Data centers are not modern factories and rarely create broad local prosperity.
They often raise local power costs while delivering few permanent jobs.
Only strict public-benefit energy rules can rebalance the deal.

The data center boom is changing small towns and electrical networks faster than most local governments can track. Global electricity use by data centers reached the hundreds of terawatt-hour range recently, and industry forecasts see consumption roughly doubling within a decade — a scale shock that would make a single new campus as consequential for a rural grid as a mid-sized steel plant once was. That stark fact reframes our question: are data centers destined to play the same civic and economic role that factories did in the 20th century, building broad local prosperity through plentiful jobs and supplier networks, or will they instead function as capital-intensive, low-employment facilities that extract affordable power and offer limited community payoff? This column argues the latter is the default outcome today. The differences among those futures depend on policy: electricity contracts, tax incentives, and whether each data center must be paired with on-site generation or with community power programs that intentionally return value to local residents. Without those bargains, the “21st-century factory” claim collapses into a short construction boom and long-term local costs.

Why the data center boom looks factory-like — and why that resemblance misleads

The raw numbers explain the analogy’s appeal. Modern hyperscale data centers require land, roads, cooling and, above all, immense, reliable power. That visible footprint resembles a factory campus: new buildings, heavy trucks, an inflow of contractors, and municipal revenues tied to construction and property. In many places, officials and economic development teams sell the project this way: quick construction jobs, higher tax bases, corporate donations and the hope of a nascent tech cluster. Those short-lived gains are real. Construction phases can employ hundreds or thousands of workers for months or years. Local suppliers benefit for the duration. Municipal budgets experience an immediate spike in receipts attributable to permit fees and developer payments. Those outcomes are the origin of the factory metaphor — a large, visible employer that forms local life.

But the similarity stops when facilities begin steady operations. Unlike a 20th-century factory that required large, skilled, and semi-skilled local labor forces for day-to-day production, many modern data centers run on automation and remote management. Once built, the permanent on-site workforce is often small relative to a site’s capital value and energy draw. The pattern emerging from recent studies shows a recurring mismatch: big infrastructure and power demands, modest operational payrolls. That means the civic bargain that once justified factory siting—dozens or hundreds of durable local jobs plus supplier ecosystems—is not automatic for data centers. The policy lesson is plain: if a community treats a data center solely as a factory replacement, it will likely end up with an expensive power bill and a handful of long-term jobs, not the broad prosperity promised.

Energy, bills and the hidden transfer of value in the data center boom

Power is the fulcrum. Data centers’ electricity demand is already substantial; reputable sector projections indicate that global data-center electricity use will increase substantially in the near term. That increase matters at the local level because grid upgrades, emergency generation and firm capacity all carry costs. Utilities and developers negotiate who pays. If those costs are socialized—through rate changes or long-term cost-recovery mechanisms—local ratepayers may face higher bills, while the data center benefits from access to cheap, reliable power. In effect, the private benefit of hosting a data center can become a public cost unless contracts are reformed.

That transfer is not hypothetical. Where clusters concentrate, utilities have warned that new, sudden loads require new substations, reconductoring, or even fossil-fuel backup to meet reliability standards during peak demand or outages. Those investments are often recovered through rate mechanisms that touch every household and small business. The visible result is an alarming local complaint: residents see new, gleaming campuses and then receive higher bills. For many small or rural communities, that outcome is politically explosive because it reverses the expected social contract. The structural fix is to rewrite the bargain: require onsite firm capacity that benefits the community; demand lower wholesale rates for residents; or tie grid upgrades directly to shared community pricing programs. Otherwise, the data center boom behaves like an energy parasite — great, concentrated demand that saps local affordability while producing limited local employment.

Jobs, taxes and the gap between promise and reality of the data center boom

Employment claims have driven many municipal approvals. Industry and allied consultants frequently produce projections that include indirect supply-chain jobs, and some national studies report figures that appear striking when read without context. Those broader estimates frequently include induced employment—the jobs supported across an entire national supply chain—and one-time construction roles. In contrast, direct, permanent on-site employment at a single hyperscale campus is often modest: dozens to low hundreds, not thousands. This distinction matters because communities vote, plan schools, and approve zoning in accordance with expectations regarding durable jobs and tax flows.

