Best of LinkedIn: Green ICT & Sustainable AI CW 32/ 33

00:00:00: This episode is provided by Thomas Allgaier and Frennis, based on the most relevant LinkedIn posts about green ICT and sustainable AI from CW-Thirty Two and Thirty Three.

00:00:11: Frennis supports ICT enterprises in the form of delivering precise ICT market and pricing intelligence that analysts subscriptions and existing databases cannot provide.

00:00:20: you can find more info in the description right.

00:00:22: so we call it The Cloud

00:00:24: Right.

00:00:26: It's... a brilliant piece of marketing on it.

00:00:28: Oh, absolutely!

00:00:28: It sounds so like weightless and ethereal?

00:00:32: Like your data or AI models are just this invisible vapor floating comfortably above us somewhere?

00:00:39: Yeah that metaphor does alot heavy lifting Because it implies this total lack of physical consequence.

00:00:46: You train a massive language model or generate a thousand images and the work just happens, well...it happens somewhere else.

00:00:52: It completely disconnects the user from the actual physical reality in infrastructure.

00:00:56: But when you look at the infrastructure being built right now to support this explosion in artificial intelligence I mean that cloud is not vapor.

00:01:05: It is concrete, it's thousands of miles of high-voltage wire.

00:01:09: It's spinning gas turbines and it is just incredibly thirsty!

00:01:14: So today we're unpacking the top green ICTs in sustainable AI trends that have been dominating the conversation among tech and digital transformation professionals over the past couple weeks.

00:01:24: Yeah...and the underlying tension across all the sources we looked at is, it's just simple physics.

00:01:30: Right AI compute demand as compounding exponentially and its slamming headfirst into physical limitations of a natural world.

00:01:38: Yeah Just wall limits.

00:01:39: Exactly Our clean power grids And our water infrastructure.

00:01:42: They cannot realistically support the curve we're on right now.

00:01:47: So were going to explore how industry trying engineer its way out of this, looking at grid constraints liquid cooling and even the software architecture fixes that are starting to emerge.

00:01:58: Yeah let's start with The Power Paradox because we all understand inherently that AI takes juice.

00:02:02: okay right but the sheer scale of the math here is staggering.

00:02:07: uh Hans Pelleig pointed out a projection that global data center power demand is set to nearly double by twenty thirty.

00:02:14: wow yeah.

00:02:15: And in the US alone We're staring down a forty-five gigawatt shortfall by twenty, twenty eight.

00:02:20: A forty five gigawatts short fall.

00:02:22: just to contextualize that is roughly the equivalent output of forty two fifty large scale nuclear reactors.

00:02:29: oh my gosh.

00:02:30: yeah It's so severe that it is actually quietly triggering the biggest nuclear power comeback we've seen in decades.

00:02:36: Which

00:02:36: is wild to think about!

00:02:37: Right,

00:02:37: because of over-arching narrative from tech sectors like AI will run on new clean energy but that narrative running years ahead what physical grid can generate and deliver today

00:02:48: And collision between corporate energy goals public policy and just raw AI demand.

00:02:54: It's creating some really fascinating contradictions, like Steve Chavez observed this dynamic unfolding in Texas he pointed out that AI load expansion which is demanding something like five times the state's peak load.

00:03:07: it's being rapidly green lit but simultaneously renewable energy development in the State Is facing all of these political and regulatory throttling

00:03:17: which regardless of your stance on energy policy, that highlights a major structural... Disconnect right you've got a massive acceleration of new electricity consumers on one side Of the ledger and an intentional slowdown of wind-and-solar deployment On The other.

00:03:31: yeah, And the physical reality is that Wind and solar are currently the technologies That can scale and deploy the fastest.

00:03:37: so when You throttle this supply?

00:03:39: And multiply the demand.

00:03:40: You're just.

00:03:41: you're creating A massive energy deficit.

00:03:43: When there's a massive Deficit companies Are forced to find the fastest bridge available Just To keep building Right.

00:03:48: Yeah mark butcher highlighted A pretty stark example of This.

00:03:51: So Amazon.

00:03:51: Amazon is building a massive future AI data center in Texas, and it's going to be powered by the GW Ranch Gas Plant.

