Best of LinkedIn: Sustainability & Green ICT CW 18/ 19

00:00:00: This episode is provided by Thomas Allgeier and Frennis, based on the most relevant LinkedIn posts about sustainability in green ICT in CW-E eighteen and nineteen.

00:00:09: Frennes supports ICT enterprises in the form of delivering precise ICT market and pricing intelligence that analysts subscriptions and existing databases cannot provide.

00:00:19: you can find more info in the description.

00:00:21: fine.

00:00:21: so imagine buying a diet soda specifically to justify eating like three extra large pizzas

00:00:28: stomach my weekend.

00:00:29: Yeah, well that is exactly the kind of math the tech industry's doing right now when it comes to AI and energy efficiency.

00:00:35: Today we're really tearing down the greenwashing.

00:00:38: We absolutely are!

00:00:39: We're looking at top sustainability in Green ICT trends.

00:00:42: It really dominated conversation across LinkedIn over these past two weeks.

00:00:47: The underlying theme here is just a massive collision between digital ambition and physical reality.

00:00:52: It's huge, we've got a massive stack of insights to get through.

00:00:55: We're gonna cover everything from the hidden energy costs of AI To you know...the deep mechanics of green software engineering

00:01:02: And sustainability reporting landscape which is honestly bit of wild west right now

00:01:06: Totally!

00:01:07: The goal for you listening right now Is cut-through noise.

00:01:11: We want extract actionable evidence based insight that can take back to your own IT infrastructure.

00:01:18: Exactly, understanding not just what is happening but you know the actual mechanics of why?

00:01:23: Right.

00:01:24: so let's start with The Elephant in the Room which is AI's explosive growth.

00:01:28: Mark Bradley shared some numbers that really set the baseline for this physical cost.

00:01:33: yeah his post was eye-opening.

00:01:35: we hear a lot about AI taking energy.

00:01:40: It's hard to grasp.

00:01:41: Bradley points out that a single chat GPT query takes about ten times the electricity of a standard Google search.

00:01:47: Ten times, just for one query?

00:01:49: Right.

00:01:49: and when you zoom-out at the macro level A single new data center in Illinois needs seven hundred megawatts power.

00:01:55: And get this three million gallons water per day.

00:01:58: That is a day!

00:01:59: Its staggering.

00:02:00: Bradley argues we have apply phenops discipline here and quote treat carbon like cash.

00:02:06: Treating Carbon Like Cash requires understanding how the spending actually works, right?

00:02:11: Because a traditional Google search is essentially a retrieval operation.

00:02:14: Right it's just looking something up

00:02:16: exactly.

00:02:17: It's looking up information in an incredibly well organized pre-existing index.

00:02:22: But generative AI query Is fundamentally different.

00:02:26: its calculating probabilities for every single word it generates in real time.

00:02:32: Its

00:02:32: firing billions of parameters to synthesize and answer yeah

00:02:36: The compute requirement is orders of magnitude higher.

00:02:39: Which makes sense on a per-query basis, but the defense you always hear from the tech giants... ...is that models are getting vastly more efficient.

00:02:46: Sure they say so

00:02:47: And they ARE right!

00:02:48: Ben Hillier highlighted Google reduced energy required for a Gemini query by thirty three times in just one year.

00:02:55: They absolutely did.

00:02:56: But Hillier also points out this brilliant counterintuitive reality here known as the Jevons Paradox.

00:03:02: Oh right, the economic principle

00:03:04: Yeah.

00:03:05: It says that as technological progress increases the efficiency of a resource, demand increases.

00:03:15: It's like adding a lane to a congested highway, it doesn't actually reduce the traffic...it just makes driving slightly easier for a minute which invites thousands of new cars until the wider highway is completely jammed again.

00:03:26: That is

00:03:26: the perfect analogy making inference The actual running-of-the AI cheaper and more efficient Just integrates into thousands of news software products Which drives up aggregate demand right?

00:03:38: And we have to look at the other side of the ledger, which is the training phase.

