Best of LinkedIn: Cloud Insights CW 07/ 08

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about cloud in calendar weeks seven and eight.

00:00:08: Frenness enables enterprises with market technology and competitive intelligence for portfolio and strategy development.

00:00:14: Welcome back to The Deep Dive!

00:00:16: We are looking at the Cloud Conversation from Calendar Week Seven & Eight of twenty-twenty six And I have to say that tone has shifted pretty significantly.

00:00:27: It's less, I don't know breathless.

00:00:29: It really has.

00:00:30: yeah usually our feed is just you know wall-to-wall hype about the latest LLM benchmarker or some new consumer aft that generates video on The Fly.

00:00:37: right

00:00:38: but looking at the curated intelligence from Frennus this week it Really feels like the adults have entered the room.

00:00:44: That's a really good way to put it honestly This sort of toy phase is fading out.

00:00:48: We're seeing a pivot toward the hard unglamorous reality Of actually running the stuff at scale.

00:00:53: Yeah we aren't talking about magic anymore.

00:00:55: No, exactly.

00:00:56: We're talking about three very concrete...very difficult things today The evolution of paying for the cloud which is Phenops?

00:01:04: Yeah!

00:01:04: The geopolitical headache of sovereign Cloud and the brutal engineering war For AI infrastructure.

00:01:12: Right.

00:01:12: so the plumbing..the politics And price tag.

00:01:16: That's a summary right there.

00:01:17: Let

00:01:17: start with money because There was a term floating around week seven and eight.

00:01:23: that actually made me sweat when I read it.

00:01:25: the cloud costs outage.

00:01:27: Oh,

00:01:28: it is a terrifying concept isn't it?

00:01:29: Truly This comes from a post by Mohammed Abdullah Abraar and just perfectly captures this new era of risk.

00:01:36: Walk me through this because usually when you hear the word outage You think silence, servers down angry customers on social media, four hundred-four errors everywhere.

00:01:45: Yeah!

00:01:45: And that's almost The good kind disaster.

00:01:47: strangely enough.

00:01:48: Right

00:01:48: Because at least we know what happens

00:01:50: Exactly Your phone blows up page duty goes off but A bra is describing a scenario where your systems are working perfectly.

00:01:56: Everything's humming along?

00:01:58: Yep, the dashboard is all green.

00:02:00: customers are totally happy but under-the-hood maybe there's a misconfiguration or runaway spoke like

00:02:06: a serverless function that scaling infinitely

00:02:08: exactly That and it's just burning through your budget at a rate of thousands of dollars an hour.

00:02:14: So you don't wake up to a pager alert.

00:02:17: You wake up To a bill thats thirty percent Or even three hundred percent higher than it should be.

00:02:22: Without a single warning and unlike the server outage you know, You can't just reboot your bank account.

00:02:28: that money is gone.

00:02:29: The hyperscaler has already built you.

00:02:33: That's why the conversation around finance Is changing so aggressively right now It no longer.

00:02:38: how do we save few bucks on reserved instances?

00:02:41: How to survive our own infrastructure Which ties into this really interesting framework proposed to handle.

00:02:48: Michael Stevenson drew a comparison that I initially thought was huge stretch.

00:02:52: Oh, the sales comparison?

00:02:53: Yeah

00:02:53: But more and more it actually made sense.

00:02:56: He compared Phenops to B-to-B Sales.

00:02:58: It sounds

00:02:58: so counterintuitive right because engineers historically hate sales.

00:03:01: Oh absolutely.

00:03:02: but Stevenson's point is brilliant.

00:03:04: he brings up Mindy DPP

00:03:07: which is a classic, somewhat complex sales qualification methodology.

00:03:12: A real acronym soup!

00:03:14: It stands for metrics economic buyer decision criteria and so on.

00:03:18: Sales people use it to ensure a deal actually closes right?

00:03:21: Sure.

00:03:22: So Stevenson argues that phenops practitioners are essentially doing the exact same job just in reverse.

00:03:27: How so

00:03:28: Well?

00:03:29: sales proves value before the check is written.

00:03:31: Phenops has to prove value after the purchase doesn't get cut internally.

00:03:36: Okay, so if you are listening to this and you're a cloud engineer You need to identify your economic buyer internally?

00:03:42: You have to!

00:03:43: If you were running a massive Kubernetes cluster who is actually paying for it Is at marketing...is that the product team right..if you can't Identify that economic buyer And you can show them metrics The M in MetiPP That prove return on investment YOU ARE IN THE DANGER ZONE.

