Best of LinkedIn: Green ICT & Sustainable AI CW 36/ 37

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about green ICT in sustainable AI from CW-ThirtysixandThirtySeven.

00:00:10: Frenness supports ICT enterprises in the form of delivering precise ICT market and pricing intelligence that analyst subscriptions and existing databases cannot provide.

00:00:19: you can find more info

00:00:22: Right, so we are jumping into another deep dive today.

00:00:25: Yeah and today We're unpacking the top trends in green ICT and sustainable AI And we're sourcing this straight from the global professional community

00:00:33: Exactly!

00:00:34: The overarching story that were seeing across all these posts is just...I mean it's wild.

00:00:38: AIs demand for resources has completely shifted Like its no longer some niche sustainability concern For a side department.

00:00:46: It is literally the ultimate bottleneck right now.

00:00:48: It's dictating Global Infrastructure Software Design & Digital Economics.

00:00:51: Yeah, and if you are managing an IT budget right now or if your designing infrastructure.

00:00:56: You are absolutely feeling this.

00:00:58: so let's start at the macro level.

00:00:59: yeah because you really can't understand AI impact without understanding The Foundation Right?

00:01:05: Electricity!

00:01:07: I want to open with this staggering statistic shared by Patrick Elfman.

00:01:11: He noted that PWC estimates global data center capital expenditure through twenty fifty is going to hit thirty one point six trillion dollars.

00:01:19: Trillion?

00:01:20: Trillion, with a T!

00:01:22: Wow

00:01:23: I mean just put it in perspective That's more money than at cost of build the railways and entire physical infrastructure or internet

00:01:30: combined.

00:01:31: It's completely staggering number but really makes you wonder what actually speed limit here.

00:01:37: are microchips real bottleneck?

00:01:39: See, you'd think so but no.

00:01:41: The chips are ultimately just... Well they're a supply chain issue and money can eventually solve that.

00:01:46: Right!

00:01:47: You could print more silicon if you throw enough cash

00:01:49: out.

00:01:49: Exactly the true constraint is access to power grid.

00:01:55: Oh for

00:01:55: sure!

00:01:56: Transmission capacity takes years sometimes decades To actually permit and build.

00:02:02: Yeah And there was really great insight shared by Obinna Icedenso about this.

00:02:07: He brought up Microsoft's twenty-year, eight hundred and thirty five megawatt deal for the three mile island nuclear plant.

00:02:14: Wait they're basically restarting Three Mile Island?

00:02:16: Yeah literally!

00:02:18: And Obena makes this brilliant point.

00:02:19: that isn't just simple procurement anymore.

00:02:22: it's Lenders are looking at these corporate tech signatures and treating them like bonds.

00:02:28: Ah, okay.

00:02:30: So the lenders financing this massive energy projects they don't even care about The physical power plants risk.

00:02:35: nope not really.

00:02:36: They just care about the ironclad Tech contract backing it.

00:02:39: we're seeing the exact same thing with Nvidia right?

00:02:41: They've got thirty six billion dollars in take or pay commitments.

00:02:44: Wow so the tech giants were basically underwriting the future of the grid

00:02:47: there.

00:02:48: but you know, that drags us straight into some really messy geopolitics.

00:02:52: Yeah and all these questions about infrastructure ownership.

00:02:54: like Alexander J. Wierser brought up this fascinating case where Google is transforming a former West Virginia coma I saw

00:03:02: that they're outfitting it with massive solar arrays in batteries.

00:03:05: yeah

00:03:05: but Alexander raises a critical point about who actually owns the intellectual property for how those green tech systems integrate into the legacy grid.

00:03:13: Oh!

00:03:14: That's the trillion dollar question right there.

00:03:16: Because whoever figures out the software and hardware integration to stabilize those massive new loads, I mean they effectively control future scalability of technology worldwide.

00:03:26: Exactly!

00:03:28: And this strain on the grid is a universal problem.

00:03:31: Steve Salanke pointed out that Australia Energy Market Operator AE AMO.

00:03:35: they're forecasting data center electricity demand to jump from five terawatt hours today To thirty four terawat hours by twenty.

00:03:42: thirty six wait

00:03:43: thirty-four.

00:03:43: That's a massive leap

00:03:45: right.

00:03:45: But then you have Julian Marzetti who rightly points out that some regions Have a huge head start like in Britain.

00:03:52: solar and wind already cover most of their annual electricity generation.

00:03:57: But see,

00:03:57: that varying grid composition it creates this massive global inequality in the actual emissions of a software you build.

00:04:05: How

00:04:05: so?

00:04:05: Well, Didier Abyssin highlighted this.

00:04:08: he showed how the carbon footprint of an AI query swings wildly just depending on physical zip code of the server.

00:04:14: Oh

00:04:14: right because they have local grid is powered

00:04:16: Exactly!