00:03:42: Hillier notes that the energy required to train frontier models has grown roughly thirty-seven thousand times in six years.

00:03:48: Thirty

00:03:49: seven thousand times,

00:03:50: wow!

00:03:51: Training GPT-IV alone took an estimated fifty-six thousand megawatt hours.

00:03:56: Okay so diet soda is the inference.

00:03:57: efficiency Yeah.

00:03:58: And three large pizzas are training costs and massive spike in aggregate demand.

00:04:02: Exactly The hangover from those pizzas lasts a very long time.

00:04:06: James Martin highlighted new MIT paper by Jennifer Turlich & John Sturman.

00:04:11: What

00:04:11: did they find?

00:04:12: They look at the actual atmospheric mechanics.

00:04:41: simply doesn't work that way.

00:04:42: Okay, let's unpack this.

00:04:44: if we're training models that require fifty six thousand megawatt hours isn't chasing per query efficiency just a massive distraction from the real problem?

00:04:53: It certainly masking larger structural issue.

00:04:55: Naveen Balani brought up this exact tension.

00:04:58: He argued that the current token economics model is completely breaking down,

00:05:01: and for you listening a token as essentially A piece of word That AI processes right?

00:05:05: And Balani's point Is at The next decade Of ai isn't going to be one by whoever Spends the most on compute.

00:05:11: it's gonna Be won By Whoever Waste the least.

00:05:13: that Makes total sense.

00:05:14: This Shear Velocity If This Growth is hitting a wall.

00:05:17: Karen VanderZand pointed out that Microsoft is currently having to deeply rethink their twenty-thirty climate goals.

00:05:24: Because they're sustainability commitments are directly threatened by their own AI deployment velocity, which brings us the physical reality of all this compute.

00:05:33: because software doesn't run in a vacuum.

00:05:36: Nope!

00:05:37: It has a physical footprint.

00:05:38: If we can't stop the AI train We have to re think where and how it runs in the physical world.

00:05:45: I want bring up Mathieu François.

00:05:47: he shared some mind-blowing research with Professor Xiaolei Ren regarding water stress.

00:05:52: Water is basically the primary mechanism we use to pull heat away from these massive GPU clusters, without it that chips melt!

00:05:59: Right and Francois and Ren found that a Lemma A-B query consumes eight point nine six milliliters of water in India versus three point two six millimeters in Germany.

00:06:08: Okay so on the surface That's less than a Three X difference.

00:06:11: volume

00:06:12: Yeah, you might think.

00:06:13: well the servers in India are just a bit less efficient at cooling.

00:06:15: But volume isn't the metric that matters here?

00:06:18: Exactly!

00:06:18: The metric that matter is regional capacity to replenish water.

00:06:23: Because of regional scarcity... ...the actual water stress impact is twenty-one times higher in India.

00:06:29: Twenty one times for exact same query.

00:06:31: Geography's primary variable not just volume.

00:06:34: This creates an incredibly difficult infrastructure bottleneck.

00:06:39: The power grid and the local water supply simply cannot move as fast, is a software company deploying new AI features.

00:06:46: Yeah And we're seeing two wildly different responses to this bottleneck popping up in our sources.

00:06:52: on one hand We have the brute force high friction approach.

00:06:54: right then negative one.

00:06:56: yeah

00:06:56: Anna Lerner-Nesbitt and Chris Adams brought attention to reports about anthropic leasing capacity at XAI's Colossus Facility in Memphis.

00:07:03: And just to be clear, we are purely reporting what was highlighted in these LinkedIn posts.

00:07:08: Yes!

00:07:09: The concern Nesbett & Adams raised is that this facility has reportedly operated dozens of unpermitted gas turbines bringing massive pollution into a frontline community.

00:07:17: But think the mechanics.

00:07:19: why a tech company would resort to burning gas locally.

00:07:22: A traditional power grid requires years of planning right?

00:07:25: It's exactly.

