Best of LinkedIn: Cloud Insights & Sovereignty CW 37/ 38

00:00:00: Provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about cloud insights in sovereignty.

00:00:06: In calendar weeks thirty seven and thirty eight.

00:00:09: Frenness has built a sovereign cloud market radar for ICT providers with weekly hot news monthly reports quarterly leadership presentations an AI podcast for field teams.

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

00:00:21: imagine waking up tomorrow.

00:00:23: You pour your coffee And you find out that The software running Your entire data center just went Up in price by like over a thousand percent.

00:00:30: Oh

00:00:30: wow, yeah that is terrifying!

00:00:32: Right but that is the exact reality hitting the European cloud market right now and I mean it is forcing.

00:00:40: So today's deep dive is zeroing in on exactly that, looking at the top digital products and services trends across the professional landscape.

00:00:48: We're really focusing on how the definitions of cloud sovereignty cost control and operating models are just being completely rewritten in real time.

00:00:55: Yeah And if you are managing digital infrastructure or navigating tech strategy today The ground under your feet is absolutely shifting.

00:01:04: It really is.

00:01:04: I mean market has matured way past the era migrating workloads and assuming the cloud provider handles the rest.

00:01:11: Right,

00:01:11: exactly!

00:01:12: We are seeing this really harsh reevaluation of what words like control actually mean especially when you start squeezing geopolitical and economic pressures into these technical architectures.

00:01:24: Yeah And we see this most clearly with well arguably The biggest board level concern right now which is digital sovereignty.

00:01:31: Absolutely Because usually When talking about control we think of physical ownership.

00:01:35: You possess the hardware.

00:01:37: It sits in your building, you decide who touches it but the market is quietly redefining sovereignty as something well far more complex than just putting your servers inside a specific geographic

00:01:48: border.".

00:01:49: Right because geography does not equal sovereignty.

00:01:51: Mityil Gavrin recently highlighted this tension.

00:01:55: he was dissecting Broadcom's European sovereign AI claims?

00:01:59: Oh yeah that was a fascinating post

00:02:02: Yeah.

00:02:02: And despite some genuinely strong engineering under the hood, Govren argues these claims fall fundamentally short of true sovereignty.

00:02:09: He makes this critical distinction having governed egress meaning you know You can filter and control that data leaving a system That is not the same as Data Residency Right!

00:02:20: Even more importantly if The Control Plane Is Something Customers Cannot Audit Or Fork The Sovereignty Is Just An Illusion.

00:02:27: Well, and we should probably define what we mean by the control plane in this context just to be clear.

00:02:32: It's

00:02:32: cool.

00:02:32: yeah

00:02:33: We are talking about the overarching management software right?

00:02:35: Yeah The stuff that provisions resources handles authentication orchestrates the environment if you cannot independently verify the code Running that orchestration layer then you really can not verify What is happening to your data

00:02:48: precisely.

00:02:49: You were basically trusting a black box And you know trust me a blackbox becomes incredibly painful when your commercial relationship sours

00:02:56: to say the least.

00:02:57: Gouverne points out that it is really hard to celebrate this version of sovereignty when your licensing costs have increased by over a thousand percent.

00:03:05: I mean saving forty percent on physical servers means literally nothing if the proprietary software licensing layered on top of it rises tenfold.

00:03:12: That math is just brutal.

00:03:14: and uh, Bert Dismitt noted a broader European trend that mirrors this exact dynamic.

00:03:20: Oh

00:03:20: right with The Sovereign Clouds.

00:03:22: Yeah he highlights that.

00:03:23: european sovereign clouds you know initiatives like France, here is S three NS and blue.

00:03:29: They are increasingly relying on underlying American technology primarily from Google and Microsoft.

00:03:35: they run under European operation in contract control sure but the foundational code is imported

00:03:40: which completely transforms sovereignty an engineering absolute into a contract term.

00:03:45: You are paying a premium and I think Desmond estimates anywhere from thirty to fifty percent just for illegal agreement rather than actual technological independence.

00:03:54: It totally reminds me of renting high security vault.

00:03:57: Okay, how so?

00:03:58: So imagine you hold the only physical key.

00:04:01: The Vault is located in your hometown And pay local guards To watch door.

00:04:06: Tim's pretty secure

00:04:07: Right.

00:04:08: But landlord who built facility Manufactured lock designed the security cameras and wrote this software running the alarm system.

00:04:16: You hold the key, sure but the underlying mechanisms of security belong to someone else

00:04:21: entirely.".

00:04:22: That

00:04:22: is a great analogy!

00:04:23: And I mean if that landlord partnership breaks down you really have to wonder if your keys still turns?

