs depreciate over multiple years, meaning capital invested today must be gradually recovered over the coming years through model API calls, enterprise software, automation services, and new products.
Patel believes market forecasts for future AI revenue may still be too low. This is because the value of AI extends beyond API calls, token fees, and subscription services, with more revenue likely stemming from new products and efficiencies created as AI penetrates other industries.
Jordan cited an example, stating that if frontier models can handle a significant amount of knowledge work, they could enter fields such as robotics, autonomous driving, drug discovery, healthcare, and life sciences. In these sectors, the value created by AI will be reflected through new drugs, automobiles, chips, and services, rather than just charging per model call.
The chip design industry is a prime example. Over the past few decades, the headcount of U.S. chip design R&D personnel has not grown significantly, yet the value created by the industry has continuously expanded, partly because engineering tools boosted developer productivity. AI could extend this productivity gain to more industries, enabling enterprises to accomplish more R&D, design, and production work with a similar headcount.
This is also the central logic behind AI infrastructure investment: investors are not betting on today's chatbot revenue, but on the incremental value generated once intelligence diffuses throughout the entire economy.
Emerging Clouds Become Key Infrastructure for AI Expansion
As frontier labs and hyperscalers continue to expand their compute demands, emerging cloud providers have become key facilitators of AI infrastructure expansion.
SemiAnalysis's ClusterMax report primarily evaluates the reliability, security, networking, storage, performance, and operational capabilities of GPU clouds. CoreWeave and Nebius were listed as top-tier vendors, while the service capabilities of Oracle and Google Cloud were also recognized.
However, technical quality and stock value are not the same thing. Although Oracle excels in cluster services, it faces risks including data center delays, pipeline permitting, debt load, and project execution. Meanwhile, some emerging cloud providers focus primarily on fundraising and expansion, underinvesting in underlying security and service quality.
This reflects that the market remains a seller's market. With customers urgently needing GPUs, suppliers can quickly secure orders and financing, which in turn reduces the pressure on them to improve service experience.
Nvidia Is Becoming the Supply Chain's Financial Backer
Nvidia's role in the emerging cloud ecosystem goes beyond that of a mere chip supplier.
SemiAnalysis estimates that the scale of off-balance-sheet guarantees and support provided by Nvidia could reach $588 billion, covering construction, storage, supply chains, and data center buildouts beyond GPU procurement. Nvidia's objective is to reduce its reliance on a few hyperscalers on the one hand, and to help more customers acquire computing power and financing on the other.
As long as end demand remains strong, this mechanism of guarantees, financing, and GPU leasing can form a virtuous cycle. Meanwhile, interdependence across the industry chain is deepening. Once AI demand slows, risks could rapidly propagate through long-term contracts, debt, and balance sheets.
Therefore, Nvidia's "backstop" serves not only as a driver of industry expansion, but also demonstrates that AI infrastructure has formed a complex financial network.
Competition Between Anthropic and Google to Shift to Commercialization
Patel is very bullish on Anthropic's long-term prospects, but he also believes that achieving an extremely high valuation requires the company to expand its business beyond model calls into more areas.
Anthropic's advantages lie not only in model capabilities, but also in higher gross margins, lower customer acquisition costs, and strong customer stickiness. If AI labs can apply models to software, pharmaceuticals, healthcare, and enterprise services, their profit margins could exceed those of traditional SaaS businesses.
Google, meanwhile, faces a different kind of pressure. In the past, Google held a significant compute advantage, but today more TPUs and infrastructure are used for external sales and supporting Anthropic, causing DeepMind's share of compute to decline.
Google is attempting to generate revenue across every segment of chips, cloud platforms, and model services, but Gemini still needs to find a sufficiently clear product advantage. Otherwise, its advantages in compute, models, and talent could continue to be eroded by OpenAI, Anthropic, and Meta.
Find out more
Comments