00:07:26: Yeah, permitting substation construction all to deliver five hundred megawatts safely and AI arms race demands that compute power now so you bypass the grid.

00:07:37: But then there's the other side of the coin, which is innovative infrastructure approaches.

00:07:44: Experts across our sources are tracking ocean-based solutions.

00:07:48: Amir Olajuan highlighted China testing underwater data centers cooled by seawater.

00:07:54: Stella Su discussed Pantherlassa building offshore AI datacenters powered entirely by wave energy.

00:08:00: And Sean Moran noted at Jacobson Start Campus Project in Portugal, achieving a water usage effectiveness of exactly zero.

00:08:07: Zero?

00:08:08: Zero!

00:08:08: By using Atlantic seawater for cooling.

00:08:10: Okay but I have to push back here.

00:08:12: Putting data centers underwater sounds like James Bond villain's lair.

00:08:15: It totally does.

00:08:16: Isn't

00:08:16: the maintenance, corrosion and connectivity of saltwater immersed tech just a logistical nightmare that completely offsets the cooling benefits?

00:08:24: The physical trade-offs are definitely high.

00:08:26: Salt Water is notoriously destructive But you have to understand the sheer desperation of thermal management right now.

00:08:33: James Hall observed that an Amazon data center in Virginia recently went down simply because it got too hot.

00:08:40: Wait, It didn't get hacked?

00:08:41: No!

00:08:42: A backhoe didn't cut a fiber cable, it just overheated.

00:08:46: Because AI chips run radically hotter than traditional cloud servers.

00:08:50: GreenOps isn't about saving carbon anymore

00:08:53: It's the sheer resilience and survival of technology.

00:08:56: Exactly

00:08:57: We're being forced into extreme environments because the traditional grid cannot handle thermal load.

00:09:03: If hardware is fighting for its life thermally we need to look at what's commanding that hardware so hard in first place.

00:09:10: The software Exactly.

00:09:12: The software, hardware only uses energy when software tells it to.

00:09:15: which brings us to our third theme Green Software and Engineering.

00:09:19: This is where we shift from physical geography To digital frugality.

00:09:23: Yeah

00:09:23: And Paul Young shared an incredible win From GitHub that illustrates this perfectly.

00:09:28: They managed to cut their agentic workflow tokens by forty three to sixty two percent.

00:09:32: That's massive!

00:09:33: For context An agent work flow.

00:09:36: When you give AI autonomy break down a big task, use tools and figure out the steps on its own.

00:09:42: Right!

00:09:43: And every time the AI thinks about step it processes tokens which burns energy.

00:09:49: GitHub achieved this massive reduction by simply moving basic data gathering at of LLM reasoning loop.

00:09:55: Oh interesting.

00:09:56: so what did they used instead?

00:09:58: Just simple command line interface commands CLI commands.

00:10:01: Using a billion parameter LLM to figure out what files are in a folder is like chartering a Boeing seven forty-seven To drive across the street to pick up groceries.

00:10:09: That's a great way to put it.

00:10:10: They also stripped out unused tool schemas didn't they?

00:10:13: Yeah, those were adding ten to fifteen kilobytes of overhead per API call.

00:10:16: So what does this all mean for the people writing the code?

00:10:19: Because it feels like we've been asking a chef to cook as zero waste meal without giving them a compost bin or proper knives.

00:10:25: We keep giving developers guilt trips instead of actual tools.

00:10:28: I love that framing and David DeCarvalho, and Johanser brought up this exact necessity.

00:10:35: Eco-design has to move past just sensitizing developers.

00:10:39: They need concrete tools.

00:10:41: Exactly, they specifically pointed to the API green score which gives a measurable metric on efficiency and they highlighted frameworks like Spring Native To make digital frugality A daily measureable reality.

00:10:54: And education is catching up too.

00:10:55: Rhett Palmera shared that The Green Software Foundation now has free self-paced courses.

00:11:01: Always great!

00:11:02: Anita Shutler discussed Blue Angel certification for sustainable software

00:11:06: Which perfectly aligns with Purvivarma's core argument.

