Latest / Best of LinkedIn: Digital Transformation & Tech / Best of LinkedIn: Green ICT & Sustainable AI CW 34/ 35

Transcript

Best of LinkedIn: Green ICT & Sustainable AI CW 34/ 35 · Best of LinkedIn: Digital Transformation & Tech

• 0:00
  Speaker 0: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about green ICT and sustainable AI from CW- ThirtyFour and ThirtyFive.
• 0:10
  Speaker 0: Frenness supports ICT enterprises in the form of delivering precise ICT market and pricing intelligence that analyst subscriptions and existing databases cannot provide.
• 0:20
  Speaker 0: you can find more info
• 0:24
  Speaker 1: in green ICT and sustainable AI trends we've seen circulating across LinkedIn.
• 0:31
  Speaker 0: Yeah, if you are a digital transformation or tech professional listening to this right now... You're living in the constant state of contradiction?
• 0:39
  Speaker 1: Oh absolutely!
• 0:40
  Speaker 0: On one hand your board is just demanding that scale AI capabilities instantly like yesterday but on other hand they expect you hit these increasingly strict corporate sustainability targets And there's literally multi-billion dollar data center projects being cancelled Not because the tech is failing, but because local town councils are just refusing to surrender their water and power into invisible AI algorithms.
• 1:06
  Speaker 1: It's really defining bottleneck of a digital age Because cloud isn't weightless right?
• 1:13
  Speaker 1: It was built on concrete copper in cooling water.
• 1:16
  Speaker 1: So today our mission is synthesize real ground level strategies Tech professionals use to balance massive resource demands with actual planetary limits.
• 1:27
  Speaker 1: Exactly!
• 1:28
  Speaker 0: We are cutting through all the fluff today.
• 1:30
  Speaker 0: Yeah,
• 1:31
  Speaker 1: we're looking at everything from uh community zoning battles and liquid-cooled hardware All the way down to the physical mechanics of why faster software code doesn't actually mean greener code.
• 1:42
  Speaker 0: Okay let's unpack this because if you look at the macro level before any tech is even deployed There's this fundamental clash between AI's physical footprint in local communities.
• 1:51
  Speaker 0: Take the power grids in Eurek right now.
• 1:53
  Speaker 1: Oh yeah, it's a mess
• 1:54
  Speaker 0: It is.
• 1:55
  Speaker 0: There was this massive panic around hyperscale capacitor and governments are taking some really drastic measures.
• 2:01
  Speaker 0: like Arwal Owen posted this great breakdown showing that Denmark actually published new emergency grid law And completely upends tech hierarchy They essentially rewrote their rules to put data centers at absolute back of line for electricity access
• 2:16
  Speaker 1: Which is wild.
• 2:18
  Speaker 1: But that legislative move in Denmark it really forces us to look at how we even arrived At this state of hyperscale panic and the first place.
• 2:25
  Speaker 0: Yeah, How did we get here?
• 2:27
  Speaker 1: Well a significant part of The grid crisis is actually driven by Just flawed forecasting at the macroeconomic level.
• 2:35
  Speaker 1: Gothier Russell Traced the European Union's Really aggressive goal To triple its data center capacity By twenty-twenty seven.
• 2:42
  Speaker 1: he traced It all the way back to its origin
• 2:44
  Speaker 0: And what does you find?
• 2:46
  Speaker 1: he found that the target was largely driven by a commercial real estate report.
• 2:49
  Speaker 0: A Real Estate Report?
• 2:50
  Speaker 1: Yeah, and it made this fundamental just glaring logical error.
• 2:54
  Speaker 1: It wrongly correlated general bandwidth growth directly with AI compute demand.
• 2:58
  Speaker 0: Oh I see so they basically assumed because data transfers going up you know more people streaming for K video or downloading larger files That we automatically need exponentially More AI processing power to handle
• 3:10
  Speaker 1: exactly which is complete non sequitur in computer science.
• 3:14
  Speaker 1: Moving data across a network requires bandwidth and basic routing infrastructure, but training the large language model that requires incredibly dense high-heat processing power.
