Best of LinkedIn: Cloud Insights & Sovereignty CW 35/ 36

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

00:00:05: In calendar weeks thirty five and thirty six.

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.

00:00:20: You can find more info in the description.

00:00:22: Yeah, and we are so glad to have those resources backing us up today.

00:00:25: definitely.

00:00:26: So if you were tuning in Today We Are doing a deep dive into the absolute top digital products And services trends across linkedin right now

00:00:35: Right?

00:00:35: And specifically looking at this massive shift where infrastructure placement just It isn't an automatic default anymore it's becoming This highly negotiated decision with A very strict price tag.

00:00:47: yeah

00:00:47: I mean if you're managing cloud budgets or enterprise architecture right now, You might feel like you've signed an apartment lease where the landlord legally owns your furniture.

00:00:55: The second you try to move

00:00:56: out.

00:00:56: Oh wow yeah that's a brutal way To put it but its true

00:00:59: Right.

00:01:00: and you wouldn't sign That contract without knowing exactly what It would cost to hire A specialized crew to sneak Your couch at back window?

00:01:07: No

00:01:07: of course not.

00:01:08: Yet for years, the entire tech industry basically signed that exact lease with public cloud providers.

00:01:14: Today we're looking at how organizations are finally staging a jailbreak.

00:01:19: We really are moving away from that era of just blindly dumping workloads into the public Cloud and you know figuring out the consequences later.

00:01:26: Exactly.

00:01:27: And that brings the financial reality of untangling yourself From a hyperscaler front-end center Because sovereignty isn't just about drawing a border around where your data physically lives anymore.

00:01:38: No, it's really about the friction of leaving

00:01:40: right

00:01:40: because The days of sovereignty Just being a residency checkbox are completely over.

00:01:45: P John De Bruin shared an incredible case study on this on LinkedIn.

00:01:49: Oh, the Airbus one.

00:01:50: Yeah exactly.

00:01:51: he highlighted how air bus is in the process of moving nine hundred applications completely off AWS.

00:01:57: wait nine hundred that Is massive?

00:01:59: It's huge.

00:02:00: they're starting with a batch of seventy and it's part Of this overarching initiative.

00:02:05: They call A minimum viable company plan

00:02:09: a minimum viable Company.

00:02:11: so structurally they are stripping their enterprise architecture back to its absolute core components

00:02:17: Exactly Just to prove they can operate independently of a single cloud provider.

00:02:21: Wow, it's essentially a simulated severance?

00:02:24: Yeah!

00:02:24: They have to measure true dependency and massive catalysts driving this kind of architectural redesign is the upcoming EU data

00:02:32: act Right...the regulation changes

00:02:34: By January twenty-twenty seven.

00:02:36: that regulation will outright ban cloud switching charges

00:02:40: Because you really cant spin an exit cost.

00:02:42: It literally becomes the ultimate measurable sovereignty metric for a procurement team.

00:02:47: Totally

00:02:47: and governments were already treating it that way, I mean they're backing up with serious capital.

00:02:51: Oh

00:02:51: yeah Anthony month posted about Germany right?

00:02:54: Yes,

00:02:54: Germany just awarded a two hundred fifty million euro tender For its sovereign AI cloud which their calling the Deutschland stack

00:03:01: And split on.

00:03:02: that contract is highly revealing.

00:03:04: It's

00:03:04: so revealing.

00:03:05: Roughly seventy percent went to a consortium of Deutsche Telekom T systems an SAP

00:03:11: To run what like Ten thousand NVIDIA GPUs natively in Munich, I think?

00:03:16: Exactly.

00:03:16: And then the remaining thirty percent went to SVA, Schwartz Digits and Codesphere.

00:03:21: So zero American hyperscalars won any piece of that tender?

00:03:25: None.

00:03:26: Google Cloud actually contested the initial bid requirements.

00:03:29: it eventually just withdrew entirely.

00:03:30: Wow so we are seeing a complete commitment to local infrastructure.

00:03:34: Oh

00:03:35: absolutely Bernd Wagner noted that Schwartz Group in the state of Mecklenburg-Vorpommern are partnering to invest up to five point six billion euros, and a massive new data center.

00:03:45: That is a staggering amount money right?

00:03:47: They're aiming for two hundred forty megawatts of power capacity by twenty thirty three.

00:03:51: And they are moving critical public administration apps, things like digital building permits directly on to stack it by twenty twenty six.

00:03:57: But

00:03:57: you know infrastructure location is really only half the battle?

00:04:00: Yeah

00:04:00: For

00:04:01: sure because having a server physically sitting in Germany doesn't automatically mean you control it.

