Data-Center Backlash Puts AI’s Infrastructure-Dependent Valuations at Risk

The election made political resistance to data centers harder to treat as a side issue in the AI trade. Alex Kantrowitz said 38 jurisdictions had voted to ban or restrict data centers, and Texas had elected a governor who campaigned against them. Estimates of how many voters considered AI crucial to their vote varied widely across polls. For Ranjan Roy, opposition to data centers was no surprise. The consequential question was what it meant for companies and investors counting on them.

The market’s reaction was difficult to read. Technology stocks rose after the election, as some investors interpreted political opposition as a check on hyperscalers’ spending. The next day, investors appeared to focus on the other side: if AI depends on data centers, constraining their construction could undermine the investment case. Stocks fell sharply, then bounced. Kantrowitz saw the reversals as evidence that investors had no settled way to value AI. Roy disagreed. The underlying issue, he said, was clear: the buildout required to make many of the sector’s projections work was unlikely to happen as forecast.

Kantrowitz laid out his theory of the case in a six-point graphic. AI, he said, is a technology with potential and a compelling story, but its earnings are not yet commensurate with that story. Trillions of dollars in public-market value are being traded on expectations attached to two companies with about $100 billion in combined revenue. The potential of the AI labs has helped drive subsidized demand for chips, but realizing that potential requires data centers. The businesses remain small relative to the expectations placed on them, and the technology is still early. In that setting, any disruption can make the trade swing.

The graphic’s progression ended on the infrastructure itself: “Datacenters in particular are the linchpin here. Once they go the story evaporates.” Data centers are the physical link between a story about future AI capability and the products and revenue meant to justify it. Roy agreed that the market’s expectations rely on unusually good execution across growth, revenue, unit economics and infrastructure. A setback in any one part can put the rest of the story under pressure.

Roy’s test for the market was not whether some construction would continue, but whether it would happen at the scale and speed needed to support the valuations investors had assumed. He argued that local resistance made that uncertain in a way earlier technology regulation had not. Data-center approvals are made project by project, rather than through a centralized federal case against a company. That makes outcomes harder to predict and, in his view, increases the chance that political constraints will actually take hold.

Kantrowitz was less sure the restrictions would prove durable. He compared the backlash with earlier waves of scrutiny against technology companies, arguing that past regulatory pressure had not changed the industry as much as its headlines suggested. Developers might expand at sites where groundwork was already in place or move to places more willing to host them. States, towns and labor unions could have reasons to compete for projects, including the near-term construction jobs they bring. Even Texas’s governor-elect, Kantrowitz noted, had shifted from opposing data centers to saying they should follow the rules. He saw the current moment as the beginning of a negotiation, not necessarily the end of construction.

The practical constraints make relocation less straightforward than picking a new site on a map. Building a data center requires land approval, a site, contractors, electricity and skilled trades. Kantrowitz said moving a project from one state to another may mean starting the process again and losing time. The result might be delay rather than cancellation. He suggested that delay could stretch peak AI capital expenditure across more years rather than bring it to an immediate stop—and could even pace the frontier’s development.

Roy separated the technology’s prospects from the investment trade built around it. He described using Claude to edit a separate video while recording, a compute-heavy task he expected to become more common. He remained confident that demand for such uses would grow. His concern was that demand would not translate in a straight line into gains for a small group of companies over the next few years. Political resistance could change who benefits and when, even if the technology keeps advancing.

That distinction shaped their discussion of compute scarcity. Companies have to allocate compute among serving current products, training the next models and testing those models for safety. Kantrowitz said none of those demands is easy to drop. Cutting inference could hurt the product and send customers to competitors. Cutting safety work would create a different kind of risk. Cutting training could weaken the promise of future models that supports much of the market’s expectation of what AI might become.

Roy said that, operationally, he would prioritize products and sacrifice training of the next frontier model. But investors are often buying the expectation of that next model and the one after it. Falling behind on benchmarks could therefore create a difficult choice between serving customers today and sustaining the market’s future-facing story.

Roy offered a more optimistic possibility: constraints might encourage more efficient computing, as they had in places where compute was less available. Kantrowitz recalled Sundar Pichai making a similar point: when companies cannot simply add compute, they have to be more creative. But efficiency might be good for the technology and still bad for the economics of companies whose valuations rely on growing infrastructure spending. Kantrowitz added that efficiency does not guarantee rising total demand if developers cannot build more data centers.

