Latest / Best of LinkedIn: Digital Transformation & Tech / Best of LinkedIn: Sustainability & Green ICT CW 20/ 21

Transcript

Best of LinkedIn: Sustainability & Green ICT CW 20/ 21 · Best of LinkedIn: Digital Transformation & Tech

• 0:00
  Speaker 1: This episode is provided by Thomas Allgeier and Frennis, based on the most relevant LinkedIn posts about sustainability in green ICT in CW- TwentyandTwentyOne.
• 0:09
  Speaker 1: Frennes 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 1: You can find more info in the description.
• 0:23
  Speaker 1: So we are thrilled you could join us for this deep dive.
• 0:26
  Speaker 1: today We're really setting our sights on a super specific rapidly evolving landscape Which is the sustainability and green ICT trends seen across LinkedIn over the past couple of weeks.
• 0:36
  Speaker 0: Yeah, thanks for having me And for everyone listening, you know, we're gonna cover a lot of ground.
• 0:40
  Speaker 0: Today.
• 0:40
  Speaker 0: were looking at software layer The transparency crisis that's happening?
• 0:45
  Speaker 0: Of course the physical hardware infrastructure That basically defining next decade tech.
• 0:49
  Speaker 1: Right, because I mean if you work in digital transformation or green ict You probably already know the conversation has totally shifted.
• 0:55
  Speaker 1: It's not just about AI hype anymore.
• 0:57
  Speaker 1: We're like slamming headfirst into the physical reality of powering all this stuff.
• 1:02
  Speaker 0: Oh absolutely it's a massive wake-up call.
• 1:04
  Speaker 0: we're trying to figure out how to power This ai boom without you know breaking The electrical grid are literally draining our local aquifers
• 1:11
  Speaker 1: exactly.
• 1:13
  Speaker 1: and um I want to start with the software layer actually because usually when we hear green tech your brain just immediately goes to, i don't know wind turbines right?
• 1:21
  Speaker 1: or massive solar farms.
• 1:22
  Speaker 0: Right?
• 1:22
  Speaker 0: The physical stuff?
• 1:23
  Speaker 1: yeah the physical stuff but the problem actually starts way before the power ever reaches the server.
• 1:28
  Speaker 1: it starts in the code.
• 1:30
  Speaker 1: there was this fascinating audit shared by Julie Schiller and she pointed out that having a verified green host for your digital product is well It's really just the baseline.
• 1:40
  Speaker 0: Right,
• 1:41
  Speaker 1: because according to her findings a massive forty two percent of digital emissions actually sit in bloated pages like unoptimized assets heavy user interfaces third-party scripts.
• 1:53
  Speaker 1: I mean what if i told you that at home page sitting on a verified green server could weigh one hundred and forty megabytes?
• 1:59
  Speaker 0: One Hundred and Forty megabytes for our homepage is just wild
• 2:02
  Speaker 1: right.
• 2:03
  Speaker 1: it's quietly burning energy tying up network bandwidth drawing power, and giving you absolutely zero additional value as a user.
• 2:11
  Speaker 0: Which basically means no hosting certificate in the world can fix fundamentally bad code.
• 2:18
  Speaker 0: And what's really fascinating here is how that directly applies to the current AI boom we're in.
• 2:23
  Speaker 0: Oh
• 2:24
  Speaker 1: totally!
• 2:24
  Speaker 0: Navin Blani laid out this framework for what he calls green efficient AI...and he argues something profoundly simple which I love.
• 2:32
  Speaker 0: He says lowest carbon inference is the one this system never actually makes.
• 2:38
  Speaker 1: Which makes perfect sense because every single time you prompt an AI, a physical processor somewhere has to spin up and do the math right?
• 2:45
  Speaker 1: The energy draw isn't abstract it's immediate.
• 2:48
  Speaker 0: Exactly so.
• 2:48
  Speaker 0: Boulogne's approach was built around four steps.
• 2:50
  Speaker 0: To dupe reuse carry in catch.
• 2:53
  Speaker 0: Before an application even sends your query into some massive large language model the architecture should be asking hey have we answered this exact question forty thousand times already?
• 3:01
  Speaker 0: Can I just
• 3:02
  Speaker 1: pull from a cache?
• 3:03
  Speaker 0: Right, can we just reuse that answer from a local cache instead of forcing a massive neural network to regenerate it completely from scratch?