00:03:59: It

00:03:59: essentially turns the engineer into an internal lobbyist For their own infrastructure.

00:04:03: That's exactly it.

00:04:05: It forces them to be business literate, you can't just go to leadership and say I need this new flashy tool because its cool and scales better.

00:04:10: You

00:04:10: have to say...I NEED THIS TOOL BECAUSE IT REDUCES OUR COST OF GOOD SOLD BY FOUR PERCENT.

00:04:15: Yes!

00:04:16: Thats the MDDP peak mindset applied to cloud

00:04:19: costs And this dovetails perfectly with what Nicholas Fandrini was arguing in his post.

00:04:25: He took a step further suggesting that Phenombs needs to basically merge with product management.

00:04:30: This is the value governance piece.

00:04:32: And honestly, this is the missing link for most enterprises right now.

00:04:36: because

00:04:36: Right?

00:04:36: Now Most dadboards just show infrastructure cost rate like hey we spent fifty thousand dollars on AWS this month which

00:04:43: Is a completely useless number in isolation.

00:04:46: right it's fifty thousand good is It bad?

00:04:49: did We make a million Dollars from that spend or Did we Make zero?

00:04:52: Precisely saundrini argues you need to track Cost per unit of Value.

00:04:57: okay break That down.

00:04:58: For me

00:04:58: it means asking What is the cloud cost per transaction?

00:05:02: Or what is the Cloud Cost Per User Journey.

00:05:05: That seems incredibly hard to instrument though, I mean if you have a complex microservices architecture.

00:05:11: how do you know that this specific database call belongs to that specific user onboarding flow?

00:05:18: It's an engineering challenge in itself but it's necessary because If You Know That Onboarding A Free Tier User Costs You Two Dollars In Compute

00:05:26: and you only ever monetize them at a dollar fifty.

00:05:29: Exactly, then you don't have an engineering problem.

00:05:31: You have business model failure And no amount of code optimization or right sizing is gonna fix that.

00:05:37: That's why product management has to own the unit economics alongside engineering.

00:05:43: We actually saw some tools popping up in the feed to help with this, which is great because it wasn't all high-level theory.

00:05:48: Robert Lobb from Avidod was showing off their Cloud Impact Analyzee tool for Azure...

00:05:53: Which is fantastic!

00:05:54: For that broad enterprise level transparency you need those heavy lifters.

00:05:58: But

00:05:58: I admit i have a soft spot for The Builder mentality we saw from Andrea Zemaitis.

00:06:02: Oh

00:06:03: the DIY serverless monitor?

00:06:05: I loved that post

00:06:06: Right.

00:06:07: It's so pragmatic.

00:06:08: he built a custom tool pings his slack every morning.

00:06:12: Like, here are your running EC two instances?

00:06:15: Here is your S three bucket count.

00:06:17: it's simple It's aggressive and puts the data right in your face before you've even had your coffee.

00:06:22: sometimes You don't need a massive fancy dashboard You just need to bot yelling at you that you left a server running over the weekend

00:06:29: totally.

00:06:30: And To be fair The cloud providers were trying to help with this internally too.

00:06:34: Victor Garcia pointed out A new feature In azure kubernetes service called end of idle nodes.

00:06:40: What does that do?

00:06:41: It

00:06:41: automatically kills infrastructure.

00:06:43: That isn't doing anything.

00:06:44: it literally scales down idle nodes to zero, its this push toward becoming efficient by default which

00:06:50: sounds great until you introduce AI into the mix because Sir Krishna Kotha had a massive warning label for everyone in week seven and eight regarding fine ops.

00:07:00: yeah they AI cost factor.

00:07:01: he says ai breaks all of old rules of costs predictability.

00:07:04: He's absolutely

00:07:05: right.

00:07:05: think about traditional software.

00:07:06: Its deterministic.

00:07:07: You write code You know roughly what it costs per execution.

00:07:12: Yeah, AI is probabilistic and continuous And more importantly It is often demand driven in a way that scales linearly with complexity.

00:07:20: So if I integrate an LLM into my customer support chat?

00:07:24: Suddenly everyone decides to ask it highly complex multi-part questions about the meaning of life.

00:07:30: instead Of just checking their order status your bill

00:07:32: explodes instantly.

00:07:34: And unlike a standard database query which takes milliseconds, an LLM inference takes actual seconds of high-end GPU time.

00:07:42: Wow!

00:07:43: Kotha's point is that this isn't discretionary spend anymore.

00:07:46: Once it's live the users control the throttle.

