Boost your AI knowledge with our five-part ebook series, designed to help you turn AI into action. This ebook is part 5 of 5 in our AI series. Explore the full collection below. 

Book 1: AI literacy is everything. Here’s your 5-step success plan

Book 2: Achieving quick wins with AI: How to turn use cases into measurable business value

Book 3: Leadership in action: Strategic AI planning and implementation

Book 4: Building your strategic AI platform

Book 5: Confidently harnessing AI: Establishing your AI governance and security framework

The rapid rise of generative AI offers trillions of dollars in economic opportunities for ambitious, forward-facing organizations. But it also comes with harrowing risks, like data breaches, hallucinated content, copyright lawsuits, and regulatory fines. ​To navigate this tightrope, organizations must establish robust AI governance and security frameworks that balance innovation with risk management. ​

Many organizations are concerned about adopting AI without firm governance standards. But frankly, it’s too late to pretend generative AI hasn’t taken the world by storm. It’s a near certainty that if you haven’t formally offered AI solutions, your employees are using shadow AI tools and exposing your organization’s data to non-enterprise tools. Your sales reps are asking for consumer-grade tools to write emails to prospective customers. Your engineers are solving coding challenges with help from a bot. And, in the name of efficiency, they’re unwittingly sharing proprietary information with the wider internet.

So how do you govern AI to enable safe innovation?

“It’s essential that business leaders install guardrails, implement security frameworks, and establish effective governance structures,” says SHI Senior Solutions Architect Aaron Richmond. “Effective governance isn’t just about stopping risk; done right, it enables safe innovation, so your teams can confidently leverage AI as a powerful tool.”

Effective governance transforms AI guardrails and security frameworks from concepts into operational reality. It’s the connective tissue bringing your AI strategy to life, ensuring it’s both ambitious and responsible. Without practical governance structures, even the best security measures remain disconnected from day-to-day operations.

NEXT STEPS

Read the full guide below to harness AI as a secure growth engine and turn uncertainty into actionable strategies for sustainable success.

Contact our experts at AI@SHI.com today.

NVIDIA and Microsoft have introduced RTX Spark, a new Windows PC platform designed to support more demanding AI models and personal AI agents directly on laptops and desktops.

Glancing at the press release, it sounds impressive:

• It powers the world’s first Windows PCs purpose-built for personal agents.
• Features one petaflop of AI performance.
• Offers industry-leading power efficiency.
• Has full-stack NVIDIA AI and graphics technology and up to 128GB of unified memory.

But what the announcement really shows is the growing momentum in a trend already well underway: that the personal computer is being re-architected for the AI era. The PC is evolving from a place where users access AI, to a place where more of that intelligence runs, responds, and acts locally. For IT leaders, the question is no longer whether endpoint AI is coming, but rather how to support it in a way that is secure, cost-effective, and operationally sustainable.

Endpoint AI has huge benefits but is evolving faster than operational readiness

It’s easy to understand the appeal of running AI locally because when more processing happens on the device, organizations can:

• Reduce latency and get faster response times.
• Have greater control over data and privacy.
• Support more personalized, always-on user experiences.
• Reduce token costs by moving suitable AI workloads onto the device.

The challenge is that endpoint AI capability is advancing faster than most operating models are being implemented.

Our data reinforces this

At SHI’s 2026 Spring Summit, the disconnect between AI demand and IT readiness was clear. Polls with the 330+ attendees showed that:

• 78% of IT leaders already have multiple AI apps on their phone.
• Only 13% say they are fully leveraging AI at the endpoint in production.
• 53% describe their IT approach as overly reactive.
• 41% say endpoints are the biggest source of digital friction.

At the same time, cost pressure is intensifying:

• 47% cite budget alignment as their top concern.

Taken together, these responses point to a widening readiness gap between growing AI expectations and the operational reality inside most IT environments. AI is already shaping how individuals work, but many organizations are still figuring out how to support it at scale across the wider environment, from infrastructure and governance through to endpoint performance and user experience.