• 3:11
  Speaker 0: And Sushu Kumar actually brought up the similar insight on LinkedIn.
• 3:14
  Speaker 0: He noted that the winners of this new AI era they won't be the companies who build biggest models.
• 3:20
  Speaker 1: Oh interesting.
• 3:21
  Speaker 1: so Who will It Be ?
• 3:23
  Speaker 0: The ones who built most efficient ones.
• 3:25
  Speaker 0: The company prioritizing smaller task optimized models over brute force compute.
• 3:31
  Speaker 1: I mean, that's like using a commercial jet engine to power golf cart.
• 3:36
  Speaker 1: Using a trillion parameter general purpose model just to generate simple internal summary or correct five lines of Python.
• 3:44
  Speaker 1: it gets you down the fairway sure But the waste is just astronomical.
• 3:48
  Speaker 0: That's a perfect analogy, and that wastes exactly what driving this sudden massive shift toward SLMs or small language models.
• 3:56
  Speaker 0: Dr Daniel Jeffrey Koch shared a recent UNESCO report which actually puts hard numbers to architectural shifts.
• 4:02
  Speaker 1: What kind of numbers are we talking about?
• 4:03
  Speaker 0: Well they found through techniques like model quantization combined with prompt optimization.
• 4:09
  Speaker 0: using those smaller models you can cut generative AI energy use by a staggering ninety percent.
• 4:15
  Speaker 1: Wait, ninety percent just from optimizing the model and prompt?
• 4:18
  Speaker 0: Ninety percent.
• 4:19
  Speaker 0: It's not a marginal gain at all.
• 4:21
  Speaker 1: That's incredible.
• 4:21
  Speaker 1: And you know let's break down quantization for a second.
• 4:24
  Speaker 1: For you guys listening because that term gets thrown around A lot.
• 4:26
  Speaker 1: Think of it like calculating a restaurant tip.
• 4:29
  Speaker 1: You don't need a supercomputer to calculate twenty point zero Zero zero one percent to the tenth decimal place just to figure out what to leave.
• 4:39
  Speaker 1: The way to rate you Just round twenty percent right
• 4:41
  Speaker 0: even simple.
• 4:42
  Speaker 1: yeah, and AI is basically doing the exact same thing during quantization.
• 4:46
  Speaker 1: It's dropping unnecessary Decimal places in its mathematical wheats To save massive amounts of memory compute power.
• 4:53
  Speaker 0: but it doesn't
• 4:53
  Speaker 1: really lose Its core reasoning ability.
• 4:55
  Speaker 0: it Is entirely transformative?
• 4:57
  Speaker 0: And we are seeing this echoed by professionals like James Martin and Aiden mere Muhammad II.
• 5:03
  Speaker 0: They're heavily advocating for what they call frugal AI and in-device AI.
• 5:08
  Speaker 1: In device, meaning running it locally?
• 5:10
  Speaker 0: Exactly.
• 5:11
  Speaker 0: Yeah.
• 5:11
  Speaker 0: Mohamedi points out that running these smaller quantized models locally on your own machine completely cuts off the data center impact.
• 5:19
  Speaker 0: I mean The data never travels over the network so that lowers costs.
• 5:23
  Speaker 0: It eliminates transmission latency And as a bonus...it keeps you proprietary data completely private.
• 5:29
  Speaker 1: And it's crucial to remember that this isn't just an AI conversation either.
• 5:33
  Speaker 1: I mean, this applies to general software engineering too.
• 5:35
  Speaker 1: Yeah Victoria Baba Eva and Wilco Bergraf both posted arguing that green software engineering is shifting from this like niche nice-to-have idea into an absolute enterprise baseline.
• 5:47
  Speaker 1: right bergraf specifically talks about expanding door metrics.
• 5:50
  Speaker 1: two includes sustainability.
• 5:52
  Speaker 0: yeah for the developers listening you know door metrics traditionally measure things like deployment frequency in lead time four changes.
• 5:58
  Speaker 0: they measured speed and stability.
• 6:00
  Speaker 1: Right, but they traditionally completely ignore the consequences of that speed.
• 6:04
  Speaker 1: so Burrgraph suggests we now have to measure the energy use water consumption The digital waste tied to those fast deployments.
