WaPo: Opinion/ Will AI require less electricity than thought?

[neillh...@earthlink.net]

The future of AI may need less electricity than we think

Analysis by Kathryn Clay

July 1, 2026 | Energy & Climate

SharePrintPrint

Tomorrow’s data centers may look very different from today’s as innovators seek to offset AI electricity demand

Artificial intelligence is often portrayed as a software revolution. In practice, its future may depend just as much on breakthroughs in physical infrastructure—from superconducting power cables and photonic networking to edge computing and advanced energy systems. Here, employees at VEIR install superconducting cable at an AI data center demonstration in November 2025. (Elayne Cronin, PhD / Elayne Cronin, PhD)

 

Key Takeaways

• AI electricity demand forecasts may be overstated because they fail to consider the likely impacts of future energy efficiency innovations.
• A new generation of companies is challenging long-standing assumptions about AI infrastructure and energy usage. Past experience suggests that major infrastructure transitions often produce innovations that dramatically improve efficiency.
• Superconductors could reduce electrical losses and increase power density. Companies including VEIR and Snowcap Compute are applying superconductivity to power delivery and computing — while Microsoft is already evaluating the technology for future AI infrastructure.
• Photonics could reduce one of AI’s fastest-growing energy burdens: moving data among thousands of processors. Companies such as Lightmatter, Ayar Labs and Celestial AI are replacing electrical interconnects with optical ones, dramatically reducing electrical losses and heat generation.
• Edge AI could reduce demand for centralized computing by performing many tasks closer to users. Advances by Qualcomm, Nvidia, Apple and start-ups such as webAI are shifting some AI workloads onto mobile devices, vehicles and industrial equipment.
• None of the above approaches is certain to succeed, however, it is likely that some innovations will bring major gains to AI energy in the coming years.

 

Intro: Technology often outpaces expectations when industry faces a resource challenge

Every forecast of AI’s future begins with the same assumption: Tomorrow’s data centers will look much like today’s, only bigger. History suggests that assumption is probably wrong. When demand collides with a scarce resource, engineers rarely accept the constraint — they redesign the system. AI may be approaching one of those moments.

Access to electricity is an important limitation in the AI data center buildout. Hyperscale data centers are already among the world’s largest electricity consumers, and AI workloads are dramatically increasing both peak power requirements and total electricity consumption. The newest AI campuses require hundreds of megawatts of power, and gigawatt-scale facilities are already moving from concept to construction. Utilities across the United States report growing interconnection queues, while technology companies are competing for scarce transmission capacity and generation resources.

 

There are many examples of technologies that initially appeared destined to overwhelm available resources before innovation changed the equation.

In the late 1980s, utilities warned that controlling acid rain would require prohibitively expensive pollution controls. Before the program entered into force, the Environmental Protection Agency (EPA) projected that even with technology advances, by 2010, the total cost to coal-fired power plant operators would be about $7.5 billion annually. By 1998, economists estimated the total cost to industry in 2010 would be between $1.1 billion and $1.7 billion.

EPA later estimated that compliance costs in 2010 were between $1 billion and $2 billion. Costs fell dramatically due to improvements in emission control technologies, cleaner fuels and emissions trading schemes.

 

In the 1990s, experts predicted that surging internet traffic would necessitate a massive expansion in communications infrastructure. Instead, companies developed fiber-optic cables and engineered waveforms to allow multiple data streams to travel the same fiber simultaneously. These innovations increased the existing network’s capacity by orders of magnitude. More recently, solar power and LED lighting followed similar paths, as rapid innovation drove costs down far faster than projections anticipated.

When markets face large technical constraints, they often produce solutions that early forecasts fail to anticipate. As AI pushes data center electricity demand sharply higher, a growing number of companies are betting that history may repeat itself — not by reducing demand for computing, but by fundamentally changing how computing is delivered.

These technologies are not merely engineering curiosities. They could influence where data centers are built, how much grid infrastructure utilities must construct, which critical minerals become strategically important, and which suppliers capture the next wave of AI investment.