Political constraints could also affect existing compute commitments. Roy speculated that if a supplier cannot build the infrastructure it promised, a customer such as Anthropic might have room to escape a commitment to spend. He cited Anthropic’s reported $518 billion in compute commitments and said he had regarded such agreements as not necessarily binding in practice. He joked that a company unable to provide the promised compute might be like a shop offering soup instead of the popsicles a customer had ordered. Infrastructure delays, Roy suggested, could even help Anthropic in the near term if they gave the company a way out of some commitments.

The infrastructure debate arrives as Anthropic’s prospective IPO is under scrutiny. A Wall Street Journal headline shown during the discussion said major investors were skeptical the company could reach a $1.5 trillion valuation. For Alex Kantrowitz, that range was inseparable from the AI story: Anthropic’s growth and other business facts supported the story, but the story was also what had made valuations as high as $1.5 trillion or $2 trillion plausible to investors. If the assumptions behind it weaken, the valuation can move with them.

Ranjan Roy reacted sharply to a separate Wall Street Journal report that Anthropic’s run rate had topped $80 billion at the end of October. He objected to the use of a leak about company financials so close to a possible IPO, calling the selective disclosure damaging and questioning the way it obscured the company’s financial story. The reported figure came from a person familiar with the matter. Roy said he did not know whether October had in fact been a strong month, and objected to the leak rather than treating the reported number as a complete account of Anthropic’s finances.

Kantrowitz asked whether the leak might signal a delayed IPO: perhaps the company wanted to put a positive number into circulation because it no longer expected to list immediately. He also cited a CNBC report that Kalshi’s odds of Anthropic announcing an IPO before December had dropped from nearly 90% to 57% after the election and market selloff. Roy said the company “kind of” had to go ahead, both for its own interests and for the industry, though his reasoning pointed in two directions. Going public amid headwinds could leave retail investors bearing risk after early investors and insiders had an opportunity to sell. But waiting could be costly too.

The valuation expectations were already moving. Roy recalled an initial $2 trillion figure associated with Anthropic’s strong second-quarter momentum, then a discussion of $1.5 trillion, followed by reporting that the range might be closer to $1.2–$1.3 trillion. Kantrowitz asked whether a number below $1 trillion should cause the company to pause. Roy said he still thought it should go public, even at a lower valuation.

His concern was that a delay would bring slower growth into view. Anthropic’s initial surge could not continue indefinitely as the company grew larger. At the same time, OpenAI was making progress in enterprise and, Roy believed, probably had more compute available. Anthropic had taken a riskier approach to securing compute, while rising construction costs and political obstacles could make that strategy harder to sustain. Waiting for those pressures to ease might give OpenAI time to strengthen its position.

Kantrowitz agreed that Anthropic had an incentive to move quickly, but questioned the logic of taking the company public on the strength of past growth if investors were not meant to extrapolate that growth forward. Roy’s answer was essentially that the market might still reward the exceptional run even as its durability became less certain. He also rejected SpaceX as a direct comparison for a high valuation: unlike Anthropic, SpaceX has Elon Musk, whom Roy described as unusually effective at raising capital and attracting retail investors.

Neither speaker treated a large valuation as guaranteed. Roy said even a valuation above $1 trillion would be a result Anthropic should be relieved by, but that figure assumes willing buyers. In a declining market, retail investors may be less inclined to buy a richly valued stock simply because it has been hyped. Kantrowitz’s own prediction was more definite: he thought Anthropic would go public on November 19 at a $1.3 trillion valuation. Roy emphasized how difficult the decision was in a volatile market. Both predictions remained uncertain, not settled outcomes.

The public opposition to data centers has also forced companies to reconsider how they negotiate with communities. Alex Kantrowitz cited a Bloomberg report that OpenAI would avoid nondisclosure agreements in data-center negotiations, ensure residents did not pay for electricity used by the projects, reduce power use at peak times, send electricity back to the grid and set local hiring goals where the workforce allowed. He pointed to earlier reporting about data centers appearing in communities under code names, arguing that this kind of secrecy was unlikely to remain sustainable.

Ranjan Roy questioned who counted as “the AI industry”: model developers, chipmakers, cloud providers or anyone with a stake in building AI data centers. Kantrowitz said the public narrative was driven particularly by companies such as OpenAI and Anthropic, along with Jensen Huang and Mark Zuckerberg. Roy thought politicians and other comparatively untainted figures might be able to speak for AI in ways those executives could not. But he did not expect a sudden change in tone to undo years of companies behaving as if community opinion did not matter.

Kantrowitz noted a practical obstacle: data centers are large and visually imposing. A mural would not make one look like farmland. He floated an intentionally improbable alternative: use the heat generated by the chips to create a public spa, with hot tubs, steam rooms and saunas. Roy first mentioned a supposed attractive data center in France, then corrected himself: the image had been a prank, not a real facility. He also raised a concern about asking people to bathe near a large power-consuming installation. Kantrowitz replied that he was talking about heat, not radioactive waste, and argued that a free local spa could serve as a social gathering place as well as a way to make a data center more welcome.