• 3:25
  Speaker 0: They are entirely different.
• 3:26
  Speaker 0: physical demands
• 3:28
  Speaker 1: One hundred percent.
• 3:28
  Speaker 1: so basing multi-billion euro infrastructure targets on that flawed correlation.
• 3:34
  Speaker 1: I mean it leads to building massive potentially unnecessary capacity That inevitably collides with local realities.
• 3:41
  Speaker 1: and
• 3:41
  Speaker 0: those local realities are pushing back hard.
• 3:43
  Speaker 0: I mean Gregor Rood shared this massive gap in public opinion.
• 3:46
  Speaker 0: his survey data showed that seventy nine percent of Americans want the country To lead the world an AI development.
• 3:51
  Speaker 1: sure
• 3:52
  Speaker 0: but only fourteen percent want a data center built in their own community.
• 3:56
  Speaker 0: Wow,
• 3:56
  Speaker 1: Only fourteen percent?
• 3:57
  Speaker 1: Yeah
• 3:58
  Speaker 0: and that gap...that sixty-five point gap it led to five major North American data center projects being either withdrawn or flat out rejected by local councils in just a single week.
• 4:09
  Speaker 1: What's fascinating here is understanding the psychological root of that resistance.
• 4:13
  Speaker 1: yeah because Dr.
• 4:14
  Speaker 1: Dagon Lemiaki broke down why this happens in himbyism, you know the whole not-in-my-backyard thing.
• 4:23
  Speaker 0: Right!
• 4:23
  Speaker 0: The complaints about the humming noise of cooling fans or just the aesthetic blight of a giant warehouse?
• 4:28
  Speaker 1: Yeah it's not just about noise or land.
• 4:30
  Speaker 1: It is really about...a loss of human agency A
• 4:33
  Speaker 0: LOSS OF AGENCY.
• 4:34
  Speaker 1: Communities feel like these tech monoliths are extracting local resources without giving anything back.
• 4:40
  Speaker 0: Okay yeah I mean, when a mega data center moves into a rural area it operates kind of like a closed fortress.
• 4:47
  Speaker 0: It vacuums up the local water table for cooling strains municipal power grid but residents don't see any local economic loop.
• 4:55
  Speaker 1: Exactly and critical infrastructure strips away community sense ownership over its own backyard.
• 5:01
  Speaker 1: political resistance is just inevitable.
• 5:03
  Speaker 0: So if Community Trust is real bottleneck how do we fix contract between AI & public?
• 5:10
  Speaker 1: Well, Naveen Balani proposed a really great concept called an AI Sighting Ledger.
• 5:15
  Speaker 1: It's essentially an eight-question framework designed to answer hard questions before a single shovel hits the dirt — it's
• 5:22
  Speaker 0: like a prenuptial agreement for tech giants in local towns?
• 5:25
  Speaker 0: Pretty
• 5:25
  Speaker 1: much!
• 5:26
  Speaker 1: Instead of standard sustainability reports that tell the community what happened after resources were used this ledger demands answers up front.
• 5:35
  Speaker 1: Who bears the financial burden for municipal grid upgrades?
• 5:39
  Speaker 1: Or, if AI bubble bursts and center closes who pays environmental cleanup.
• 5:45
  Speaker 0: That makes so much sense.
• 5:46
  Speaker 0: it shifts power dynamic entirely!
• 5:47
  Speaker 1: It does.
• 5:48
  Speaker 1: And Evan Salman noted that Canada is already setting a national framework for responsible data-centered development that's explicitly centered on this kind of community input.
• 5:56
  Speaker 0: Okay So we can't just build endlessly.
• 5:58
  Speaker 0: We have to build smarter Which transitions us into the Data Center hardware itself.
• 6:04
  Speaker 0: Since space, power and community approval are so scarce operators have to rethink the physical design of the centers that do get approved.
• 6:11
  Speaker 0: Oh completely!