00:04:06: Nevin Mulvey pointed out that fine print this kind of sovereignty

00:04:10: The operational control gap

00:04:12: Exactly.

00:04:13: He examined the secure setup that pairs Z-scalars set up with Schwartz digits infrastructure in German data centers.

00:04:19: Right, so on paper legally it absolutely checks the data residency box.

00:04:24: The data is on European soil.

00:04:27: But Mulvey raises the real critical question Who actually has access to that environment at two a.m.. During a major incident?

00:04:35: Oh That's such a good point right.

00:04:37: if a server and Frankfurt crashes but the diagnostic tool roots telemetry logs back to a U.S.-controlled control plane, and it's read by support staff in another time

00:04:47: zone."

00:04:48: Then the residency might be German but operational control is heavily external.

00:04:52: Exactly!

00:04:53: Where do encryption keys live?

00:04:55: If an external vendor goes down can you keep lights on?

00:04:58: So if a server can fully comply with its location or if dependency still runs through an outside vendor your sovereignty is kind of an illusion Which

00:05:06: honestly sounds like absolute nightmare for FineOps teams trying to budget this year.

00:05:10: Oh

00:05:10: yeah Nicholas Fundrini introduced this concept of the sovereignty premium, which I thought was fascinating.

00:05:15: Yeah because phenops isn't just about scanning dashboards for the cheapest raw compute power anymore

00:05:22: Not at all.

00:05:22: financial operations teams now have to deliberately calculate The cost of jurisdictional control you know, the economic exposure of vendor lock-in.

00:05:31: And the literal price of exit readiness!

00:05:34: Right.

00:05:34: so sometimes paying a higher premium for compute isn't a sign of operational inefficiency.

00:05:39: it's a deliberate calculated investment in resilience

00:05:43: which naturally shifts how we think about the destination of these workloads.

00:05:47: How so?

00:05:47: Well, if organizations are willing to pay a premium for control that automatic default-to-the public cloud begins to crack especially when you introduce massive resource requirements.

00:06:06: Yeah, Michael Gandhi shared a statistic from Broadcom's Private Cloud Outlook twenty-twenty six report that put some heavy numbers behind this.

00:06:13: What do they find?

00:06:14: They found that fifty-six percent of enterprises are now running or actively planning to run production.

00:06:19: AI inferencing on private cloud infrastructure.

00:06:22: Wow so more than half.

00:06:24: yeah.

00:06:24: and meanwhile the public clouds share for that exact same AI workload dropped from fifty six percent down to forty one percent in just a single year.

00:06:34: That is a massive drop But it makes sense.

00:06:36: Customers want the IEI models physically close to their proprietary data stores

00:06:40: secured under their own internal protocol exactly

00:06:43: without paying astronomical fees, to constantly move that data back and forth over the Internet?

00:06:49: eight let me push back a bit here.

00:06:50: there

00:06:50: go.

00:06:50: for

00:06:51: if we are talking about broad repatriation pulling these massive workloads back on premise aren't we just throwing away a decade of digital transformation?

00:07:02: Like, are we seriously moving the servers back into the basement and abandoning The Cloud.

00:07:06: It's not a rejection to the cloud model itself.

00:07:08: it is really just a correction in the blind cloud-first mantra.

00:07:11: Okay that's good distinction.

00:07:13: David Lenthicombe has discussed this dynamic extensively.

00:07:16: The re-patriation trend is very real but stems from an economic realization.

00:07:20: They egress fees

00:07:22: Exactly!

00:07:23: Enterprises spent Billions of dollars on data.

00:07:25: egress fees just the cost have taken their own data out of the cloud.

00:07:29: So it's basically realizing you don't need to rent a massive commercial moving truck Just to go pick up a single gallon of milk.

00:07:36: That is a perfect analogy.

00:07:38: A significant portion of the optimization efficiencies that public clouds promised ended up simply being absorbed as hyperscale or profit margins.

00:07:47: Right, we adopted cloud first without asking if it made mathematical sense for predictable high utilization workloads.

00:07:54: AI models are the ultimate high-utilization workload.

00:07:58: they completely break the old deployment calculus

00:08:01: especially in heavily regulated spaces.

00:08:03: Jingwa Gua highlighted how the defense sector used to operate on a very simple binary right?

00:08:08: Yeah, highly sensitive data state on premise.

00:08:11: Non-sensitive applications went to the public cloud.

00:08:13: But Frontier AI demands a level of specialized compute and capital investment that most organizations cannot simply rack and stack in their own basements overnight.

