Assertion Supported AI assessment confidence: 95% certainty 4/5 debate potential 1/5

Rao: Vast majority of energy in digital systems goes to data movement

Naveen Rao · Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea · Feb 7, 2026 · at 1:02:46

Unconventional AI CEO Naveen Rao answers an audience question about the fundamental bottlenecks facing current computing architectures.

0:00 / 0:06exact quote · 6.8s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“Well, yeah, I mean if you look at an existing digital system, the vast majority of energy actually goes into data movement.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

In existing digital computing systems, the vast majority of energy is consumed by data movement rather than computation.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

Peer-reviewed computer architecture literature and empirical hardware benchmarks confirm that in standard digital computing and AI systems, data movement accounts for the vast majority (typically 60% to over 85%) of total system energy consumption. Foundational hardware characterizations, including Horowitz (IEEE ISSCC 2014), show that DRAM reads require over two orders of magnitude more energy than 32-bit floating-point arithmetic operations. Furthermore, system-level evaluations across standard workloads (e.g., Boroumand et al., ACM ASPLOS 2018; Mutlu et al., 2020–2023; LLM inference benchmarks 2023–2026) show that memory access, interconnects, and caching consume between 62.7% and 90% of dynamic energy.

Mark Horowitz (IEEE ISSCC 2014) established that a 32-bit DRAM read consumes ~640 pJ compared to ~0.9 pJ for a 32-bit FP addition and ~3.7 pJ for multiplication in 45nm CMOS.
Boroumand et al. (ACM ASPLOS 2018) demonstrated across standard consumer and machine learning workloads (TensorFlow Mobile, Chrome, VP9) that an average of 62.7% of total system energy is spent strictly on data movement.
Mutlu et al. (2020–2023 PIM syntheses) and modern LLM inference characterizations (2023–2026) report that data movement across memory hierarchies accounts for 60% to 90% of total dynamic system energy in AI and data-intensive computing.

Checked against

[1] semanticscholar.org [2] nationalacademies.org [3] nationalacademies.org [4] rackcdn.com [5] cmu.edu [6] ethz.ch [7] uni.wroc.pl [8] researchgate.net

energy data movement vs computation computer architecture percentage DRAM SRAM "data movement" "62.7%" OR "62%" OR "60%" "energy" "workloads" ASPLOS Mutlu Mark Horowitz Computing energy problem ISSCC 2014 energy data movement arithmetic data movement consumes energy computation 60 percent 80 percent IEEE ACM "energy spent on data movement" "computation" IEEE OR ACM "accelerators" energy breakdown of modern GPUs data movement vs compute "energy is consumed by data movement" "computation"

assessed as of 2026-09-05

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

data movement digital system

⚑ Spot an error? Report it

More from Naveen Rao

Prediction Not checkable as stated

Rao: Trillion-dollar GPU data centers are physically impossible to power

“None of this is physically realizable, right? That's what's really kind of annoying to me. I'm like, okay, you're talking about, you know, a trillion dollars of GPUs. Where the fuck are you gonna get the power for that? I'm sorry, right? Like you're just, you'…”

Naveen Rao Feb 7, 2026 ▶ 37:35 Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea

Opinion

Rao: xAI's GPU clusters broke rules and are not repeatable

“And the things that have been realized I mean, like the XAI clusters and stuff like that, You know, valiant efforts. No doubt about it. And it's amazing to push through and get it done, but it broke a lot of rules and is doing a lot of stuff very inefficiently…”

Naveen Rao Feb 7, 2026 ▶ 37:47 Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea

Opinion

Rao: Circular AI mega-deals are justified by coming demand

“I know we talk about it looks bad or whatever, but the bet is that the demand will come, and I think that's actually a good bet. So I see it as like, all right, fine, it looks, it doesn't look great in the short term, and I can see why people are irked by it b…”

Naveen Rao Feb 7, 2026 ▶ 45:16 Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea

Opinion

Rao: True AI requires fundamental hardware substrate innovation

“My personal take is that we need innovation on the hardware substrate to actually build true intelligence. You're not going to get there with this sort of rudimentary thing that we have.”

Naveen Rao Feb 7, 2026 ▶ 8:41 Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea

Opinion

Rao: Nervana Was Sold to Intel Too Early

“And so, you know, we sold too early. Sold to Intel too early.”

Naveen Rao Feb 7, 2026 ▶ 19:25 Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea

Prediction Not checkable as stated

Rao: Unconventional AI Could Open a Computing Paradigm Lasting 80 Years

“I think we have the opportunity to completely open up a new paradigm that could exist for another 80 years.”

Naveen Rao Feb 7, 2026 ▶ 21:40 Inside the $4.5B Startup Building Brain-Inspired Chips for AI · Sourcery with Molly O'Shea

Rao: Vast majority of energy in digital systems goes to data movement

More from Naveen Rao

Turn any episode into a week of clips.