• 6:12
  Speaker 1: like if your CICD pipeline pushes an update in record time But that updates spikes the applications CPU cycles by thirty percent across a million devices.
• 6:22
  Speaker 1: Well under a green ops mindset That is a failed deployment.
• 6:26
  Speaker 0: Yeah, it's a totally different way to look at success.
• 6:28
  Speaker 0: And that operational shift is exactly why major partnerships are forming to track this stuff in real time.
• 6:35
  Speaker 0: Ashi's Dash highlighted a recent collaboration between Infosys and Soft.
• 6:39
  Speaker 0: They literally partnered to embed green IT measurement directly into enterprise operating models.
• 6:45
  Speaker 1: Wow, so no more guessing?
• 6:46
  Speaker 1: Exactly!
• 6:47
  Speaker 0: By replacing estimated carbon emissions with actual real-time usage data from the servers they're seeing up to forty seven percent CO² reductions just through cloud optimization in right sizing workloads.
• 6:57
  Speaker 0: They are proving that co-deficiency is massive financial win not just an environmental one.
• 7:02
  Speaker 1: Okay but wait Let me play devil's advocate here for a second.
• 7:06
  Speaker 1: if we optimize the software that heavily Don't you run into a pretty dangerous trap like if we make the code that efficient aren't developers and companies just going to use way more of it?
• 7:16
  Speaker 0: You're hitting right on the Givens paradox,
• 7:17
  Speaker 1: right?
• 7:18
  Speaker 1: The chevons paradox where making a resource more efficiently used actually Increases the total demand for rather than decreasing.
• 7:25
  Speaker 1: It like highway expansion Adding three lanes to a congested highway doesn't usually solve the traffic.
• 7:31
  Speaker 1: It just encourages more people to drive, and year later all five lanes are completely
• 7:36
  Speaker 1: gridlocked.".
• 7:36
  Speaker 0: Exactly!
• 7:37
  Speaker 0: And Eger Treschoff actually shared a recap recently from a digital infrastructure event in The Netherlands that perfectly illustrates this... He noted even though society as whole —and server hardware specifically— is becoming incredibly energy efficient Dutch data center electricity consumption has quadrupled since Quadrupled.
• 7:58
  Speaker 1: Wow!
• 7:59
  Speaker 1: So if we make AI cheaper and faster to run by using these highly efficient small language models, developers might just run ten times as many queries which effectively wipes out all of our green software gains
• 8:11
  Speaker 0: Precisely.
• 8:12
  Speaker 0: Which brings us why measurement and transparency are suddenly the biggest battlegrounds in tech sector right now.
• 8:18
  Speaker 0: The Jevons Paradox thrives in dark You know.
• 8:22
  Speaker 0: If you don't know true cost for traffic We can't manage lanes.
• 8:25
  Speaker 0: That
• 8:25
  Speaker 1: makes a lot sense.
• 8:26
  Speaker 0: And to understand how wild this transparency issue really is, we have a look at what Mathieu François presented recently at CERN.
• 8:34
  Speaker 0: He introduced something called the one token model which attempts to measure the actual physical impact of AI
• 8:42
  Speaker 1: inference."
• 8:50
  Speaker 0: Exactly, the everyday use.
• 8:52
  Speaker 0: So Francois exposed this terrifying reality about how cloud architecture fundamentally abstracts physical reality away from the user.
• 9:01
  Speaker 0: He demonstrated that the exact same AI query running The exact same model producing the exact Same output Generates three point six grams of co two equivalent if the Cloud provider processes it in France okay?
• 9:13
  Speaker 0: Three point six Grams.
• 9:14
  Speaker 0: but If the cloud's load balancer routes that exact same query to a server in India, it generates one hundred and eleven grams of CO₂.
• 9:22
  Speaker 0: Wait!
• 9:22
  Speaker 1: One hundred and Eleven?
• 9:24
  Speaker 1: That is a one-to-thirty ratio for the exact same digital action.
• 9:27
  Speaker 1: Yep
• 9:27
  Speaker 0: based entirely on invisible geographic routing because when you hit enter on an AI prompt The Cloud Providers API Is basically searching for the fastest or cheapest available compute node globally at specific millisecond.
• 9:39
  Speaker 0: You as user have zero visibility into where this math actually happening.