00:07:49: It requires a level of financial discipline That most IT departments simply do not have yet.

00:07:53: Which brings us to the second massive headache we tracked across these two weeks.

00:07:57: Even if you had money for all those compute Where are actually allowed to run?

00:08:01: Ahh...

00:08:01: The sovereign cloud debate in Europe.

00:08:03: It is getting heated.

00:08:04: it's becoming a bit of a culture war, honestly You have this tension between the need for American innovation The hyper scalers like AWS and Microsoft And the European imperative For digital independence.

00:08:16: Heated Is putting it mildly.

00:08:19: Benjamin Herman had a post that essentially told everyone to take A collective deep breath.

00:08:24: he called it a binary Screaming match.

00:08:27: Any spot on?

00:08:30: total US dependency versus hermit kingdom Europe.

00:08:34: Right, and Herman argues that reality is shades of gray.

00:08:38: you can have sovereignty without totally disconnecting from the global internet.

00:08:42: You don't need to build an airtight firewall around the continent To be safe.

00:08:46: But how do we actually measure safe?

00:08:48: Yeah.

00:08:49: Because you and Peterite released some data that I found fascinating, his sovereign cloud compass.

00:08:54: This was a stand-up piece of analysis for me.

00:08:56: Peterite didn't just opine on the philosophy of it.

00:08:59: he brought hard data.

00:09:00: yeah He closed sixty eight data gaps in this compass model

00:09:03: And The results were surprising.

00:09:05: European providers like Stackit and OVH Cloud hit hundred percent On His sovereignty criteria.

00:09:11: i mean i'd expect That from the locals but tCloud public which is dorsche telecom also improves significantly.

00:09:17: What's holding the other providers back from hitting that hundred percent?

00:09:20: Documentation Really,

00:09:21: just paperwork?

00:09:22: Essentially yes Peter writes.

00:09:24: big takeaway was that barrier isn't necessarily technology anymore it is transparency.

00:09:29: The vendors simply aren't documenting their data chains clearly enough to definitively prove sovereignty in an auditor.

00:09:36: So tech itself might actually be compliant but they can't proof on paper

00:09:40: Which makes useless for a regulated industry like banking or healthcare.

00:09:45: But the Americans are definitely listening to this feedback.

00:09:48: We saw that interview with Rodrigo Gattini from AWS, which is shared by Dennis Gorda

00:09:52: right?

00:09:53: He was talking about The AWS European sovereign cloud and I noticed Vettini use a very specific phrase.

00:09:58: he called it sovereign buy design.

00:10:00: What does that mean in practice versus just saying hey we have a data center in Frankfurt.

00:10:05: It's a very bold architectural claim from AWS, it means strict physical and logical separation.

00:10:11: The claim is that this European cloud could continue to operate completely autonomously even if Europe were physically cut off from the global network

00:10:19: like a survival bunker for data.

00:10:20: effectively yes they are trying to assure european governments that Even If A major geopolitical crisis severs the transatlantic cables or if US Cloud Laws Suddenly Change the European infrastructure keeps humming along.

00:10:34: But then you have Bozena Slaminsky, who argues even that isn't enough.

00:10:39: She introduced a concept I think is going to stick around for a long time – the thermodynamic tax!

00:10:45: I loved this post.

00:10:46: it sounds like science fiction but its pure physics.

00:10:48: Walk us through it….

00:10:49: She argues compute sovereignty isn't just about software or where hard drives sit physically.

00:10:54: It's all about hardware and energy itself.

00:10:57: Right.

00:10:57: so if your rinting chips from an American company You aren't truly sovereign

00:11:01: Exactly.

00:11:02: And if you are importing the energy to run those massive data centers, You aren't sovereign either!

00:11:08: The thermodynamic tax is the inherent cost of relying on foreign infrastructure.

00:11:13: Slamansky's point is that If YOU don't own the atoms —the actual silicon and electrons powering them—you're just a tenant no matter what the SLA contract says.

00:11:23: That focus on owning the atoms transitions us perfectly into our third theme from our sources... ...The AI Infrastructure War Because right now, the biggest companies on earth are fighting a brutal war over who owns the silicon.

00:11:37: And their strategies are completely diverging.

00:11:39: we saw really clever take from Sandy Carter regarding Google.

00:11:43: she called them The Switzerland of AI.

00:11:45: I like that analogy neutral ground.

00:11:47: it

00:11:47: fits perfectly because while Microsoft is basically married to open AI Google's playing the field with their vertex AI platform.