As endpoint AI grows, IT planning gets more complex

AI at the endpoint is no longer a standalone device conversation. It’s part of an integrated AI infrastructure strategy, where performance depends on having the right mix of device, network, and backend infrastructure working together seamlessly to support:

• Better employee experience – performance, responsiveness, and usability of AI tools.
• Security and governance – how AI accesses data, systems, and workflows.
• Cost management – hardware refresh cycles, memory requirements, energy use, and the balance between local and cloud AI spend.
• Operational efficiency – support models, automation, and device lifecycle management.

AI readiness cannot be treated as an infrastructure decision alone. Many organizations are focused on private cloud, backend capacity, and centralized AI platforms, but value still depends on how well AI performs, integrates, and can be supported at the endpoint.

What organizations need next for endpoint AI readiness

AI at the endpoint will continue to evolve quickly. Hardware innovation is accelerating. Software ecosystems are adapting. And user expectations are shifting just as fast. But success will not be defined by access to the latest technology. It will be defined by how well organizations can translate that capability into a sustainable, secure, and cost-effective operating model.

That is the gap most organizations are now trying to close. And as a new generation of AI-capable devices reaches the market, that gap is becoming harder to postpone.

Our take on what NVIDIA’s announcement means for your endpoint AI strategy

NVIDIA’s announcement adds momentum and visibility to a direction the market is already heading in, and while it does not mean an immediate refresh for every organization, it does make AI-ready endpoint strategy a planning issue now.

The issue is not simply whether AI-ready devices are available. It is whether the wider environment is ready to support them. Organizations need a joined-up strategy from the data center to the end user, rather than treating AI PCs as a standalone refresh decision. We help customers align infrastructure, data, security, networking, endpoint readiness, governance, support, and lifecycle strategy so AI investments deliver value in the real world.

What we’re seeing across customers is that the groundwork for AI at the endpoint is often incomplete, especially in terms of use-case clarity, governance, support readiness, and lifecycle planning. Before investing in the next generation of devices, organizations should focus on:

• Understanding real AI use cases – where local AI will deliver measurable value, not just novelty. We offer this through our AI & Cyber Labs.
• Assessing endpoint readiness – whether current devices, memory profiles, and performance baselines can support emerging workloads, in our Next-Gen Device Lab.
• Establishing governance early – defining how AI agents access data, systems, and workflows.
• Shifting support models – moving from reactive IT to proactive, AI-aware operations.
• Creating a cost strategy – aligning refresh cycles and investment to business outcomes, not hype. We give you clarity on this with our Intelligent Refresh Program and FinOps/ITAM services.

AI-driven endpoint and infrastructure demand is also colliding with a constrained hardware market. Memory-intensive workloads are increasing device requirements just as supply pressures drive up costs, forcing organizations to rethink refresh cycles and move toward more intelligent, data-driven lifecycle strategies.

The organizations that gain the most won’t necessarily be the ones that refresh first. They will be the ones that build a clear operating model for AI at the endpoint, so when new devices arrive, they are ready to scale them with confidence.

NEXT STEPS:

Speak to an SHI expert about how we help organizations operationalize AI across the full environment, including through our Intelligent Refresh Program.

Learn more about how AI-enabled endpoints transform work in this blog.

When organizations can measure something, there is always a temptation to turn it into a competition. This is showing up in AI adoption too, as some organizations use token-based scoreboards to encourage experimentation and track usage, often called ‘tokenmaxxing’. But in FinOps-for-AI, usage alone is not the goal.

Recently, the tech world has been talking about internal dashboards that give the highest rankings to people who burn the most AI tokens. Tokens are the unit many models use to meter usage, so they map quickly to cost. While token scoreboards can feel exciting to some leaders, they’re mildly terrifying to anyone responsible for budgets. Reports about internal token leaderboards also describe how easily the metric can be inflated if people leave agents running or chase rank for rank’s sake.