00:03:59: The gas plant?

00:04:00: Yeah!

00:04:00: And air permits for this site show at projecting up to thirty-three million tons of CO₂ emissions per year.

00:04:06: Wait...thirty three million

00:04:08: tons?!

00:04:10: That's roughly nine times larger than Amazon's entire reported Global Scope II emissions last year….

00:04:17: Just one facility.

00:04:18: ...tied into single facilities operation.

00:04:21: So how I mean, how does a hyperscaler with these massive public climate pledges?

00:04:26: Connect to a facility like that without completely destroying their carbon ledger?

00:04:31: Well it comes down to the complex mechanics of carbon accounting because Amazon doesn't actually own the gas plant.

00:04:37: A third-party Pacifico energy Does I see so?

00:04:40: because Pacifico is the entity physically burning?

00:04:43: The gas those thirty three million tons fall under Pacifico scope one direct emissions.

00:04:48: Oh, wow right Amazon just buys the electricity generated by the plant which categorizes it as a scope to indirect emission for them.

00:04:55: sneaky

00:04:56: and Scope two emissions can be offset using market-based accounting And renewable energy certificates or you know RECs.

00:05:02: So

00:05:02: it's essentially a financial abstraction.

00:05:04: It's like claiming your on a strict diet because your friend technically purchased the junk food even though you're the one eating it.

00:05:10: That's

00:05:10: a perfect way to put it!

00:05:11: The atmosphere is still absorbing the carbon, but hey...the corporate spreadsheet is balanced.

00:05:16: Yeah..The physical atmosphere doesn't really recognize accounting categories.

00:05:20: But not every region is solving this power crunch by spinning up new gas plants.

00:05:27: We are seeing fundamental shift in geographic thinking too.

00:05:31: Okay, like what?

00:05:32: Well instead of dragging power across the country to a data center developers are starting to drop data centers directly onto stranded

00:05:40: power.

00:05:40: Oh that makes sense!

00:05:41: Andrew Sagequist actually shared some brilliant insights on this happening in Australia.

00:05:45: oh yeah

00:05:47: he noted that in twenty-twenty five New South Wales curtailed twenty nine percent of their wind and solar generation.

00:05:53: wow

00:05:54: meaning over two thousand megawatts clean energy was generated.

00:05:58: but they literally had just throw it away.

00:06:00: Yeah, and for those dealing with digital infrastructure curtailment is a crucial concept to grasp.

00:06:06: Right?

00:06:07: Can you explain that bit?

00:06:08: sure.

00:06:08: so electrical grids have to maintain perfect balance of supply-and-demand.

00:06:12: keep us steady frequency.

00:06:14: You can't just Pump unlimited electrons into a wire.

00:06:17: Right, it's not a bucket

00:06:18: exactly.

00:06:20: if demand is low and the wind Is blowing hard that excess power will literally melt lines or trip breakers.

00:06:26: Yikes.

00:06:27: so grid operators have to shut.

00:06:29: The winter binds off.

00:06:30: they curtailed the power.

00:06:32: But the point here is that this flips the narrative?

00:06:35: It's not A power shortage story anymore its a geographic matching Story.

00:06:39: right If you co-locate an AI data center right next to those renewable nodes.

00:06:45: That curtailed energy stops being a grid liability.

00:06:48: Exactly, becomes incredibly cheap.

00:06:50: clean fuel the data center just access this massive controllable sponge.

00:06:55: Yeah

00:06:55: and George Whitkin noted that China is actually executing This exact strategy at a national scale.

00:07:00: Oh really yeah.

00:07:01: in their latest five-year plan They're explicitly mapping Their major compute clusters to Western regions That have vast amounts of disconnected excess renewables?

00:07:10: I mean thats Just highly pragmatic engineering.

00:07:13: But the ultimate example of this geographic mismatch comes from South Africa.

00:07:18: This was via an analysis from Obina Aziadinso, Escom which is the public utility there actually has a six gigawatt surplus power right now.

00:07:29: Wait...a

00:07:30: surplus?

00:07:31: The megawatts exist but they are entirely bottlenecked by physical wires.

