onsemi Data Center Energy Reporting Challenges Blanket Moratoriums

onsemi wants data center operators to disclose how efficiently they turn grid electricity into AI computing, despite growing political support for construction moratoriums. CEO Hassane El-Khoury argues that communities need comparable operating data before deciding whether a proposed facility wastes power or deserves approval.

The proposal arrives during a conflict between two blunt measurements. Developers announce facilities by electrical capacity, often in hundreds of megawatts or even gigawatts. Residents experience the projects through utility bills, water demands, noise, land use, and new power infrastructure.

New York has chosen a temporary pause for new hyperscale facilities while it develops stricter standards. El-Khoury calls blanket restrictions shortsighted. His alternative, onsemi data center energy reporting, would judge facilities by useful computing output relative to the energy drawn from the grid.

That sounds more precise than approving or rejecting projects by size alone. It also raises difficult questions. AI workloads vary, useful compute has no universal unit, and the company proposing the metric sells the power chips needed to improve it.

onsemi Wants Efficiency Disclosures, Not Blanket Bans

The immediate change is onsemi’s call for a public efficiency standard that goes beyond a facility’s advertised electrical capacity.

El-Khoury presented the argument during a tour of onsemi’s manufacturing facility in East Fishkill, New York. According to the original energy disclosure report, he said governments should demand information about energy waste instead of rejecting AI infrastructure as a category.

The distinction matters because a project’s announced capacity says little about what happens after electricity reaches the property. A proposed one-gigawatt campus tells regulators how much power the development expects to consume at maximum scale. It does not reveal how much electricity ultimately reaches processors, memory, networking equipment, and storage.

Several conversion steps sit between the grid connection and an AI chip. Each step changes voltage, converts alternating current to direct current, or distributes power across racks and boards. No conversion is perfect, so some energy becomes heat.

That heat creates another burden. Cooling equipment consumes additional electricity, while some cooling designs also require substantial water. A loss inside the electrical system can therefore produce a second operational cost when the cooling system removes the resulting heat.

El-Khoury wants policymakers to compare raw electricity consumption with the useful computing delivered by a facility. In principle, this would reward operators that produce more AI work from the same grid connection.

The proposal shifts the debate from scale toward performance. A large facility with efficient power delivery, optimized servers, and high equipment utilization might outperform a smaller but poorly used site. Capacity alone would not reveal that difference.

Current industry metrics only partly solve the problem. Power usage effectiveness, or PUE, compares a facility’s total electricity consumption with the electricity used by its information technology equipment. A PUE closer to 1.0 indicates that less energy supports cooling, lighting, and other overhead.

PUE does not measure whether processors are doing valuable work. A building filled with idle servers can record an attractive PUE because nearly all its electricity reaches IT equipment. The metric does not tell the public how effectively that equipment completes AI training or inference tasks.

El-Khoury’s proposal therefore goes beyond ordinary facility efficiency. It asks operators to connect grid demand with computing output. That is a more revealing goal, but it is also much harder to standardize.

The timing reflects onsemi’s commercial position. The company supplies power semiconductors that act as fast electrical switches in data centers, vehicles, and industrial systems. Better conversion can reduce electrical losses throughout what the industry calls the power tree, the chain that carries electricity from a facility connection to individual chips.

onsemi also has a direct interest in convincing governments that better hardware can relieve part of the energy problem. More rigorous disclosure would give efficient power components greater visibility during procurement and permitting decisions.

That interest does not invalidate the proposal. It does mean policymakers should separate the useful idea of disclosure from any single vendor’s preferred measurement or technology.

Why AI Data Center Energy Reporting Is Becoming Urgent

The pressure comes from a widening gap between the speed of AI construction and the quality of public information about its local costs.

U.S. data centers consumed about 176 terawatt-hours of electricity in 2023, or roughly 4.4% of national electricity use. A federal energy demand study projected that their share could reach between 6.7% and 12% by 2028.

The wide range is important. It shows that even national researchers cannot reduce future demand to one reliable number. Hardware shipments, AI adoption, server utilization, cooling systems, and operating practices can all change the result.