Adoption matters. The question is what you reward while you are chasing it.

AI tokens measure usage, not business value

FinOps teams care about unit economics because they help connect consumption to cost, accountability, and business outcomes.

Tokens are simply a meter of compute usage for many LLM experiences. A token meter is helpful because it is measurable and ties to spend. But a meter is not the thing you are trying to achieve.

If you measure AI success as “maximum tokens,” you are essentially rewarding consumption. It is the same logic as celebrating the department that used the most ride-share credits, printed the most pages, or booked the most flights. You will get more of the thing you reward, but not necessarily more of what you want. We’ve been here before, recently. It wasn’t that long ago since the guidance was “move everything to the cloud.” We’ve learned so much about measuring cost per outcome, governance, and reporting through cloud. We don’t have to learn those lessons again.

The FinOps Foundation’s AI guidance is clear. AI introduces new usage patterns and cost drivers, and teams need to track and review AI costs and usage while aligning them to business outcomes. In other words, cost visibility is essential, but it should be pointed toward value, not volume.

Why organizations use token leaderboards to drive AI adoption

Let’s be fair to the impulse. If you are trying to make AI a normal part of work, you often hit these blockers:

• People do not know what to use it for
• People worry about doing it wrong
• People assume it is for engineers
• People avoid experimentation because it feels unproductive

A scoreboard can feel like a quick cultural shortcut. Some coverage of the broader token leaderboard trend frames it as an attempt to normalize experimentation and accelerate comfort with new tools across a company.

From a change-management angle, that can work.

From a FinOps angle, it comes with a predictable side effect: if you pay for tokens, and you reward token burn, you will buy a lot of tokens.

Why token-based AI KPIs create waste and distorted incentives

Any metric tied to recognition or reward will shape behavior, regardless of whether it reflects real business value.

Several reports about internal token leaderboards point out that usage-based rankings can be padded by running agents longer than needed or by using models in inefficient ways, because the scoreboard is tracking activity, not effectiveness.

If you build a token Olympics, do not be shocked when people start training for the wrong event.

And this is where the FinOps lens matters: AI costs are unusually good at sneaking into places budgets are not prepared for. AI adoption spreads fast, crosses team boundaries, and can show up as just a little usage in dozens of places until it becomes a big new line item all at once. The FinOps Foundation describes AI spend as involving new stakeholders and new usage metrics, and calls for disciplined tracking, allocation, and governance.

A better FinOps-for-AI approach to measuring adoption

If you want to gamify AI adoption, great. Just do it in a way that rewards what you want.

Here are four scoreboard upgrades that align better with FinOps principles and still let people compete.

1. Compete on value per token, not tokens per human

Tokens can stay in the picture, but they should be the denominator, not the headline.

Examples of better scoreboard metrics:

• Cost per support ticket resolved or deflected
• Cost per proposal delivered
• Cost per marketing asset shipped
• Cost per engineering hour saved in a sprint

This fits the FinOps guidance that AI cost management should connect real-time cost monitoring to business outcomes and introduce meaningful KPIs for AI workloads

2. Reward efficient AI model choice, not just usage

In many organizations, the difference between a manageable AI bill and a surprise invoice is not whether AI was used. It is how it was used.

Celebrate behaviors like:

• Choosing an appropriate model tier for the task
• Reusing context and caching outputs when possible
• Keeping prompts tight and purposeful
• Avoiding unnecessary re-tries and runaway loops

These are all consistent with the FinOps theme that AI cost control requires visibility, optimization, and practical guardrails around usage and allocation.

3. Use showback and cost attribution before public AI scoreboards

A public leaderboard is social pressure. Sometimes that is motivating. Sometimes it creates weird incentives.

FinOps typically starts with transparency and ownership:

• Show costs by team, by use case, and by environment
• Use showback before chargeback
• Make it easy for teams to see what changed and why

The FinOps community emphasizes that AI introduces allocation complexity and that showback and attribution are foundational capabilities for governance.

