Data-center delays: AI demand, community costs and hardware risks

After the US midterms, AI infrastructure stocks fell sharply. Silicon Valley 101 invited guests from energy, data centers and investing to ask whether markets fear construction delays or are reassessing genuine AI demand. The program also interviews residents near data centers.

The opening previews differing views: regulation could add project costs, voter opposition should not be underestimated, and demand is central to the industry outlook. Residents also question projects and seek compensation.

The program previews local interviews: some residents oppose data centers in their communities, while others seek compensation for housing and infrastructure costs. The hosts then begin the full discussion.

The host says changes to data-center approvals, taxes and community conditions could affect supply chains for GPUs, optical communications, storage and power equipment. Slower approvals constrain the speed of computing supply; they do not necessarily prove demand has disappeared.

Ethan Xu thinks the risk of substantial data-center delays over the next one or two years has increased after the election, with the scale depending on local policy. He also worries some projects could be canceled due to opposition, not just postponed.

Ethan says developers may need deeper negotiations with communities over grid upgrades, electricity prices, water, roads, hospitals and schools. He expects some places with high demands and no agreement could reject projects.

He thinks Ohio, Georgia and other states may also face policy changes, and will watch how Texas balances residents and developers. Tougher proposals from politicians responding to voters could increase market volatility.

Ethan expects months of bargaining could produce a compromise in Texas less severe than markets fear, but says the legislative and negotiation cycle itself is unfavorable for construction. He thinks stock volatility could persist until a plan becomes clearer.

He expects a long legislative cycle to leave developers uncertain while they wait. Even if final rules are milder than the most extreme proposals, projects and related stocks could remain volatile before a consensus forms.

Asked about added costs, Ethan roughly estimates that electricity and water compensation could add 5%–10%; heavier taxes could raise costs 10%–20% in a worse case. He emphasizes that the terms and range remain uncertain.

Ethan thinks companies may find it relatively straightforward to discuss paying additional electricity and water costs; removing incentives or adding sales and property taxes could increase the burden substantially. His range is not a uniform quote.

Christina Xu separates regulation into federal, state and local levels: different agencies and procedures govern permits, tax incentives and electricity prices. She thinks the biggest impact on projects may be when they can enter service.

Christina says rising electricity bills are a credible community concern; water impacts, she thinks, depend on cooling technology. She cites one 140-megawatt facility in an investment using recirculated cooling, which she says uses about as much water as ten households. It is one example.

On employment, Christina thinks AI may displace some white-collar jobs, while data-center construction also needs blue-collar workers. She says developers may have to pay for supporting infrastructure and taxes.

Christina describes an invested Texas data-center project that needs many construction workers and supporting infrastructure such as roads. She sees benefits in blue-collar jobs and tax revenue, and thinks political incentives for a complete shutdown are limited.

Citing her reading of economic data, Christina says AI investment contributes to growth, so she expects policymakers will be reluctant to halt construction across the board for long. She sees some recent opposition as election-period sentiment.

Ethan says some technology companies have begun paying for power-grid and water-system upgrades. He proposes an example where a company expands a water system, reserves part of the new capacity for residents and uses the rest to cool its data center, potentially helping gain community acceptance.

Ethan stresses that his water example is hypothetical: a company spends $30 million on an expansion adding five million gallons a day, half for residents and half for the data center. He uses it to illustrate a possible sharing arrangement.

In Ethan’s view, a developer funding some utility upgrades could benefit both sides and help secure a local license to operate. He treats infrastructure investment as part of negotiation, rather than equating compensation only with direct payments.

Bruce Liu asks whether the jobs are temporary construction work or lasting employment. If only a few operating jobs remain after construction workers leave, residents may be reluctant to bear the project’s effects.

Ethan acknowledges the distinction: building a large data center may take two or three years and average roughly 2,000 workers on site, whereas operation may require only dozens. He says that differs from a car factory employing thousands long term.

He nevertheless expects US data-center construction could continue for five to ten years, with workers moving between projects. Thus jobs may not stay in one location, but the industry construction cycle could last longer. Christina also describes higher pay when hiring electricians.