The exchange was comic, but the underlying problem was not. OpenAI’s reported commitments address some of the objections residents might raise—electricity costs, peak demand, local hiring and secrecy. They do not settle whether a community wants a data center nearby, or whether the project’s benefits outweigh its local costs. Even the most polished outreach has to contend with the physical presence of the infrastructure.

A report from METR, an AI research nonprofit, suggested that AI use inside Anthropic may have doubled the company’s pace of AI development. Alex Kantrowitz read from a report by The Information describing the rate of improvement in Anthropic’s models as roughly twice as fast in October as between March 2024 and October 2025. METR estimated that about 80% of the acceleration came from Anthropic’s internal use of AI.

The findings came with an important limitation. Anthropic had given METR employee-like access, but it also exercised rights to redact information it considered proprietary. The report’s redaction summary said Anthropic had removed a majority of the evidence METR found most informative, while allowing METR to publish its highest-level conclusions about risk. METR said some of the redacted information was difficult to describe without revealing what Anthropic had asked it to withhold.

Kantrowitz considered the reported acceleration plausible. He recalled Anthropic saying the year before that releases would arrive faster and the technology would improve more quickly—and said the company had followed through. He pointed to new products such as Dot and Muse as examples of work that had not been possible the previous year. For him, the METR result fit a wider pattern of faster development and raised the question of whether people were taking seriously how quickly the technology was improving.

Roy did not dismiss the conclusion. He said recent improvements in video editing had been striking: users could ask AI to add motion graphics, remove jump cuts and improve audio through prompts, capabilities he said had not existed in the same way two months earlier. His objection was to the value of a report that announces a worrying acceleration while withholding much of the evidence needed to assess it.

The question of whether Anthropic is approaching recursive self-improvement came up separately from METR’s estimate. The report, as described in the discussion, concerned the pace of Anthropic’s research and the contribution of internal AI use; it was not presented as a finding that the company was near RSI. Kantrowitz treated the estimate of faster progress as credible. Roy’s concern was that reports about acceleration and risk did not provide enough usable information to judge what was happening or what should be done.

That skepticism also shaped Roy’s view of the safety discussion. He questioned the emphasis on agents going rogue, arguing that the more immediate danger might be malicious actors using AI. He referred to a prior conversation with security expert Alex Stamos that had left him more concerned about those risks. In Roy’s view, a report that says development is accelerating, while leaving much of its evidence redacted, could reinforce a sense of danger without giving the public information it can use to decide what to do.

The disagreement was not over whether AI was improving quickly. Roy said the progress was real; Kantrowitz believed the report was credible. The tension was between evidence of rapid internal progress and limited public visibility into how that conclusion had been reached. Roy also noted the timing: Anthropic was discussing its rate of progress as it approached an IPO in a difficult market. That context sharpened his skepticism, though he acknowledged he might be reading the report through the lens of the company’s financial position.

On election night, Ranjan Roy watched prediction markets call many close races before television networks officially did. Prices in some contests rose to 97 or 99 cents before the networks declared a winner. To him, that speed was striking. Prediction markets aggregate traders’ views into a real-time signal; unlike polls, which take time to conduct and publish, prices can move as new information arrives.

But fast calls do not by themselves establish that markets are more accurate or more informative than polling models. Kalshi CEO Tarek Mansour said prediction-market favorites won 479 of the first 487 races, or 98%. Nate Silver countered that polls and models had pointed traders in the right direction, and that better election models were more confident and got there earlier. Chris Hayes agreed with Silver, arguing that betting markets moved toward the models rather than discovering the result independently.

Alex Kantrowitz described the dispute as a question of whether markets were adding information or reflecting work already done by pollsters and modelers. Roy saw room for the market and the polls to influence one another. He raised the example of Annie Andrews’s South Carolina race, where, he said, internal polling had shown her three points ahead a month earlier. Her campaign kept those numbers private, Roy suggested, partly to avoid a surge of attention and resources that could have changed the story. The example illustrated a feedback problem: poll results can guide traders, market prices can shape coverage, and coverage can then affect what people believe about the race.

Roy remained uneasy about prediction markets as platforms. Even if the prices provide useful information, he sees them largely as gambling products and worries about people risking money they cannot afford to lose. Kantrowitz agreed that “the gamification of everything” was undesirable, while maintaining that live prices can be a useful data point. The discussion left the central question open: prediction markets may call races quickly, but it did not establish whether their information is independent of the polls and models that may be guiding traders.