• 6:12
  Speaker 1: The physical architecture has to change
• 6:14
  Speaker 0: Right.
• 6:15
  Speaker 0: Allison Freeman highlighted a massive leap in this direction with Dell's new IR- seven thousand liquid cooled rack.
• 6:22
  Speaker 0: This single cabinet handles a massive of four hundred and eighty kilowatts.
• 6:26
  Speaker 0: Of power supporting one hundred and forty-four GPUs,
• 6:30
  Speaker 1: which is incredible density It is.
• 6:32
  Speaker 0: it cuts cooling energy by up to seventy four percent And requires seventy five percent.
• 6:37
  Speaker 0: fewer racks perform the exact same workload
• 6:40
  Speaker 1: and that's all because of liquid cooling Because water or specialized cooling fluid absorbs heat continuously and directly from the chip unlike just blowing chilled air around.
• 6:49
  Speaker 1: But you know we also need to rethink data storage.
• 6:52
  Speaker 0: Okay, how so?
• 6:53
  Speaker 1: Well, Curry Lentor and Christine Suber pointed out a surprising comeback in the industry.
• 6:58
  Speaker 1: IBM's climate-controlled diamondback tape storage... Wait!
• 7:01
  Speaker 1: Tape storage like
• 7:02
  Speaker 0: cassettes?!
• 7:02
  Speaker 0: Basically yeah.
• 7:03
  Speaker 0: That sounds like we are dragging nineteen eighties tech into a hypermodern AI facility.
• 7:07
  Speaker 0: How does moving backwards solve an energy crisis?
• 7:10
  Speaker 1: I know it sounds counterintuitive But modern spinning hard drives need constant electricity just to keep data accessible.
• 7:18
  Speaker 1: Magnetic tape however sits completely dormant on a shelf.
• 7:21
  Speaker 1: It draws zero volts until robotic arm physically grabs it.
• 7:25
  Speaker 1: Oh, yeah for unstructured data?
• 7:28
  Speaker 1: It cuts costs by eighty four percent and energy by ninety seven percent.
• 7:33
  Speaker 1: And they can run in standard data centers up to forty five degrees Celsius and ninety percent humidity.
• 7:38
  Speaker 0: So you don't even need to hyper cool the storage rooms exactly like treating modern data centers Like living ecosystems that need to adapt their environments.
• 7:48
  Speaker 0: Carl Ray brought up a fascinating application of this, talking about using engineered timber in data center construction instead of traditional concrete.
• 7:56
  Speaker 1: Right because concrete is one the largest carbon emitters.
• 7:58
  Speaker 1: Exactly!
• 7:59
  Speaker 0: Engineered Timber actually sequesters Carbon.
• 8:02
  Speaker 0: You turn shell of the data center into a carbon sink rather than a carbon emitter.
• 8:07
  Speaker 0: And by the way if you want to stay on top how these infrastructure shifts impact your tech stacks make sure subscribe.
• 8:13
  Speaker 0: so catch future editions
• 8:16
  Speaker 1: Definitely.
• 8:16
  Speaker 1: And you know, keeping that hardware running longer matters too.
• 8:20
  Speaker 1: Karen van der Zanden shared an analysis of thirty eight hundred government servers.
• 8:24
  Speaker 0: Okay
• 8:25
  Speaker 1: and it proved That a nine year old server can often perform just as well As the two-year-old one for standard workloads.
• 8:32
  Speaker 0: Wait!
• 8:32
  Speaker 0: A Nine Year
• 8:33
  Speaker 1: Old Server?
• 8:33
  Speaker 1: Yes There are massive energy and cost savings just waiting to be unlocked simply by ignoring standard accounting depreciation cycles.
• 8:42
  Speaker 0: So we're throwing out perfectly good silicone.
• 8:45
  Speaker 0: Just because a spreadsheet says it's old
• 8:47
  Speaker 1: pretty much.
• 8:48
  Speaker 0: But I mean you can optimize the hardware all-you want.
• 8:50
  Speaker 0: if the software is bloated You're still wasting energy right?