Communities must still make decisions before that uncertainty disappears. Utilities need generation, transmission, and substation plans years before a large customer begins operating. Regulators must decide who pays for upgrades if projected demand changes or a planned campus never reaches full capacity.

Residents face an information disadvantage during that process. Contracts between utilities and large customers can contain confidential terms. Project announcements emphasize investment, jobs, and maximum capacity, while operating data often arrives later or remains aggregated.

The federal government has begun addressing the measurement gap. In March 2026, the U.S. Energy Information Administration launched three voluntary pilot studies covering Texas, Washington state, Northern Virginia, and Washington, DC.

The agency identified 196 companies operating data centers across those regions. Its pilot energy survey asks participating companies about electricity use, energy sources, site characteristics, servers, and cooling systems.

Those categories are broader than a single efficiency ratio. They can help researchers distinguish computing demand from supporting infrastructure and identify differences among facilities.

However, the pilots ask each participating company to report information for at least one facility. They do not yet create a complete national inventory. Voluntary participation also leaves the government dependent on industry cooperation.

EIA has indicated that it plans to develop a mandatory survey after completing the pilots. That effort would give policymakers a stronger baseline for evaluating onsemi data center energy reporting or another standardized framework.

The reporting debate has also become a household affordability issue. Large facilities can require new power plants, transmission lines, substations, and backup systems. Whether ordinary customers absorb those costs depends on utility rate design, contracts, state regulation, and the accuracy of demand forecasts.

Efficiency disclosure cannot resolve every cost-allocation dispute. It can make those disputes more concrete. Regulators could compare projected consumption with measured operation, examine peak demand, and determine whether promised efficiency materialized.

Developers would also face pressure to define their claims. “Efficient” would need a denominator, a reporting period, and a consistent boundary around the facility. Without those details, the term remains a marketing description rather than an auditable result.

For knowledge workers following rapidly changing infrastructure commitments, a searchable AI knowledge base can help connect permits, utility filings, company statements, and later operating disclosures. The challenge is not finding one number. It is preserving enough context to compare promises with performance.

New York Chose a Pause While onsemi Argues for Measurement

The central conflict is targeted transparency versus precautionary limits, not chipmakers versus environmental protection.

New York made the conflict tangible in July 2026. Governor Kathy Hochul imposed a temporary statewide moratorium on new hyperscale data centers while officials develop stronger environmental and community standards.

The policy pauses new facilities above the state’s threshold rather than waiting for operators to prove their efficiency after construction. New York’s moratorium announcement framed the measure as protection against higher utility bills, water pressure, noise, emissions, and grid strain.

State officials also cited nearly 12 gigawatts of data center load requests in New York’s interconnection queue as of May 2026. More than eight gigawatts reportedly entered the queue during 2025.

An interconnection request is not the same as an operating facility. Some proposed projects change size, move, or never receive construction approval. Yet utilities must investigate credible requests because even a fraction of that demand can affect long-term planning.

The moratorium gives the state time to create rules before approving another wave of large loads. Its logic is precautionary. When information is incomplete and the potential cost is large, a pause limits exposure while regulators build a framework.

El-Khoury’s criticism targets the breadth of that response. He argues that policymakers should identify the particular harm they want to prevent and measure it directly. A facility should not be treated as unacceptable merely because it supports AI or consumes substantial electricity.

That argument has practical force. A blanket pause treats operators with different technologies, energy contracts, cooling methods, and community agreements similarly. It can also direct investment toward neighboring states without improving national energy efficiency.

The counterargument is equally practical. Efficiency standards do not automatically protect residents from infrastructure costs, local pollution, land conflicts, or water withdrawals. A facility can perform well on computing output per kilowatt-hour and still create unacceptable local burdens.

New York’s approach also recognizes an enforcement problem. A promised efficiency level offers limited protection unless regulators can inspect operating data, verify calculations, and impose consequences for underperformance.

The strongest policy may eventually combine both approaches. Governments can require environmental review and cost protections before construction, then demand standardized operational disclosure after a facility comes online.

Under that model, reporting would not replace permitting. It would create a feedback loop. Measured results from operating sites would improve future forecasts, rate cases, permits, and community agreements.

This is where onsemi data center energy reporting becomes more than a corporate talking point. If standardized independently, it could help regulators distinguish efficient projects from facilities that consume reserved capacity without producing the promised economic value.

