Best of LinkedIn: Green ICT & Sustainable AI CW 44/ 45

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

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

00:00:18: You can find more info in description.

00:00:21: So today we're really looking at a very real possibility you know, draining municipal water supplies and risking

00:00:28: local grid blackouts.

00:00:29: Right?

00:00:29: Yeah

00:00:30: exactly all just to generate AI emails.

00:00:34: we're cutting through the whole corporate sustainability hype to look at the hard data The physical bottlenecks and how tech professionals are actually scrambling to fix the massive resource demands of modern AI.

00:00:46: Because it's so easy to get lost in like, the marketing brochures about green data centers.

00:00:50: but for you listening the reality is that the physical infrastructure required to support these models as hitting some severe ceilings geographically and regulatoryly

00:01:00: It really does.

00:01:01: So we're unpacking the mechanics of power distribution, hidden costs data center cooling and basically the strict new rules dictating how our digital future actually gets built.

00:01:11: Right but before even look at physical concrete steel in those datacenters I mean We really need to understand invisible workload burning through all that power

00:01:20: The software footprint.

00:01:22: Exactly it starts with a softwares

00:01:23: And numbers were dealing with at the software layer are just frankly staggering.

00:01:28: Yeah they realy.

00:01:29: Edward

00:01:30: Chin and Bjorn Soren-Gigler highlighted some recent data from the UN University in The International Energy Agency, And they project that datacenter electricity use is going to jump from four hundred and forty eight terawatt hours In twenty twenty five To roughly nine hundred fifty terawat hours by twenty thirty

00:01:48: Which I mean it's basically doubling the power draw just five years.

00:01:51: It's insane.

00:01:53: And the most crucial detail in those numbers isn't even the initial training of these massive models.

00:01:57: Right,

00:01:57: it's the inference!

00:01:58: Exactly between eighty and ninety percent of AI energy consumption actually goes into inference using it.

00:02:05: Wow because training at LLM is like building a massive factory right?

00:02:09: It takes this huge concentrated amount of energy once but inference that's the factory running twenty four seven.

00:02:17: Oh, okay.

00:02:17: Is the energy consumed?

00:02:18: every single time you query a model or ask it to summarize document and generate line of code?

00:02:24: It just compounds!

00:02:25: So the problem isn't just the sheer size of the models, it's the sheer volume of usage.

00:02:30: Which actually reminds me this brilliant historical parallel that Alex Wilmez brought up regarding William Stanley Jevons and The Steam Engine back in eighteen sixty-five?

00:02:38: Oh!

00:02:38: The Jevon's Paradox.

00:02:39: Yeah exactly which basically tells us.

00:02:42: making a technology more fuel efficient doesn't decrease resource consumption.

00:02:46: It

00:02:46: usually explodes

00:02:47: Right, because efficiency lowers the cost of operation.

00:02:50: When they made steam engines more efficient burning coal became cheaper and more profitable.

00:02:54: So that didn't use less coal?

00:02:55: No!

00:02:56: They just built hundreds new factories and embedded steam engines into literally everything.

00:03:00: And we are watching exact same rebound effect play out with compute power

00:03:03: right now.

00:03:04: Exactly.

00:03:05: A paper from Bank for International Settlements which Julio C. Serrano shared warns this AI rebound is happening RIGHT as we speak.

00:03:13: Because if we optimize a model, so it requires half the compute power.

00:03:16: developers don't just you know self rate The energy savings and go home.

00:03:20: they embed that AI into refrigerators customer service bots in every single piece of enterprise software.

00:03:26: So our total energy consumption skyrockets anyway.

00:03:29: right an Abhijit Sunil raised A great point about this.

00:03:33: If Enterprise architects Don't start tracking the total volume Of their ai usage any efficiency gains at the silicon layer will Just be completely swallowed up by that increased demand.

00:03:42: But tracking total volume sounds great in theory, but tying abstract software operations to a physical carbon footprint is notoriously difficult?

00:03:50: Yeah I mean if i'm a software engineer I can't easily see how my specific lines of code are impacting a power grid like a thousand miles away.

00:03:57: Exactly

00:03:58: you need frameworks that bridge that gap.

00:04:00: our will Owen highlighted Avila stars C three framework which Is designed to do exactly this for massive enterprise environments.

00:04:07: okay How does that work?

00:04:08: It maps emissions directly to specific system components like SAP memory, CPU and storage.

00:04:14: Oh wow!

00:04:15: To put that into perspective for you if your outside the enterprise software space SAP isn't just a simple application it often runs on these massive in-memory databases.

