AI’s Buildout Is Concentrating Capital, Control, and Local Costs

The current AI buildout is forcing three linked decisions. Companies must decide where to place capital as spending shifts toward chips, networking, cloud capacity, and inference. Governments must decide how much control to exert over frontier models. Communities must decide what physical infrastructure they will accept to support it.

John Coogan located IBM’s 25% stock decline in the first of those decisions. The five-year chart displayed during the discussion showed a 64.88% gain in IBM’s share price. But the company had reset expectations around its server business after identifying a shift away from mainframe spending and toward the physical AI buildout.

Coogan’s argument was not that IBM lacks AI assets or profitable businesses. It was that the largest incremental AI budgets are currently flowing to GPUs, memory, networking, hyperscale cloud computing, and frontier-model inference—categories where IBM is not a major winner. AI demand can be robust while IBM still loses share of a customer’s total technology budget.

IBM built its earlier dominance through integrated corporate systems: reliable hardware and software, proprietary dependencies, high switching costs, and extensive support relationships. The PC era separated hardware and software layers, and IBM’s later strategy under Lou Gerstner moved toward integration, outsourcing, and consulting rather than manufacturing every technology component itself.

That history still explains the company’s current mix.

IBM acquired Red Hat for $34 billion in 2019 and spun off its traditional managed-infrastructure outsourcing operation in 2021. Coogan said the Red Hat acquisition had begun paying off, while the z17 mainframe cycle had been unexpectedly strong. Red Hat OpenShift, IBM’s enterprise Kubernetes platform for orchestrating workloads across multiple computers, remains a meaningful asset for companies deploying AI in complex enterprise environments.

But Coogan’s assessment was that IBM faces many other companies offering AI capabilities throughout the stack. IBM’s institutional relationships, software margins, and infrastructure business may remain valuable, yet they do not automatically place the company on the highest-growth spending path. The immediate problem, in his telling, is that customers are prioritizing the hardware and cloud infrastructure required to build and run frontier AI systems rather than the mainframe spending IBM had just identified as shifting.

DeepMind chief Demis Hassabis has called for a U.S.-led standards body to test frontier AI models for national-security threats. He argued that cybersecurity risks are already visible and that biological and nuclear risks may emerge as capabilities improve. His proposed testing regime would be “dynamic, adaptable, and rigorous,” with the United States taking the lead because of its economic and technical position.

The proposal, as John Coogan summarized it, would establish a federally overseen body funded by AI companies; define and update thresholds for what counts as a frontier model; require qualifying labs to submit models for testing up to 30 days before release; and assess cyber, biological, nuclear, deception, autonomy, and guardrail-bypass capabilities.

It would also require cybersecurity measures, personnel vetting, model cards, watermarking, and safety research; use national labs, federal agencies, and independent auditors; create confidential evaluations that labs cannot simply train against; and require remedies for serious vulnerabilities found after release. The framework would apply to frontier models deployed in the United States, including foreign and open-source models, while exempting smaller systems. Hassabis also called for coordination among frontier labs if testing identifies sufficiently serious risks, with the eventual goal of shared international standards.

The policy dispute was less about whether government should prepare than about what useful preparation looks like. Jordi Hays argued for concrete scenarios that give lawmakers a practical basis for action: what might happen to trucking employment, for example, and what choices would government face if displacement occurred. Simply warning that AI will improve at hacking does not tell Washington much, he said, because hacking is already illegal and companies already have an incentive to improve their defenses.

Coogan preferred pre-agreed triggers to forecasts. Rather than debate a speculative timeline, he argued, policymakers could define actions tied to measurable conditions. If unemployment passed 10%, for example, a government could already know whether it intended to send means-tested stimulus payments, create a jobs program, or use another intervention. He cited the pandemic response—when unemployment reached 15% and checks were sent—as an example of a visible condition linked to a concrete response.

This difference exposes an operating question in Hassabis’s proposal. Hassabis argues that frontier testing is needed; the hosts wanted more specificity about how a system would operate. Tyler, another panelist, said the proposed institution resembled AISI, the Center for AI Safety and Innovation under the Commerce Department, but with more authority. AISI is currently opt-in, Tyler said; a more actionable proposal would specify how to strengthen it, what policies it should enforce, and what operational role it should have.

