Spectro Cloud

06/02/2026

AI + San Fran = 🫶
In a couple of weeks, we're heading to Moscone Center for our first time at Databricks' Data + AI Summit, and we're excited to bring the latest on PaletteAI to a room full of the data and AI teams actually putting this stuff into production.

If you're also attending, come find us at booth 426 for live demos and a chat with our experts about turning a (maybe messy?) AI stack into a platform your AI and infra teams can both rely on.

06/01/2026

NVIDIA DSX™ 🤝 PaletteAI
There's so much behind our name listed in NVIDIA's announcement as one of the NCP and software vendor ecosystem partners adopting DSX OS components.

The fit between PaletteAI and NVIDIA DSX is architectural, and it's intentional.
DSX is the full-stack playbook for how AI factories get designed, built, and operated: silicon, software, power, cooling, networking, all co-designed as one system.

PaletteAI is the platform to run it in production, giving teams the lifecycle management, multi-tenancy, and self-service access controls to actually operate that infrastructure at scale.

We see in NVIDIA the same vision we've been building toward with PaletteAI for the last two years: AI factory infrastructure that works as one system, from design to operations.

🔗 NVIDIA'S announcement: https://okt.to/DKkd2i

05/29/2026

"You can't run modern AI on yesterday's infrastructure."
Mic-drop moment from Justin Swagler at the National Restaurant Association Show. Plus, it's just... true.

Things like agentic managers reacting to a lunch rush in real time and computer vision telling you if a burger is undercooked before it goes out aren't future-state ideas anymore.

Brands are doing them. Like right now. As you read this post.

Guess what separates the ones that scale from the ones stuck in pilots? Yup, you got it. The infrastructure underneath.

Justin and Deborah from AWS got into all of it in their session, including how Yum! Brands is pulling it off across 40,000+ locations.

Truly great session.
And you can watch it here 👉 https://okt.to/LZGka7

05/29/2026

If you're running a neocloud or an MSP right now, you've probably noticed that demand isn't the problem. Basically every enterprise on the planet wants AI infrastructure... question: how to turn a pile of GPUs into a business that actually makes money?

McKinsey puts GPU rental gross margins at 14-16% after labor, power, and depreciation. Which is, awkwardly, worse than most non-tech retail. 😬

Which is to say: if you're just renting out raw GPU capacity, the economics aren't necessarily in your favor. A.k.a: the real money is in selling something more on top of it.

Our Global Head of Partner Programs Keith Wilson wrote a blog laying out the five things that separate the AI service providers who'll still be around in 2028 from the ones who won't:
1️⃣ Multi-tenancy customers will actually trust
2️⃣ Open choice, because no two customers want the same stack
3️⃣ A service catalog of repeatable blueprints, so onboarding doesn't turn into bespoke engineering every time
4️⃣ GPU utilization and metering you can run a P&L on
5️⃣ Sovereignty and deployment flexibility from day one

He also covers how PaletteAI is built for exactly this kind of operating reality, including the bits that matter for partners specifically (things like white-labeling the interface, so it's your brand customers see, not ours).

For the full breakdown 👉 https://okt.to/WOY3Nk

05/28/2026

One for the techies out there: if you're building or scoping multi-tenant AI infrastructure right now (cause who isn't, right? 🤓), the orchestration-to-fabric handoff is probably the part keeping you up at night... but we have something to help fix your sleep.

Bedside book recommendation for you is: our tech brief explaining the integration reference behind our PaletteAI platform + Aviz Networks + NVIDIA Spectrum-X multi-tenant AI factory architecture.

(To be clear, we don't mean it'll put you to sleep, ok? We mean it'll help with the actual headache.)

The full doc walks through how the three pieces actually wire together: PaletteAI compute pool labels driving Aviz ONES API calls, Aviz ONES pushing SONiC config to Spectrum-X Ethernet switches, Red/Blue tenant isolation getting enforced end-to-end. It has it all: screenshots, API flow, IP scheme, ping output, the whole lot.

We know this is the kind of detail nobody really broadly shares, because this handoff is usually annoying to solve and most vendors would rather not talk about it. But, well, we thought it was worth talking about.

And shoutout to our friends at Aviz Networks for the great partnership that makes this architecture possible.

For the full technical reference, check out 👉 https://okt.to/LyRQU0

05/27/2026

If you've ever tried to get GPUs running on Kubernetes for AI workloads from scratch, you already know NVIDIA's documentation can be... a lot. 😅
Our Principal AI Architect Pedro Oliveira wrote a hands-on guide a while back that breaks down how the NVIDIA GPU Operator actually works under the hood.

And yes, the GPU-on-K8s landscape has evolved a lot since he wrote it, with DRA, Dynamo, and a whole new layer of orchestration tooling entering the picture... but the Operator is still the foundation underneath all of it.

So we're resharing this guide for anyone who might be getting started with GPUs on K8s, or maybe just debugging the Operator and want to understand what's actually happening inside it.