How to put AI cost guardrails in place without slowing innovation

The goal is not to stop experimentation. It is to keep experimentation from becoming accidental production.

Practical guardrails that align with FinOps-for-AI best practices include:

• Quotas by environment (dev, test, prod)
• Defaulting to lower-cost models, with exceptions for specific needs
• Tagging or attributing AI usage to a use case, team, or product
• Budget alerts and anomaly detection for AI usage spikes

These map directly to the FinOps Foundation’s AI recommendations around tracking, quotas, tagging, and ongoing review of AI costs and usage.

The goal of FinOps for AI: predictable spend and measurable value

The point is straightforward: A company can be all in on AI and still be disciplined.

In fact, disciplined organizations usually move faster because they can scale what works without fear. When teams can see spend clearly, attribute it cleanly, and tie it to outcomes, they stop debating AI in theory and start managing it as a normal part of operations.

That is the heart of FinOps: collaboration, ownership, and decisions guided by business value. Microsoft’s FinOps guidance makes the same point: as organizations adopt cloud-native solutions like AI, they still need data-driven decisions, cross-functional accountability, and a clear focus on business value.

So if you want a leaderboard, make it one that nudges the organization toward value:

• Highest measurable impact
• Best cost efficiency
• Most reused components
• Most business process cycle time reduced

You can still have badges. You can still have fun. You can still create momentum. Just do not accidentally create a game where the winning strategy is to spend more!

A practical FinOps-for-AI operating model

If you want a practical playbook, here is one:

Step 1: Make usage visible
Track AI costs and usage in a way teams can understand, ideally tied to organization units or use cases.

Step 2: Add lightweight guardrails
Quotas, tagging, and basic policies that prevent runaway spend while teams explore.

Step 3: Shift the metrics to outcomes
Move from how much did we use to what did we get. Start managing cost per outcome.

Step 4: Scale what works and optimize the rest
Treat AI like any other spend category: allocate, govern, and continuously improve.

NEXT STEPS:

Assess your AI spend before it scales out of control.

Speak to an SHI expert about building FinOps practices for AI, and get the visibility, guardrails, cost attribution, and cost-to-value metrics needed to support responsible AI adoption.

You’re in a room where time matters.

Screens are filled with numbers and maps. People are calling things out. A few systems get extra attention because they’re the ones you can’t afford to get wrong.

Every decision depends on one assumption: the information in front of you can be trusted.

That’s what Artemis II was built to test— a crewed mission that launched on April 1, 2026, designed to confirm that spacecraft systems operate as expected in deep space before NASA moves to the next phase.

Federal agencies face that same kind of moment, just without the rocket.

Across government, data now shapes how benefits are delivered, how funding is allocated, how risks are assessed, and how AI tools are used. Leaders know the data exists — the question is whether they can act on it. Which numbers are official? Who’s responsible for them? And when something changes, how quickly does everyone else find out?

These are practical questions. And answering them has pushed data governance into daily mission work.

What makes governance feel urgent right now

AI adoption in government is moving quickly, raising the stakes on data quality, traceability, and access. In a July 2025 report, GAO found that reported AI use cases across selected agencies nearly doubled from 571 in 2023 to 1,110 in 2024, and generative AI use cases jumped about ninefold (32 to 282).

At the same time, expectations for how agencies manage and publish data assets continue to mature. OMB’s Phase 2 Evidence Act guidance (M-25-05) reinforces requirements for data inventories, metadata, and management practices that support access and evidence-building while maintaining appropriate safeguards.

Put simply: more use, more visibility, and less tolerance for uncertainty.

Flight plan: four governance moves that help agencies stay on course

If governance is going to accelerate missions, it has to show up where the work happens — in access decisions, shared definitions, accountability, and safeguards that don’t arrive after the fact. These four moves are built for that.

