Latest / Best of LinkedIn: Digital Transformation & Tech / Best of LinkedIn: Green ICT & Sustainable AI CW 42/ 43

Transcript

Best of LinkedIn: Green ICT & Sustainable AI CW 42/ 43 · Best of LinkedIn: Digital Transformation & Tech

• 0:00
  Speaker 1: This episode was provided by Thomas Allgaier in Frennus, based on the most relevant LinkedIn posts about green ICT and sustainable AI from CW-Forty Two and Forty Three.
• 0:12
  Speaker 1: Frenness supports ICT enterprises in the form of delivering precise ICT market and pricing intelligence that analyst subscriptions an existing databases cannot provide.
• 0:21
  Speaker 1: you can find more info
• 0:24
  Speaker 0: Right and today We are extracting the top green ICT and sustainable AI trends that are currently circulating across The professional tech community.
• 0:33
  Speaker 0: Yeah,
• 0:33
  Speaker 1: the mission of this deep dive is well to really cut through the hype
• 0:37
  Speaker 0: for you.
• 0:38
  Speaker 0: Exactly
• 0:39
  Speaker 1: we want to uncover the actual physical software And the regulatory realities of deploying Sustainable AI.
• 0:45
  Speaker 1: so it's basically your shortcut To being highly will inform for your next digital transformation strategy meeting.
• 0:52
  Speaker 0: Because there's a lot going on behind the scenes that people miss?
• 0:55
  Speaker 0: For sure,
• 0:56
  Speaker 1: I mean we constantly hear about how much power AI uses right but we rarely discuss the massive physical constraints of just cooling it down.
• 1:04
  Speaker 0: Yeah to really understand the footprint of AI you have to start at the foundational physical layer The data center.
• 1:11
  Speaker 0: where is this intense This tug of war between power and water unfolding right
• 1:17
  Speaker 1: now.
• 1:18
  Speaker 1: It's a brutal reality check.
• 1:19
  Speaker 0: it is Avanish Pandar shared some data that just breaks down the Brutal math of evaporative cooling, its roughly one point eight liters Of water per kilowatt hour.
• 1:29
  Speaker 1: Wow
• 1:30
  Speaker 0: wait so what does That look like at scale?
• 1:31
  Speaker 0: well
• 1:32
  Speaker 1: for A standard ten megawatt IT load?
• 1:34
  Speaker 1: That facility Consumes about a hundred and fifty-eight thousand cubic meters of Water year
• 1:38
  Speaker 0: Which Is What?
• 1:39
  Speaker 0: Like A small Lake
• 1:40
  Speaker 1: it's Roughly sixty Three Olympic size swimming pools.
• 1:43
  Speaker 0: oh my God.
• 1:44
  Speaker 0: And the alternative, like if you want to avoid that is dry coolers.
• 1:47
  Speaker 1: Right just blast it with air.
• 1:48
  Speaker 0: Exactly That brings your on-site water usage effectiveness Your WUE close to zero.
• 1:54
  Speaker 1: But There's always a but Always.
• 1:56
  Speaker 0: It requires massive fan energy and huge physical footprint.
• 2:00
  Speaker 1: So its basically a zero sum game.
• 2:01
  Speaker 1: You're trading municipal water for grid power
• 2:04
  Speaker 0: Completely And Nathaniel brought up really good point about this.
• 2:07
  Speaker 0: Its actually Water Quality Not necessarily the source that dictates cooling viability.
• 2:13
  Speaker 1: Oh Really?
• 2:13
  Speaker 1: So it doesn't have to be fresh drinking water.
• 2:15
  Speaker 0: No, exactly!
• 2:16
  Speaker 0: ASHRAE actually flags issues like water hardness and ammonia because if the chemistry is wrong...it destroys the cooling loops.
• 2:22
  Speaker 0: Oh
• 2:22
  Speaker 1: that makes sense.
• 2:24
  Speaker 1: It scales up or corrodes the pipes?
• 2:26
  Speaker 0: Yeah so as long as a chemistry has treated you're fine.
• 2:29
  Speaker 0: AWS's already using reclaimed water at twenty-six data centers which proves..you know You don't always need to tap fresh groundwater.
• 2:36
  Speaker 1: Okay let's unpack this Because optimizing just the data center building without looking at what the servers are doing, it's kind of like putting a hyper-efficient hybrid engine into a tank and calling it eco friendly.
• 2:49
  Speaker 0: The head is great way to put it right?