 

Questioning assumptions about how data centers are built and operated may be key.

A growing number of innovators are questioning assumptions that have long shaped the design of data centers. This report highlights three examples. They are not the only ideas under development, nor will they necessarily succeed. But they indicate that engineers are beginning to go beyond simply making AI infrastructure more efficient incrementally — by designing improved GPUs, for example — and instead considering how to transform the underlying architecture of AI infrastructure. The three assumptions are:

Assumption 1: Electricity must flow through copper.

Copper has been the obvious choice for electrical components for many decades. It is inexpensive, abundant and malleable enough to be manufactured into a variety of forms. But the economics of data centers may change the calculation, making exotic materials more competitive.

Assumption 2: Data must move electrically.

Since the earliest days of computing, engineers have relied on electrical signals traveling through copper to move data between processors, memory and storage. That architecture has served the industry remarkably well, but, as AI systems scale to unprecedented size, approaches that were once too expensive may now make business sense.

Assumption 3: AI must run in giant data centers.

Cloud computing established the practice of performing the largest computing tasks in the largest data centers. That remains true for training frontier AI models. As everyday devices employ increasingly powerful AI processors, some companies are pursuing ways to get hundreds of millions of existing devices to shoulder more of the work.

More companies are questioning each of these assumptions. If even some of the proposed technologies prove commercially successful, they could reshape the economics of AI infrastructure — not by reducing demand for computing but by delivering more computing with less electricity. Each of these companies has its own approach, but all reflect the belief that the future of AI will be shaped as much by changes in the architecture of AI infrastructure as they are by improvements in AI models.

Analysis: Exploring three bets on a different AI future

 

Assumption 1: Electricity must flow through copper.

Traditional copper cables — the mainstay of electrical systems for over a century — generate heat as electricity flows through them. That wasted energy becomes an additional cooling load inside the facility.

A class of materials called high-temperature superconductors (HTS) eliminates virtually all electrical resistance when cooled with liquid nitrogen, allowing substantially more electricity to move through much smaller cables.

VEIR A superconducting cable is being wound at a manufacturing facility of VEIR. VEIR supplies superconducting cables to the AI data center market. (VEIR / VEIR)

A leading company pursuing a strategy to exploit the HTS advantage within data centers is VEIR, a venture-backed company headquartered in Woburn, Massachusetts. The company says its superconducting cables can carry roughly 10 times the power density of conventional conductors while dramatically reducing electrical losses.

 

Tim Heidel, VEIR’s co-founder and CEO, explained in an interview that conventional copper power systems are becoming a bottleneck for next-generation AI data centers, and that high-temperature superconductors offer a way to deliver dramatically more power in less space. Heidel said that his technology can lower data center electrical demand that can speed the deployment of these facilities. “In the race to win AI, speed to power is everything right now.”

A Microsoft investment in VEIR through its climate innovation fund shows that hyperscalers are beginning to evaluate superconductors as more than an academic curiosity. In February, Microsoft said it is investigating HTS technology to reduce transmission losses, increase power density and “rethink traditional power designs” as AI drives rapidly growing electricity demand. Researchers have explored superconducting computer chips for decades, but the technology remained largely confined to laboratories because it requires cryogenic cooling.

 

Operating superconducting circuits at temperatures only a few degrees above absolute zero demands specialized refrigeration systems that themselves consume significant amounts of electricity. Until recently, that energy penalty outweighed the potential benefits for most commercial applications. But the unprecedented power demands of AI are changing the equation. At the scale of today’s hyperscale AI campuses, operators may be willing to devote substantial energy to cryogenic cooling if it enables dramatically larger gains in computing performance and overall energy efficiency.

One company pursuing that vision is Snowcap Compute, a Silicon Valley start-up developing superconducting processors for AI and other high-performance computing applications. Superconducting circuits have essentially no electrical resistance under cryogenic conditions, allowing them to operate with far less energy than traditional silicon-based processors. Snowcap says its processors can be manufactured using existing semiconductor fabrication techniques, although they rely on superconducting materials such as niobium titanium nitride, a specialized compound derived from niobium and titanium.