00:07:36: They need to build fourteen thousand five hundred kilometers of new high voltage transmission lines over the next decade just to move that power toward a compute.

00:07:45: demand actually is.

00:07:47: And building high-voltage transmission agonizingly slow.

00:07:51: Yeah, you're dealing with land rights environmental impact studies the The physical supply chain of steel and copper.

00:07:57: It's like having a massive reservoir water but only is standard garden hose to drain it.

00:08:01: yeah You know?

00:08:02: You have the volume But you completely lack the throughput

00:08:04: right which actually brings us to the next inescapable physical constraint Of AI.

00:08:09: okay when you do manage To push all those megawatts of power through millions of tiny silicon transistors you generate resistance And in physics Electrical resistance equals heat.

00:08:20: Massive concentrated amounts of heat.

00:08:22: Yeah, thermodynamics never loses.

00:08:24: Never!

00:08:25: And the way we've traditionally managed heat in data centers is fundamentally breaking down under these new AI workloads.

00:08:32: Philippe Giibushi laid out a technical reality check

00:08:35: here.

00:08:35: How did he say?

00:08:36: So... A standard conventional air-cooled server rack tops at around ten to fifteen kilowatts per hour.

00:08:42: Okay But the new AI clusters being deployed today, they're pushing past a hundred kilowatts

00:08:47: per rack.

00:08:48: Yeah and air is simply a terrible thermal conductor.

00:08:51: it physically cannot move that density of thermal energy away from the silicon fast enough to stop the chips from melting right.

00:08:58: so The industry is just being forced into liquid cooling direct-to-chip liquid cooling where cold plates sit right on the processors That can capture about sixty to seventy percent of the heat.

00:09:07: okay.

00:09:08: But the engineering endgame here is immersion cooling.

00:09:11: Immersion

00:09:12: Cooling?

00:09:12: That's where the entire server blade is literally submerged into a tank of liquid, right?

00:09:17: Which sounds completely counterintuitive!

00:09:19: Like how do you drop running Enterprise electronics in to a bath without immediately short-circuiting a multi million dollar

00:09:24: rack?!

00:09:25: Well because it isn't water... Oh okay They use highly engineered non conductive dielectric fluids So often synthetic oils or fluorochemicals.

00:09:35: And because the fluid doesn't conduct electricity, the electronics run perfectly fine but liquid is vastly denser than air so it absorbs the heat immediately.

00:09:45: Immersion cooling captures up to one hundred percent of the server heat.

00:09:49: That's

00:09:49: incredible!

00:09:50: Yeah and it drops the facility power usage effectiveness the PUE below one point zero.

00:09:55: three...and

00:09:55: just for context a perfect PUE as one-point-zero right meaning Every single watt entering the building goes to processors, and zero watts are wasted on overhead like fans or chillers.

00:10:06: Exactly!

00:10:07: Because a conventional air-cooled facility usually sits around what?

00:10:10: One point two to one point seven Yeah...usually.

00:10:13: So hitting one point zero three is massive efficiency leap.

00:10:15: It changes entire operational model of a facility.

00:10:18: Tim Krasofferson actually highlighted NVIDIA's new liquid cooled Rubin architecture which was designed.

00:10:24: take advantage.

00:10:25: Oh, nice!

00:10:26: Yeah and it almost entirely eliminates on-site water consumption.

00:10:30: Right For a standard fifty megawatt facility.

00:10:33: switching to this architecture saves roughly four million dollars a year.

00:10:37: just in cooling overhead

00:10:39: Just in cooling?

00:10:40: Wow.

00:10:41: And by the way if you find yourself needing to stay ahead of rapid infrastructure shifts like this make sure you hit subscribe so you catch our future deep digs.

00:10:49: Good call because we are definitely seeing these shifts accelerate at the facility scale.

00:10:53: oh yeah Michael Lisniak broke down the details of Google.

00:10:57: as planned, two-point seven gigawatt Wyoming campus.

00:11:00: Two point seven gigawatts!

00:11:02: It's practically a city unto itself.

00:11:04: but the crucial engineering detail is that the campus is designed to take zero municipal water.

00:11:09: Really?

00:11:10: Zero?

00:11:11: Zero.