00:11:55: they're hosting everyone Anthropic Meta Lama Mistral come one Come all.

00:12:00: So Google's play is basically, we don't know which foundational model will win this race and frankly We Don't Care as long you pay us to run the compute.

00:12:07: Exactly!

00:12:08: They are selling the shovels in The Gold Rush.

00:12:10: It's a pure volume play.

00:12:12: but then You look at Oracle And the story gets even more interesting.

00:12:15: Yeah Jason Langley shared A deep dive referencing Sanjay Basu.

00:12:18: That really challenged my view of oracle because Oracles Is usually seen As the legacy player right?

00:12:24: The database folks?

00:12:25: they were very late To the cloud party

00:12:27: Very late, a decade behind AWS.

00:12:30: But Langley argues that was actually a latecomer advantage.

00:12:34: How does being late help in tech?

00:12:36: Because they missed the entire first wave of cloud architecture ten years ago.

00:12:41: They didn't make the same foundational mistakes that AWS and Azure made.

00:12:45: They built their Gen-II Cloud specifically for this moment from the ground up.

00:12:50: What's actual technical difference though?

00:12:52: It is called off box virtualization.

00:12:54: Okay, let's break that down for the non-engineers listening.

00:12:56: What does it mean?

00:12:57: Sure

00:12:58: in a traditional cloud server like the early EC.

00:13:01: two instances on AWS The software that manages the network and security at the hypervisor runs On the exact same main CPU as your application.

00:13:10: they share resources.

00:13:11: So if the management software gets busy doing background tasks my app slows Down

00:13:15: exactly.

00:13:16: we call it noisy neighbors or jitter.

00:13:19: For a standard web server hosting a blog, that's fine.

00:13:21: You don't notice.

00:13:22: but for training a massive AI model... It

00:13:24: is a disaster!

00:13:25: ...complete disaster.

00:13:26: you need consistent maximum throughput with zero interruptions.

00:13:30: An

00:13:30: oracle does it differently?

00:13:31: Yes they moved all of the management software off the main board and onto separate physical card A SmartNIC.

00:13:39: So the main CCU is one hundred percent yours.

00:13:42: The network is non-oversubscribed.

00:13:44: That's why you see heavyweights like NVIDIA and Cohere using Oracle infrastructure right now, they aren't trying to host a website.

00:13:51: They're trying to build a supercomputer oracle accidentally built the perfect AI cloud just by being late.

00:13:58: that

00:13:59: Is fascinating.

00:13:59: so this so called legacy player?

00:14:01: It actually the most modern in terms of bare metal performance for AI.

00:14:05: it's quite the plot twist

00:14:06: But not everyone has going into the public Cloud for this.

00:14:08: are they?

00:14:09: because Michael McGrory From Rakuten shared some insights from the telco world that point in the exact opposite direction.

00:14:16: Right,

00:14:16: The Telcos are completely bucking trend.

00:14:18: They're moving their AI to private cloud

00:14:22: Which feels like a huge step backward.

00:14:23: to what?

00:14:25: Why are they doing this?

00:14:25: Two main reasons Cost certainty and latency.

00:14:28: We talked about the cloud cost outage earlier.

00:14:30: Yeah, the unpredictable bills.

00:14:32: Telcos operate on extremely thin margins.

00:14:34: They literally cannot afford variable explosive bills from hyperscalers.

00:14:40: So they buy the hardware outright, but also it's just physics

00:14:43: now.

00:14:43: so

00:14:44: if you are running a five G network orchestrating autonomous vehicles or real time factory robotics You can't afford The twenty millisecond delay of sending data out to a public cloud region and back right?

00:14:57: They need the AI inference to happen Right there on premise in what they call that AI Factory.

00:15:03: This leads us perfectly into our final theme from week seven and eight, architecture.

00:15:07: And modernization.

00:15:08: it really feels like the dogma of cloud.

00:15:11: first this idea that you should move everything to the public Cloud no matter what is officially dead.

00:15:15: Oh

00:15:16: wait he's completely dead.

00:15:17: It's being replaced by concepts like cloud smart or workload right?

00:15:20: Mimish Gupta laid out a specific design in his post That has becoming The gold standard for high resilience

00:15:25: one plus two model Right One Public Cloud Plus Two private data centers.

00:15:29: That's

00:15:29: the one.

00:15:30: how does that split actually work in practice?

00:15:32: You use the public cloud for bursting so when your traffic goes crazy during a Super Bowl ad or Black Friday and for disaster recovery but you're highly sensitive workloads, the steady state stuff like payment processing.