00:04:28: Exactly.

00:04:29: Harry Milonis points out Dora Compliance, the Digital Operational Resilience Act.

00:04:36: Right which is huge in Europe right now.

00:04:37: Massive.

00:04:38: and he argues that market has fallen into a trap by confusing provider sovereignty with architectural exitability.

00:04:44: Okay let's break it down because those sounds similar but they are very different.

00:04:48: so Provider Sovereignty guarantees the hyperscaler cannot legally or physically access your workload.

00:04:54: Correct

00:04:54: But architectural accessibility means you actually possess a mathematically viable, tested strategy to move that workload somewhere else if the provider fails.

00:05:03: And Dora compliance demands that exit strategy.

00:05:07: It is simply not enough to have a legal document shielding your data, right?

00:05:11: Because if you're architecture relies on AI reasoning loops they are completely hard coded into a provider's proprietary managed service.

00:05:19: You cannot leave.

00:05:20: you can't execute an exit test Compliance as illegal state but excitability Is a hardcore engineering constraint yet

00:05:28: tangling yourself in the providers proprietary authorization logic just leaves you mathematically stuck.

00:05:34: Exactly, and... This requirement for exitability is forcing providers to build entirely new infrastructure models.

00:05:42: Ivo Pinto shared some insights regarding the AWS European Sovereign Cloud that they kind of reveal the sheer scale this engineering challenge?

00:05:50: Oh

00:05:50: yeah, because a lot architects just assume it's like an EU region where you can easily spin workloads up

00:05:55: in which is fundamental misunderstanding underlying cryptography.

00:05:59: It isn't a new region but completely new partition.

00:06:02: Wait

00:06:02: explain the difference there.

00:06:03: who might mix those?

00:06:04: Sure, so a region is just a physical grouping of data centers.

00:06:08: A partition though is a distinct logical boundary with its own root of trust and identity management.

00:06:15: the commercial AWS network literally does not possess the cryptographic keys to see into the sovereign partition.

00:06:22: Oh wow!

00:06:22: So if I am a network architect i cannot use my standard cross-region tools.

00:06:27: VPC peering transit gateways automated S-three bucket replication, none of that crosses the partition line.

00:06:33: The boundary breaks all of it.

00:06:34: Separate billing separate identity centers your commercial direct connect gateway cannot reach the sovereign side.

00:06:40: That's wild.

00:06:41: Yeah You can not just copy an image across that line you have to manually push It through a secure Gateway.

00:06:46: the mental model is not extending Your existing cloud?

00:06:49: You are basically standing up a secondary fully isolated cloud that Just happens to use the same API syntax.

00:06:55: well looking at a massive complex IT estate trying to duplicate operations across two entirely separate cryptographic partitions sounds like an absolute administrative nightmare.

00:07:05: It really is,

00:07:06: which raises the question of whether every single piece of corporate data actually needs that level of isolation?

00:07:13: it definitely doesn't and honestly overreacting as dangerous here rick martyre and joe leemans both argue that sovereignty's a spectrum

00:07:23: applying blanket, statewide restrictions just because a fraction of your data is highly sensitive.

00:07:29: It destroys operational agility completely.

00:07:32: you have to assess the requirements workload by work load.

00:07:36: Like, The public cloud handles the standard applications specific regions handle data residency requirements and then maybe an air-gapped sovereign partition houses the core IP in financial records.

00:07:48: The goal is choosing right mechanism for each situation.

00:07:51: but this strict reevaluation of where workloads actually belong Is kind colliding with a separate crisis in infrastructure world

00:07:59: Right?

00:07:59: Oh yeah virtualization reset.

00:08:01: Organizations are auditing their environments for sovereignty and simultaneously running into this massive virtualization shift which is driving a sudden renaissance of the private cloud.

00:08:11: Yeah, we touched on the Broadcom and VMware situation earlier.

00:08:14: But Jonathan Webber detailed exactly how this is changing the operating model.

00:08:18: Broadcom is killing at The License Included Model for VMWare Cloud Foundation On Hyperscalers.

00:08:25: Okay

00:08:25: meaning what?

00:08:26: Exactly!

00:08:26: So previously if you ran VMware on Azure or legal cloud... ...the license was just a bundled line item on your monthly cloud bill but starting in late twenty-twenty five that ends it becomes entirely.

00:08:38: bring your own license or BYL.

00:08:41: Ah, so Broadcom wants a direct commercial relationship with the end user?

00:08:44: Exactly but by shifting to BYL costs and compliance burden lands squarely back on the customer's engineering teams, you have to track your core entitlements across hyperscalar hardware that you don't even physically own.