00:11:11: It relies entirely on its own deep groundwater wells and utilizes a closed-loop cooling system, meaning they aren't constantly evaporating water into the atmosphere to reject heat.

00:11:22: But wait that completely contradicts the narrative we see in mainstream news?

00:11:28: If hyperscalars are utilizing closed loop systems or deep private wells why is there constant headlines about AI draining community water supplies?

00:11:36: Well it's a mix of aggregate data, local infrastructure limits and frankly psychological bias.

00:11:42: Okay bias.

00:11:43: Yeah Adrian Monday provided a really fascinating perspective on the biased side.

00:11:48: he pointed out that an average Google Data Center uses about four hundred fifty thousand gallons water per day

00:11:54: which sounds massive to a layperson.

00:11:56: It does, but a single golf course in Palm Springs uses about three hundred and seventy thousand gallons per day.

00:12:02: Wait they are virtually identical in daily consumption!

00:12:05: That is wild

00:12:06: And an aggregate.

00:12:07: across the United States Golf courses use roughly thirty times the direct water of all U.S.

00:12:12: data centers combined

00:12:13: Thirty times?

00:12:14: Yeah

00:12:15: Monday argues that grass is green, so it visually reads to us as natural you know environmentally innocent.

00:12:21: Whereas a data center is this brutalist concrete box filled with blinking lights.

00:12:25: So it visually codes as this industrial threat hoarding resources?

00:12:29: That makes so much sense.

00:12:31: But beyond the perception problem What's the actual physical friction happening in these local communities?

00:12:36: because if aggregate water supply isn't the existential threat Why do local water boards push back so hard?

00:12:43: Well, Shelley Ren clarified the actual mechanics of this bottleneck and it's really interesting.

00:12:48: What is the

00:12:48: issue?

00:12:48: The issue for a town hosting data center usually isn't total volume water in their aquifer or reservoirs.

00:12:56: The issues are physical.

00:12:57: pipe and pumping capacity of municipal utility Delivery.

00:13:01: infrastructure has to be scaled to accommodate facilities.

00:13:04: absolute maximum peak demand.

00:13:07: Okay so its flow rate problem Exactly Its like stadium at halftime There might be plenty of water in the city tower, but if fifty thousand people flushed the toilet at that exact same second... The pipes lose all pressure.

00:13:18: Precisely!

00:13:19: Many local utility networks just do not have a pipe diameter to handle data center-pulling peak volume without dropping the water pressure for rest of town.

00:13:27: So it's plumbing problem and no scarcity?

00:13:30: Yes exactly.

00:13:32: And before we move off physics heat It is worth noting.

00:13:35: waste heat doesn't need liability

00:13:37: True

00:13:38: Peddery.

00:13:38: Nicky pointed out that Finland has actually passed a law requiring all datacenders over one megawatt to reuse their waste heat by December, twenty-twenty seven.

00:13:49: Really?

00:13:49: How do they do that?

00:13:50: They pumped the hot water directly into local district heating networks to warm residential homes.

00:13:55: Oh

00:13:55: this is brilliant!

00:13:56: It forces developers treat thermal energy as secondary commodity.

00:13:59: Exactly So we're engineering better ways power and cool hardware But hardware is only half the equation, right?

00:14:05: Yep.

00:14:06: We'll look at the logic layer like what does this software actually doing to demand all of these compute in first place?

00:14:11: Yeah and the software bloat is expanding just as fast as the hardware.

00:14:14: Oh yeah Pascal Jolie shared a deeply alarming metric about shift toward agentic AI.

00:14:20: Right And Just To Clarify Agentic AI represents a shift from simple query response model Where you ask question and get paragraphs back into an autonomous model.

00:14:30: You give an AI agent high level goal and it writes code, tests that code.

00:14:35: Browses the web encounters an error rewrites the code and just iterates in a continuous loop until it achieves the goal.

00:14:42: And Jelly notes these agetic tasks consume roughly one thousand times more tokens than standard chat exchange

00:14:49: A THOUSAND TIMES!

00:14:50: Yeah

00:14:51: We should probably define what token is here because directly translates to electricity.

00:14:55: Definitely A Token essentially means piece of word.