Ethan says workers may move to another project in the same state after finishing a two- or three-year build. Jobs at one site end, but industry demand may continue. He compares this mobility with real estate or highway construction.

Ethan says skilled construction workers can earn high wages. Christina cites a project hiring senior electricians at around $300,000 a year, an example from that project rather than a figure for all positions.

Christina says community hearings often require several rounds addressing pollution, electricity, water, noise and traffic. She gives examples of roads and supporting infrastructure at Texas projects; Ethan says his team also meets residents and negotiates item by item.

Christina says developers may build roads, cover local costs and in some cases consider emergency power. These matters often require multiple hearings and negotiations rather than a single decision.

On costs, Christina estimates a one-gigawatt data center might require $50 billion–$60 billion, with computing equipment accounting for roughly half to 60%. Roads and utilities also cost money, though their spending structure differs from equipment.

Christina thinks one reason major model companies design their own chips is to reduce computing-equipment costs, a large share of the total. The host notes that infrastructure and compensation costs must still be assessed case by case.

Ethan says terms vary by state and even by community. Possible costs include grid and water upgrades, higher electricity costs, sales and property taxes, and hospitals, schools or training facilities; his earlier 10%–20% figure is a rough worse-case estimate.

Christina also notes voter concerns about corporate tax incentives and school-district spending. She and Ethan think that even if project costs rise, providers with existing capacity could gain pricing power if less new capacity becomes available.

The program introduces the Texas gubernatorial election and the incoming governor’s position on new approvals, as well as restrictive congressional proposals concerning data centers and advanced AI. The host stresses these are proposals whose passage is uncertain; state and local authorities also have different powers over utilities, land and hearings.

The program distinguishes federal proposals, state administrative approvals and local land-use decisions. Different levels affect different aspects, and a proposal is not an enacted law. The guests think changes at these levels make projects harder to predict.

Rob Li thinks markets had partly anticipated Democratic control of Congress, but the Texas governor’s race may have caused investors to reassess local permitting risk. He calls this concern about policy change “Gina risk.”

Rob says Texas is important for data-center construction and the governor has some administrative authority, so markets will watch her statements. He does not equate the election result with every project stopping.

From the incoming governor’s policy statements, political background and union support, Rob speculates that her eventual stance may be more centrist than a total construction ban. It is his judgment, not a settled policy outcome; he still sees delays or cancellations as real risks.

Rob reviews the governor-elect’s family and political background and speculates she may be closer to Democratic centrists. He acknowledges that a father’s views do not necessarily represent his daughter’s, so background is only one part of his judgment.

Rob also points to unions and employment: he thinks supporters want to protect jobs, and data-center construction requires electricians and others, so a total ban could conflict with some supporters’ interests.

He compares the governor-elect’s remarks on regulation with stricter proposals by other politicians, and sees her focus as community benefit. Rob infers that a final arrangement may require compromise among investment, voters’ concerns and jobs.

Rob says the governor-elect emphasized regulation and community benefits rather than the outright ban he associates with the most aggressive proposals. A result milder than markets’ worst expectations could change sentiment, he thinks, but policy takes time.

Even if policy ultimately lands on a compromise, Rob still expects some projects may be delayed, canceled or relocated. If markets previously priced an extreme outcome, a milder development could change expectations.

From valuation changes he infers that some investors fear not just delays but the entire AI narrative. He contrasts semiconductor and consumer-stock valuations to explain his reading of the market and says pricing should be distinguished from actual demand.

Bruce says opposition to data centers is not confined to one party: local voters in both may worry about electricity, water and the environment. Voters can change policy through elections, and the industry should not underestimate grassroots opposition.

Bruce says local opposition can cross party lines. He thinks markets may previously have assumed federal positions would dominate local policy, while the election brought local constraints into sharper focus.

Bruce thinks data centers can bring jobs and local economic benefits, but says these were not adequately explained during the election. He also acknowledges that water, environmental and electricity-price concerns are real and cannot be resolved simply by announcing long-term benefits.

He cites the Richmond, Virginia area as an example where jobs and economic activity from data centers can be described with data, but thinks those benefits were not communicated well during the campaign.