00:08:22: And leading AI labs aren't handing over their core proprietary models To run locally onto customers private hardware either.

00:08:29: So the binary is just dead.

00:08:31: It's now all workload by work load calculation.

00:08:34: Yep

00:08:34: An engineering team might start out in the public cloud to quickly prototype and build a new AI capability because the tools are just readily available.

00:08:42: But, The moment that application scales and its mission criticality rises... ...the economics and security mandates basically force migration on-premise infrastructure.

00:08:53: And managing this kind of fluidity requires an entirely new architectural philosophy.

00:08:59: Brian White framed this shift as the fourth cloud concept.

00:09:02: Right, utilizing IBM's infrastructure portfolio

00:09:07: We have to stop debating between public, private or hybrid silos.

00:09:10: The fourth cloud is about granting the enterprise a unified management layer To run applications wherever it makes the most contextual

00:09:17: sense And he divides into five distinct pillars Right Run Reach Understand Scale and Control.

00:09:23: Exactly.

00:09:24: But the control in scale pillars are massive bottlenecks Especially regarding data storage Because

00:09:29: AI requires an unimaginable volume of historical training data.

00:09:33: Right, and if you try to keep petabytes of deep historical data sitting in public cloud storage.

00:09:39: the retrieval an API request fees will bankrupt your IT department before the model even finishes training.

00:09:45: Which is why some of the oldest technology in The Data Center is quietly winning the cold-storage war right now!

00:09:51: Wait...you mean tape?

00:09:52: Literal Tape!

00:09:54: Greg Schwenk pointed out that Public Cloud Deep Archive tiers look incredibly cheap If You Only Look at Per Gigabyte Storage Cost.

00:10:01: But when you need to retrieve that data the API fees skyrocket.

00:10:05: Exactly, so he highlighted the IBM Diamondback Tape Archive system.

00:10:09: it actually offers an eighty five percent lower total cost of ownership compared to public cloud deep archive.

00:10:16: we are talking about literal magnetic tape drives literal

00:10:19: tape robotic arms fetching magnetic cartridges.

00:10:22: that is wild

00:10:23: but The Engineering Is Modern.

00:10:25: they're packing sixty one petabytes raw storage capacity into a single standard nineteen inch server act.

00:10:31: That's insane density and it operates using the standard S-three compatible API, right?

00:10:36: Yep.

00:10:36: Meaning developers interact with the physical tape library using the exact same code they use to talk to modern cloud storage.

00:10:43: Oh, so they don't have to rewrite their software workflows?

00:10:48: To interface with legacy hardware

00:10:50: exactly.

00:10:51: but I mean try to picture this sheer complexity of this architecture.

00:10:54: We have sixty one petabytes of data sitting on robotic tape drives in a single rack while sensitive AI inferencing runs on private sovereign clouds.

00:11:04: and Germany the initial prototyping is happening on AWS or Google Cloud.

00:11:09: The human brain just cannot track building security and optimization for that kind of fragmented ecosystem,

00:11:15: right?

00:11:16: Which means engineering teams are basically being forced to hand the operational credit card over to AI.

00:11:21: yeah at a scale of multi-cloud infrastructure has completely exceeded human capacity.

00:11:25: you have to use AI to manage the underlying AI infrastructure.

00:11:28: which Quick sidebar, if you're finding this deep dive valuable as we get into the automation site of things make sure to hit subscribe so that you can catch our future Deep Dives.

00:11:38: We cover a lot ground in these sessions.

00:11:41: So

00:11:41: how are engineering teams realistically automating this sprawling fine ops

00:11:45: headache?

00:11:47: Well were moving from monthly human spreadsheet reviews To continuous agentic AI governance.

00:11:53: Okay what does it actually look like?

00:11:55: Anoreg Jossiula demonstrated really powerful prototype of an autonomous FlanOps agent.

00:12:01: He used Amazon Bedrock, AWS Lambda and CloudTrail.

00:12:05: And what does it do differently?

00:12:07: Instead of just triggering a generic alert when budget threshold is crossed this large language model actually reasons through the cloud logs.

00:12:14: Wait how does that actually reason for billing swipe though?

00:12:17: So in ingest the CloudTrails events which are basically ongoing diary.

00:12:21: every API call made into environment Okay If its spots sudden cost anomaly like teams suddenly spinning up fifty expensive GPUs, the LLM maps that action to a specific user identity.

00:12:33: Oh wow!

00:12:34: Yeah it analyzes the infrastructure change generates a plain English root cause summary and posts suggested fix directly to the engineering team's slack channel.