• 8:53
  Speaker 0: The fastest lever a dev team has to pull his software optimization.
• 8:57
  Speaker 0: No one good art shared a brilliant analogy from and nature commentary on this.
• 9:01
  Speaker 0: Most people don't need a Formula One car when a reliable Toyota Corolla will do.
• 9:06
  Speaker 1: that is so true.
• 9:07
  Speaker 1: Why use massive frontier models for simple tasks?
• 9:10
  Speaker 0: Exactly why burn all that energy?
• 9:12
  Speaker 0: just to summarize the PDF And James Martin provided the hard data on this frugal AI approach.
• 9:18
  Speaker 0: He looked at the Frugal AI switchboard in a model called deep-seek v three.
• 9:22
  Speaker 1: I saw this
• 9:23
  Speaker 0: Yeah.
• 9:24
  Speaker 0: DeepSeq V-III is fifty four times smaller, has four times less environmental impact and is nine hundred times cheaper than a frontier model.
• 9:32
  Speaker 0: And for data analysis it only loses about two percent accuracy.
• 9:35
  Speaker 0: You
• 9:36
  Speaker 1: drop two percent performance For a nine hundred time cost reduction.
• 9:39
  Speaker 1: that is staggering
• 9:41
  Speaker 0: Right!
• 9:42
  Speaker 0: Robert Key shared an example of this.
• 9:44
  Speaker 0: applied to coding He introduced an app called Honey, which forces AI coding agents to be concise.
• 9:50
  Speaker 1: Oh because coding agents always generate that massive token bloat?
• 9:54
  Speaker 0: Exactly!
• 9:55
  Speaker 0: They explain everything in five paragraphs.
• 9:58
  Speaker 0: The Honey App trims token output by thirty-nine percent and code by fifty three percent while maintaining a one hundred percent test pass rate.
• 10:06
  Speaker 1: That is brilliant But you know where and when you compute matters just as much as what you compute.
• 10:11
  Speaker 1: Pochensu who goes by Steven helped build a hackathon project called EcoRouter.
• 10:17
  Speaker 1: And Diego Espinoza launched an open-source workbench, called TRET.
• 10:21
  Speaker 0: They
• 10:21
  Speaker 1: both actively route AI tasks to the most carbon efficient models and grids available in real time.
• 10:27
  Speaker 0: Oh, wow.
• 10:28
  Speaker 0: So if the grid in one region relies on coal because it's nighttime It just instantly shifts The compute job across the world to a region where solar power
• 10:37
  Speaker 1: is peaking.
• 10:38
  Speaker 1: Exactly surge pricing for this sun and wind.
• 10:40
  Speaker 1: that's
• 10:40
  Speaker 0: incredible.
• 10:41
  Speaker 0: But let me ask you this doesn't making code run faster automatically make it greener?
• 10:45
  Speaker 0: like If my team optimizes A function to finish in ten milliseconds instead of fifty?
• 10:50
  Speaker 0: isn't That using less power?
• 10:52
  Speaker 1: Well This raises an important question about how we measure efficiency because Michael Hearnberger shared a research correcting that exact assumption.
• 10:59
  Speaker 1: Profiling shows code blocks accounting for just twenty percent of execution time can actually consume forty-percent of the energy.
• 11:07
  Speaker 0: Wait, really?
• 11:08
  Speaker 0: How does it
• 11:08
  Speaker 0: work?!
• 11:09
  Speaker 1: Because time doesn't equal thermal output.
• 11:12
  Speaker 1: Those rapid memory access patterns and state transitions cause massive spikes in voltage.
• 11:17
  Speaker 1: We need to start measuring joules not just milliseconds.
• 11:21
  Speaker 0: Here's where this gets REALLY interesting Because if we are optimizing software and hardware, We actually need to accurately measure what we're saving.
• 11:30
  Speaker 0: The financial scale of this industry is just staggering!
• 11:32
  Speaker 1: It really is.
• 11:33
  Speaker 1: Yeah.