Move 1: Build a data council that makes decisions

Many agencies already have a data council. The difference between a helpful council and a slow one usually comes down to two things: purpose and authority.

When a council exists mainly to review documentation and share updates, the same patterns show up: meetings happen, but decisions don’t; each office brings its own priorities; and no one leaves knowing what actually changes next.

A council that moves work forward is grounded in a clear, agency-level objective and led in a way that benefits the whole agency. It brings together mission owners, IT, security, and data owners, and it treats data and AI as connected work.

More importantly, it makes decisions that remove confusion: which datasets are considered official; who owns them (and what that ownership actually means); what access looks like; and what gets prioritized first.

When those decisions are made early, teams stop wasting time arguing about which number is real and start working from the same map.

Move 2: Treat inventory as the start — then make data usable

Inventories are necessary. They’re also where a lot of momentum dies.

Agencies often complete an inventory to satisfy a requirement, then struggle to turn it into something mission teams can use. The common breakdowns are predictable: poor prioritization, limited usability, and no clear ownership tied to keeping information current.

A catalog becomes valuable when it’s built for non-specialists, not just data teams. That means:

• Plain-language descriptions that explain what the data is and why it exists
• Clear ownership so people know who to contact and who updates it
• Access flags that prevent sensitive data from being distributed casually, while still making it findable to people who truly need it

Usability also depends on traceability. When teams understand where data came from, confidence in dashboards and AI outputs increases because accountability is no longer just implied.

Move 3: Focus data quality where it matters most

Data quality is where good intentions go to die if the goal is to fix everything.

A more realistic approach is to focus on high-impact data first — the datasets tied to outcomes agencies can’t afford to get wrong, such as grants, benefits, and eligibility decisions. That focus keeps governance practical and ties quality efforts directly to mission results.

This isn’t a theoretical issue. A 2025 IBM Institute for Business Value survey found data quality remains a top operational priority, with organizations estimating significant financial losses tied to poor data.

When teams trust the data they’re using, the impact is immediate. Decisions move faster because inputs aren’t constantly questioned. Duplicate reporting and manual reconciliation drop because definitions and sources are consistent. Audits become easier to support because the underlying data is easier to explain. And AI adoption feels less risky because teams aren’t feeding models with outdated or unclear inputs.

Move 4: Build safety into the process from the start

In federal environments, it’s no longer realistic to separate governance from security. Governance determines who can access data, how it’s shared, and how sensitive information is protected.

AI makes that inseparability more obvious. People tend to trust AI outputs quickly, and if those outputs are based on poorly understood data, they can spread confusion just as quickly.

The issue is often timing. Controls are introduced after data has already been shared, after AI tools are already producing outputs, or after audits force a response, leading to rework and stalled progress.

Building safety early looks like:

• Classifying data and defining access rules from day one
• Involving privacy and security teams during development
• Treating safeguards as part of the workflow, not a separate approval layer

This is how agencies protect mission speed over the long term — by reducing late-stage slowdowns.

Where SHI helps agencies turn intent into action

Many agencies aren’t stuck because they lack policies. They’re stuck because execution is hard in real environments — siloed departments, unclear ownership, limited time to work through years of accumulated data, and modernization constraints that make large-scale change unrealistic.

SHI helps agencies build momentum without disrupting mission systems by working within what already exists and layering governance in practical, incremental ways.

That typically looks like:

• Turning decisions into action. Defining ownership “swim lanes,” decision rights, and council structures that keep data, AI, and mission leaders aligned.
• Making data usable. Converting inventories into navigable catalogs with practical metadata and clear traceability.
• Prioritizing what matters. Focusing on high-impact datasets first so progress is visible and sustainable.
• Building safety in early. Establishing classification and access rules upfront so AI doesn’t amplify sensitive or unclear data.
• Connecting governance to AI readiness. Aligning governance work to what agencies are trying to do now — scale analytics and AI responsibly, with teams moving together.

