Sunday, November 22, 2026Vol. III · No. 326Subscribe
The Mining, Energy & Technology Wire

The Next Chip War Is Fought in Joules

Power, not silicon supply, now sets the pace of AI. Two very different designs, one simulated CPU and one compute-in-drive SSD, bet that the winners will be whoever wastes the least energy moving data.

The Next Chip War Is Fought in Joules
PhotographPower, not silicon supply, now sets the pace of AI. Two very different designs, one simulated CPU and one compute-in-drive SSD, bet that the winners will be whoever wastes the least energy moving data.Photo: Blake Wisz / Unsplash

In a Palo Alto simulation, a processor that does not physically exist spends just 6% of its energy on fetching, decoding and issuing instructions. The conventional core it is measured against spends 27%, according to Architect Labs, which built the comparison. Chips are no longer sold mainly on how fast they compute. They are sold on how little they waste, and the waste is mostly in moving data.

That is the argument here: with electricity, not wafers, now setting the pace of the AI build-out, the next contest among chipmakers is over joules per task. The winners will be the designs that stop hauling bytes around.

The grid sets the price of compute

The demand picture explains the pivot. The IEA's World Energy Outlook 2026 projects data centre electricity demand will double from about 500 TWh today to 1,000 TWh by 2030, and reach 1,500 TWh by 2035, Eco-Business reports. Servers built for AI drive most of the growth to 2030. Separately, the agency found that within four years AI data centers would consume an extra 350 terawatt-hours every year, more than the annual electricity consumption of the UK, the Taipei Times reported, relaying the Guardian.

Much of that power will be dirty. Coal and gas supply about 60 per cent of what data centres use today, and sector emissions are on track to rise 70 per cent by 2030, per the IEA. The agency expects about half the electricity for China's AI data centers to come from coal by the end of the decade, and the US to lean on gas for about 60 percent of its own by 2030. In August, plans existed for about 180 gigawatts of new gas plants linked to US data centers, the Taipei Times said, almost as much new gas capacity as the EU has installed in total.

Buyers feel this directly. Schneider Electric's Peter Weckesser told Reuters, in a report carried on radio station WBOW's site, that higher European power prices could push AI investment elsewhere because energy costs feed directly into the cost of computing. When the electricity bill is the product's cost base, a chip that does the same job on less of it is worth real money.

Where the joules go

Architect Labs gives the clearest picture of where chips burn them. Leaving out clock distribution, its comparison core spends 57% of its energy on memory, the company says. Frostbite, the design its AI systems discovered, designed and verified, cuts memory to 38%, Architect Labs says. It keeps data in small local memories beside the logic that uses it, which strips out much of the area and power a conventional chip devotes to large caches.

On a set of 25 workloads, 15 of them held out until the final evaluation, Architect Labs reports a 4.59x performance-per-watt multiple. The held-out workloads score 3.83x. The spread is wide. Tom's Hardware's table puts conv2d at 23.12x and recursive Fibonacci, the worst case, at 1.15x. Frostbite also pays for sharing work, spending 26% of its energy on its on-chip network and scheduler. The savings come from relocating costs, not abolishing them.

Phison attacks the same problem from the storage side. At Supercomputing 2026 in Chicago it showed an SSD with an NPU and onboard memory running at 130 GB/s, TechPowerUp reports. The drive scans and processes a dataset in parallel and sends back only the relevant results rather than shipping everything across the host. Phison says that could cut data transfers by 10 to 100 times for suitable workloads.

One design moves compute toward data inside a chip, the other inside a drive. Both treat the trip, not the arithmetic, as the expensive part.

The honest objections

The first objection is that the evidence is soft. Frostbite exists only as a simulated 7 nm design, and Architect Labs says so plainly: "These are, of course, still pre-silicon power numbers." The company cannot name its comparison core. "For legal reasons, we cannot disclose the exact CPU." reads a footnote in its post. Its power model was first off by up to almost 250% before an automated loop pulled it to within 5% of the circuit design, Tom's Hardware reports. TechPowerUp says the Phison drive's controller, interface and form factor are unknown and advises treating such products as proofs of concept. Architect Labs says it will test against multi-core arrangements with large caches and larger out-of-order CPUs on SPEC and PARSEC in the coming months. Until then, a lead over one undisclosed core proves less than it seems to.

The second objection is about scale. Data centres still account for only about 1.7 per cent of global electricity demand, and by 2035 they add less new demand than electric vehicles or factories making batteries and solar panels, the IEA says via Eco-Business. Why should a share that small reorder chip design?

Because the demand is lumpy, not diffuse. More than 80 per cent of it sits in China, the EU and the US, and it arrives at specific substations. In September Thailand paused 49 projects under construction and 117 awaiting approval while it wrote rules on electricity, water, siting and community impacts, according to the IEA. A regulator cannot freeze a global average. It can freeze one campus. The relevant constraint is local, and local constraints reward efficiency per rack.

The IEA also notes, as quoted by Eco-Business, that AI "could help to optimise the electricity system and bring down costs." The technology may help the grid even as it strains it. That does not change which side of the meter the chip designer sits on.

If this read is right

If joules per task becomes the metric buyers pay for, the pressure falls on whatever is hungriest. For Architect Labs, that means conventional cache hierarchies. It says it expects every class of chip, from datacenter CPUs to AI accelerators to network switches, to go through similar architectural shifts. The company's inference accelerator for robots and drones, Redwood, still runs only on an FPGA, so the claim is untested. A rival, Tendrils Compute of Cambridge, U.K., raised funding this month for a processor it says will have "thousands of cores operating in parallel."

The proof will come from silicon, not slides. Until a chip ships and a meter is attached, every efficiency number in this piece is a forecast. The utilities, the regulators in Bangkok and the buyers in Europe are already reading the meter.

Original reporting and analysis by the Stake & Paper editorial team. See linked sources within the article.

Share this story

More from Stake & Paper

Was this article helpful?

ClaimWatch

Mining claims intelligence — from query to report, in minutes.

Every unpatented mining claim across all twelve BLM states. Leadfile audits, due diligence, site selection, regional prospecting, entity investigations, and AOI monitoring — delivered as complete report packages.

4.4M+
Claims Tracked
12
BLM States
7
Report Types
Request a Sample Report
Stake & Paper AM

One morning brief. The whole energy sector.

Original analysis, the day's most important wire stories, and market data — delivered before your first cup of coffee. Free.