AI Strategy & Platform · Mid-Market
Cisco just teamed up with Palantir and NVIDIA on a framework for AI strategy that's genuinely right. It's also built for a scale almost nobody has. Here's what it actually takes to run, and what the same idea looks like at a size that isn't a sovereign nation.
Quick Answer
Cisco's new deal with Palantir and NVIDIA gives the largest enterprises and governments a sovereign, full-stack AI factory built around intelligence, cost, and control. Good framework. But forward-deployed engineers, custom eval infrastructure, and a private reference architecture take a scale most companies simply don't have. OnStak applies the same three ideas at mid-market scale, in language that already fits how we talk about this, production readiness, cost governance, and operational assurance, delivered on Cisco AI infrastructure as an engagement rather than a multi-year build.
What Cisco, Palantir, and NVIDIA Just Announced
This week, Cisco announced it's teaming up with Palantir and NVIDIA to extend Cisco's Secure AI Factory as a preferred foundation for Palantir's Sovereign AI OS, built on NVIDIA's open Nemotron models. Compute, networking, storage, security, observability, and AI workflows, validated together as one architecture, with forward-deployed engineers on hand to actually stand it up.
Read: this is for governments and the largest regulated enterprises, the kind of organization that wants to own its data, its models, and its infrastructure, full stop, and has the budget to make that happen.
The Framework Behind It: Intelligence, Cost, Control
Honestly, the deal itself is less interesting than the logic behind it, since it holds up well beyond one vendor's announcement. Every AI strategy has to balance three things: intelligence, cost, and control. A model tuned to your actual workload will usually beat a bigger one with a flashier benchmark score. Cost isn't what you pay per token, it's how much useful work you get per dollar. And control comes down to a simple question, who owns the data, the security posture, and the environment the model runs in.
That's the right way to think about AI strategy. We've just been saying it under different names all year.
“
That's exactly the value prop we need to deliver to mid-market customers, who do not have the critical mass to adopt such a solution.
Fabio Gori, Chief Product & Marketing Officer, OnStak
Why the Delivery Model Doesn't Fit Mid-Market
It helps to think about what actually goes into standing up a private, sovereign AI factory. A full-stack reference architecture across five infrastructure layers. A post-training pipeline to specialize open models on your own data. Custom eval harnesses to prove those models work on your workflows. Forward-deployed engineers living on-site to keep it running. That's less a purchase and more an ongoing capability, and it tends to favor organizations with hyperscale budgets and a dedicated AI research bench.
That's the gap most mid-market companies find themselves in. Few are actually choosing between "sovereign AI factory" and "no AI strategy at all." Most are somewhere in between, wanting the same discipline around intelligence, cost, and control the largest players just got validated for, without quite the balance sheet or headcount to build it entirely in-house, and often without a complete AI readiness evaluation to know where to even start.
Two Ways to Apply the Same Framework
None of this stops mattering just because a company has 2,000 employees instead of 200,000. Intelligence, cost, and control still matter. They just have to show up differently, delivered by a partner already sitting inside your stack, rather than a factory you have to build and staff yourself.
🏢
Sovereign AI FactoryBuilt For Hyperscale & Sovereigns
IntelligenceCustom post-training and eval pipelines, run by a dedicated AI research team.
CostOwned infrastructure amortized across massive, sustained workload volume.
ControlA fully private, on-prem reference architecture the customer builds and runs.
DeliveryA multi-year capital project with forward-deployed engineers on-site.
⚡
OnStak Mid-Market ModelBuilt For Mid-Market & Mid-Enterprise
IntelligencePilot to Production: getting a working model running on real workflows.
CostTokenomics: cost discipline designed in, not optimized after the fact.
ControlAgentic Ops: governed, auditable control via AI Correlation Fabric and AI Assurance.
DeliveryAn engagement on Cisco AI infrastructure, live in a matter of months.
Mapping the Three Vectors to What We're Already Saying
Almost everything OnStak has published this year ladders back to one of three conversations, and it turns out all three answer the same question Cisco's framework is asking, just at a different scale.
Intelligence → Pilot to Production
Applied intelligence only counts once it's actually running the business, not once it's nailed the demo. Gartner puts the share of organizations with agents in production at 17%. Getting from a working model to something trusted in production is the hard half of "intelligence," and it's the half most vendors quietly skip.