• 9:43
  Speaker 1: Right, and because the power grid in France leans so heavily on low-carbon nuclear while the grid in India might lean heavier on coal at that specific hour your digital carbon footprint just swings wildly.
• 9:56
  Speaker 1: And I'm guessing the water consumed by the power plants to run those specific GPUs?
• 10:00
  Speaker 1: That simply doesn't appear in the standard API response or your company's sustainability report?
• 10:05
  Speaker 0: Not at all.
• 10:06
  Speaker 0: it is completely hidden.
• 10:07
  Speaker 1: And you know, that lack of transparency is everywhere right now especially when we look at the physical infrastructure housing these massive foundation models.
• 10:16
  Speaker 1: Boris Gamazaychikov and Asim Hussein actually recently analyzed anthropics announcement That their AI model Claude will be running across Elon Musk's Colossus data center campuses.
• 10:26
  Speaker 0: Yeah, I saw that analysis.
• 10:28
  Speaker 1: And looking strictly at the data shared in their posts Hussain did the math and estimates.
• 10:32
  Speaker 1: this specific infrastructure deal raises Claude's carbon intensity per kilowatt hour by about fifteen percent.
• 10:39
  Speaker 0: The underlying driver of that fifteen-percent increase really comes down to those specific energy sources being utilized just to get them online so quickly.
• 10:48
  Speaker 1: Exactly!
• 10:49
  Speaker 1: Gamazaitchikov pointed out that first colossus site utilised gas turbines situated behind.
• 10:55
  Speaker 1: The urgency for power was just so high that they brought in physical gas turbines rather than waiting for standard grid interconnections.
• 11:04
  Speaker 1: And the sources note these turbines have reportedly operated a way to impact local air quality.
• 11:10
  Speaker 0: Right, the localized impacts.
• 11:11
  Speaker 1: Yeah potentially becoming massive source of not emissions and Memphis which is already struggling with high asthma rates.
• 11:18
  Speaker 1: but what's incredibly frustrating from reporting standpoint?
• 11:21
  Speaker 1: how accounting rules actually handle this?
• 11:24
  Speaker 0: Right.
• 11:24
  Speaker 0: You're referring to the scope emissions classifications?
• 11:27
  Speaker 1: Yes, because despite upcoming California laws requiring companies to disclose their direct-scope one and indirect-scoped two emissions, Gama Zhechkov's Suspects Anthropic might legally classify these specific data center emissions as scope three since anthropic doesn't own the building or the turbines they just rent the compute.
• 11:45
  Speaker 1: They can claim that they don't have operational control.
• 11:48
  Speaker 0: The classic loophole
• 11:49
  Speaker 1: Exactly.
• 11:50
  Speaker 1: That means the vast majority of their models' actual physical footprint could remain buried deep in their supply chain reporting, basically hidden from immediate public
• 12:00
  Speaker 1: scrutiny.".
• 12:01
  Speaker 0: And that lack of holistic end-to-end accountability is exactly why we saw Anna Lerner Nezbitt publicly push back on some of the sweeping claims coming out of NVIDIA recently….
• 12:12
  Speaker 0: An executive there stated there's a growing consensus that AI climate benefits outweigh its costs... pointing to things like, oh AI optimizing power grids or discovering new battery materials.
• 12:23
  Speaker 0: Sure
• 12:23
  Speaker 1: the utopian view
• 12:24
  Speaker 0: right but Nezbit pointed out that The actual scientific research literature does not support a blanket statement Like That at all.
• 12:31
  Speaker 1: she cited the work of researchers like Dr.
• 12:33
  Speaker 1: Sasha Lucione.
• 12:34
  Speaker 1: Right who has been pioneering?
• 12:35
  Speaker 1: The quantification of ai's environmental impact for years?
• 12:39
  Speaker 1: and the conclusion isn't that Ai is inherently bad.
• 12:41
  Speaker 1: It's just that the benefits are highly highly conditional.
• 12:44
  Speaker 0: exactly Dr.
• 12:45
  Speaker 0: Lucioni's work proves that AI can absolutely be a net positive for the climate, but only if very specific stringent conditions are met.
• 12:55
  Speaker 0: You need transparent full lifecycle emissions reporting and more importantly you need to power these sprawling data centers on genuinely additional renewable energy.
• 13:05
  Speaker 1: right not just buying paper offsets.