 

“We’re not chasing Moore’s Law. We’re stepping beyond it,” co-founder Anna Herr wrote when the company was launched in 2025. (Moore’s Law refers to the long-standing observation that computing power roughly doubles every two years while costs per computation fall.)

As superconductor innovators are asking whether electricity can move more efficiently, photonics companies pose a more radical question: What if some data never had to move as electricity at all?

 

Assumption 2: Data must move electrically.

As AI clusters grow larger, moving data is becoming almost as important as processing it. Every time GPUs exchange information, electrical signals travel through copper connections that consume power, generate heat and eventually require amplification. At the scale of modern AI superclusters, that communication overhead has become one of the fastest-growing sources of electricity demand inside the data center.

 

The challenge is becoming increasingly important because interconnect technology is improving far more slowly than AI computing power. Lightmatter, one of the leading companies developing photonic networking technology for next-generation AI systems, notes that while frontier AI models have expanded roughly 240-fold in only three years, and AI clusters have grown tenfold, interconnect bandwidth has improved only about twofold — creating an increasingly severe bottleneck. “This is what the next era of AI infrastructure looks like,” said Nick Harris, Lightmatter founder and CEO, in a June 2 press release.

Photonics transmits information as pulses of light through optical waveguides (e.g., fibers) instead of electrical signals through copper. Because light can travel farther with much lower loss, using photonic interconnects could reduce both power consumption and heat generation, while allowing larger AI clusters to function as a single computing system.

The next revolution in AI may come not from more powerful processors, but from moving information more efficiently. By replacing copper with optical pathways, photonic chips could reduce one of the fastest-growing sources of energy demand in AI data centers. (Ann Wang / Reuters)

 

Lower networking power also reduces cooling requirements. Every watt that is not dissipated by high-speed electrical interconnects is a watt that does not need to be removed by cooling systems, allowing photonics to reduce both direct electricity consumption and the secondary energy required to keep AI hardware within operating temperatures.

 

Because optical interconnects allow GPUs to exchange information much more quickly, they also reduce the amount of time expensive AI accelerators spend waiting for data. Several developers argue that the result is not only lower energy consumption but faster training of frontier AI models because computing resources remain productive rather than idle.

Lightmatter has developed photonic interconnects and optical “engines” designed to connect AI chips with substantially lower power consumption than conventional electrical links. Rather than replacing GPUs, the company’s technology focuses on reducing the energy required to move data among them — an increasingly important challenge as AI systems scale into tens of thousands of accelerators.

 

Other companies are pursuing similar goals using different architectures. Ayar Labs, a Silicon Valley start-up, has developed optical input/output (I/O) “chiplets” that replace short copper connections with fiber-optic links. The company states that its technology improves data bandwidth per watt substantially. Ayar has attracted strategic investments from Nvidia, Advanced Micro Devices and Intel, reflecting growing industry confidence that optical interconnects will become a core component of future AI systems. Another company, Celestial AI, is developing what it calls a “photonic fabric” that uses light to connect processors and memory more efficiently, while reducing the energy required to move data across increasingly large AI clusters.

A third group of companies is questioning an even more fundamental assumption that is not just about how AI data center components are connected, but, rather, where AI runs in the first place.

 

Assumption 3: AI must run in giant data centers.

Some innovators believe that the best way to reduce data center electricity demand is to avoid sending so much work to data centers in the first place. Large language model training and many enterprise applications seem likely to require centralized computing for the foreseeable future. That is not necessarily true for most day-to-day work performed by AI systems. The AI work most users actually see — answering questions, generating text, recognizing images or translating languages — is called “inference.” Every inference task performed locally means less electricity consumed inside a remote AI campus.