Recent audits and state studies show the dissonance. Detailed, state-level reviews show that while capital investment and construction increase short-term employment and local revenues, recurring operations generate far fewer local jobs than advertised—and tax deals often blunt the fiscal gains. When generous exemptions or equipment tax breaks are granted to attract projects, net long-term revenue may decline, leaving local budgets to shoulder infrastructure and service costs. Critics maintain this represents a poor return on public incentives. Supporters counter that national supply chain impacts and corporate tax remittances justify incentives. The sensible policy path recognizes both truths: count direct permanent roles separately from temporary construction work and design incentive packages that scale with demonstrable, long-term local benefits. Without that discipline, the data center boom will deliver a construction windfall and then a thin, steady stream of local benefits—far from the mass-employment factories once offered.

A practical bargain: SMRs, community power, and rewriting the deal for the data center boom

If data centers will not reliably create large-scale local employment, then communities must demand a different kind of return: shared, affordable, and resilient energy. Small, small reactors (SMRs) and other firm, local generation options have moved from abstract ideas to concrete corporate experiments. Major cloud providers have announced agreements and investments to pair data centers with advanced nuclear or long-term firm capacity procurement. When implemented, such arrangements can insulate grids, reduce marginal power costs for local customers, and provide a revenue or capacity stream that can be repurposed for public goods. The crucial policy requirement is simple: require every major new campus to demonstrate community power benefits that lower household or municipal rates or to dedicate a concrete portion of firm output to local public use.

There are multiple practical forms this bargain can take. An SMR or co-located firm generator, jointly owned with the utility, can deliver lower-cost, uninterrupted power to a local tariff band for residents and municipal facilities. Alternatively, a community benefit agreement can lock a data center into long-term payments to a local energy trust that subsidizes electricity for low-income households and public schools. Another option is a capacity-sharing rule: a percentage of off-peak power is sold back at cost to local businesses. Each design needs clear, enforceable metrics and independent auditing to prevent creative accounting. The point is not to stop data centers, but to align their massive energy footprint with durable, local public value, rather than abandoning communities to bear the external costs while tech firms capture private gains.

Reclaiming the civic bargain in the data center boom

The data center boom will not, on its own, become the 21st-century factory that spreads prosperity widely. The default pattern is one of enormous capital investment, intense short-term employment, and small operational payrolls — coupled with long-term pressures on energy systems and local affordability. Communities can accept that reality or they can legislate a different bargain: require firm generation commitments, community-priced electricity, and incentive structures that scale with demonstrable local benefits. These are not theoretical fixes. There are emerging corporate and utility experiments pairing data infrastructure with SMRs and long-term power contracts — moves that, if replicated under public oversight and with binding community returns, would produce something closer to a factory’s civic role. The policy test for the next wave of siting decisions is this: can a locality negotiate a binding, transparent package that turns raw kilowatts into local wealth, not local cost? If the answer is no, then the towns that host these campuses will have traded a short boom for a long bill. It is time to make the data center boom pay for the people it affects.

The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.

References

Energy Information Administration (IEA). 2025. Energy and AI: Energy demand from AI. International Energy Agency.
JLARC (Joint Legislative Audit and Review Commission). 2024. Data Centers in Virginia. Commonwealth of Virginia.
Lawrence Berkeley National Laboratory (Shehabi et al.). 2024. United States Data Center Energy Usage Report. LBNL-2001637.
Northern Virginia Technology Council (NVTC). 2024. The Impact of Data Centers on Virginia's State and Local Economies.
PwC / Data Center Coalition. 2025. Economic Contributions of Data Centers in the United States 2017–2023. PricewaterhouseCoopers.
Financial Times. 2024. Google orders small modular nuclear reactors for its data centres.
Food & Water Watch. 2026. The Illusion of Big Tech's Data Center Employment Claims.
Congressional Research Service. 2026. Data Centers and Their Energy Consumption (CRS product R48646).