00:11:09: She emphasizes that sustainability cannot be a side initiative.

00:11:13: It must be a core cloud architecture principle

00:11:15: right alongside cost and performance.

00:11:17: Exactly, but here's the catch to prove to your leadership That using spring native actually saved energy you need bulletproof data.

00:11:25: And now leads directly into our final theme The fragmented world of sustainability report.

00:11:30: yeah You can't reduce what you haven't measured.

00:11:33: Pascal Jolie shared a post detailing at baseline AI emissions assessment He just completed for octa.

00:11:38: He was mapping the carbon footprint of their LLM chatbots, right?

00:11:42: Yeah.

00:11:43: Yes and his biggest takeaway is that most sustainability teams have zero visibility into enterprise AI usage.

00:11:51: They are completely flying blind

00:11:53: Totally flying blind And part reason because the methodology still so raw.

00:11:57: Dr Sasha Lucione's literature review found that AI environmental reporting Is currently relying heavily on course proxies

00:12:05: Because the actual life cycle of AI is still poorly defined, so they're just guessing emissions based on money spent.

00:12:12: Exactly!

00:12:13: And when a metric is poorly-defined it leaves massive loopholes.

00:12:17: Vinayak Setpute and Shreya Gupta broke down.

00:12:20: nine tactics companies use beyond traditional greenwashing.

00:12:23: Tactics like greenhushing?

00:12:24: Yes

00:12:25: staying silent about genuine wins to avoid audits on other initiatives...and green crowding

00:12:30: Right, where you hide in slow-moving industry coalitions.

00:12:33: You sign a pledge to be carbon neutral by twenty fifty put the badge on your website and then do absolutely nothing.

00:12:38: Exactly!

00:12:39: The only antidote of these tactics is hard evidence.

00:12:42: And the market's responding?

00:12:44: Bjorn Stengel and Brianna Weiss mentioned.

00:12:46: IDC just mapped over thirty sustainability software vendors.

00:12:49: Charlie Appleton highlighted partnership between Greenpixie and Generate Zero To combine AI usage data with scope one through three carbon tracking

00:12:57: And Jay Ruckelshaus pointed out that the IFRS is shifting ESG disclosure rules to focus on substance over

00:13:04: form,

00:13:05: meaning disclosures must connect directly to financial performance and value creation.

00:13:10: Right now though it sounds like everyone is grading their own homework using entirely different rubrics

00:13:16: Pretty much.

00:13:17: When do we finally get a standardized test for carbon accounting so you can actually compare company A to company B?

00:13:22: That's exactly what Kate Goldenring or Svia Segment are pushing shared emissions standards.

00:13:28: Until we have interoperable data, as Bjorn Soaring-Gigler mentioned climate intelligence won't be truly actionable.

00:13:35: We're basically building the airplane while flying it!

00:13:37: We really are.

00:13:38: and that brings me to a final thought.

00:13:39: I want to leave the listener with today weaving together Marosa Zalabak's point about AI ethics

00:13:44: The question of who benefits?

00:13:45: And Who is harmed?

00:13:47: Yes pair that with Nelson Switzer rule of forty four regarding climate tech scaling.

00:13:52: The challenge for you listening is this, as you look at your own organization's digital transformation don't just ask if your new AI tool or cloud architecture is highly efficient.

00:14:02: Ask if the compute power is genuinely creating enough human-or business value to justify its physical footprint on earth?

00:14:11: Are we building efficient systems of things that we do not need in first place?

00:14:15: That is a phenomenal question to end on.

00:14:17: Yeah, if you enjoyed this episode new episodes drop every two weeks.

00:14:22: also check out our other editions on cloud digital products and services artificial intelligence an ICT in tech insights health tech defense

00:14:30: Tech.

00:14:31: thanks for listening everyone.

00:14:32: thank You so much for taking the steep dive with us today.

00:14:34: don't forget To subscribe And we'll catch On The next one.