Edge AI is a concept that takes advantage of the “latent compute” of billions of processors that already exist but remain underutilized much of the time. Rather than transmitting every request to a hyperscale facility, local devices — such as increasingly powerful smartphones, PCs, vehicles and industrial equipment — perform AI tasks and communicate with cloud systems only intermittently.

 

Industrial robots increasingly rely on AI running locally on factory floors rather than sending every decision to distant data centers. This "edge AI" approach reduces latency, improves reliability and allows machines to react in milliseconds. (Jeff Chiu / AP)

Edge AI will not eliminate hyperscale data centers. But widespread adoption of it could dramatically reduce the number and size of data centers that would otherwise be required.

Companies pursuing this strategy include Qualcomm, whose AI processors enable on-device inference in mobile and industrial applications; Nvidia, whose edge computing platforms target factories, robotics and autonomous systems; and Apple, which increasingly performs AI workloads directly on iPhones and Macs. Each of these approaches reduces network traffic by shifting computation away from centralized facilities.

 

Start-ups are pushing the idea further by envisioning networks of personal devices that collectively provide computing resources when idle, effectively treating consumer electronics as a distributed computing platform.

Rather than assuming AI must run in hyperscale cloud data centers, Austin-based start-up webAI is betting that much of it can run on hardware that organizations already own. Its software compresses and distributes AI models across devices ranging from laptops and smartphones to local enterprise servers — allowing AI applications to perform inference where the data resides instead of sending every request to a remote data center. The company argues that this “sovereign AI” approach reduces latency, enhances privacy, lowers operating costs and eases pressure on centralized computing infrastructure.

“We have this web of models, and we can target the right model for the task,” David Stout, webAI co-founder and CEO, said in an interview. “The energy reduction is so tremendous that you could even run most of these tasks on your consumer hardware. So, no data center required.”

Lookahead

None of these technologies has yet demonstrated commercial success. Superconductors remain expensive. Photonics must compete against decades of investment in conventional interconnects. Edge AI raises difficult questions about software deployment, security and model management. Past experience suggests that many promising infrastructure technologies fail: In fact, even revolutionary technologies can fail commercially because incumbent technologies continue to improve.

The question is not whether every one of these approaches succeeds, but whether one or more fundamentally changes the economics of AI infrastructure.

AI is unlikely to become less compute-intensive. But it may become dramatically more energy-efficient than today’s projections assume. The history of technology suggests that when demand collides with physical constraints, engineers rarely respond by accepting the constraint. Instead, they change the system. Whether through superconductors, photonics, edge computing or approaches yet to be imagined, the next big breakthrough in AI may not be a smarter model — rather it may be a more efficient way to power, connect and deploy one.

Recommendations

Don’t treat current electricity forecasts as fixed. Infrastructure planning should incorporate scenarios in which advances in networking, power delivery and edge computing reduce the long-term electricity intensity of AI workloads.

Broaden competitive intelligence beyond chipmakers. The next AI winners may be companies solving power delivery, networking and systems architecture rather than building larger processors.

Build flexibility into capital plans. Data centers designed around modular power systems, optical networking and distributed computing architectures will be better positioned to adopt emerging technologies.

Evaluate where inference actually belongs. Not every AI workload needs to be performed in a hyperscale facility. Organizations should assess whether some inference can be shifted to edge devices or private infrastructure to lower operating costs, improve latency, and reduce cloud dependence.

Monitor the critical minerals supply chain. Technologies such as superconductors and silicon photonics rely on specialized materials — including niobium compounds and indium phosphide — that could create new supply chain dependencies as deployment accelerates.

Separate short-term planning from long-term strategy. Utilities and developers should continue planning for substantial growth in AI electricity demand over the next several years. At the same time, they should recognize that technological breakthroughs could materially alter infrastructure requirements over the longer term, just as they have during previous periods of rapid technological change.

Stress-test long-term capital investments against multiple technology scenarios. Data centers designed around today’s assumptions may prove less valuable if breakthroughs in networking or distributed inference reduce future electricity requirements.

Nour Wood contributed to this report.