The third AI stack is a political ambition, not an industrial reality
China’s open-source push wins users, not hardware supremacy
Europe and Korea must focus on interoperability and skills, not full-stack rivalry

Between August 2024 and August 2025, Chinese open-model developers accounted for roughly 17% of global Hugging Face downloads—surpassing U.S. developers for the first time. That single data point captures a truth and a misreading at once: China has achieved scale in open-source and community adoption, but scale in downloads is not the same as an independent, competitive compute and platform stack. The talk of a third AI stack—a European or Korean alternative that will sit beside the U.S. and China—mixes aspiration with wishful thinking. Open-source diffusion buys influence and the breadth of users. It does not automatically produce the high-performance silicon, the global cloud infrastructure, or the developer lock-in that underpin dominant AI platforms. If decision-makers and instructors in Europe and Korea make strategy from comfort, they will misread the extent of capital, talent, and time required to translate downloads into durable industrial power.

Why the "third AI stack" is a mirage

When we refer to the third AI stack, we mean a self-sustaining, region-led combination of hardware, software standards, and large-scale cloud services that can engage global customers and developers independently of U.S. or Chinese ecosystems. That is a very high bar. Building it requires three mutually reinforcing assets at an industrial scale: advanced silicon, a large installed base of high-end cloud nodes, and widely adopted developer tools and runtimes. Europe and Korea have pockets of strength—excellent researchers, healthy public funding, leading semiconductor firms in Korea—but they lack the simultaneous depth across all three pillars. Export controls and supply-chain frictions have altered incentives. The U.S. controls high-end GPU exports and has layered licensing reviews, slowing the direct flow of the world’s fastest inference hardware; this constrains some Chinese compute access but does not erase the huge lead that proprietary GPU frameworks and optimized software stacks already enjoy. Those controls alter incentives but do not shrink the technical gap overnight.

Capital matters more than rhetoric. To catch up after a decade of lag requires orders of magnitude more investment than typical national AI grants provide. Training the largest models needs sustained petaflop-years of compute and the advanced interconnects and memory hierarchies that only a few vendors can supply. Korea’s chipmakers can and do make excellent GPUs for consumer and embedded markets; competing at datacenter-class AI compute needs sustained foundry access, advanced packaging, and validated software toolchains that are expensive and time-consuming. Likewise, European cloud operators can focus on sovereign clouds and regulation-friendly offerings, but winning global developer mindshare requires low friction, competitive pricing, and a thriving community-driven ecosystem—things that privilege incumbents with broad scale. In short, the gap is structural, not simply financial or political.

Finally, we should separate user-facing openness from platform control. China’s success in open-model downloads shows how well governments and firms can seed ecosystems when proprietary hardware is constrained. But hosting, inference, and production deployments still congregate where performance, reliability, and developer convenience converge. Until an alternative stack meets those baseline expectations on cost and latency, enterprises and researchers will keep one foot in the U.S. or Chinese ecosystems, and hedging will win over wholesale migration.

How China’s open-source play reshapes the “third AI stack” narrative

China’s open-weight push is clever and consequential. By freely releasing model weights and encouraging derivative works, Chinese firms have multiplied global usage and created an optics advantage in the Global South and among developers who prize openness. That movement is not trivial: an ecosystem of lighter models, efficient condensation techniques, and developer tools can make AI useful inside constrained settings. This becomes a form of platform-building that competes on accessibility rather than raw exaflops. The result is a two-track competition: one axis is the top-tier compute race; the other is broad-based software adoption. China leads the latter. That matters for standards, formats, and the flows of talent and data that shape long-term software interoperability.

But this competitive win has limits. Open models need inference capacity to power commercial services at scale. When the highest-performance accelerators are absent or restricted, the economics and latency of large-scale services change. China’s response has been to develop indigenous chips and high-volume data center capacity. Early chips are serviceable and serve many workloads well, but independent analyses suggest they lag top-tier Nvidia chips in raw throughput and capability—sometimes by margins that matter for the most demanding training and inference uses. The engineering deficit combines node-level performance gaps with deficits in software maturity: compiler toolchains, distributed training orchestration, and optimized kernels are hard to replicate quickly. Engineers can reduce gaps over time, but these gaps interact with supply chains and standards in ways that compound the incumbent's advantage.