• 2:51
  Speaker 1: I mean we have look at actual output in this system
• 2:54
  Speaker 0: And were seeing industry wake up that Paul Chernuk pointed out.
• 2:58
  Speaker 0: design teams are shifting from just optimizing their buildings in isolation
• 3:03
  Speaker 1: To what exactly
• 3:04
  Speaker 0: to optimizing power, compute and economics together as one coupled system.
• 3:08
  Speaker 0: So the new metric that's emerging is tokens per
• 3:11
  Speaker 0: megawatt.".
• 3:11
  Speaker 1: Oh I like that.
• 3:13
  Speaker 1: so it's basically fuel-to-token efficiency?
• 3:16
  Speaker 0: Yes exactly.
• 3:18
  Speaker 1: And we actually have real world proof that these new designs are hitting production right now.
• 3:22
  Speaker 1: Like Geaseley KR posted about the SP-Zero One campus over in Granada Spain
• 3:28
  Speaker 0: Right The closed loop one.
• 3:29
  Speaker 1: Yeah It uses a close loop design with this tiny WUE of point oh one three liters per kilowatt hour which
• 3:37
  Speaker 0: is just incredibly efficient.
• 3:39
  Speaker 1: Which the cool thing, you don't even need a massive facility to do this.
• 3:43
  Speaker 1: Srinivas Varadarajan shared his hyper local approach of rural AI micro data center in Karnataka India.
• 3:50
  Speaker 1: Okay It runs completely air-cooled zero cooling water and it still maintains a PUE below one point Three.
• 3:56
  Speaker 0: so it completely proves that AI infrastructure can be distributed and sovereign Even hot climates absolutely.
• 4:03
  Speaker 0: And then on the other end of the spectrum you have Alistair Alexander showcasing the CSCS Alps Supercomputer in Switzerland.
• 4:09
  Speaker 1: Right, they use The Lake right?
• 4:11
  Speaker 0: Yeah!
• 4:11
  Speaker 0: They draw up to seven hundred fifty liters a minute from Lake Lugano run on local Hydro and their using it train the open source Aperidus model.
• 4:19
  Speaker 1: That is wild.
• 4:20
  Speaker 1: But if the physical data center is the engine... ...the AI models?
• 4:23
  Speaker 1: are the software dictating how hard that engine revs?
• 4:26
  Speaker 0: Definitely
• 4:27
  Speaker 1: You just can't optimize fuel-to-token efficiency if you don't know how much energy different models actually consume.
• 4:33
  Speaker 0: And that has been the problem for a long time, The industries finally like cracking open in black box of these proprietary AI models to measure their actual energy and carbon costs.
• 4:43
  Speaker 1: Right!
• 4:44
  Speaker 1: A big part was the Sustainable AI Group releasing CLEAR which stands for Closed Model Latent Energy Estimation Range.
• 4:52
  Speaker 1: Andrew Winston, along with folks like Boris Gomez-Ikoff and Heather Clancy were discussing the resulting data.
• 4:57
  Speaker 1: And it is staggering!
• 4:59
  Speaker 0: It really does
• 5:00
  Speaker 1: Wait.
• 5:00
  Speaker 1: if a typical agentic session uses about twenty seven times the energy of normal chat sessions are we basically leaving lights on in an entire skyscraper just to read single book?
• 5:11
  Speaker 1: Like do you need an autonomous agent for everything?
• 5:14
  Speaker 0: What's fascinating here?
• 5:15
  Speaker 0: that this data totally validates your point.
• 5:17
  Speaker 0: The variance is extreme.
• 5:19
  Speaker 1: How extreme.
• 5:20
  Speaker 0: Well, the data shows that least efficient models use over thirty times more energy than most efficient ones for the exact same task.
• 5:26
  Speaker 0: Oh wow!
• 5:26
  Speaker 0: Yeah and if you combine a heavy model with one of those agentic tasks constantly loops in the background The energy used can literally multiply thousandfold
• 5:35
  Speaker 1: A THOUSANDFOLD?
• 5:36
  Speaker 1: That is just insane.
• 5:37
  Speaker 1: By the way If want to make sure catch our future deep dives on how these tech trends evolve Make sure hit subscribe so they drop right into your feed
• 5:45
  Speaker 0: Because things are moving fast.
• 5:47
  Speaker 0: Companies already acting this clear data Chelsea M and Sam Brundrich shared that Etsy is actually the first to adopt the clear methodology.
• 5:55
  Speaker 0: Oh nice,
• 5:55
  Speaker 1: good for them.