00:04:24: so holding terabytes of global supply chains data in ActiveRAM requires constant mass flow electricity.

00:04:32: Just keep the data accessible

00:04:34: Yeah.

00:04:35: And by mapping that specific memory footprint to carbon emissions, the C-III framework allows developers to see exactly where lazy database queries are burning both carbon and company money.

00:04:45: That's fascinating.

00:04:46: The industry is also starting to tackle this dynamically at the hardware execution layer too.

00:04:52: Nina Benoit pointed out a company called NeuralWatt.

00:04:55: They're entirely rethinking how GPUs handle these workloads By dynamically reconfiguring GPU every one hundred milliseconds.

00:05:02: Wait really?

00:05:04: Every one hundred milliseconds.

00:05:05: Yeah,

00:05:06: it fundamentally changes the incentive structure.

00:05:09: instead of a GPU running at maximum voltage non-stop regardless with load It analyzes the instantaneous compute demand and adjusts its power limits multiple times per second.

00:05:19: Oh that makes sense!

00:05:20: Its like difference between keeping your foot slammed on gas pedal at all time versus you know coasting down hill when don't need engine torque

00:05:27: Exactly.

00:05:28: And by doing this they managed to slash energy used per token by seventy percent.

00:05:32: Seventy percent

00:05:34: And even more interestingly, NeuralWatt changed their pricing model.

00:05:38: They charge for AI by the kilowatt hour instead of by token.

00:05:41: Oh that's huge!

00:05:42: Because Pricing By The Kilowatt Hour forces the provider to actually care about efficiency.

00:05:48: If you charge it by the Token You just want more volume.

00:05:51: But if charged by energy consumed Deep Code Optimization becomes your fastest path To wider margins

00:05:57: Which is exactly what Scott Hamilton argued.

00:06:00: Strategic hardware reuse and optimizing the code itself completely outclasses the old industry habit of just throwing raw brute force cloud capacity at bloated software.

00:06:09: But I mean, code optimization eventually hits a physical ceiling right?

00:06:13: It does.

00:06:14: yeah

00:06:14: no matter how clean your database queries are The silicon processing them generates an immense amount of heat.

00:06:20: So we really have to look at the physical reality of the data center infrastructure

00:06:24: And how we measure its efficiency which is a whole other issue.

00:06:27: Right Because the gold standard for measuring that physical efficiency has always been PUE or power usage effectiveness.

00:06:35: But as Nicola Spadoni explained, we're reaching a point where PUE is masking some major infrastructure flaws?

00:06:41: Yeah...for sure!

00:06:42: For those who aren't familiar, PUE is basically a ratio of total facility power versus the power used strictly by IT equipment.

00:06:50: Think about data center like car.

00:06:51: PUE measures how much gasoline is actually turning the wheels versus how much gas has burned just running the radiator fans to stop the engine from melting down.

00:06:59: That's a great analogy!

00:07:01: So, A PUE of two point oh means for every one kilowatt hour your servers use you are burning another full kilowatt-hour on overhead like air conditioning.

00:07:09: but a PUE Of One Point Two Is Highly Efficient.

00:07:12: You're only spending zero point two kilowatt hours To cool that same unit of compute

00:07:16: Exactly

00:07:16: And real quick, if you find these kinds of deep dives into technical metrics and infrastructure mechanics valuable make sure you subscribe to the show so you catch our future breakdowns on tech trends.

00:07:26: What going back to PUE?

00:07:28: The fatal flaw is that it operates entirely in a vacuum.

00:07:32: what do mean?

00:07:32: well It

00:07:33: doesn't care If your highly efficient one point two facility Is powered by coal plant or wind farm.

00:07:38: Right!

00:07:39: It ignores the massive carbon debt Of concrete steel used for construction And above all else, it completely ignores the consumption of local water.

00:07:47: Which is arguably the most terrifying bottleneck we face right

00:07:50: now.

00:07:50: It really does.

00:07:51: Shall we rent detail this tension perfectly?

00:07:53: If you want a zero-water data center so relying entirely on massive fans and dry cooling You have to consume significantly more electricity To push that air.

00:08:04: So your PUE suffers

00:08:05: Right.

00:08:06: Your peak PUE might jump up to one point three or one point five.

00:08:10: But if you want to push your power efficiency down into a pristine one point one or one point two, You have to use evaporative cooling.

00:08:18: Which uses water?

00:08:20: A lot of it!

00:08:21: It evolves running local municipal water over cooling media and blowing the hot server air through it.

00:08:27: The water absorbs heat and evaporates out of building.