It would also expose whether a moratorium addresses the right unit of risk. States may discover that grid timing, local generation, cooling water, or customer cost guarantees matter more than a facility’s nameplate capacity alone.

The Hard Part Is Defining Useful AI Compute

A compute-per-energy standard is attractive because it sounds objective, but AI workloads resist simple comparison.

AI training and AI inference perform different jobs. Training builds or updates a model using large datasets and repeated mathematical operations. Inference uses a trained model to answer a prompt, classify information, generate media, or take another action.

Even within inference, one output can differ sharply from another. A short classification request uses different resources than a long reasoning session. Text, images, audio, and video also impose different computing and memory demands.

Counting tasks would therefore create distorted comparisons. An operator could appear more efficient by processing many easy requests instead of fewer demanding ones.

Counting tokens, a common unit for pieces of text processed by a model, would not fully solve the issue. Model architectures handle tokens differently. Longer context windows increase memory traffic, while reasoning systems can perform internal computation that is not visible in the final response.

Chip performance benchmarks offer another option. Regulators could compare completed standardized workloads per unit of energy. Benchmarks make hardware configurations easier to evaluate under controlled conditions.

Yet a benchmark is not an operating record. Real facilities run mixed workloads, experience changing utilization, and reserve spare capacity for traffic spikes or failures. A system optimized for one benchmark can perform differently under customer demand.

Utilization creates another complication. An efficient server that remains idle most of the day can waste more total energy than a less efficient server kept consistently busy. Facility disclosure must therefore include both component performance and operational use.

The measurement boundary matters too. A company might report energy consumed inside the data center while excluding electricity used by nearby generation, networking infrastructure, or water treatment. Another operator might include those systems.

Time also changes the result. Renewable generation can vary by hour, and regional grids rely on different resources during peak periods. Annual averages can hide whether a facility adds demand during the most constrained hours.

A credible framework needs several complementary measures rather than one score. These could include total electricity consumption, peak load, IT energy, cooling energy, server utilization, completed standardized workloads, water use, and carbon intensity by reporting period.

The European Union offers a relevant precedent. Its Energy Efficiency Directive established reporting obligations for qualifying data centers, while the European Commission has worked on a common sustainability rating scheme. The EU’s data center framework collects information before policymakers settle on a final rating method.

That sequence is instructive. Governments can require consistent raw data first, examine what the information reveals, and then decide which composite metrics deserve regulatory weight.

The United States could follow a similar path through EIA’s survey work. Facility-level confidentiality would still require careful treatment, particularly when operating data reveals customer activity or commercially sensitive capacity.

Public reports may need aggregation or delayed release. Regulators, however, could receive more detailed submissions under confidentiality protections. Independent auditors could verify calculations without exposing customer workloads.

The standard-setting process should include operators, utilities, semiconductor companies, community groups, efficiency researchers, and independent measurement experts. onsemi can contribute technical knowledge, but it should not control the denominator that determines whether its customers look efficient.

Efficient Power Chips Help, but They Do Not Settle the Debate

onsemi’s technology addresses real electrical losses, although chip-level gains cannot answer every community objection.

Data center power delivery involves repeated conversions. Electricity may arrive as high-voltage alternating current, pass through transformers and switchgear, move into direct-current distribution, and step down again before reaching processors.

Power semiconductors control those transitions. Silicon carbide and gallium nitride devices can operate at high voltages, high switching frequencies, or elevated temperatures in selected applications. Their characteristics can reduce the size of some components and improve conversion efficiency.

onsemi has promoted solid-state transformers for future AI facilities. A solid-state transformer uses power electronics for voltage conversion and control, replacing some functions handled by conventional electromagnetic transformers.

The company says higher-voltage direct-current architectures can reduce conversion stages. Its power conversion plan describes an industry shift toward 800-volt distribution later in the decade.

Removing a conversion stage can avoid part of the energy loss that would otherwise become heat. Higher distribution voltage can also reduce current for the same power level, which can lower resistive losses and conductor requirements.