At the same time, Bruce says a person familiar with local districts repeatedly warned him not to dismiss resident opposition. He thinks the industry needs both to explain benefits and address real costs rather than pretend resistance is absent.

He expects political actors may ultimately seek compromise, but says it takes time for residents to experience tangible benefits. If planned data centers are not ready on time, chip suppliers’ earlier earnings forecasts may suffer in the short to medium term.

Bruce offers his own political assessment of Democratic factions, thinking some newly elected politicians may seek a more centrist approach. Even so, policy communication and distributing benefits take time.

He says semiconductor earnings forecasts depend on checking construction plans with customers and their customers. If projects no longer proceed on schedule, earlier delivery and earnings assumptions could weaken.

From an investment standpoint, Bruce is unsure whether markets have fully priced those risks. He says optimistic long-term prospects can coexist with near-term earnings pressure, and investors still need to manage exposure.

Bruce says short-term risk management requires not just judging technology’s long-term prospects but considering whether investors have already priced the earnings effect of project delays.

Bruce says the key is communicating with local voters and managing near-term risk; he remains optimistic about long-term economic benefits. He recounts hearing from a lawmaker that the direction was to support AI development while requiring safety measures, possibly including an emergency shutoff.

Bruce distinguishes possible federal safety rules from local data-center permits. His account of an emergency shutoff describes his understanding of a possible future bill, not a uniform requirement already in force.

Asked about the hardest permits, Ethan identifies grid connection, air pollution and land approvals. Noise, visual impacts and taxes could also enter negotiations, and policy may change over the coming months.

Ethan adds that noise, visual impacts, local taxes and other requirements can affect project timing. He expects residents, companies and politicians to keep discussing those conditions in the coming months.

Ethan says some resident requests concerning grid upgrades and electricity bills are reasonable and should be addressed. At the same time, he worries expanding approvals could produce excessive regulation and delay. This is his assessment of regulatory risk.

He also mentions disputes over land and wildlife effects. Ethan worries that expanding permit requirements could lead to overregulation, while saying reasonable grid and electricity-price concerns should be addressed.

Ethan thinks appropriate regulation could benefit communities and companies, but worries excessive delay could hurt technology companies, stock markets and retirement accounts holding related assets. That is his forecast of potential economic effects.

The program traces the chain: data-center plans affect orders for chips, storage, optical communications and power equipment; cloud companies use capital spending to produce future revenue; emerging cloud providers may also rely on long-term customer contracts, equipment collateral and external finance.

The host stresses that construction limits do not automatically mean a collapse in demand. If demand remains strong, existing capacity could become scarcer; delays could also postpone cloud revenue, hardware orders and project cash flow, affecting financing.

On moving construction abroad, one guest says existing projects still provide a time buffer and companies are unlikely to move all plans quickly because of short-term policy changes. He compares relocation with supply-chain moves and says decisions involve long-term costs.

The guests think proposals to move construction to Southeast Asia or space may also serve as negotiating leverage, and do not mean companies have decided on mass relocation. Actual site selection still requires assessing energy and local policy.

Christina says some cloud providers are seeking power in Europe and elsewhere, and some startups will use overseas computing, but large model-training and inference companies must still weigh domestic deployment. She has not seen a clear mass migration.

Christina mentions emerging cloud firms seeking overseas power, but says large model companies must consider domestic training and inference needs. Overseas opportunities exist, she thinks, but do not by themselves show a wholesale exodus.

She and another guest say moving across borders takes time and infrastructure. A restriction in one place does not mean all projects can immediately be transferred elsewhere.

Rob says demand, not supply alone, is central to whether the industry faces a crisis. If demand is unchanged while supply grows more slowly, owners of existing capacity may gain short-term pricing power; if demand disappears, the whole chain suffers.

Rob says providers that already have data centers could find it easier to sell existing capacity to strong demand when supply grows slowly; the logic changes if demand fails to materialize.

Rob uses cloud companies’ long-term growth expectations to show the other side: scarcity might support near-term prices, but delayed new facilities could prevent firms meeting expansion goals. He thinks markets have not reached a clear conclusion about the trade-off.

He cites optimistic long-term guidance from a large cloud company as context and says such targets assume sufficient new capacity. The issue is whether forecasts can be achieved, not a guarantee of a particular growth rate.