00:08:31: So that is just physically gone from the local watershed.

00:08:34: Gone Ren pointed out that for a one gigawatt hyperscale data center in the hot climate, evaporative cooling can demand a peak of ten million gallons water every single day.

00:08:44: Wait... Ten Million Gallons?

00:08:47: A Day!

00:08:49: Every Single Day.

00:08:50: And The bottleneck isn't just finding you nearby lake.

00:08:52: Think about local municipal water treatment plant trying to process and pipe ten million extra gallons per day through city's network during like July heat wave.

00:09:01: You are directly competing with the local communities drinking and agricultural supply.

00:09:06: Exactly!

00:09:07: This extreme thermal challenge forces facility designers into a really difficult corner regarding how they build their cooling systems.

00:09:14: But, with AI hardware evolving so fast isn't optimizing of facilities' cooling for today's GPUs a massive trap?

00:09:22: That is exactly what Dilemma PS Lee highlighted.

00:09:24: It was

00:09:25: this tension between building tightly matched infrastructure versus an adaptive data center.

00:09:30: Because if I'm building a massive facility today, i might perfectly tailor my liquid cooling manifold to the exact thermal specs of this year's AI chips right?

00:09:39: Right.

00:09:39: Which gives you peak efficiency.

00:09:40: on day one.

00:09:41: It's like buying a perfectly tailored suit.

00:09:44: But when next years GPUs roll out and they run say thirty percent hotter that suit no longer fits.

00:09:49: You have to ripout millions of dollars of plumbing To retrofit the building.

00:09:52: Exactly So.

00:09:53: Lee argues That Building Adaptable Cooling Zones which function more Like an adjustable belt is far more economically resilient.

00:10:02: They might run slightly less efficiently on day one, but they hold their value longer because it can handle fluctuating thermal profiles without a massive retrofit every two

00:10:11: years.".

00:10:12: And we see that demand for resilience extending to the lifecycle of actual hardware sitting in those racks too?

00:10:18: The traditional decommissioning process just involves shredding old hard drives and melting down servers.

00:10:25: But Jeff Bittner pointed out that true sustainability requires R-IIvIII certification.

00:10:31: Write the r-II standard for responsible recycling, The v-III update heavily emphasizes functional reuse of a raw material recovery

00:10:40: Because you measure success by individual unit save not pound of metal scrapped.

00:10:45: Yes it requires facilities to rigorously test and wipe hardware To prove its dead before they are legally allowed to shred it

00:10:51: Which is great.

00:10:53: Mitch Black brought up a fantastic micro-example of this with the company called RefreshTech.

00:10:57: They buy back aging Chromebooks from school districts, securely wipe drives and refurbish them for continued use.

00:11:03: That completely circumvents the carbon cost manufacturing new hardware

00:11:07: Exactly!

00:11:08: And it keeps toxic e-waste out of landfills while cycling capital into schools.

00:11:12: It's a win-win.

00:11:13: But even all these efficiencies as we hit Absolute ceilings on local power grids and water networks in traditional hubs like Virginia or Silicon Valley.

00:11:23: Yeah,

00:11:24: hyperscalers are aggressively hunting for alternative geographic locations.

00:11:28: But geography fights back.

00:11:31: dropping a one hundred megawatt AI cluster into a new environment comes with immense physical friction.

00:11:37: it does but some we're getting really creative.

00:11:40: Chukunord and Eric Obi highlighted to start up named pantalassa.

00:11:43: Oh the ocean among.

00:11:44: Yes, they're trying to bypass land grids entirely by building wave-powered subsea AI nodes off the coast of Washington state.

00:11:53: The mechanics are fascinating right?

00:11:55: They submerge sealed server capsules into the ocean.

00:11:58: Yeah

00:11:59: and the kinetic motion of waves generates electricity while natural ambient temperature provides passive cooling.

00:12:07: So it completely removes need for massive air conditioning units

00:12:10: Exactly

00:12:11: But sounds brilliantly self contained until you look closer.

00:12:15: LMR and police raised some severe concerns about the ecological blind spots here.

00:12:19: Oh for sure.

00:12:20: I

00:12:20: mean if you scale these subsea capsules up to hyperscalar capacity.

00:12:24: aren't we just continuously dumping concentrated server heat directly into fragile marine ecosystems?

00:12:30: Yeah, your literally heating

00:12:32: plus the acoustic noise generated by wave turbines and internal cooling pumps, we risk severely disrupting marine wildlife.

00:12:40: It feels like we are simply moving our environmental footprint out of sight where it's harder for the public to see the

00:12:45: damage.".