00:04:17: So running a query on Francero's Grid which has heavily nuclear

00:04:21: costs

00:04:22: up to thirty-two grams CO² per kilowatt

00:04:24: hour Okay.

00:04:25: But if you run that exact same code in Virginia, which is a massive data center hub.

00:04:30: It's four hundred and eighteen grams.

00:04:32: wait That's more than tenfold difference just based on where the compute happens to land.

00:04:35: yeah.

00:04:36: And that geographic disparity translates directly into economic disparity.

00:04:41: Daniel Jeffrey Koch summarized a UNDP report with this really stark warning.

00:04:45: What did it say?

00:04:46: It warned that developing nations are currently on track to capture less than ten percent of AI is projected, fifteen point seven trillion dollar economic boom because you know competitiveness now completely tethered having cheap reliable low carbon power.

00:05:02: Though Angel Melguiso pushes back on that narrative a bit, he argues that Latin America could actually capture a one trillion dollar opportunity.

00:05:10: He points to Brazil's fiscal push under Lula where they're tying digital infrastructure directly the skills development and sustainability

00:05:18: And tech giants are clearly seeing the same thing.

00:05:20: Ruchali Gad highlighted how Google is actually combining AI deployment with a hundred and fifty megawatts solar project in India.

00:05:27: Oh,

00:05:28: the one with the climate smart agriculture?

00:05:29: Yeah

00:05:30: they're helping like seventy thousand farmers reduce methane water use all in exchange for building out their tech infrastructure which

00:05:37: totally explains this massive shift that Chris Wedding and Jeremy Lunar observing in venture capital.

00:05:42: right now Climate Tech has just suddenly flushed with cash again.

00:05:46: mass mutuals just launched A huge fund.

00:05:49: But it's not altruistic, right?

00:05:50: It's just a pure sustainability thesis?

00:05:53: No...not at all!

00:05:54: It is the survival mechanism.

00:05:55: AI desperately needs power And Boris Kamazechikov warned that if we throw traditional fossil fuels of this command Ai-driven gas plants could spike US power emissions by twenty percent.

00:06:07: Yikes Which forces us to look past just the macro grid and actually step inside buildings themselves.

00:06:13: Yeah let talk about physical data centers

00:06:16: Because if electricity is the fuel for this boom, heat is inevitable exhaust.

00:06:21: We're transitioning from power grid into managing that heat.

00:06:24: Right!

00:06:24: And Managing that heat isn't just an airflow problem anymore it's basically death of air cooling.

00:06:29: For decades you blew cold air down a server aisle and were fine.

00:06:34: Not any more.

00:06:35: Russ Dailey and Muhammad Quater noted next gen AI racks are pushing past one hundred kilowatts per rack.

00:06:41: That's insane.

00:06:42: Moving air physically cannot carry that much thermal density away.

00:06:45: Right, the air molecules are just too spread out.

00:06:47: so facilities are aggressively shifting to liquid cooling.

00:06:50: it's dropping the industry average power usage effectiveness or PUE.

00:06:55: Oh yeah, PUE

00:06:56: It's dropping from about one point five six down around One Point One Five.

00:07:01: and raw energy savings there is massive like thirty-five gigawatt hours a year for ten megawatt load.

00:07:07: But you know I have jump in here because Recapusala makes really compelling argument that PUE is actually becoming obsolete.

00:07:15: Wait, obsolete?

00:07:16: But it's the gold standard for facility efficiency.

00:07:19: Well she argues that only measures the facility not actual computing efficiency.

00:07:24: You could have a pristine PUE of one point one five meaning your pumps and chillers are super efficient.

00:07:30: but if GPUs inside just sitting idle due to thermal bottlenecks you're still wasting massive energy

00:07:35: I see.

00:07:36: so we need transition tracking compute per watt instead.

00:07:39: Exactly And blindly relying on PUE creates these huge regulatory blind spots.

00:07:43: Myco actually modeled Spain's AZA proposed grid rule, which would mandate a PUE at one point one five to maintain the grid connection.

00:07:50: Let me guess it completely penalizes older infrastructure.

00:07:53: Nailed It.

00:07:54: His model show works perfectly for optimized AI facilities But fails conventional data centers that still need to run standard air-cooled racks For non-AI stuff.

00:08:02: Right!

00:08:03: You can't just regulate all of this diverse infrastructure with one blunt metric?

00:08:07: No you really cant.

00:08:08: and hey While we're on the topic of cooling, We have to talk about the whole thirsty AI myth.

00:08:13: Oh!

00:08:14: The headlines claiming generative AI is drinking all municipal water supplies?

00:08:18: Yes

00:08:18: it's a bit more nuanced than that A Denilson Belisario points out.

00:08:22: in Brazil leading collocation centers are achieving near zero or zero point.