Those gains become valuable at AI scale. A percentage point of efficiency at a very large campus represents meaningful electricity and cooling demand. The benefit can continue throughout the facility’s operating life.

onsemi has strong commercial reasons to emphasize this mechanism. In its second-quarter 2026 results, the company described AI data centers as its fastest-growing business and said related revenue should more than double during the year.

The company reported total quarterly revenue of $1.604 billion, up 9% from the prior year. It also highlighted a broader role in Nvidia’s MGX ecosystem, a modular server platform used by system builders.

Its quarterly results show why the reporting proposal matters strategically. If operators must document conversion losses, purchasing efficient power hardware becomes easier to justify.

That does not make the hardware unnecessary. It means readers should recognize both the engineering value and the incentive behind the policy argument.

Power conversion also represents only one layer of data center efficiency. Processor design, memory movement, networking, cooling, software scheduling, model architecture, and utilization can consume or save more energy elsewhere in the system.

A more efficient power tree can even contribute to a rebound effect. Lower energy use per computation reduces the cost of running AI, which can encourage customers to request more computation. Total electricity demand can rise even as each task becomes more efficient.

This effect does not erase the savings. It prevents policymakers from treating component efficiency as a guaranteed reduction in absolute consumption.

Data center developers must also manage when they consume power. A highly efficient campus that reaches peak demand during a grid emergency can create more system stress than a flexible facility that reduces workloads during constrained hours.

Reporting should therefore capture demand flexibility. Operators can shift some training jobs, use stored energy, adjust cooling, or schedule nonurgent work when electricity is more available. Not every inference workload can wait, but many computing tasks have scheduling flexibility.

The public value of disclosure lies in revealing these differences. Regulators could identify facilities that merely purchase efficient components and those that coordinate hardware, software, and grid demand effectively.

What onsemi Data Center Energy Reporting Must Prove Next

The proposal will matter only if independent rules turn an appealing efficiency concept into comparable, enforceable evidence.

The first signal to watch is EIA’s move from voluntary pilots to a mandatory national survey. The pilot already covers energy sources, electricity consumption, server characteristics, and cooling systems.

A mandatory survey would establish a common factual baseline. If it includes sufficiently detailed operating measures, it would strengthen the case for performance-based regulation. A narrow or heavily aggregated survey would leave local officials searching for other evidence.

The second signal is New York’s blueprint for rules after its temporary moratorium. Officials must decide which facilities can proceed, who pays for grid upgrades, how environmental effects are reviewed, and what operators must disclose.

If New York adopts measurable performance thresholds alongside community protections, it would narrow the gap between El-Khoury’s proposal and the state’s precautionary approach. If it relies mainly on project size, the conflict will remain.

The third signal is whether hyperscale operators publish comparable compute-efficiency data. Electricity and water disclosures are becoming more common, but useful AI output remains difficult to compare among companies.

A credible industry response would use common workload definitions, disclose measurement boundaries, report utilization, and accept independent verification. Selective case studies would weaken the argument that voluntary transparency can substitute for stricter limits.

Readers should also watch the difference between planned and operating capacity. Announced gigawatts attract attention, but actual grid demand depends on construction schedules, equipment delivery, customer demand, and facility utilization.

Reported efficiency should be tracked across time rather than celebrated at launch. A facility’s first months can look different from steady operation. Hardware refreshes, workload changes, and rising occupancy can alter performance.

Communities deserve more than a single corporate score. They need enough information to answer practical questions about bills, water, noise, emissions, reliability, and local benefits.

Developers also need predictable standards. A transparent framework can reduce political uncertainty by defining the evidence required for approval and continued operation.

onsemi’s intervention exposes a real weakness in the current debate. Megawatts describe the size of an electrical connection, not the quality or public value of the computing behind it.

Its proposed remedy remains incomplete. Compute output is difficult to normalize, efficiency can increase total demand, and better power chips do not resolve every local impact.

Still, measurement and precaution do not have to be opposing policies. Governments can protect communities before construction while requiring operators to prove performance afterward.

The next test is whether onsemi data center energy reporting becomes an independently governed standard or remains an argument from a supplier positioned to benefit. Readers tracking the issue should compare the coming federal survey, New York’s final rules, and hyperscaler disclosures. Those three signals will show whether the industry is ready to replace broad efficiency claims with evidence.