The definition of “frontier” is central. A regime could leave ordinary AI applications alone: a recommender system at Netflix might use AI but not qualify as a frontier system. Yet determining which systems cross that line requires benchmarks that are regularly updated, difficult to game, and credible across competing labs.

Foreign and open models pose an even harder enforcement problem. Coogan questioned whether a review process imposed on major U.S. labs would simply slow the actors easiest to regulate while advanced open models from outside the country continued to circulate. He cited Kimi K2 and GLM as examples of the concern: a closed model delayed for months could be overtaken by, or distilled into, a more widely available alternative.

Hays suggested that compute may be easier to regulate than model weights. Coogan speculated that enforcement could focus on large hosting providers, inference platforms, or major distribution sites such as Hugging Face and GitHub. That would put control at the infrastructure layer rather than solely with the model developer—a prospect Coogan expected to be controversial among open-source advocates.

There is also a competition question embedded in safety regulation. A demanding review process might slow leading closed labs and create room for open-source alternatives. Or it might concentrate power among organizations with compliance budgets, legal teams, and government relationships. Coogan compared that possibility to biotech, where small companies may be acquired by large pharmaceutical companies before launch because the larger firms have the infrastructure to navigate regulatory approval.

Coogan welcomed Hassabis making his position explicit, but found the practical path unresolved: who defines “frontier,” how quickly testing can operate, how foreign and open models can be governed, and whether safety rules will make regulatory capacity another advantage held primarily by the largest labs.

New York Governor Kathy Hochul signed an executive order placing a one-year pause on new AI data centers while the state develops a regulatory framework and conducts environmental impact assessments. A post shown during the discussion described the measure as the nation’s first yearlong moratorium on new AI data centers.

The state plans to examine energy demand, water use and water quality, air quality, and effects on the electric grid. Coogan questioned why some of those issues would not already be covered by general environmental rules, but the order treats AI data centers as a distinct category requiring dedicated review.

If it stands, the policy would make New York the first state to impose a broad moratorium on large-scale AI data centers. Technology-industry critics argue that restricting construction could cost communities jobs and weaken the U.S. position in the AI race. A similar proposal in Maine was vetoed by Democratic Governor Janet Mills amid concern that it would block a major project in a town still recovering from the closure of a paper mill.

Hochul’s Republican challenger, Bruce Blakeman, opposed the pause on the grounds that local governments—not the state—should decide whether to approve projects promising significant economic benefits. Jordi Hays said he was not aware of a substantial data-center boom in New York. Coogan emphasized that the order’s actual reach will depend on how the state defines an “AI data center”: by electricity use, installed GPUs, or another standard.

Hays referenced Ken Griffin’s argument that data centers will be built somewhere, and that refusing to build them locally shifts associated revenue and investment elsewhere. Coogan’s immediate point was that projects might go to other states before other countries; Hays’s broader concern was that New York could become a model for wider domestic restrictions.

The siting dispute is not only about industrial policy. It is about the physical experience of living near the infrastructure. The source showed both a generic, windowless facility immediately beside a suburban neighborhood and a Gensler rendering for a proposed Phoenix data center with textured concrete, rust-colored panels, landscaping, and palms. The visual contrast sharpened the claim that local resistance is shaped not only by a site’s water and power demands, but by whether its architecture treats neighboring residents as an afterthought.

Hays called the suburban placement “objectively” rough for the people living next to it. The Phoenix concept did not settle the underlying environmental questions, but it illustrated the alternative the hosts had in mind: a facility designed to resemble a campus or civic-scale building rather than a blank industrial box.

Coogan cited a Wall Street Journal report on architects attempting to reduce local backlash by making data centers look more like tech campuses or art museums. Gensler design director Jeffrey Diamond argued that data centers are buildings like any other and should not be designed to look worse. Coogan’s view was that a more deliberate façade should cost very little relative to the overall project while materially changing public perception.

Hays set a more demanding standard. If a data center were clean, did not drive up energy costs, used no water through a closed-loop system, and looked like the Phoenix rendering, he said, he would not merely accept it in his town—he would want it there.

Architecture can address one source of resistance. It cannot answer concerns about electricity demand, water, emissions, grid capacity, or proximity to homes. New York’s policy therefore turns on whether the state can review and mitigate those local burdens without sending the jobs, revenue, and investment that supporters associate with data-center construction to another jurisdiction.