• 11:34
  Speaker 1: Sumit Habu highlighted a Wall Street Journal analysis uncovering three trillion dollars in future AI infrastructure commitments.
• 11:43
  Speaker 1: Three trillion?
• 11:44
  Speaker 1: And companies like Meta are using joint ventures like their twenty-seven billion dollar Louisiana project to basically finance this off of the direct balance sheets.
• 11:53
  Speaker 0: Which makes tracking risk incredibly difficult.
• 11:55
  Speaker 0: Brock Lambert's actually warned about it, he pointed out that most green AI dashboards showing one hundred percent policy compliance is just participation
• 12:05
  Speaker 1: trophies.
• 12:06
  Speaker 0: They just mask real operational risks.
• 12:09
  Speaker 0: they track effort not exposure
• 12:11
  Speaker 1: and if we look at planetary cost its just as daunting.
• 12:14
  Speaker 1: Federico Bugliaro-Gogia shared a UN University report stating datacenters could use nine hundred and forty five terawatt hours of power, at nine point three trillion liters water by twenty thirty.
• 12:26
  Speaker 1: That
• 12:26
  Speaker 0: is hard to even visualize
• 12:27
  Speaker 1: right an inference taking eighty to ninety percent that energy plus.
• 12:31
  Speaker 1: Ludovic Subran from Allianz research notes the true data center carbon footprint as likely fifty seven percent higher than international energy agency estimates.
• 12:39
  Speaker 0: but context everything here.
• 12:42
  Speaker 0: Valdeer S brought in a really nuanced point via CBS report.
• 12:46
  Speaker 0: He noted that U.S.
• 12:47
  Speaker 0: data center water use is actually less than what's used to irrigate golf courses.
• 12:52
  Speaker 0: Oh,
• 12:52
  Speaker 1: really?
• 12:52
  Speaker 1: Golf courses use more!
• 12:54
  Speaker 1: Yes,
• 12:54
  Speaker 0: proving the true impact is highly dependent on local geography and power grid.
• 12:59
  Speaker 0: Right
• 12:59
  Speaker 1: – A facility in desert running on coal entirely different from one in a water-abundant region.
• 13:07
  Speaker 1: Boris Kamazaevichov made the point that we aren't even tracking the full life cycle yet.
• 13:11
  Speaker 1: What
• 13:11
  Speaker 0: are you missing?
• 13:12
  Speaker 0: The
• 13:12
  Speaker 1: safety testing phase, open AI's GPT-Five Point Six automated red teaming alone took three hundred and eight megawatt hours.
• 13:20
  Speaker 1: Whoa!
• 13:21
  Speaker 1: That is a quarter of the energy used to pre train GPT Three entirely governance just has.
• 13:26
  Speaker 1: look at the whole picture.
• 13:27
  Speaker 0: Yeah it absolutely does do.
• 13:29
  Speaker 0: so.
• 13:29
  Speaker 0: wrapping all this up the ultimate takeaway for your listening Is that scaling AI isn't just a software challenge anymore.
• 13:35
  Speaker 0: It was profound physical, communal and environmental challenge.
• 13:39
  Speaker 0: You cannot just brute force the processing.
• 13:41
  Speaker 1: No you can't.
• 13:42
  Speaker 1: And I think it leaves us with one really provocative thought for the future as AI models gain agentic autonomy The ability to endlessly query a loop on their own.
• 13:53
  Speaker 1: What happens when autonomous digital intelligence naturally demands more physical resources like water land and power than the planet's localized grids could even provide?
• 14:03
  Speaker 0: That's terrifying.
• 14:04
  Speaker 1: Well, we have to start hard coding physical planetary limits into the foundational logic of AI itself.
• 14:10
  Speaker 0: that is something We are all going to have to mull over.
• 14:12
  Speaker 0: if you enjoyed this episode new episodes drop every two weeks.
• 14:16
  Speaker 0: Also check out our other editions on cloud insights and sovereignty digital products And services ai in agentic systems health tech ICT and Tech Insights defense tech and HealthTech.