Oh, and Pedro also covers how our Cluster Profiles capability lets you deploy validated GPU Operator configurations across all your clusters as code, which is what turns this from a one-cluster setup into something you can actually run consistently at scale. Which, you know, is pretty cool, if you didn't know about it. 😎

Catch the blog here 👉 https://okt.to/UhJLt1

The GPU Operator is a Kubernetes Operator that provisions and manages NVIDIA GPUs on top of Kubernetes. This ultimately exposes the GPUs as resources available to be used by your Kubernetes nodes.

05/26/2026

Friendly reality check: your data scientists want a new model to work on, not a new ticket to wait on.
And no, they don't want to think about GPU drivers, inference engines, or container images. They just want to pick a model and start prompting it.

The gap between those two things is usually something around weeks of tickets, manual configuration, and back-and-forth with whoever owns the cluster. Multiply that across a team experimenting with a dozen models, and the bottleneck now is not your GPUs, but the whole workflow to get anything onto them.

Well, we built Model as a Service in PaletteAI to close that gap.

Here's how it works:
- Platform teams define the rules once: which model sources are allowed (Hugging Face, NVIDIA NIMs), which inference engines map to which Profile Bundles, and how GPU quotas are enforced per tenant and project.

- After that, data scientists browse the catalog, pick a model, and PaletteAI handles the matching, validates the infrastructure is compatible, and deploys to an existing Compute Pool. No tickets, no bespoke config, no surprise misconfigurations at runtime.

The point here isn't just speed.
(Though going from selection to inference in minutes is a fair brag, isn't it? 😎)

Think of it as a win-win: platform teams keep governance, RBAC, multi-tenancy, and lifecycle management intact, while AI teams get the self-service experience they actually wanted.

You can check out the full datasheet if you want more details about how it works: https://okt.to/nUpZ2R

PaletteAI provides a "Model as a Service" (MaaS) solution designed to help platform teams offer governed, self-service access to AI models. It enables teams to deploy pre-trained or custom models from Hugging Face and NVIDIA NIMs in minutes, rather than weeks.

05/25/2026

Finding out your multi-tenant isolation doesn't actually work is a pretty bad way to learn it. Especially when you're learning it on hardware that took months to procure, in production, with real tenants running real workloads. Tenant traffic leaks, GPU allocation conflicts, and fabric misconfigurations are all expensive things to discover after the fact.

The fix is to validate the architecture before the hardware ever shows up.

That's what NVIDIA DSX Air is built for.
It's a cloud-hosted simulation platform that builds digital twins of AI data center environments, running in a browser, where teams can design a multi-tenant topology, test it, and intentionally break it before any physical kit is on the line.

Andreea Munteanu walks through how PaletteAI, Aviz ONES, and NVIDIA Spectrum-X Ethernet come together for multi-tenant AI infrastructure, and how DSX Air gives teams a sandbox to verify it all holds up.

If you're designing multi-tenancy for a sovereign AI cloud, an AI Grid deployment, or an internal platform shared across teams, check out the full blog here 👉 https://okt.to/NcyBMV

Deploying multi-tenant AI infrastructure is hard enough to get right in production. Doing it wrong is expensive: tenant traffic leaks, GPU allocation conflicts, and isolation failures are discovered after the fact — all on hardware that took months to procure

05/22/2026

GPUs alone don't make an AI factory.
You also need a fabric that keeps GPU traffic deterministic, Kubernetes clusters that actually align with GPU and DPU topology, multi-tenancy that holds across every layer, and lifecycle automation that doesn't fall apart at scale. Missing any of those means expensive hardware sits underused while teams fight the infrastructure.

That's the gap we're closing with Aviz Networks.

Aviz ONES brings GPU-aware fabric orchestration and EVPN/VXLAN segmentation across NVIDIA Spectrum-X Ethernet networks. Our PaletteAI handles full-stack lifecycle automation for Kubernetes fleets, NVIDIA AI Enterprise integration, and blueprint-driven AI environment deployment.

Together, it turns NVIDIA-powered infrastructure into a governed, repeatable AI factory platform, aligned with NVIDIA's Enterprise and Cloud Provider Reference Architectures.

The full solution brief is here 👉 https://okt.to/5sAmQH

Joint solution from Aviz Networks and Spectro Cloud combining GPU-aware fabric orchestration with full-stack fleet lifecycle automation for NVIDIA AI Factory infrastructure.

05/21/2026

Choosing a K8s platform is not a small decision.
If you're evaluating options for edge, for cloud, or for both, the GigaOm Radar reports are one of the more honest reads out there. They look at the stuff that actually matters once you're past the demo: how platforms hold up at scale, how flexible they really are, what the ecosystem looks like, and where the costs hide.

Yes, we're in there as a Leader... in both reports, for the third year, and we're not going to pretend we're not pleased... but the (other) reason we keep pointing people to these reports is that they give you a side-by-side view of vendors without the marketing layer on top.

So if you're in the middle of an evaluation, or just trying to make sense of the space, grab a copy 👉 https://okt.to/9tx263