00:15:45: that runs twenty four seven

00:15:47: Like Visa

00:15:48: Exactly!

00:15:49: That lives into two private datacenters.

00:15:51: And he mentioned this concept of client DC affinity.

00:15:54: Yeah thats'a fancy engineering way of saying If a user is logging in from India, pin their session to the India data center.

00:16:02: don't bounce their traffic around the globe to U.S.

00:16:04: East One, it reduces latency and makes the overall system much more

00:16:08: stable.".

00:16:08: It's just

00:16:09: solid old-school engineering making a comeback?

00:16:11: It

00:16:11: really is!

00:16:12: Speaking of Old School Engineering I have bring up this story from Leftris, Georgia.

00:16:16: This was highly entertaining.

00:16:17: He talked about the classic three tier architecture interview question.

00:16:20: Oh...this painfully relatable for anyone who has hired engineers recently.

00:16:24: So he

00:16:24: asked candidates design standard three tier you know, web server application layer database.

00:16:31: And apparently they all fail?

00:16:33: They fail because they just draw a diagram that memorized from an AWS certification course.

00:16:39: Here is the box for the web server.

00:16:40: here's a box of data base

00:16:42: and thats not good enough.

00:16:43: Thats not engineering thats art class.

00:16:45: Wow Brutal.

00:16:47: So what he actually looking for The

00:16:49: why Real engineering is explaining the trade-offs of design.

00:16:54: Why did you choose an application load balancer in ALB instead a network load balancer in NLB?

00:17:00: Because I need layer seven routing for my microservices!

00:17:03: Exactly, if can explain that an ALB looks at actual content to make routing decisions while an NLb just blindly look at IP address... You don't really understand architecture.

00:17:16: We know how to click buttons on cloud console.

00:17:18: Right, we're moving away from these certification collectors to needing people who deeply understand First Principles.

00:17:24: And sometimes those first principles take you all the way back to the oldest iron in the building.

00:17:28: Mary Simmons had a fantastic post about The Mainframe... She

00:17:31: did!

00:17:32: Her quote was legendary.

00:17:34: Physics

00:17:34: doesn't care about your cloud-first strategy.

00:17:36: It's

00:17:36: so true.

00:17:37: gravity is gravity.

00:17:38: she argues that for some massive enterprises banks big insurers It's their sovereign private cloud, which just happens to be a mainframe.

00:17:51: Because of data gravity?

00:17:52: Right!

00:17:53: Think about it.

00:17:53: if you have petabytes of core transaction data sitting securely on the main frame moving all that data across the network into public cloud and run in new AI model is slow its insecure incredibly expensive so

00:18:07: sometimes actually smarter.

00:18:10: bring the A.I..to-the-data

00:18:11: exactly reality check.

00:18:13: Just because the hardware is old doesn't mean architecture was wrong for that specific workload.

00:18:18: So, if we look at week seven and eight of twenty-twenty six as a whole what's through line here?

00:18:23: We have scary fine ops bills European data bunkers Oracle winning AI deals with Smart Nectis.

00:18:29: The

00:18:29: overarching story is maturity.

00:18:31: Toy phases over.

00:18:32: We aren't just playing with shiny new tools anymore.

00:18:34: We are facing hard realities of economics geopolitics And physics.

00:18:38: It really about governance

00:18:40: Moving from height to architectural realism, whether it's realizing that AI costs need a completely new financial model or you can't just wish away data residency laws.

00:18:51: Or sometimes the private data center just flat out beats the public cloud.

00:18:56: The industry is sobering up and honestly thats when real lasting work gets done.

00:19:01: You

00:19:01: know what makes me think about Slominsky thermodynamic tax again?

00:19:05: If sovereign Cloud requires local energy and these AI models demand such massive amounts of power, we might see a future where the real cloud war isn't over silicon at all.

00:19:16: It's an energy grid war like will your cloud provider eventually need to also be your energy provider just to guarantee uptime?

00:19:24: Now that is a terrifying but very realistic thought-to-hand on.

00:19:27: it

00:19:27: has been busy two weeks That's for sure.

00:19:29: absolutely.

00:19:29: if you enjoyed this episode new episodes drop every two weeks.

00:19:32: Also check out our other editions on ICT in tech digital products and services, artificial intelligence sustainability in green ICT defense tech and health.

00:19:40: Thanks for listening.

00:19:41: everyone keep questioning your architecture.

00:19:44: thanks for joining us!

00:19:45: And don't forget to subscribe.

00:19:46: see you on the next deep dive.