00:12:46: That is a very fair critique because every single geographic stream carries a hidden cost...

00:12:51: Like

00:12:52: take the proposal Tech Jung Mahat shared regarding massive forty to fifty-five megawatt green AI hub in the Himalayas of Nepal.

00:13:00: Okay,

00:13:01: using mountain air?

00:13:02: Yeah!

00:13:02: The pitch relies on utilizing cheap hydropower and free freezing mountain air for cooling.

00:13:07: But mountains also mean incredibly complex high risk terrain And hydro power in Nepal is predominantly run-of river right

00:13:16: Exactly Like a massive dam with reservoir that stores potential energy year round.

00:13:21: Run-of River depends entirely on natural seasonal flow water.

00:13:25: So when the winter dry season hits and the rivers drop, the turbines slow down.

00:13:29: Your facility suddenly doesn't have enough

00:13:31: power Right forcing you to import dirty grid electricity from neighboring regions.

00:13:36: And added that it's your geographic danger.

00:13:38: Natural disasters.

00:13:39: Yeah

00:13:40: Mahat noted that the recent Boti-Koshi flood which was triggered by a glacial collapse completely devastated local infrastructure.

00:13:48: Building billion dollar compute clusters in zones prone to sudden landslides earthquakes of glacial floods requires catastrophic risk insurance that basically ruins the economic viability of the project.

00:13:59: Because geography dictates the terms?

00:14:01: Exactly!

00:14:03: But we are seeing major structural investments in locations that can blend diverse energy sources to mitigate that risk, like Yasser Al Barami shared an update on a four hundred megawatt project in Salalah Oman backed by three billion dollar initial investment.

00:14:19: By balancing solar, wind and large-scale battery storage they are leveraging Oman's natural grid advantages alongside its strategic subsea cable access.

00:14:28: And we also see pragmatic compromises happening in less ideal environments too.

00:14:32: Medet Rakhimchenov discussed a fifty megawatt data center campus planned for Ekibastu's Kazakhstan by twenty twenty seven.

00:14:39: But that relies heavily on coal doesn't it?

00:14:42: It does.

00:14:43: It relies largely on an existing coal-based power supply, which is obviously a massive step backward for carbon emissions.

00:14:49: However they are engineering a waste heat recovery system to feed thermal energy directly into the local municipal heating infrastructure.

00:14:57: Okay so providing local utility from a carbon heavy setup?

00:15:00: Exactly but when space and power are completely locked down by government The engineering has become flawless

00:15:07: Like in Singapore.

00:15:08: Yes

00:15:09: Torsten Pot highlighted Equinix's new SG-VI facility in Singapore.

00:15:14: Because Singapore literally paused all a new data center development to rethink their strategy, right?

00:15:18: They did and SG- VI is basically the blueprint for compliance now built under stricter government standards.

00:15:25: It's designed to run significantly hotter at twenty seven degrees Celsius which drastically lowers The power required for cooling.

00:15:31: Oh

00:15:31: smart

00:15:32: it features direct to chip liquid cooling integrates solar directly onto the facade And strictly targets A one point two zero PUE.

00:15:40: Well, this shift in Singapore feels like a precursor to a massive global regulatory wall that developers are hitting.

00:15:46: Because you can't just build wherever you want anymore

00:15:48: No!

00:15:49: You really cant.

00:15:49: The era of writing voluntary corporate sustainability reports and promising to be green by twenty forty is officially dead.

00:15:57: Governments and local communities are actively stepping with strict rules And red tape.

00:16:02: We could see the transition clearly in EU right now.

00:16:06: J Ruckels Hall highlighted a new EU proposal that takes effect in August, twenty-twenty

00:16:10: seven.

00:16:11: What does it mandate?

00:16:12: Any data center operating at five hundred kilowatts or larger will be legally required to publish an annual sustainability label.

00:16:19: they'll be graded from A to G on their PUE of.

00:16:22: there are water usage effectiveness and Their heat reuse capabilities.

00:16:26: now It's forcing complete public transparency exactly like the energy rating printed on a kitchen appliance

00:16:32: Exactly.

00:16:33: And Spain is taking this step further aggressively tying green compliance directly to grid reliability.

00:16:40: Fredit Chloéar Sabater shared details on a really fascinating revised Spanish decree,

00:16:45: right?

00:16:45: The twenty thirty one deadline?

00:16:46: yeah

00:16:47: the government is giving data centers until twenty thirty-one to prove they are matching eighty percent of their power with hourly renewables.