Rob offers a hypothetical: if a company expected fast growth but insufficient data centers force slower growth, the market may reassess its stock. Near-term pricing power and long-term capacity constraints can coexist.

He returns to demand: if customers keep buying computing capacity, supply problems may be resolved through adjustment; if businesses see no return and stop buying, the investment case changes.

Bruce is skeptical of the term “circular financing,” noting that one party’s spending is another’s income. He thinks actual revenue must support investment, while acknowledging that data centers are capital-intensive and need capital markets.

Bruce uses the circulation of money to argue that companies investing in and buying from each other does not automatically prove a bubble. He says verifiable revenue growth matters and demand cannot be judged from transaction relationships alone.

Bruce partly attributes credit-market pressure to higher and more volatile Treasury yields, and says concentrated issuance can temporarily congest markets. For him the question is not whether every company can borrow, but what price borrowers of different credit quality must pay.

He describes higher interest rates feeding through to credit markets: borrowers may still obtain financing but pay more. Large bond issues concentrated in time can also temporarily absorb market capacity.

Bruce says creditworthy issuers may still complete deals after a funding bottleneck eases. Financing terms differ, so he looks at credit spreads and underlying assets rather than making a blanket judgment.

He says a strong large technology company and an emerging cloud provider dependent on a new project may face different financing costs. Approval prospects and when cash flow begins affect the risk premium lenders demand; borrowing also depends on expected returns.

Using hypothetical interest rates, he says lenders may demand a larger risk premium from an emerging cloud company facing greater uncertainty. The company must compare borrowing cost with expected project returns before deciding.

Bruce thinks AI infrastructure capital spending may exceed current revenue for a period, with capital markets needed to fund the gap. Availability and price of financing still depend on the project and market conditions.

The host raises SemiAnalysis forecasts for data-center and chip supply. Ethan disputes its higher estimate of US capacity coming online in 2027 and thinks the US is more likely to face data-center shortages than excess facilities in 2027–28.

Ethan says he agrees with some of the research firm’s semiconductor analysis but not its conclusions about US data-center capacity and chip supply over the next two years.

Ethan says some projects lack all required approvals, while new grid reviews and community negotiations may increase delay risk. He sees possible local cases globally of facilities without chips, but not as the main trend.

Ethan thinks substantial planned Texas capacity is still in grid approval or review, so announced plans cannot all be assumed construction-ready. New reviews, costs and community requirements may further affect completion.

He therefore disagrees with predictions of many empty US facilities in the near term. Local mismatches elsewhere may occur, he says, but should not automatically be treated as a global trend.

On whether existing capacity immediately improves profits, Ethan notes that many orders were signed at fixed prices and providers must consider long-term reputation, so not all contracts can be assumed to reprice. He expects regulatory and public-opinion uncertainty may keep markets volatile.

The guests ask whether improved model efficiency reduces hardware demand. One investor says the result depends on demand growth after prices fall: a 90% drop in unit compute cost with demand growing less than tenfold has different implications from demand growing more than tenfold.

The guests do not rule on a particular single-rack estimate because they do not know all its assumptions. They instead discuss whether better model efficiency lowers computing cost per task.

One investor says total hardware demand may still rise if cheaper computing stimulates enough demand; with a weaker response, it may fall. The key is how demand responds to price changes.

Rob says the growth rate of enterprise AI spending over the next two years remains difficult to know even when cloud and chip companies give optimistic forecasts. He suggests watching actual usage visible to data platforms and clouds, while noting real-time demand cannot prove long-term projections.

Rob cautions that cloud providers or suppliers forecasting 35%–40% annual growth do not prove enterprise customers have decided to raise spending at that rate. Many companies themselves may not know future budgets, he says.

He suggests tracking enterprise usage visible through Snowflake, Databricks and major cloud platforms, while stressing these are current or near-term signals rather than proof of spending two years ahead.

Ethan is more optimistic about demand, seeing large cloud backlogs as a signal. Bruce warns that orders can be canceled and wants evidence closer to end users; they differ on whether backlogs adequately validate demand.

