Hyperscale Builds Surge While Productivity Gains (Maybe) Begin to Materialize

Hyperscale Infrastructure Spending Defies Persistent Skepticism

The scale of capital flowing into AI infrastructure has reached a point where even seasoned observers struggle to contextualize the numbers. Google's decision to lift its 2025 forecast by $10 billion, bringing total spend to approximately $85 billion, represents just one data point in a broader pattern of escalating commitments. Collectively, hyperscaler capital expenditure is projected to reach $350 billion in 2025 and climb toward $400 billion in 2026, figures that would have seemed implausible just three years ago.

Perscient's semantic signature tracking expectations of continued hyperscale growth rose by 0.5 over the past week to reach a z-score of 3.4, reflecting intensifying media coverage of these massive infrastructure projects. Yet media framing remains bifurcated. The semantic signature tracking questions and skepticism about hyperscale builds likewise strengthened by 0.3 to a z-score of 2.5, suggesting that coverage is balancing bullish real-world capex projections against skeptical interpretations about whether they will ever be justified.

The tension between these narratives found sharp expression when IBM CEO Arvind Krishna walked through some napkin math on Big Tech's AI data center spending and raised pointed doubts about profitability. Building and operating a 1-gigawatt AI data center costs approximately $80 billion, and with companies planning roughly 100 GW of capacity, the implied $8 trillion in total spending would require approximately $800 billion in annual profit just to service the interest. Krishna's assessment that there's "no way" to turn a profit at current costs stands in stark contrast to the continued acceleration of commitments.

JPMorgan's analysis offers a different perspective, projecting global data center and AI infrastructure spend to hit $5 trillion while noting that demand for compute "remains astronomical." The bank argues that unlike past episodes of speculative excess, today's AI cycle is being largely financed by profitable, cash-rich firms with free cash flow margins near 20 percent, more than double their late 1990s levels. This financial foundation distinguishes current spending patterns from the debt-fueled buildouts of previous technology cycles.

The comparison to 1990s telecom overbuilding continues to resonate in media coverage, even as the semantic signature comparing AI capex to fiber construction during the dot-com boom declined by 0.5 to a still-astonishing 3.1. Alphabet, Microsoft, Amazon, and Meta invested nearly $200 billion in CapEx in 2024, a figure expected to climb by over 40 percent in 2025 as they rush to build the computational power needed to train and deploy next-generation AI models. In 2025 alone, the "Magnificent Seven" tech firms are on track to invest a record $364 billion in data center construction and upgrades, spending that is now contributing more to US GDP growth than all consumer spending.

AI Productivity Gains Show Early Signs of Reaching Corporate Bottom Lines

The question of whether these infrastructure investments will generate returns connects directly to emerging evidence of AI's impact on corporate performance. The semantic signature tracking AI improvements hitting corporate bottom lines rose by 1.1 over the past week to reach a z-score of 0.4, representing the largest one-week change among all tracked signatures. This marks a transition in media framing toward recognizing tangible AI returns.

The EY US AI Pulse Survey provides concrete evidence of this shift. Among 500 business executives at US companies that experienced AI-driven productivity gains, just 17 percent turned around and cut jobs. The majority instead reported reinvesting gains into existing AI capabilities (47 percent), developing new AI capabilities (42 percent), strengthening cybersecurity (41 percent), investing in R&D (39 percent), and upskilling employees (38 percent). As one EY analyst noted, "There's a narrative that we hear quite frequently about companies looking to take that benefit that they're seeing and put it into the financial statements."

Still, the semantic signature tracking the view that productivity gains from AI haven't materialized remained elevated at 2.0 (although it declined by 0.1). The persistence of this skeptical framing alongside rising bottom-line impact coverage suggests a potential transitional moment in media attention. BCG's survey of C-suite executives found that meaningful enterprise-wide bottom-line impact from AI use continues to be rare, though respondents who attribute EBIT impact of 5 percent or more to AI use, representing about 6 percent of respondents, report pushing for transformative innovation.

The banking sector offers a particularly instructive case study. U.S. banks including JPMorgan Chase and Wells Fargo have described AI as a force multiplier, with executives increasingly framing it as a near-term productivity lever. AI is transitioning from an innovation lab experiment to a day-to-day operating system at these institutions, contributing to a gradual softening of the most pessimistic assessments as the semantic signature tracking the view that big corporate AI experiments have been failures declined by 0.1 to 1.2.

An OpenAI report revealed that AI productivity gains are not evenly distributed across all users but concentrated among those who use the technology most intensively. Workers who engage across approximately seven distinct task types report saving five times as much time as those who use only four, suggesting that organizational approaches to AI deployment may matter as much as the technology itself.

US-China AI Competition Intensifies as Both Nations Pursue Divergent Strategies

The domestic debate over AI returns unfolds against an increasingly prominent geopolitical backdrop. The semantic signature tracking the view that big AI capex is needed to compete with China strengthened by 0.54 to 3.08, reaching its highest recorded level. The semantic signature tracking the view that the US must win the AI race rose by 0.17 to a z-score of 0.54, reflecting intensifying media focus on competitive dynamics.

The framing of this competition varies between the two nations. China and the United States are racing toward different ends, with the United States tending to define the competition in terms of the race toward Artificial General Intelligence. By Trump's AI czar David Sacks' estimate, "China is not years and years behind us in AI. Maybe they're three to six months," though no one can be certain of the precise gap.

The performance gap between the best Chinese and U.S. AI models has shrunk dramatically, from 9.3 percent in 2024 to 1.7 percent in February 2025. This narrowing has occurred despite export restrictions, as Chinese companies have increasingly found ways to use scale in engineers, less advanced chips, and data to gradually build domestic capabilities. That's apparent in how Chinese AI models rival OpenAI at a fraction of the cost.

Nvidia CEO Jensen Huang offered a nuanced assessment of the competitive position, affirming a US lead of "generations ahead" on chips but warning against complacency. He highlighted China's advantages at the infrastructure layer, noting that building a data center in the US takes about three years from groundbreaking to operation, while China can build very large facilities far more rapidly. Media coverage continues to favor US prospects while acknowledging China's competitive position, with the semantic signature tracking the view that the US will win the AI race rising to 0.9, while the signature tracking the view that China will win remained stable at 0.6. The semantic signature tracking language expressing the view that AI is a pillar of US exceptionalism rose by to 1.3.

As we have been writing on these pages for some time, World War AI is well underway, and these narratives are very much at its heart.