• 13:07
  Speaker 0: Exactly, physically funding and building new wind and solar farms to match your demand.
• 13:12
  Speaker 0: not just buying paper renewable energy certificates but artificially offset dirty power you're pulling from the grid.
• 13:18
  Speaker 0: when companies make these net positive claims without proving those conditions it basically shuts down critical debate rather than contributing.
• 13:26
  Speaker 1: But it seems like regulation is finally attempting to force that transparency, right?
• 13:30
  Speaker 1: Pulling these invisible costs out of the shadows.
• 13:33
  Speaker 1: Because Boris Dundasov posted about EU AI Act and how its beginning It really is.
• 13:42
  Speaker 1: it forces transparent green AI reporting, meaning companies will actually have to measure and disclose the carbon footprint in raw processing power used by their models.
• 13:51
  Speaker 1: And over In The
• 13:51
  Speaker 1: U.S.,
• 13:52
  Speaker 1: Nolan Goddard successfully campaigned To get specific language regarding water and carbon costs of AI inserted directly into California's Generative AI Executive Order.
• 14:01
  Speaker 0: And you know...the inclusion Of Water in that executive order Is a vital pivot Because this leads us Into the final and perhaps most immovable layer This whole equation.
• 14:10
  Speaker 0: If software optimization is layer one and transparent reporting as layer two, we ultimately slam into the physical constraints of the planet.
• 14:18
  Speaker 0: Hardware?
• 14:19
  Speaker 0: Exactly!
• 14:20
  Speaker 0: The AI boom requires hardware...and hardware requires massive amounts of physical space power & water just to keep that silicon from melting.
• 14:28
  Speaker 1: you cannot run a virtual machine without a very real, physical and hot
• 14:32
  Speaker 0: server.
• 14:33
  Speaker 0: And cooling those servers brings us to pretty sobering reality highlighted by both Danielle Lemon and Connor Love.
• 14:40
  Speaker 0: Data centers don't actually need to consume massive amounts of water!
• 14:43
  Speaker 0: No
• 14:43
  Speaker 1: they do not as love points out because the tech industry historically just chose the absolute cheapest laziest cooling designs available and aggressively scaled them across a trillion dollar industry.
• 14:55
  Speaker 0: Yeah The Path Of Lease Resistance for maximizing profit margins was evaporative cooling.
• 15:00
  Speaker 0: You just run massive amounts of local water over cooling pads, fans blow the server heat through the pads The Water evaporates and the servers stay cool.
• 15:08
  Speaker 0: It is super cheap to build but it consumes water relentlessly.
• 15:12
  Speaker 0: But love highlights that completely waterless cooling exists today Immersion Cooling where the servers are literally physically submerged in a specially engineered non-conductive liquid That absorbs the heat without evaporating.
• 15:24
  Speaker 0: that exist right now.
• 15:26
  Speaker 1: But hold on, if emerging cooling completely solves the water evaporation issue why isn't every major hyperscaler just doing that?
• 15:34
  Speaker 1: Because for example love-sighted Metta's upcoming Louisiana data center which will reportedly dry as much water daily at seventeen thousand people pulling directly from same local aquifer families and farmers rely.
• 15:47
  Speaker 1: If technology is solved Why build a water guzzling facility in twenty twenty six?
• 15:53
  Speaker 1: Is it just too expensive to retrofit or deploy?
• 15:56
  Speaker 0: Well, yeah.
• 15:57
  Speaker 0: It requires a fundamental redesign of the server architecture and building itself which demands massive up-front capital.
• 16:03
  Speaker 0: However...it is absolutely happening at scale in regions where local policy and market incentives actively force innovation.
• 16:10
  Speaker 1: Oh really?
• 16:11
  Speaker 0: Like why?
• 16:11
  Speaker 0: We'll take The Nordics for example.
• 16:12
  Speaker 0: John McDonald highlighted a company called T loop entering the Market in Finland And they aren't just building a traditional data center.
• 16:18
  Speaker 0: They're building what they categorize as A Data Energy Center.
• 16:21
  Speaker 1: Okay, so how does the data energy center differ from What we are building here in
• 16:25
  Speaker 1: U.S.?
• 16:26
  Speaker 0: It's designed for ground up As a circular asset For surrounding community.