00:08:58: That sounds incredibly stressful

00:08:59: it is.

00:09:00: if you miscalculate the CPU cores underlying your cloud instances You risk falling out of compliance in facing some pretty severe financial penalties which

00:09:08: just adds a layer of administrative overhead that forces IT leaders To justify why they are paying for the virtualization platform The first place

00:09:15: right like.

00:09:15: Is this Even Worth It?

00:09:16: Colin Gallagher at how Tottenham-Hotspur handled this exact scenario, and it is a great real world example.

00:09:24: The football club actually managed to cut their VMware licensing costs by eighty-five percent.

00:09:29: Eighty five percent, that is huge!

00:09:31: How did they pull that off?

00:09:32: They achieved it by building a modern hybrid cloud platform with HPE GreenLake.

00:09:37: Oh right because HPE green lake operates By bringing Cloud Light consumption models to on-premise hardware

00:09:44: Exactly...by shifting their workloads directly onto this Modern Hybrid infrastructure..they just stripped out the massive hypervisor licensing tax That legacy virtualization was imposing

00:09:55: Well escaping licensing fees.

00:09:58: But David Lenthicum laid out the pure economics of workloads we are running today, and it shows that this shift back to private infrastructure is also heavily being driven by artificial intelligence.

00:10:09: Yeah!

00:10:09: The math on enterprise AI infrastructure is just brutal isn't?

00:10:13: It's staggering – AI systems cost ten-twenty times more to operate than traditional software system.

00:10:18: Wow…

00:10:19: Ten to twenty times.

00:10:20: I mean, hyperscalers are fantastic for bursting compute like running a short pilot testing.

00:10:25: A new model paying by the hour but The moment you move a sustained two hundred four seven enterprise AI workload into production on a hyperscaler that burst pricing model becomes viciously expensive because

00:10:39: You can't halt your AI adoption But you also cannot afford to run continuous inference at a premium markup

00:10:46: right?

00:10:46: So the alternative is dedicated hardware For scheduled, always-on machine learning workloads.

00:10:52: dedicated bare metal private clouds drop the cost per unit dramatically.

00:10:56: And Bare Metal means you are running directly on physical server without a virtualization layer eating up compute resources?

00:11:02: Exactly!

00:11:03: Get predictable performance consistent latency because don't have noisy neighbors sharing your service space and get flat costs structure.

00:11:10: The most pragmatic path to advanced AI operations is often just a rack of dedicated GPUs in private or sovereign environments.

00:11:17: Well,

00:11:17: before we look at the specific tactics companies are using to pay for all these new AI workloads... Just a quick reminder if you want stay ahead architectural shifts like this make sure you subscribe so that you catch future editions of Deep Dive!

00:11:29: Yes definitely Subscribe.

00:11:31: So we've established that AI compute as an order of magnitude more expensive than traditional systems?

00:11:37: If organizations are moving workloads to private clouds just to manage the bleeding, then the old methods of simply reviewing a cloud dashboard at the end of month or completely obsolete.

00:11:48: So completely we're entering an era of AI fine ops.

00:11:52: Amenia Murni draws this really great historical comparison that contextualizes what engineering teams are facing right now?

00:11:59: Okay What's

00:12:01: He says?

00:12:01: the AI agent cost explosion we see today is repeating the exact same bill shock pattern that hit early cloud adopters back in twenty fifteen.

00:12:09: Oh man, I remember that!

00:12:11: Back then startups were waking up to horrifying AWS bills because engineers just left massive database instances running all weekend for some minor test

00:12:19: Exactly.

00:12:20: The ease of provisioning vastly outpaced the maturity financial controls.

00:12:25: Yeah

00:12:25: Today, it is happening with agentic workflows.

00:12:28: Teams ship an AI agent that works flawlessly in a controlled demo environment Then goes into production and the invoice arrives.

00:12:36: These are context windows.

00:12:37: Yes

00:12:38: Modern language models use Context Windows which represent total amount of text they can process at once.

00:12:44: An inefficient agent might re-read entire conversation history on every single sequential step effectively taying to process the exact same data dozens of times.

00:12:55: Or,

00:12:55: The architecture forces the system To call a massive expensive frontier model just to perform basic data writing.

00:13:02: you know something that A small highly compressed open source model could handle for literal pennies Right?

00:13:08: or it burns thousands of reasoning tokens meaning the model is spending compute power to generate internal hidden logic steps Just to arrive at a decision.

00:13:16: That was already completely obvious from the system prompt.