00:08:27: zero three Water usage effectiveness.

00:08:30: Wait how do they cool those loads without evaporating the water?

00:08:33: Closed loop systems.

00:08:35: The water absorbs heat runs through a heat exchanger cools down and circulates right back, doesn't evaporate away.

00:08:41: Belizario even compared it to the local pulp in paper industry which uses up to sixty thousand liters of water just to produce a single tonne of pulp.

00:08:49: Wow So you really can decouple compute from water depletion if you engineer it right?

00:08:54: Totally.

00:08:54: And we're seeing with some radical designs.

00:08:56: taking this even further, Ankar Badra and Carolyn Pistone highlighted using the ocean as the ultimate heat sink.

00:09:02: Oh

00:09:02: like Samsung's floating data centers.

00:09:04: Exactly!

00:09:05: Norway just activated this wave-powered seawater cooled offshore platform.

00:09:12: It is completely silent zero mechanical fans.

00:09:15: That's incredible.

00:09:16: And even the physical materials for these buildings are changing.

00:09:20: Carl Rave suggests using engineered timber as a structural alternative to concrete, to lower embodied carbon and Dietmar Barrow emphasizes the need to transition to low GWP refrigerants.

00:09:32: Low global warming potential refrigerant?

00:09:34: Right!

00:09:34: And Jerome Tordiot notes that locking in these electrical and architectural decisions early on is exactly how companies hit those hyperscaler efficiency levels.

00:09:42: But life cycle doesn't end there right.

00:09:44: Hardware circularity is huge now.

00:09:46: Mervyn Foo is championing this idea of a circular compute passport to track embodied carbon from mining all the way to

00:09:52: recycling.".

00:09:53: And Didier Apples showcased how proper ITAD processing, you know refurbishing assets can actually extend a device's life by up to six years.

00:10:00: But

00:10:00: proving all those to market IS grueling!

00:10:03: Haley Willis Chittick talked about massive backstage work required just to publish a corporate sustainability report showing real water-free cooling games...

00:10:11: Because it comes down to operations ultimately….

00:10:13: Brett Ridley give a great warning about this.

00:10:16: Having neatly color-coded emergency binders doesn't mean your data center team actually knows how to swing a mechanical chiller under pressure.

00:10:24: Right, real operational maturity versus just checking boxes Exactly.

00:10:28: By the way if you are enjoying this deep dive make sure hit subscribe so that we catch all our future additions.

00:10:33: We love bringing these stuff.

00:10:35: So we've optimized the macro power grid, We overhauled physical buildings.

00:10:40: What about software layer?

00:10:41: Yeah what about actual code running on machines?

00:10:44: This is next frontier.

00:10:45: right trimming digital fat But measuring softwares carbon footprint is shockingly hard.

00:10:52: Oh yeah Lilabeth Bustos-Linaris built SOMA AI And she found out that just misclassifying an AI model tier can throw your carbon estimates off by four to five times

00:11:03: And regional changes can swing emissions by twelve X, right?

00:11:05: Like we talked about earlier.

00:11:06: Exactly!

00:11:07: Which is why open source solutions are stepping up.

00:11:10: Sven Bang, Anita Schutler and Wilco Berggrath introduced Nolanda's token dashboard.

00:11:16: What does that do?

00:11:16: It shifts away from flat rate billing By tracking API costs in CO-II privately using UIDs.

00:11:23: So it tracks the anonymous identifier not individual developer to find systemic waste.

00:11:27: Oh thats

00:11:28: smart.

00:11:28: Nathan Bailey built something similar The Carbon Tracking MCP.

00:11:32: It estimates CO-II directly from local coding logs without even making network calls.

00:11:36: Totally private and once you measure the waste, You can route around it.

00:11:39: Ben Hillier launched the TRET workbench which automatically routes tasks to most carbon efficient model

00:11:45: Because using a frontier model for basic task is just overkill.

00:11:49: They use ten to a hundred times more power.

00:11:51: It's like chartering a seven forty-seven to go the grocery store.

00:11:55: Edenmere, Muhammadi highlighted tool called Ponytail that trims over engineered AI agent code before tokens even accumulate

00:12:02: An empty Citicorp.

00:12:03: RHman is discussing the ICS EFS machine learning model to detect Java code smells to support greener software.

00:12:10: Yeah, this whole frugal AI movement it's so real.

00:12:13: The laundry chart hurry showcased compact AI running on a Raspberry Pi.

00:12:17: It cut energy from two point one-one watt hours on a cloud GPU down to point zero one nine watt hours.

00:12:23: Wow.

00:12:24: And Robert Keyes noted that green PT integrated deep-seq v. four point one flash.

00:12:29: It's a five hundred and fifty two billion parameter model But it achieves high performance by only activating a fraction of those parameters

00:12:36: Which first shot.