Cost → Tokenomics
MIT found that 95% of GenAI pilots show no measurable P&L impact. The 6% that do win didn't cut their token spend, they redesigned the workflow around the model. Cost discipline is something you design for up front, not a line item you go back and optimize later.
Control → Agentic Ops
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mostly over weak risk controls. Control can't be a checkbox you tick at the end. It has to be built into how agents are correlated, governed, and audited from the start.
17%of orgs have agents actually in production · Gartner
95%of GenAI pilots show no P&L impact · MIT
40%+of agentic projects cancelled by 2027 · Gartner
What This Looks Like Without a Factory
This is the gap OnStak works in today, for mid-market and mid-enterprise customers, running on Cisco AI infrastructure, without ever asking the customer to stand up a research lab first.
The OnStak Delivery Model
1
Discover. Map what's actually there and figure out which use cases have a real shot at production.
2
Prove. Test the approach in your real environment, on your real data, not a cleaned-up demo.
3
Build. OnStak AI Correlation Fabric pre-correlates your data so any model, open or frontier, actually performs on your workflows.
4
Operate. OnStak AI Assurance, the Yes Layer, keeps governance and audit evidence built in from the start, not bolted on afterward.
Same three vectors as the Cisco framework. A fraction of the infrastructure lift.
From an OnStak Banking Engagement
$47M in value identified through AI readiness work.
2.8× expected ROI from the resulting engagement.
32% lower total cost of ownership versus the prior approach.
Three Questions Before You Chase a Factory You Don't Need
Before you start benchmarking sovereign AI stacks built for a completely different scale, it's worth asking whether infrastructure is actually your problem, or whether it's the operating model wrapped around it.
Do you genuinely need air-gapped, on-prem control? Or would governed, auditable control be enough, control that can still run in the cloud, at the edge, or on-prem depending on the workload?
Can you actually staff a standing AI engineering team? Post-training, custom evals, forward-deployed engineering, these are ongoing jobs, not a one-time project.
Is the model really the constraint? Most stalled AI initiatives get stuck on data readiness and governance design long before the model becomes the bottleneck.
Where We'll Be Talking About This Next
We'll be at Splunk .conf26, September 14–18, at Booth O11, talking through exactly this, what intelligence, cost, and control actually look like at mid-market scale instead of sovereign-nation scale. It'll also be the first place we show where AI Correlation Fabric and AI Assurance are headed next, as we bring both into a single platform.
The framework is right.
Now bring it to your scale.
Meet OnStak at Splunk .conf26, September 14–18, Booth O11, to see how intelligence, cost, and control come together on the OnStak AI Portfolio, without the factory.
Let's Talk →
FAQ
What exactly is driving this demand?+
OnStak applies the same three ideas at mid-market scale, in language that already fits how we talk about this, production readiness, cost governance, and operational assurance, delivered on Cisco AI infrastructure as an engagement rather than a multi-year build.
What is the "critical mass" problem in enterprise AI?+
It's the gap between wanting full control over your AI infrastructure, data, and models, and actually having the scale, capital, and dedicated engineering headcount to build and run that infrastructure yourself. A sovereign, full-stack AI factory takes a standing capability that most mid-market companies, and plenty of large enterprises, just don't have.
Can mid-market companies get the same AI control as large enterprises?+
Yes, just delivered differently. Rather than standing up a private full-stack AI factory, mid-market organizations can get governed, auditable control over their data and models through a partner that's already embedded in their infrastructure, without having to own the capability themselves.
How is OnStak different from a sovereign AI stack like Cisco, Palantir, and NVIDIA's?+
That collaboration is built for governments and the largest regulated enterprises that need a fully sovereign, on-prem AI factory. OnStak applies the same intelligence, cost, and control thinking at mid-market and mid-enterprise scale, as an engagement on Cisco AI infrastructure rather than a capital-intensive build.
How does OnStak apply intelligence, cost, and control for mid-market?+
Intelligence, by getting AI from pilot to production on real workflows. Cost, through tokenomics discipline that's designed in rather than fixed after the fact. And control, through OnStak AI Correlation Fabric and OnStak AI Assurance, which build governance and audit evidence in from day one.
Keep Reading