00:14:59: If you take the word hamburger, The AI might process that as three separate tokens and for every single token an AI model generates it has to perform billions of mathematical matrix multiplications across its neural network.

00:15:11: Yeah

00:15:12: Every calculation requires a physical electrical pulse in a GPU.

00:15:15: exactly

00:15:16: so.

00:15:16: multiply That by one thousand four agentic tasks And the compute budget becomes astronomical.

00:15:20: But there are immediate practical software engineering fixes terrain this end which is good news

00:15:26: right?

00:15:26: Simon R pointed out that simply constraining output length, literally just depending answer in three bullet points to your prompt reduces the energy use of that query by ten to thirty percent.

00:15:36: That is such a simple fix!

00:15:38: Right.

00:15:39: furthermore using low-energy verbs can make a massive difference.

00:15:43: Low energy verbs?

00:15:44: Yeah.

00:15:44: so asking a model to summarize a document instead of asking it analyze can cut energy consumption by up to thirty percent.

00:15:52: Oh, because it triggers a fundamentally different semantic pathway in the model's architecture.

00:15:57: Exactly.

00:15:58: Analyze... forces the LLM to weigh complex, abstract connections and synthesize new insights across billions of parameters.

00:16:07: But summarized is largely an extraction task so it's computationally much lighter.

00:16:12: It's like forcing a Michelin star chef To explain the agricultural history of wheat The physics of yeast fermentation And cultural significance of grain.

00:16:21: Just get them make you a turkey sandwich.

00:16:23: Yes

00:16:24: We are burning massive amounts of electricity on AI Fluff.

00:16:27: We really are, and developers are starting to realize that inefficiency is costing them real money.

00:16:32: Oh absolutely!

00:16:33: Robert Key has highlighted an update.

00:16:35: in a tool called GreenPT which provides API level compression model They use rule sets developed by the open source community specifically models nicknamed caveman & ponytail To compress output before LLM wastes compute generating it.

00:16:50: Then

00:16:50: it strips out all the conversational pleasantries.

00:16:52: Exactly, you don't need AI to generate... Sure!

00:16:55: I would be happy with that Python script.

00:16:57: here is code.

00:16:59: It just outputs raw data arrays.

00:17:01: Nice

00:17:02: And results are wild.

00:17:04: They can cut output tokens by upto seventy percent and reduce generated code.

00:17:10: That's

00:17:10: huge.

00:17:10: You get the exact same functional answer, but you've dramatically reduced the API cost and compute time in the resulting CO-II emissions.

00:17:17: And

00:17:17: if we want to take that software hardware integration into the absolute extreme Nolwin Goddard flagged some ongoing research into neuromorphic computing.

00:17:26: Oh!

00:17:27: Neuromorphic?

00:17:28: Fascinating.

00:17:29: Yeah this is hardware structurally designed to mimic the spiking neural pathways of human brain.

00:17:34: It uses event driven processing.

00:17:36: Okay how does it compare with what we use now?

00:17:38: Well, think of conventional chips like a light bulb left on twenty-four hours per day.

00:17:42: Right?

00:17:43: Just constantly running a clock cycle whether it's doing useful work or not.

00:17:46: Okay Neuromorphic computing is like a motion sensor light.

00:17:50: Yeah It only fires an electrical spike when there are specific data to process.

00:17:54: Wow!

00:17:55: And because that event driven architecture draws power in the milliwatts right?

00:18:00: Yes,

00:18:00: mill one

00:18:00: Whereas conventional AI hardware pulls three hundred and seven hundred watts per chip Exactly.

00:18:06: It is a massive paradigm shift in how we process information, particularly at the edge.

00:18:12: So we optimize this software?

00:18:14: We shrink the prompts and build hyper-efficient neuromorphic chips.

00:18:18: That solves that problem right?

00:18:19: Yeah!

00:18:20: Not exactly Because it brings us to our final thing The physical life cycle of the hardware itself, right?

00:18:25: And this is a critical blind spot in IT procurement and sustainability reports.

00:18:29: Really yeah.

00:18:31: Naveen Balani shared an important warning about the Hardware refresh cycle.

00:18:35: Often A cloud platform team will justify ripping out old servers and buying new ones by pointing to the newer more energy-efficient chips.