00:13:19: Yeah, infrastructure does not get expensive merely because you use it—it gets expensive when you deploy without cost as a core architectural

00:13:26: constraint.".

00:13:27: Which is exactly what Reeves Smith argues….

00:13:30: AI fine-offs can no longer function.

00:13:36: One user prompt can trigger retrieval augmented generation steps, multiple agent reasoning loops external API fetches validation retries.

00:13:50: It's highly complex

00:13:51: and you cannot manage that complexity by merely tracking the cost per token.

00:13:56: Token cost only tells you what compute was consumed.

00:13:59: Finops teams have to measure the costs per business outcome,

00:14:01: right?

00:14:02: Yeah!

00:14:02: You have to ask whether a complex string of agentic reasoning is actually worth computing power required...to generate that specific answer.

00:14:09: But

00:14:09: let me push back on this because I was looking at the Gartner report Nitin Badoria referenced and it exposes massive failure in how the industry currently handles this.

00:14:17: Oh…the implementation staff.

00:14:19: Out of every right-sizing recommendation generated by automated fine ops tools, only about one in five actually gets implemented by the engineering teams.

00:14:28: It's terrible!

00:14:29: Right what is the point of measuring costs per outcome if eighty percent of the known waste is simply ignored?

00:14:36: Well because a dashboard without architectural authority is just noise.

00:14:40: it just creates the illusion of progress Management.

00:14:43: teams see hundreds of optimization tickets in JIRA, and they feel like they are actively doing phenops while the actual monthly bill remains completely flat.

00:14:52: Yeah because implementing a right-sizing recommendation requires an engineer to actually take a system offline rewrite configuration code run regression tests

00:15:02: Exactly!

00:15:04: And unless leadership incentivizes that downtime The recommendations will always be ignored.

00:15:09: engineers or busy shipping new features.

00:15:11: Right.

00:15:12: So to break that cycle, you have to change the fundamental premise of this conversation.

00:15:16: Nicholas Fondrini proposes a framework called Zero-Based Phenops.

00:15:20: Oh

00:15:20: I like his concept!

00:15:21: It's great.

00:15:21: A traditional phenops team looks at a ten million euro cloud run rate and asks how to optimize margins for that baseline.

00:15:28: Zero based phenops entirely throws out the baseline.

00:15:33: If that ten million euros were sitting in cash on the table today, would we invest it in this exact architecture again?

00:15:40: It forces continuous economic justification.

00:15:43: You cannot just tweak the edges of a poorly designed system.

00:15:46: you have to justify why the system exists and its current form in the first place

00:15:50: And that level of justification is mandatory when AI API costs can spiral out-of control in a single afternoon.

00:15:57: The irony here, though is that to manage these spiraling costs and govern incredibly complex multi-model environments the industry's just turning to AI.

00:16:06: To Manage the A.I

00:16:08: Yeah fighting fire with fire.

00:16:10: Ice Pelman predicts that a gateways are gonna replace traditional cloud management platforms or CMP as the critical control layer for the enterprise.

00:16:18: Yeah, think back to how organizations handled the multicloud explosion.

00:16:22: suddenly IT departments were managing AWS Azure Google Cloud and on-premise infrastructure all simultaneously.

00:16:29: was total operational.

00:16:30: chaos

00:16:31: right in CMP's emerged is an abstraction layer that governs security and route workloads between those disparate environments.

00:16:37: And pelicans argues.

00:16:38: we are on that exact same trajectory Right now with AI models

00:16:42: because a modern Enterprise isn't just using open AI anymore.

00:16:45: They're querying anthropic for coding tasks, using Mistral for local data processing.

00:16:51: Maybe deploying specialized open source models for internal

00:16:54: documentation.".

00:16:55: Right so the architectural challenge is no longer identifying this singular best model – The Challenge is dynamically routing a user's request to most cost-effective model capable of handling that specific task…

00:17:07: And an AI gateway arbitrates costs versus performance in real time?

00:17:11: Exactly!

00:17:12: It uses semantic routing to analyze a prompt... If the prompt is simple, the gateway routes it to a cheat model.

00:17:18: if the prompt requires complex logic The Gateway escalates into premium models

00:17:23: and caches previous answers to avoid redundant compute And governs security prompts across all those providers simultaneously.

00:17:29: It's brilliant!

00:17:30: We are seeing these agentic workflows actively deployed on the operations floor already.

00:17:35: Yeah

00:17:35: Ashwith V shared real-world example regarding agentic SRE site reliability.

00:17:40: engineering Teams using AI agents automate root cause analysis.

00:17:44: Right, so when a system goes down the agent correlates historical logs and investigates the incident context before human even opens the ticket.