Bruce says backlogs can be large but customers may cancel or adjust, so cloud firms may not be the first to see final-demand changes. He thinks markets will also watch model-company user and revenue growth.

The guests mention large cloud-company backlogs, but Bruce emphasizes that contracts can change, so markets need to keep monitoring customer usage.

Christina says annualized revenue at frontier-model companies is not the only measure; inference platforms and industry-wide token use also matter. She says use grew rapidly at some startups she knows, but that is her observed sample.

Christina adds that beyond model-lab revenue, usage of more inference providers and open models should be tracked. She cites changes at some portfolio companies without treating those cases as market-wide statistics.

On recursive self-improvement, Christina says some teams apply it to safety, drug development and low-level computing, while still seeking more compute. She thinks competition depends on customers and data as well as sufficient computing capacity.

The guests discuss model price cuts: lower prices may expand use, but markets may worry about revenue growth if usage does not rise enough to offset them. They offer competition and financing as possible motives, not established facts.

The guests differ on whether model-provider price cuts are necessary. Christina thinks the cuts may be strategic without necessarily proving weaker demand; others worry that usage may not grow enough to offset lower prices.

The discussion also covers cheaper models and open-source alternatives. One guest says businesses may choose open models for data security and privacy as well as price.

The host asks whether other US states can take the projects. The guests say firms will assess alternatives, but local approvals, energy and construction resources mean relocation takes time.

Ethan says moving data centers to other states requires more than cheap power: experienced construction labor, fiber backbones, local rules and political stability matter. Many companies are preparing alternatives, but quick relocation is difficult.

Ethan explains that remote regions with ample power may still lack skilled workers willing to move for years; a new fiber backbone may also take years. Local government and resident attitudes can change.

Companies planning interstate alternatives must therefore assess legal and community stability. Ethan thinks backup plans are worthwhile but not an immediate solution.

Christina adds that power permits take time, and self-supply also requires generating equipment and infrastructure. She says a long gap can separate a completed shell from actual power connection, so planned capacity is not equivalent to usable compute.

Christina says self-generation or behind-the-meter power requires actual generating capacity, equipment and permits. Planned industry capacity is not the same as capacity connected and available.

On emerging-cloud financing, Christina sees substantial differences among companies. Equipment allocations, facilities and long customer contracts can influence lending terms; companies must weigh the certainty of long contracts against potentially higher on-demand prices.

She explains that emerging cloud providers may use chip allocations, customer contracts or facilities to obtain finance. Whether lenders accept short-term on-demand revenue affects how companies design contracts.

In her view some firms want to retain high on-demand prices but also need long-term contracts to support financing, so bargaining power and capital structure matter. Emerging cloud companies cannot all be treated alike.

The host asks whether markets are also pricing AI safety regulation. Bruce says market pricing is hard to separate precisely, but expects possible bipartisan safety measures; as he understands it, the direction remains supporting development while requiring attention to safety.

Bruce says markets may not have priced restrictions much stricter than the direction he currently understands. He expects more safety-regulation debate but does not think current policy aims to obstruct AI across the board.

Ethan thinks local permit and community risks for data centers may be harder for technology firms to control than model companies’ internal safety processes. He also expects pre-release safety testing could moderately extend development as models become more capable.

Ethan says frontier labs already check model safety before release and can build safeguards into research and development, while local project approvals involve factors outside any company’s sole control.

The guests discuss corporate self-restraint and public regulation as possible approaches to a dedicated AI safety body. Bruce says his latest understanding remains that policy supports development while requiring substantive safety measures.

The guests think testing internally, then externally, before release could be a reasonable slowdown as models become more capable. Future policy discussions may change; their views reflect information currently available.

On hardware, Bruce worries data-center delays could cause upstream suppliers to cut order and earnings-per-share forecasts even if end demand remains strong. He sees the speed and size of cuts, not just valuation, as a risk to watch.

Bruce offers a scenario: if some data centers assumed in forecasts are not ready on time, equipment suppliers may need to reduce delivery and earnings expectations. He presents this as a scenario, not a universal order cut already underway.

He thinks markets have reacted to some risks, but the actual scale of future cuts remains uncertain. Better macro conditions could support valuations, while changed earnings expectations could still pressure stocks.

