Hyperscale Skepticism Rises as Energy Constraints Bite

The Great AI Infrastructure Debate Continues

Perscient's semantic signature tracking the density of language consistent with companies becoming more skeptical of big AI investments climbed to a z-score of 1.9, among the highest readings we monitor.

The numbers, as we have observed in multiple recent Pulses, tell a sobering story. According to research commissioned by DDN in partnership with Google Cloud and Cognizant, more than half of AI projects have been delayed or canceled within the last two years, with enterprises citing the sheer complexity of AI infrastructure as the primary obstacle. Two-thirds of the 600 IT and business decision-makers surveyed at large US enterprises admitted that their AI environments have become too unwieldy to manage effectively.

The semantic signature tracking language questioning hyperscale builds registered at 2.5, though it declined by 0.4 from the prior week, suggesting that while skepticism remains pronounced, the intensity of debate moderated following commentary at CES 2026 and Davos. At the World Economic Forum, Nvidia CEO Jensen Huang pushed back forcefully against growing concerns about AI spending sustainability, framing current capital expenditures as "the largest infrastructure build-out in human history" rather than speculative excess.

Yet spending commitments continue to escalate. CreditSights projects that capital expenditures for the top five hyperscalers will climb from approximately $256 billion in 2024 to $443 billion in 2025 and roughly $602 billion in 2026. Third-quarter earnings triggered another upward revision in consensus estimates, with projected 2026 capital spending now at $527 billion, up from $465 billion at the start of earnings season. As one social media observer noted, Taiwan Semiconductor's announcement that it expects the next three years of capital expenditure to exceed the previous three years combined signals that "the AI factory of the global economy just told you this cycle is nowhere near over."

The disconnect between investment and returns remains the central tension. Our semantic signature tracking language consistent with productivity gains from AI not materializing rose by 0.2 to a z-score of 1.8. The Duke CFO Survey findings paint a stark picture: despite US tech giants spending $380 billion on AI-driven infrastructure in 2025, most chief financial officers still cannot point to measurable returns. When asked about AI's impact over the past twelve months, the vast majority reported "no change" across operational metrics.

S&P Global Ratings observed that healthy skepticism toward AI is emerging in the North American technology sector, driven partly by debt-fueled spending for massive infrastructure projects. The credit agency noted that this "heightens the risk and amplifies the impact if enterprise productivity gains do not materialize fast enough to justify these investments." This concern finds echoes in the cancellation of 41 data center projects in just the last six weeks, compared to only 15 from June through November 2025.

The semantic signature tracking language characterizing AI capital expenditure as a risky bet with uncertain payoffs declined by 0.3 but remains elevated at 1.8. As enterprises move from proof of concept to production-scale deployment, they are discovering that their existing infrastructure strategies were not designed for AI's distinctive demands.

Google's Gemini Surge Putting Pressure on Peers

While enterprise skepticism about AI infrastructure mounts, the competitive dynamics among AI providers underwent a meaningful shift over the past year, with Google's Gemini emerging as the most significant challenger to ChatGPT's dominance since the generative AI boom began. Our semantic signature tracking language asserting that Google or Gemini is winning the AI race registered at 2.7, the highest among all competitive positioning signatures we monitor, though it declined by 0.9 from the prior week following intense coverage around market share gains.

The data confirms substantial redistribution of traffic. Industry tracking from Similarweb shows that Gemini commands roughly 20% of global LLM traffic, while ChatGPT's share among generative AI chatbot websites dropped from 86.7% in January 2025 to 64.5% in January 2026. As one market analyst observed, Gemini has climbed from 5.3% of global generative AI web traffic to 22.0% over the past year, with most of that share coming directly from ChatGPT's decline.

The Apple partnership represents perhaps the most consequential development in this competitive realignment. Reports indicate that Apple will integrate Google Gemini into Siri beginning in February, with iOS 26.4 bringing the new capabilities to iPhones. This arrangement grants Google access to over 1.5 billion daily requests through the iPhone ecosystem. According to Bloomberg's Mark Gurman, Google would supply Gemini models for Siri and future Apple Intelligence features, initially running on Apple's Private Cloud Compute servers.

The selection process proved telling. As Gurman reported, "OpenAI hadn't been in the running for a while and Apple decided not to work with them. And they picked Gemini because the gap closed with Anthropic—not OpenAI—and was cheaper." Apple is reportedly spending over $1 billion annually for Gemini access, though this remains modest compared to Google's payments exceeding $20 billion to maintain its position as the default search engine on Apple devices.

Our semantic signature tracking language positioning OpenAI as the AI leader remained essentially flat at a z-score of just 0.0, reflecting a decline in media framing of OpenAI as the clear frontrunner. The Wall Street Journal reported that OpenAI declared a "code red" in December following Gemini's surprising growth and will begin testing advertisements in ChatGPT as it pushes for fresh revenue streams.