00:12:42: That is incredible.

00:12:44: And does all this reasoning for just few dollars per month in compute costs?

00:12:48: We are seeing major providers build this natively too.

00:12:51: Diana Romos noted that AWS's own native Phenops agent is moving into public preview.

00:12:56: Yeah, which radically accelerates the investigation phase of cost anomalies.

00:13:00: But she also noted a major point of hesitation in the market

00:13:04: Which is cloud

00:13:05: practitioners are highly cautious about allowing AI to autonomously execute optimization recommendations.

00:13:11: Oh yeah Of course because an AI agent can identify and expensive cluster of servers but it lacks the human engineering context.

00:13:19: Exactly, It doesn't know that those servers are intentionally over-provisioned because a massive marketing campaign is launching in twelve hours.

00:13:26: Right You still need a Human In The Loop to understand why the workload was configured this way.

00:13:30: But automation is brilliant for repetitive grunt work like cleaning up digital mess.

00:13:35: we leave behind.

00:13:36: Yeah Jex Rathenam detailed project where he automated cloud hygiene at a massive enterprise scale.

00:13:41: What was the use case?

00:13:42: He

00:13:43: utilized AWS step functions to hunt down and clean up over ten thousand orphaned EBS snapshots.

00:13:49: Basically, those massive abandoned virtual hard drive backups that teams create before an update and then forget to delete

00:13:57: exactly just quietly bleeding The monthly budget.

00:14:00: And the primary technical hurdle there isn't writing a simple delete script.

00:14:05: It's navigating the cloud providers own guardrails.

00:14:08: right.

00:14:08: yeah If a script just blasts ten thousand delete requests at once, AWS will instantly trip its API throttling limits and shut the process down to protect the broader network.

00:14:19: Right, so Rothenam had to build in adaptive retry configurations and exponential back-off mechanisms.

00:14:25: Exactly!

00:14:26: The script essentially has to pace itself slowing down and speeding up just to keep the automated cleanup running silently in the background across fifteen different global regions.

00:14:34: Surveilless automation is really the only sustainable way to maintain that level of cloud hygiene.

00:14:40: but handing keys over to automated scripts introduces a terrifying vulnerability.

00:14:44: Oh

00:14:44: absolutely

00:14:45: Because if we have AI agents monitoring and actively altering the infrastructure, I mean who is monitoring monitors?

00:14:51: Harry Malonis offered a very stark warning on this.

00:14:54: He stated that automation does not eliminate risk.

00:14:57: it concentrates it.

00:14:58: That's great way to put it

00:15:00: Right.

00:15:00: If you automate infrastructure changes without deterministic hard-coded governance You are rapidly accelerating your legal liability.

00:15:10: A compromised automated playbook, or a rogue CICD pipeline doesn't pause to ask for a manual compliance review.

00:15:17: It executes at cloud speed.

00:15:19: if the financial institution is operating under Dorae... You

00:15:21: do up the EU's Digital Operational Resilience

00:15:24: Act Exactly which forces firms to prove they can survive a vendor outage.

00:15:29: Or under the EU AI act a runaway automated script Can instantly shatter their global compliance posture.

00:15:36: So regulatory authority basically has to be hard-coded into the execution layer itself, preventing the automation from making changes outside of legal boundaries.

00:15:44: Which

00:15:45: brings up the physical reality of how these automated systems interact.

00:15:48: because if we have phenops agents and AI models running across a highly fragmented multi cloud world How do these disparate environments actually talk?

00:15:57: And more urgently, how do we prevent a compromise agent from moving laterally across those connections to wreak havoc?

00:16:04: Well Sid Nagg pointed out that the physical internet backbone connecting these clouds is undergoing a major shift right now.

00:16:11: AWS and Azure are actively collaborating to build private high-speed interconnects directly between their networks Which is

00:16:19: huge!

00:16:19: For years if you wanted AWS & Azure to communicate securely You either had to route traffic over the public internet?

00:16:27: Or you had to lease space in a neutral physical data center and manually string fiber optic cross connects between their respective cages.

00:16:34: Yeah, now the hyperscalers are treating cloud-to-cloud connectivity as a native infrastructure requirement.

00:16:40: Enterprises can provision dedicated high speed fiber links directly through Cloud Native APIs.

00:16:45: But just plugging a faster private cable between two clouds does not make a secure multi-cloud architecture.

00:16:51: No!

00:16:51: And that was Nag's core argument.

00:16:53: True multi-cloud requires integration layers far beyond just a private network pipe.