00:17:53: Which saves so much time.

00:17:55: but Ashroth includes really critical caveat regarding where this automation stops.

00:18:01: AI should only be deployed were reasoning creates tangible engineering value.

00:18:06: In infrastructure, you have to preserve absolute determinism and tight execution limits.

00:18:12: Yeah You definitely do not want a hallucinating AI agent possessing the permissions To execute a multi-million dollar infrastructure change at two in the morning just because it misread a CPU utilization log.

00:18:22: Oh my gosh that would be disaster

00:18:24: A total disaster which is exactly why using AI to manage infrastructure complexity actually places a massive premium on deep human engineering talent.

00:18:33: You need engineers who really understand how to constrain the agents.

00:18:36: Yet,

00:18:36: market seems deeply confused right now about price and source this talent.

00:18:40: Highly

00:18:40: confused Andres Torres-Duran posted very sharp critique of current hiring trends.

00:18:45: He highlighted job descriptions that essentially bundle an entire IT department into single requisition.

00:18:50: Oh I've seen those Companies are demanding cloud engineer A site reliability engineer, a platform engineer, DevOps security and fine ops all wrapped into one human being.

00:19:03: Yes!

00:19:04: And according to Torres Durand the real punchline is that they expect to pay competitive salary for Latin America basically treating the latin talent pool like a discount aisle.

00:19:14: Well in his argument centers on the fact that geographic arbitrage applies currency exchange and local living costs.

00:19:22: But it absolutely does not apply to engineering complexity,

00:19:25: right?

00:19:25: The architectural difficulty of a system does not shrink based on the engineer's zip code.

00:19:30: in AWS environment doesn't magically become easier to architect just because the engineer building it is sitting in Costa Rica.

00:19:37: A production incident at two in the morning is just as critical if the pager goes off south of Texas.

00:19:41: Right

00:19:41: exactly.

00:19:42: The failure modes have a Kubernetes cluster are mathematically identical in Mexico as they aren't Silicon Valley.

00:19:47: Yeah, if an organization wants one human bone to hold accountability for infrastructure provisioning reliability deployment pipelines security patching and writing production code they are not hiring a cheap junior operations role.

00:20:01: No!

00:20:02: They're consolidating multiple senior level disciplines into one budget line.

00:20:06: An Ahmed Bilal reinforced what a Senior Cloud Engineering Role actually entails.

00:20:11: today it is no longer about memorizing a checklist of individual command line tools.

00:20:17: Right, senior engineering demands deep systems understanding.

00:20:20: You have to understand how the entire system behaves as a living organism.

00:20:24: Like where are network bottlenecks?

00:20:26: How does state replicate during a database failure?

00:20:29: How is application recover from an availability zone outage?

00:20:32: Exactly

00:20:33: Specific tools and syntax change every six months.

00:20:36: Systems thinking and architectural design patterns are permanent.

00:20:39: We have covered massive ground in this deep dive.

00:20:42: Yeah, we started with the realization that sovereignty is fundamentally redefining The mechanics of cloud control.

00:20:47: yeah and we explored how the extreme economics Of AI are driving a sudden return to dedicated private clouds?

00:20:54: We examined the painful architectural reality of AI fine-ups And we concluded would be immense human complexity required To actually govern egentic operations.

00:21:05: it's a lot.

00:21:06: But the common thread through all of these shifts can be summarized by an insight from Marcus Oskertag regarding the shared responsibility model.

00:21:14: Oh, this is a great way to frame it!

00:21:15: In standard cloud training you know we are taught that The line between what the provider secures and What the customer secures?

00:21:22: Is a fixed boundary.

00:21:23: Oskirtag points out That It's not A Fixed Line.

00:21:26: its actually a slider

00:21:27: And depending on the architectural services You select that Slider moves dramatically.

00:21:31: Exactly If you choose serverless functions, the cloud provider takes on almost all of operational burden.

00:21:41: But if you move back to sovereign private clouds and bare metal environments we discussed today You are physically pushing that slider All the way back To maximum responsibility.

00:21:50: And where

00:21:51: this slider rests Dictates engineering skills The headcount you must hire And total cost Of operating your entire environment.

00:21:57: Yeah when make these technical decisions You aren't just buying infrastructure.

00:22:01: You are choosing an operating

00:22:03: model.

00:22:03: And remember that vault analogy from earlier, before you secure your data make absolutely sure you know who forged the key.

00:22:11: very true if you enjoyed this episode new episodes drop every two weeks.

00:22:16: also check out our other editions on digital products and services defense tech AI and agentic systems green ICT in sustainable AI and health tech.