• 5:57
  Speaker 0: Yeah they set a target to hold their cloud emissions intensity completely flat through twenty-twenty seven
• 6:02
  Speaker 1: which Is going to be tough if there's scaling AI
• 6:05
  Speaker 0: exactly.
• 6:06
  Speaker 0: but
• 6:06
  Speaker 1: There are architectural ways to fix this too.
• 6:09
  Speaker 1: Like, Aporva S shared an example of structural efficiency with Wiley's RE tool.
• 6:13
  Speaker 0: Oh!
• 6:14
  Speaker 0: The reading tool right?
• 6:15
  Speaker 0: Yeah
• 6:15
  Speaker 1: it uses eighty-eight percent less generative AI energy than a standard LLM build.
• 6:20
  Speaker 1: That
• 6:20
  Speaker 0: is huge drop.
• 6:21
  Speaker 0: how do they do that?
• 6:22
  Speaker 1: Because It reads the full text once extracts all claims and then just reuses them instead of, you know forcing the LLM to reread entire literature database for every single user prompt.
• 6:34
  Speaker 0: That is so smart and it proves that even basic algorithmic choices matter.
• 6:39
  Speaker 0: Dr.
• 6:39
  Speaker 0: Naseeb Kochman shared a study showing that quicksort uses less energy than merge sort on data sets with lots duplicates
• 6:46
  Speaker 1: Just from The Math
• 6:47
  Speaker 0: Just from the math and how it interacts with memory.
• 6:50
  Speaker 0: The crazy part was Strassen matrix multiplication.
• 6:53
  Speaker 0: It's theoretically more efficient, but it surprisingly uses more energy than conventional methods.
• 6:58
  Speaker 1: Wait really why?
• 6:59
  Speaker 0: Because of the recursive memory overhead... ...the constant state switching physically burns more watts even if the math takes fewer steps.
• 7:07
  Speaker 1: That just blows my mind
• 7:09
  Speaker 0: Right.
• 7:09
  Speaker 1: Well now that we can actually measure the massive footprint Of infrastructure in software We are seeing some real world friction
• 7:16
  Speaker 0: Big time.
• 7:17
  Speaker 1: The honeymoon phase for AI data centers is over.
• 7:20
  Speaker 1: Host communities and regulators are aggressively negotiating terms now.
• 7:24
  Speaker 0: Yeah, they're not just rolling over anymore.
• 7:26
  Speaker 0: Joe Spiker shared a perfect example of this.
• 7:29
  Speaker 0: Singapore has instituted Southeast Asia's first binding rules For Data Centers.
• 7:34
  Speaker 1: Binding Rules
• 7:35
  Speaker 0: Exactly
• 7:35
  Speaker 1: So Not Suggestions Exactly
• 7:37
  Speaker 0: Okay.
• 7:38
  Speaker 0: They require new data centers to hit as strict PUE Of one point two five or better Just To Get A Permit.
• 7:45
  Speaker 1: Here's where it gets really interesting, though.
• 7:47
  Speaker 1: It sounds like communities are treating these digital factories like actual heavy industry now...
• 7:52
  Speaker 0: They absolutely are!
• 7:54
  Speaker 1: ...they aren't just taking the tax revenue and smiling anymore they want to deed to the town.
• 7:58
  Speaker 1: well
• 7:58
  Speaker 0: Exactly Beth Simone Novick cited some great cases of this.
• 8:03
  Speaker 0: in West Virginia a committee directed four point five million dollars of community energy funds.
• 8:08
  Speaker 0: straight-to-local workforce an emergency prep.
• 8:10
  Speaker 0: Wow.
• 8:11
  Speaker 0: And in Colorado City, Texas data center builders actually had to fund repairs.
• 8:15
  Speaker 0: To a failing local water system.
• 8:17
  Speaker 1: so it's basically the cost of entry down
• 8:19
  Speaker 0: exactly plus.
• 8:20
  Speaker 0: Massachusetts and Pennsylvania now require local approvals.
• 8:23
  Speaker 0: California mandates Water use disclosures.
• 8:26
  Speaker 0: Spain is requiring new centers to run on eighty percent renewable energy.
• 8:30
  Speaker 1: The scrutiny just getting tighter.
• 8:31
  Speaker 1: everywhere We're also seeing this massive increase in scrutiny on corporate reporting.
• 8:36
  Speaker 1: you know, just to be impartial here and look at what the sources are saying.
• 8:40
  Speaker 1: Mark Butcher analyzed AWS's thirty-four page Australian Senate submission And he noted that it contained absolutely no absolute figures for energy use or grid capacity.