• 16:30
  Speaker 0: So instead of venting The massive amounts Of heat generated by GPUs Uselessly into atmosphere They capture that waste heat In water loops and feed it directly Into local district heating grid.
• 16:41
  Speaker 0: Wow The byproduct of training an AI model actually warms the homes and businesses in this surrounding city.
• 16:48
  Speaker 0: It turns a massive parasitic energy consumer into a stabilizing asset for local
• 16:53
  Speaker 0: grid.".
• 17:05
  Speaker 1: This is a massive facility.
• 17:06
  Speaker 1: It's designed to scale up to six hundred megawatts of IT capacity,
• 17:11
  Speaker 0: that's huge!
• 17:12
  Speaker 1: it's basically the power draw of a medium-sized city.
• 17:15
  Speaker 1: but their approach to powering isn't just plugging into the Spanish grid.
• 17:19
  Speaker 1: its an entirely hybrid ecosystem.
• 17:21
  Speaker 1: they are combining on site solar generation massive industrial battery storage systems eco LNG and low carbon hydrogen.
• 17:30
  Speaker 1: It's a localized bespoke energy mix designed specifically to support high-performance AI workloads without crashing the regional power infrastructure.
• 17:39
  Speaker 0: And that Spanish facility perfectly validates the argument Carl Raeben and Anita Falls have been making recently.
• 17:45
  Speaker 0: Power availability in sustainable design are no longer just like ESG checkboxes That our company fills out for a quarterly sustainability report,
• 17:53
  Speaker 1: right?
• 17:53
  Speaker 1: it's existential.
• 17:54
  Speaker 0: Exactly they're the ultimate competitive advantages of technology sector.
• 17:58
  Speaker 1: Yeah, Falls made a brilliant point that securing power is no longer just a back-end utility decision handed off to a facility manager.
• 18:06
  Speaker 1: It actively dictates where when and how data centers get built.
• 18:10
  Speaker 1: if you cannot secure the power or If you can't cool The facility efficiently enough to get permits from the local water board You simply cannot scale your AI operations.
• 18:20
  Speaker 1: Your product roadmap literally hits a physical concrete wall
• 18:25
  Speaker 0: And big tech clearly recognizes this physical bottleneck.
• 18:29
  Speaker 0: Nina Benoit reported that the major players, Amazon Google, Meta and Microsoft are all actively packing The Data Center Innovation Initiative or DCII.
• 18:38
  Speaker 1: Oh I've heard of that!
• 18:39
  Speaker 0: Yeah they're aggressively funding startups trying to solve next-generation energy storage advanced electrical distribution systems an industrial scale alternative cooling.
• 18:48
  Speaker 0: They know That To keep their software revenue growing...they have to completely reinvent the physical infrastructure holding it up
• 18:55
  Speaker 1: Because the digital world is entirely reliant on the physical
• 18:58
  Speaker 0: Precisely.
• 18:58
  Speaker 0: Yeah, and you know if we pull all these threads together from across the last two weeks From Julie Schiller's bloated hundred forty megabyte code to The invisible geographic routing of the one token model All the way down into physical constraints Of water and power in Louisiana In Spain It leads To a fascinating paradigm shift.
• 19:17
  Speaker 0: And how?
• 19:18
  Speaker 0: We really need to think about building technology
• 19:20
  Speaker 1: Definitely
• 19:21
  Speaker 0: for everyone listening especially those view architecting solutions in the ICT space.
• 19:26
  Speaker 0: i want to leave You with this final thought Malovar.
• 19:28
  Speaker 0: What if the ultimate metric for a successful, sustainable digital transformation isn't actually how efficiently your AI runs?
• 19:35
  Speaker 0: What If The Highest Level Of Mastery Is How Elegantly Your System Architecture Avoids Using AI Altogether When It Isn't Strictly Necessary.
• 19:46
  Speaker 0: valuing the void of compute,
• 19:50
  Speaker 1: that is a great thought to end on.
• 19:51
  Speaker 1: If you enjoyed this episode new episodes drop every two weeks.
• 19:54
  Speaker 1: also check out our other editions on cloud digital products and services artificial intelligence an ICT in tech insights health tech defense Tech.
• 20:02
  Speaker 1: thank You so much for joining us on this deep dive.
• 20:04
  Speaker 1: don't forget to subscribe And we'll catch you next time.