AI's hidden power problem: The little-known company Weeteq trying to cut the electricity wasted inside data centres

You may have often asked AI to plan your week, explain a medical report or debug your code - but what you don't see is a chip in a building somewhere consuming power to provide you with the results. There's a hidden power bill behind every result, answer you get from AI.

According to Glasgow-based deep tech company Weeteq (CEO Dr Taner Dosluoglu), power, not computing, is now the bottleneck in AI data centers. The industry has spent years perfecting the chips. The system that feeds them electricity hasn't kept pace with the development.

The problem

AI chips swing from idle to full load in a fraction of a microsecond. Today's typical power path steps voltage down from 48 volts to 12, then to 0.8. Those conversion stages can't react fast enough. So engineers play it safe and run chips at a higher voltage than they strictly need, to stay stable. Weeteq calls it a hidden tax.


It's built into every AI data center today and quietly drives up power consumption.

Jennifer Scanlon, CEO of UL Solutions, told Bloomberg that worn or damaged electrical equipment can raise the risk of arc flashes that could damage AI chips.

What's the fix on the table?

Weeteq's answer is DIVA.

DIVA is an acronym for 'Dynamic Current and Voltage Control for ASICs'.

Instead of one system doing everything, it splits the job in two. One path delivers everyday power efficiently. A second, ultra-fast path handles sudden spikes, responding at more than 10 amps per nanosecond. With spikes handled locally, the chip no longer needs that extra voltage cushion.

The numbers, as Weeteq reports:

• Voltage at the processor drops from 0.8 to 0.7 volts
• Power use in that domain falls by about 30%
• A compute tray goes from 15 kilowatts to 10
• A full rack goes from 120 kilowatts to 80
• Output stays the same, measured in token throughput

That's 40 kilowatts saved per rack, with less cooling needed, more equipment per space, and a lower total cost of ownership.

ALSO READ: A Minnesota couple built a business around rice harvested from lakes and rivers. Now it eyes $50,000 more in annual revenue

Scale makes it matter. Data centers are heading toward megawatt-scale racks, and the industry is shifting from 48/54-volt to 800-volt DC systems. DIVA is designed to work across that range. At that size, a kilowatt saved on one rack, multiplied across thousands of GPUs, stops being a rounding error.

"DIVA eliminates the artificial boundary between higher voltages and on-chip power domains. The architecture pairs a high-efficiency bulk path at 91% efficiency with a high-bandwidth fast transient path at 81% efficiency to actively compensate for voltage droops in real time," writes Weeteq on its page.

DIVA can work with existing IVR and digital LDO solutions. On-chip and in-package integrated voltage regulators can initially be used alongside the proposed parallel path delivery system, which matters for customers with existing investments in those technologies.

The reality check

Weeteq's figures show where the waste hides: the final stage, where power meets silicon.

Will a smarter power trick save AI data centers tons of electricity?

AI runs inside huge buildings called data centers, packed with powerful chips called GPUs. Everyone talks about how fast these chips are. Very few people talk about how they get their electricity.

ALSO READ: Is your used kitchen oil the future of aviation? Kelantan innovators turn UCO into sustainable jet fuel

Here's the problem,

Power reaches these chips in steps, a bit like water passing through a series of pipes, with the voltage (the electrical "pressure") lowered at each stage. But AI chips change how much power they need in a split second, jumping from relaxed to full sprint almost instantly. The power system can't always keep up.

While playing it safe, engineers give the chips a little more voltage than they really need, like carrying an umbrella every day just in case. That extra cushion keeps the chips stable, but it wastes electricity. With thousands of chips at work, power wastage adds up. The company Weeteq calls this a hidden tax.

Instead of one system doing two jobs, it splits the work. One path is steady and efficient, handling everyday power. A second, ultra-fast path jumps in the instant a sudden spike hits. Think of a marathon runner for the long haul, with a sprinter waiting on standby.

Because the spikes are covered, the chip no longer needs the extra cushion. Weeteq says the voltage can drop from 0.8 to 0.7 volts, cutting the chip's power use by about 30%.

For a whole rack (a cabinet full of servers), that means going from 120 kilowatts to 80, saving 40 kilowatts for the same amount of work. Less power also means less heat, so less cooling is needed.

So why does this matter now?

With data centers growing fast and electricity becoming their biggest limit, saving 40 kilowatts per rack, multiplied across thousands of racks, is a big deal.

Amber Villegas-Williamson, principal consultant at the Uptime Institute in the UK, which advises electricity suppliers and data centres on standards and reliability, told The Star that AI does create very unusual power demand. It’s like over-revving your car wears out the engine faster than keeping a constant speed.


Sources:

Weeteq

Codebase

FAQs

1. How much electricity will AI data centers consume by 2030?
   The International Energy Agency's 2026 assessment projects global data center electricity consumption will increase from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours by 2030. AI-focused data center consumption is projected to triple over that period
2. Why do AI data centers consume so much electricity?
   AI data centers use powerful processors, including GPUs, to train and run artificial intelligence models. These processors require substantial electricity, while cooling systems and other infrastructure add to overall energy demand.
3. Could more efficient AI chips reduce electricity costs?
   More efficient chip power delivery could reduce electricity consumption and associated cooling requirements for comparable workloads. However, actual savings would depend on equipment configuration, electricity prices, utilisation and deployment at scale.