00:18:42: makes sense Right the operational math the daily electricity bill looks great.

00:18:46: on paper

00:18:47: it does but that ignores embodied carbon.

00:18:50: Embodied Carbon Is the total sum Of All the energy it took to mine raw silicon and rare earth metals smelt copper, manufactured motherboards in a fab ship heavy metal across ocean.

00:19:04: If outgoing server fleet is retired at only half of its life span massive carbon of incoming hardware completely wipes out any operational energy savings you get from new chips.

00:19:16: Wow.

00:19:17: Upgrading a data center to be more power efficient actually puts you deeper into the carbon red if you throw away perfectly functioning gear,

00:19:23: so true sustainability and IT often means extending the lifespan of equipment that we already have

00:19:28: or looking backward at modernized versions much older technologies.

00:19:33: Carlos Sandoval Castro shared perhaps the most surprising alternative in these sources because when tech professionals think cutting edge data storage They do not usually think of magnetic tape.

00:19:45: No, tape storage feels like a total throwback to the mainframe era in the nineteen seventies.

00:19:50: Right

00:19:50: Like cassettes.

00:19:51: But modern tape storage is having massive renaissance for archiving unstructured AI data.

00:19:57: It really is.

00:19:58: Castro highlighted IBM's climate control diamond back system.

00:20:02: It offers eighty-four percent lower costs and ninety seven percent lower energy consumption compared to spinning disk storage.

00:20:09: Ninety Seven percent Lower Energy.

00:20:11: Yeah,

00:20:12: And the key innovation isn't just a tape itself.

00:20:14: it's environmental resilience because historically Tape required heavily climate controlled clean rooms.

00:20:20: Right

00:20:20: very sensitive.

00:20:21: But this new generation of hardware can operate in environments up to forty-five degrees Celsius.

00:20:26: Which is over a hundred and ten degrees Fahrenheit?

00:20:28: Exactly,

00:20:29: you do not need a dedicated children room

00:20:31: because it's cold storage.

00:20:32: It just sits there quietly holding petabytes data using almost zero power until robotic arm physically retrieves the cartridge To read yeah its brilliant low tech no power solution too.

00:20:46: very modern problem

00:20:47: really.

00:20:47: as so looking back at ground we've covered today The overarching theme is just the sheer scale of physical engineering required to keep our digital world running.

00:20:57: We moved from forty five gigawatt power shortfalls and transmission line bottlenecks through the fluid dynamics of immersion cooling in municipal water pipes, navigated the token bloat.

00:21:09: agentic software ended up at a hidden carbon cost.

00:21:13: server upgrades magnetic tape drives.

00:21:15: It really proves that the digital transition is at its core a heavy industrial transition.

00:21:21: Yeah, it's very physical.

00:21:22: We are moving massive amounts of physical atoms just to process bits and that leads to one final perspective.

00:21:28: What

00:21:29: was I?

00:21:29: well?

00:21:29: we've spent this entire deep dive Examining how the industry is trying to make deacenters greener and more efficient.

00:21:35: But lexicido shared a recent peer-reviewed study published in nature That looks at the demand side of the equation.

00:21:41: Oh meaning what the AI is actually being used for.

00:21:44: Exactly, it's the ultimate secondary effect!

00:21:47: The study found that the AI enabled productivity gains being utilized by the fossil fuel industry.

00:21:52: so using AI to locate new reserves and accelerate oil and gas extraction those gains are projected to emit three to thirteen times more carbon then the data centers themselves are emitting.

00:22:04: Three to thirteen times more, yes that completely dwarfs the direct physical footprint of the compute.

00:22:10: we just spent all this time analyzing it.

00:22:12: does or were you spending all these engineering brilliance optimizing the data center?

00:22:16: Yeah Just use a resulting AI to dig up more carbon.

00:22:18: That is heavy thought.

00:22:19: to leave with

00:22:20: It really was well.

00:22:22: If you enjoyed this episode, new episodes drop every two weeks.

00:22:25: Also check out our other editions on Cloud Insights and Sovereignty, Digital Products & Services, AI and Agenic Systems, HealthTech, ICT and Tech Insights, DefenseTech and HealthTech.