00:16:54: but operating during this transition period exposes facilities two s wrap curtailment

00:16:59: and Srap is an automated power reduction protocol managed directly by.

00:17:05: Think of SRAP like getting bumped from an overbooked flight, except you receive zero compensation.

00:17:12: If the Spanish grid faces a sudden demand spike or a transformer goes down.

00:17:16: data centers operating under these transition rules are very first consumers to be forcibly and automatically disconnected.

00:17:26: It is simply the non-negotiable cost of plugging into their grid.

00:17:29: This forces developers to spend tens of millions on robust onsite battery backups if they want a guarantee uptime for their clients.

00:17:37: And we are seeing grid access weaponized to force compliance in multiple markets.

00:17:42: In the UK, Brian Burns and Automass from Gartner discussed The Data Center Covenant which is tying faster grid access directly to environmental and social requirements

00:17:51: Which makes sense.

00:17:52: Yeah An earmux center noted as similar trended Portugal where our previously voluntary water efficiency commitments are shifting into strict regulatory laws.

00:18:00: But you know, the resistance isn't just coming from federal lawmakers—the grassroots backlash from local communities is becoming a massive bottleneck!

00:18:08: No absolutely.

00:18:09: James Martin looked at data from recent U.S.

00:18:11: midterms and voters actively mobilizing against unchecked tech development in their backyards.

00:18:17: out of forty local ballot measures related to datacenters thirty-nine were voted on to specifically restrict or halt development.

00:18:24: That's huge.

00:18:26: Communities are looking at the ten million gallons of water and a massive strain in local electrical substations, just saying no!

00:18:33: Because this localized resistance up to forty-one percent plan data center capacity could face significant developmental pauses or outright cancellations.

00:18:43: I mean...a forty one percent disruption is catastrophic for hyperscalers trying to deploy AI chips.

00:18:50: And this public friction is happening globally.

00:18:52: Martin H. Morissette highlighted a recent situation in Finland involving Google's massive one point three billion euro project, Google was forced to publicly apologize after they reportedly cleared over five hundred hectares of forest before the required environmental impact assessments had even been finalized.

00:19:11: when a hyperscaler with limitless capital has to halt construction and apologized to local municipality It reveals a fundamental shift in leverage.

00:19:21: Totally,

00:19:22: as Mateus Hamas pointed out at the DenaCon conference The tech industry has to adapt.

00:19:28: The open source software community was built entirely on the concept of inspectable verifiable code.

00:19:34: Right

00:19:35: Now society is demanding that same rigorous verifiability for the physical infrastructure layer.

00:19:40: .The era of self-declared sustainability Is ending.

00:19:44: every claim must be independently audited.

00:19:46: So we are looking at an industry wrestling with highly inefficient software workloads, hardware that is literally boiling municipal water supplies.

00:19:55: Geographic experiments fraught with ecological risk and in increasingly hostile regulatory environment can disconnect your facility a moment's notice.

00:20:04: Which brings us to profound underlying paradox regarding

00:20:08: why were

00:20:09: building this infrastructure the first place.

00:20:11: The ultimate irony

00:20:13: Yes Nolwin Goddard shared data illustrating the incredible theoretical capability of AI to solve the very problems it is exacerbating.

00:20:21: Because the models we are spending billions to train possess unparalleled analytical power?

00:20:26: Exactly!

00:20:27: According to Goddart, if directed properly, AI has a potential cut between three point two and five point four billion tons of CO₂ emissions annually by twenty thirty-five.

00:20:38: Just by optimizing things.

00:20:40: Yeah

00:20:40: by perfectly optimizing national power grid distribution, predicting renewable energy generation down to the minute and radically streamlining global logistics in supply chains.

00:20:50: So we have built a tool perfectly suited to untangle the complexities of grid management and carbon reduction?

00:20:56: But currently outside of a handful of pilot programs The largest technology companies are not aiming this massive intelligence at infrastructure problems.

00:21:04: No they're not!

00:21:05: The vast majority is being utilized to generate boilerplate code create synthetic images, and optimize consumer advertising algorithms.

00:21:15: Right so as you look at your own organization's tech stack I want to leave you with this final thought.

00:21:20: if we are willing to drain reservoirs clear-cut forests and push global power grids to the brink of collapse in order to build this incredibly powerful resource hungry intelligence why aren't we demanding that it be directed?

00:21:35: If you enjoyed this episode, new episodes drop every two weeks.

00:21:42: Also check out our other editions on cloud insights and sovereignty digital products and services AI and agentic systems health tech ICTN Tech Insights defense tech and HealthTech.