• 8:51
  Speaker 0: Just ratios probably
• 8:52
  Speaker 1: Yeah mostly.
• 8:53
  Speaker 1: In addition AWS is facing unproven allegations in a Virginia court right now from a former sustainability manager.
• 9:00
  Speaker 1: The claim Is That The Company Withdraws Water All Year Despite Claiming Zero Usage.
• 9:05
  Speaker 1: Ninety Seven Percent Of The Year.
• 9:07
  Speaker 0: Right.
• 9:07
  Speaker 0: And it's not just AWS, there is similar scrutiny on hardware providers like NVIDIA.
• 9:12
  Speaker 0: Akitan Joshi cited some criticism around NVIDia's partnership with the EarthShot Prize.
• 9:17
  Speaker 1: Okay what was that push back then?
• 9:19
  Speaker 0: Well, an analysis estimated that NVIDIA's Scope III product use emissions so the power of their chips draw.
• 9:25
  Speaker 0: when customers used them could hit twenty-one megatons of CO₂ equivalent for a single year.
• 9:30
  Speaker 1: Wait!
• 9:31
  Speaker 1: Twenty one megatones?
• 9:32
  Speaker 0: Yeah which actually exceeds the total eighteen megatonds in emissions avoided by the EarthShot prize to date.
• 9:39
  Speaker 0: and they noted supply chain emissions are up seven hundred twenty five percent since twenty twenty.
• 9:44
  Speaker 0: Just
• 9:44
  Speaker 1: wild numbers.
• 9:46
  Speaker 1: And again, corporate buyers are looking at these numbers and pushing back.
• 9:50
  Speaker 1: Richard Tarvitan highlighted a WRI roundtable that the Climate Week NYC... Oh
• 9:54
  Speaker 0: yeah!
• 9:54
  Speaker 0: The buyer consortium Yeah
• 9:57
  Speaker 1: Major buyers basically agreed to leverage their joint buying power To demand product level AI footprint data.
• 10:03
  Speaker 1: They want know exactly what they're buying into
• 10:05
  Speaker 0: Which is forcing whole new operational mindset.
• 10:08
  Speaker 0: It's driving rise of green ops which basically bridges finance engineering in sustainability.
• 10:14
  Speaker 0: So
• 10:14
  Speaker 1: it was like FinOps, but for carbon.
• 10:16
  Speaker 0: Exactly!
• 10:17
  Speaker 0: Georgina M., Naveen Balani and Julia Neosh who actually highlighted the Spruce open source project.
• 10:23
  Speaker 0: they pointed out that the aim is shifting entirely to waste prevention
• 10:27
  Speaker 1: instead of just paying for the waste later right?
• 10:29
  Speaker 0: And simple defaults work wonders here.
• 10:31
  Speaker 0: one organization Just made a smaller efficient model The Default For Their Dev's.
• 10:35
  Speaker 0: if you wanted the larger One You had To actively justify
• 10:39
  Speaker 1: It.
• 10:39
  Speaker 1: Oh That Is Such A Smart Behavioral Trick.
• 10:41
  Speaker 0: It Works.
• 10:42
  Speaker 0: It cut their AI costs by sixty percent and they're emissions by eighty percent.
• 10:45
  Speaker 1: Wow Okay, I want to leave you our listener with a final provocative thought tomorrow over this builds on report from James Martin And the association green IT.
• 10:56
  Speaker 0: The stat is sobering.
• 10:57
  Speaker 1: it really is.
• 10:58
  Speaker 1: AI accounted for twenty two percent of the European sustainable budget last year.
• 11:03
  Speaker 0: right
• 11:04
  Speaker 1: but it Is projected to hit one hundred and fifty six percent?
• 11:09
  Speaker 1: By twenty thirty That is seven times the sustainable level.
• 11:14
  Speaker 1: So, the ultimate limit to AI might not be silicon or even money.
• 11:19
  Speaker 1: it might simply be the unforgiving laws of thermodynamics and consent for communities hosting the grid.
• 11:25
  Speaker 0: It's a physical ceiling we're going hit very soon.
• 11:28
  Speaker 0: Thanks for joining me on this one.
• 11:30
  Speaker 1: If you enjoy this episode New episodes drop every two weeks.
• 11:33
  Speaker 1: Also check out our other editions On Cloud Insights & Sovereignty digital products and services, AI and agentic systems health tech ICT in Tech Insights defense tech and HealthTech.