Another realistic implication: China’s open strategy has bought it a different kind of influence. If Chinese models, libraries, and runtimes become the de facto norms in many emerging markets, they shape developer expectations and institutional procurement. Europe and Korea can still set rules about data governance, privacy, and ethical uses. But norm-setting without compatible infrastructure is incomplete. The countries that combine standards with market-simulacra—deployable clouds, certified hardware, and developer incentives—will have more leverage. The open-source approach gives China a broad base from which to project standards through use; it is not yet a claim on high-end performance, but it is a claim on adoption and default formats.

Policy choices for educators, administrators, and policy makers in a US–China duopoly

Educatorshave to adapt curricula to the new terrain. Teaching neural network architecture without hands-on experience on modern accelerators is increasingly theoretical. Universities and vocational programs should prioritize access to a mix of cloud credits that expose students to both leading commercial stacks and popular open-source stacks used in real-world settings. That means negotiating partnerships to rotate cloud access, funding GPU hours for labs, and building coursework that evaluates compromises between performance and cost. Policymakers should treat computer access as infrastructure—like labs or telescopes—rather than an ephemeral grant. Subsidizing access to testbeds, investing in software toolchains, and underwriting open benchmarking projects will have outsized returns compared to one-off research grants. When students learn to optimize for latency and cost on real hardware, they graduate with industrially relevant skills rather than abstract knowledge.

Administrators and procurement officials face a choice matrix. They can double down on national sovereignty—buying local hardware and mandating local deployment—or pragmatically hedge by instrumenting multi-cloud and multi-stack interoperability. The latter is the wiser posture. Insisting on a pure “third AI stack” that isolates education systems and public services risks lock-in to immature platforms that will carry higher long-run operating costs. Instead, public institutions should require interoperability layers, fund open benchmarks, and sponsor translator tooling so models and datasets can port across stacks. That approach safeguards the regulatory agency while keeping operational performance within tolerable bounds. It also creates an exportable product: governments that can demonstrate safe, portable, and efficient AI deployments will have a stronger voice in global standards negotiations.

We must also face hard political-economic questions. China’s decision to prefer domestic stacks is a deliberate infant-industry strategy. If those domestic engineers succeed, the global configuration shifts; if they fail, China will still have bought time for its software standards to entrench in pockets of the world. Either outcome imposes costs and benefits on education systems. For universities and think tanks, the practical implication is clear: invest in cross-sectional analysis and applied deployments now. Trial projects that benchmark similar services on different stacks, publish the methods and results, and teach students to reason about trade-offs. In short, convert geopolitical uncertainty into a pedagogical advantage.

From wishful thinking to workable strategy

The third AI stack continues as a compelling political slogan. It promises national autonomy and dignity. It will not, however, spring into being because leaders declare it. The evidence is simple: open-model downloads are a form of influence but not an immediate substitute for exascale compute, top-shelf hardware, or mature software stacks. Europe and Korea can and should aim for partial sovereignty—investing in secure clouds, supporting open toolchains, and training engineers across multiple stacks—but the clearer path to impact lies in pragmatic interoperability, targeted industrial investment, and educational reform. We should treat compute and developer access as public goods, not prizes to be hoarded. That means real budgets for testbeds, transparent benchmarks, and curriculum revisions that privilege operational competence. If policymakers act with that realism, educators will graduate learners who can move between stacks rather than waiting for a mythical third column to appear. The world will be safer and more plural if alternative stacks grow by winning customers with clear trade-offs, not by political proclamation. The opening statistic matters because it highlights where influence is currently accumulating; our job now is to convert that influence into durable, responsible capacity—practically, cheaply, and with an eye on standards, not slogans.

The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of the Swiss Institute of Artificial Intelligence (SIAI) or its affiliates.

References

Chin, C. L. (2026). Standards are the new frontier in the US–China AI competition. East Asia Forum.
CSIS. (2025). The limits of chip export controls in meeting the China challenge. Center for Strategic and International Studies.
Financial Times. (2025). China leapfrogs the US in the global market for 'open' AI models. Financial Times.
Hugging Face / Stanford HAI analysis. (2025). China's diverse open-weight AI ecosystem and its policy implications. Stanford HAI briefing.
International analysis of chip performance. (2025). The H20 problem: inference, supercomputers, and US hardware. Independent analysis.
Reuters. (2026). NVIDIA AI chip sales to China stalled by US security review. Reuters.