Google's advantages extend beyond raw market share. According to DataCamp analysis, Gemini outperforms ChatGPT in coding and research tasks, scoring higher in ecosystem integration that allows seamless workflow incorporation. As one technology observer noted, "Gemini is poised to dominate consumer AI. It's already in my email, calendar, drive, maps and browser."

Meanwhile, our semantic signature tracking language asserting that Anthropic or Claude is winning the AI race registered at 2.3, remaining elevated and essentially flat week-over-week, suggesting that Anthropic maintains strong positioning in specialized enterprise and safety-focused segments even as the consumer market shifts. The Information reported that Google Gemini API calls more than doubled to 85 billion by August, while Google has sold 8 million paid seats of Gemini Enterprise.

The market is fragmenting in meaningful ways. Data compiled by Microsoft shows that DeepSeek claims 3.7% market share with particularly strong adoption in developing nations, while Grok has reached 3.4% market share benefiting from X integration. DeepSeek's usage is estimated at two to four times higher in Africa compared to other regions, gaining traction in markets underserved by Western AI platforms. Fortune analyst Daniel Newman characterized 2026 as a "make-or-break year" for Apple, noting that while the company has the user base and distribution to be patient in chasing new trends, this represents a critical juncture.

Energy Infrastructure Emerges as the Decisive Constraint in AI Competition

As competitive positioning among AI providers intensifies, energy infrastructure has emerged as the factor most likely to determine winners. Our semantic signature tracking language asserting that the AI winner will be determined by which country builds energy capacity registered at 3.2, the highest z-score in our entire dataset, declining only by 0.1 from the prior week.

The industry's most prominent voices have converged on this assessment. At Davos, Elon Musk stated that while AI capabilities will likely surpass human intelligence by the end of 2026, the critical bottleneck is energy production: "It's clear that we're very soon—maybe even later this year—we'll be producing more chips than we can turn on." Jensen Huang echoed this concern, declaring that "the amount of energy that we have limits what we can get done" and identifying power efficiency as his top priority.

The scale of demand is staggering. By 2026, US data center grid-power demand is expected to rise by 22% to approximately 75.8 gigawatts for IT equipment, with projections showing this nearly tripling by 2030. Eric Schmidt provided context: "There's one calculation that we need another 90 gigawatts of power in America. Ninety gigawatts is ninety nuclear power plants in America. Not happening, we're building zero." Schmidt added: "To understand how big gigawatts are, think cities per data center. That's how much power we need."

The political implications are becoming impossible to ignore. The New York Times reported that while the Trump administration pushes AI data center development, the GOP itself remains cool to specific projects, with concerns that "in the short run, at least, growing demand for energy by AI and the development of data centers will increase household and business costs of electricity." Our semantic signature tracking language calling for regulation of AI's impact on energy and water rose by 0.2 to 0.9.

Residential electricity prices are forecast to rise another 4% on average nationwide in 2026 after increasing about 5% in 2025, according to the federal Energy Information Administration. The impact of data centers on local communities is likely to feature in mid-term election campaigns. Democratic Socialist Senator Bernie Sanders and conservative Governor Ron DeSantis have found common ground as leading skeptics of the data center boom, signaling a brewing political reckoning over the industry's impact on electricity prices, grid stability, and employment.

The Trump administration's response has been to propose an emergency auction forcing tech companies to fund new power plants, with the nation's largest grid operator PJM holding emergency power auctions for tech companies to bid for long-term contracts. OpenAI unveiled its own Stargate Community plan aimed at "paying its way on energy" and ensuring that its operations do not raise electricity costs for surrounding communities.

Community opposition has proven more effective than many anticipated. According to MorePerfectUnion data, $98 billion in planned AI data center development was derailed in a single quarter last year by community organizing and pushback. Reuters reported that plans for just over 150 gigawatts of new data center power capacity have been filed nationally across 24 different states, but the path from proposal to operation remains fraught.

Our semantic signature tracking language framing big AI capital expenditure as necessary to compete with China declined by 0.6 to 0.9. However, the competitive dimensions remain salient. As one analyst observed, "China has a lot of electricity, so if they get the chips, we know who has the advantage." Data shows that China generates 40% more electricity than the US and EU combined, and the country's ability to build infrastructure quickly without the pressures of public opposition may prove decisive.

The constraint on AI progress has shifted from "can we build better models?" to "can we power and cool the infrastructure to run them?" Digital sovereignty is moving AI governance from ethics discussions to enforceable levers like market access, cloud procurement, chip supply, and grid capacity. Leaders at the World Economic Forum discussed how AI tools could help boost the performance of existing grid infrastructure, unlocking hidden capacity. But for now, every megawatt that can be brought online will be monetized, and the race to secure power may determine who leads the next phase of AI development.