Ethan thinks GPU and optical-module deliveries could be deferred by project delays, while storage and some power equipment are already tight and may be affected differently. Another guest warns that easing shortages could weaken pricing power and pressure related stocks.

Ethan says delayed projects may defer planned GPU and optical-module deliveries. Storage, if already severely constrained, could remain undersupplied even if demand falls.

The guests use transformers to discuss long lead times for power equipment. One investor cautions that storage-stock valuations may rely on rising prices, so a change in price expectations could matter even if quantities remain tight.

An investor thinks storage-company shares may be driven more by expectations of price increases; a shift in supply and demand that weakens pricing power could create risk even if shipment volumes change little.

Bruce says he would watch upstream earnings or order cuts over the next few quarters. Another guest would track actual gigawatts brought online by major cloud firms each year. Ethan expects 2027 and 2028 deployments may be below original plans.

Bruce says negative free cash flow at major cloud firms is related to high capital spending and does not by itself prove demand deterioration. He is more concerned with whether equipment suppliers actually see order cuts than with market fears alone.

Another guest suggests tracking actual facilities brought online. Ethan says announced plans should be distinguished from gigawatts actually commissioned each year and expects some plans may be postponed.

Ethan thinks negative media narratives about data centers have intensified public opposition; another guest links cost-of-living concerns to war, tariffs and data centers. These are their interpretations of opinion and costs; the discussion also acknowledges that real community concerns affect votes.

Ethan believes recent reporting on data centers has been negative and has influenced public perceptions. He wants quantitative analysis that discusses benefits and costs together.

He mentions examples of higher electricity bills in some areas, while others emphasize household cost pressures. Guests offer their own judgments about the largest contributors; the program does not present a single consolidated accounting.

Bruce thinks war and tariffs also contribute importantly to living costs and that data centers have received too much blame in public debate. This is his interpretation of inflation and media narratives.

The host notes signals of continuing growth in model development and usage, contrasting with negative market sentiment. The guests say research teams’ experience may differ from how stocks price future risk.

The guests think Silicon Valley researchers and small-town residents who do not use AI may experience very different changes. The host says the latter’s voting power cannot be underestimated and the industry needs to understand host communities.

The guests say Silicon Valley’s everyday AI use does not represent every community. Residents who do not directly use these products still vote locally, and technology firms need to address their actual costs.

Ethan again says media narratives can change public opinion. The host then introduces resident interviews so people can describe their experiences directly rather than only through industry guests.

The program turns to residents. Monterey Park resident Steven questions whether a data center can benefit a city without increasing noise, pollution and resource burdens, and disputes technology companies’ account of AI demand.

Steven says that if projects really brought revenue without noise, property-value, air or water problems, cities should be more willing to accept them. He doubts those conditions can be met.

The program says 88% of voters in a Monterey Park referendum favored permanently banning data centers. Mayor Elizabeth Yang says resident opposition and forceful letters from the developer’s lawyers led the city to put the decision to voters; she thinks the city can prosper without data centers.

Elizabeth Yang says residents wrote in opposition after the application, and she did not receive a clear response to her invitation for the developer to answer questions at a community meeting. She says the developer’s lawyers warned the city could face a lawsuit.

She believes giving voters the decision allowed residents to express their views directly. She also says the city developed for many years without a data center and does not think rejecting one would cause its economy to collapse.

Abilene, Texas resident Steve says housing, rent and traffic pressures rose during data-center construction. He created a website to collect feedback and thinks communities bearing these costs should receive benefits such as road repairs, school improvements or direct compensation.

Steve created Save Abilene to gather anonymous resident reports about rent, housing, traffic and noise. He says some residents report that construction has affected commutes and housing choices.

Steve calls residents reluctant “investors” because they bear changes to quality of life. He wants the developer or government to share benefits through tax rebates, road and school improvements or other measures.

The program concludes that delays may affect orders, earnings and financing but do not by themselves prove AI demand has disappeared. Silicon Valley focuses on technological progress while small-town residents care about electricity, water and lasting jobs; how those costs and benefits are shared will also affect policy and investment.