00:16:57: Right, if an autonomous agent operating in Azure needs to analyze a database sitting in AWS it needs the network path but also require trusted unified identity layer

00:17:09: and enforceable cross cloud security policies.

00:17:11: end-to-end observability

00:17:13: because of an agent gets compromised is going use that convenient high speed native interconnect spread instantly.

00:17:20: oh yeah Helen, you discussed this thread.

00:17:23: Focusing on the need for deep east-west visibility inside the environments where these agentic AI systems operate.

00:17:30: Traditional

00:17:30: perimeter security, like the firewalls designed to keep external threats out was never built to handle autonomous agents that already have internal credentials.

00:17:38: Agents.

00:17:38: they can invoke APIs and move laterally across a distributed network at machine speed?

00:17:43: It's like traditional security is putting a massive reinforced steel door on the front of a bank vault to stop robbers.

00:17:49: That acts as great for the perimeter.

00:17:52: But securing a genetic AI with east-west visibility means you need security cameras inside the vault itself, tracking every single micromovement that tellers may.

00:18:02: Just in case a trusted employee suddenly decides to start quietly pocketing cash.

00:18:07: That is exactly it!

00:18:09: You have monitor and govern internal API connections to sever lateral movement – the millisecond A nonhuman identity begins acting erratically

00:18:18: And the timeline to implement these controls is vanishing fast.

00:18:21: Yeah, Alexandra Bailey and C warned that frontier AI models have already demonstrated the capability to autonomously conduct end-to-end system compromises.

00:18:31: He estimates that enterprises have a window of roughly six months to prepare before automated agentic attacks become a routine reality.

00:18:38: Six months to rearchitect internal security.

00:18:41: Does native cloud security even have the tools to handle that kind of lateral threat

00:18:45: right now?

00:18:47: capability gaps, specifically looking at Google Cloud.

00:18:50: He acknowledged that Google's native application security suite is highly robust like

00:18:54: using cloud armor and recap TGA enterprise?

00:18:57: Yeah but in a fragmented multi-cloud environment relying solely on need of tools creates a logistical nightmare.

00:19:03: right if you are running workloads on Google AWS an Azure You suddenly have to maintain three entirely different security postures

00:19:12: And try to reconcile three conflicting sets of telemetry data during an active breach.

00:19:17: That sounds impossible.

00:19:19: And beyond the operational friction, there are massive compliance gaps.

00:19:22: Kurios highlighted the strict new requirements rolling out under PCI DSS.

00:19:27: four point zero point one specifically regarding the integrity of payment page grips.

00:19:32: Oh this is a perfect example why old tools were failing.

00:19:35: Traditional cloud firewalls only monitor traffic actively hitting server But this new PCI DSS rule targets a completely different vector.

00:19:44: It targets malicious code that hijacks the payment form directly inside the customer's web browser before credit card data ever travels to Exactly.

00:19:57: Because no major cloud provider offers a native tool to satisfy that specific client-side requirement and our prices are just being forced to adopt specialized third party security layers, just to remain legally compliant.

00:20:10: The industry is really caught in the paradox right now.

00:20:13: We desperately want the agility and speed of the Cloud

00:20:16: But the compounding weight of multi-cloud policy fragmentation, the rise of autonomous AI threats and strict regulatory mandates means that operational burden on managing it all has never been heavier.

00:20:28: It is a delicate balancing act of cost control in automation

00:20:32: Yeah which basically brings everything full circle.

00:20:35: We started by looking at how infrastructure placement no longer a default but highly negotiated decision driven by true price exit.

00:20:43: We're seeing sovereignty premiums and massive AI data sets drive workloads back to private clouds, even physical take-drives.

00:20:50: And simultaneously we are handing the management of this fragmented mess over to AI Finops agents while racing to build internal security cameras to stop those same agents from going rogue across high speed cloud interconnects.

00:21:02: So I want leave you our listener with a final thought to mull over.

00:21:07: We are currently designing AI agents to autonomously manage our cloud billing, analyze logs and clean up our storage footprint.

00:21:14: OK.

00:21:14: If exit costs are now the ultimate sovereignty metric, and if we're increasingly trusting AI to optimize our enterprise spending how long until these agents are authorized to dynamically negotiate their own infrastructure placement?

00:21:27: Oh wow!

00:21:27: Right... How long until an AI model realizes a lease is too expensive and executes its own cloud exit strategy in real time moving itself into cheaper data center entirely without human intervention?

00:21:38: That's terrifying but very real possibility.

00:21:43: If you enjoyed this episode, new episodes drop every two weeks.

00:21:46: 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.