00:12:37: Iran he says is going to be crucial for web sustainability.

00:12:40: Going forward,

00:12:41: but you know as we build all these tools Allison Freeman issued a stark warning.

00:12:45: She shared an AI impact calculator but said that sustainability is a systemic issue.

00:12:50: It's not individual user burden.

00:12:52: Right,

00:12:52: you can't just blame the person typing their prompt.

00:12:54: Exactly Because Chinmai Anand summarized The Jevenz Paradox here perfectly.

00:12:59: Google's Gemini text prompt uses Just .- two four watt hours.

00:13:03: Super efficient

00:13:04: Yes sounds great

00:13:05: But IEA data shows Total Data Center Electricity Surging seventeen percent.

00:13:09: The efficiency Is completely outpaced by sheer demand

00:13:12: Which brings up Drishtramali's ultimate question Is this AI deployment genuinely net positive for the climate?

00:13:18: That's a million dollar question, but honestly none of these matters.

00:13:21: if accounting frameworks are flawed.

00:13:23: Oh that is wild!

00:13:24: Let us transition to the Accounting and Regulation side because Mark Butcher exposed massive flaw in spend-based carbon reporting.

00:13:31: So it was like getting a speeding ticket just as Kelly Blue Book value of your car went up.

00:13:36: A company bought identical servers two years in a row…but because global memory prices surged The spend-based reporting recorded a fake, four hundred and seven percent emissions increase.

00:13:46: Oh my God!

00:13:47: Did they get fined?

00:13:48: Yes it cost them over two hundred sixty six thousand pounds in UK ETS charges for emissions that literally never existed.

00:13:56: That completely undermines everything.

00:13:58: Miguel Fernandez touched on this too noting that financed emissions depend on really grueling unglamorous underlying data like concrete and steel records.

00:14:07: Right

00:14:07: if that date is wrong the AI analysis is useless.

00:14:11: Manuel Levier-Paul highlighted that ADVASE and EUScore.Tech are partnering to publish independent environmental scores for IT providers,

00:14:18: which we desperately need because Christophe Cluzot noted an ITU report showing that despite all these climate commitments digital companies' emissions are still rising.

00:14:27: so to navigate this Laxitivora outlined four core ESG learning areas carbon markets regulation governance and data And Arpit Sharma curated a top-twelve list of reports on AIR's role in ESG.

00:14:43: Because regulation is moving incredibly fast now, Cora Braga explained Brazil's redata law... which ties tax incentives directly to renewable energy and water commitments.

00:14:53: And

00:14:53: Kari Hayden-Pendley pointed out that AI governance is moving from proposal to actual enforcement globally.

00:14:59: Yeah, Cal Ennis noted that META actually left the RE- one hundred initiative because of their gas investments for AI.

00:15:04: He just couldn't meet the Renewable Threshold anymore

00:15:06: Exactly!

00:15:07: In Pascal Jolie questioned what's gonna happen to hugging faces sustainability leaderboards now that NVIDIA acquired them.

00:15:12: Will it stay transparent?

00:15:14: Who knows.

00:15:15: But ultimately you need human alignment.

00:15:17: Gakia Chaban highlighted NTT data's carbon literacy training for global IT staff.

00:15:21: Because real change requires the whole team, John Berter's look review emphasized psychological safety in engineering teams.

00:15:28: you have to be able to admit failures when building these complex systems.

00:15:31: Yeah and Karen Vanerzane has shared her new role at TNO Building Data Ecosystems To actually turn this research into commercial innovation.

00:15:39: Odoso-Odaro is pushing for this same systems based approach to AI governance.

00:15:44: It's amazing to see the community mobilizing, Nolan Godard and Nina Benoit were discussing

00:15:49: A.I.'s

00:15:49: resource strain at NY Climate Week

00:15:51: And in a Green IO London event, Navin Balani Kevin Leslie & James Hall focused on Agentec AI and greenups.

00:15:58: Dave Masalink and Joanna Massraff were driving green software foundation events around local versus frontier models.

00:16:05: Over in Senegal, Ada Kau Eoyol is promoting green coding.

00:16:09: Harold McBean is out there publishing KSA MEP checklists for Saudi data centers.

00:16:14: Bernhard Heinrich was at the Global Decarbonization Forum

00:16:17: David Edwards is backing a Green Compute Edge challenge in Western Australia And Dr.

00:16:21: Sasha Lutioni spoke with The Allian conference about measuring AI's true footprint.

00:16:26: just incredible momentum and you know it leaves us with this final thought.

00:16:30: We are entering an era where computing power is no longer constrained by the speed of processors, right?

00:16:35: It's constrained by physical limits on earths electrical grid or water tables.

00:16:40: if code eat world has to learn how to digest efficiently so will next trillion dollar company be one that builds biggest model on a fraction

00:17:04: of what.