Data governance doesn’t accelerate missions because it adds process. It accelerates missions because it reduces uncertainty.

When teams agree on what data means, who owns it, how it can be used, and how it’s protected, leaders spend less time reconciling conflicting answers — and more time acting on decisions that hold up under scrutiny.

NEXT STEPS

Want to see what data governance looks like in your agency’s environment? We’re here to help you work through it — reach out to our team to start the conversation.

Looking for more on how federal teams are approaching AI and modernization? Read our latest perspective.

At the recent SHI Spring Summit, one theme came up repeatedly: organizations are still leaving too much money on the table across cloud, SaaS, and software licensing.

Recent industry research backs that up. Flexera’s 2026 State of the Cloud Report found that wasted cloud spend rose to 29 percent in 2026, the first increase in five years. Zylo’s 2026 SaaS Management Index found that only 54 percent of SaaS licenses are actively used.

This waste is not just the result of missing tools or inattentive teams. The deeper issue is how technology spend is managed.

Why CFOs struggle with cloud, SaaS, and software spend

Fundamentally, it is that most organizations still manage SaaS, cloud, and software renewals in silos, with different data, different workflows, and different owners.

CFOs are being asked to protect margins while funding growth, AI, and modernization. When costs move fast and ownership is unclear, finance is left reacting after decisions have already been made. Without a clear view across the estate, it becomes much harder to shape spend early, before commitments are made and value is lost.

Where cloud, SaaS, and software spend is wasted

Technology waste rarely shows up as one dramatic line item. More often, it appears in less visible ways, including:

• Unused licenses
• Premium SaaS tiers assigned to the wrong people
• Idle non-production cloud environments
• Duplicate tools
• Contracts renewed without enough time or evidence to challenge pricing

On their own, these issues may look manageable. Together, they put real pressure on margin.

Some technology spend should increase because it supports resilience, customer experience, or speed. The problem is that many organizations still struggle to separate high-value spend from low-value drag. When that line is blurry, every budgeting cycle becomes harder than it needs to be and it becomes more difficult to confidently fund the priorities that matter most.

Why one-time cost optimization efforts do not last

Most organizations have had a savings win somewhere. A renewal review uncovers shelfware. A SaaS rationalization exercise removes unused apps. A cloud optimization sprint cuts the monthly bill. Those wins matter, but they are often treated as isolated projects instead of part of an ongoing discipline.

Without clear ownership, repeatable workflows, and regular review points, the same waste comes back. SaaS sprawl returns. Cloud costs drift upward. Software renewals become reactive again. Finance gets a short-term win, but not a lasting change.

Why technology spend management breaks down

If waste keeps reappearing across cloud, SaaS, and software licensing, the real issue is not the technology itself. It is the operating model around it.

Decentralized purchasing, incomplete inventories, fragmented usage data, weak ownership, and tactical renewals all point to the same gap: the organization has not built a durable way to govern technology spend across its lifecycle.

That matters to CFOs because weak operating models do more than create waste. They reduce forecasting confidence, weaken negotiating leverage, and make it harder to assign accountability. They also make it harder to move money toward the priorities the business really cares about.

What better technology spend management looks like

A better model starts with a simple shift. Technology spend cannot be managed well as a series of one-off interventions. It must be managed as an ongoing business process. That means moving from administrative renewal behavior to continuous decision-making.

In practice, that means better visibility, defined lifecycle workflows, and shared accountability. Finance, IT, procurement, engineering, and application owners need a shared view of what the business owns, what it uses, what it costs, and who is accountable for it. When people are working from the same information, spend becomes easier to manage and easier to justify. This becomes even more important as organizations invest in areas where costs move quickly and value is harder to measure, requiring stronger governance and earlier decision-making.

The goal is not to drive every cost as low as possible. It is to separate spend that is creating value from spend that is surviving on inertia. That is a better financial discipline, and a more useful way to protect margin.

How to create value across the technology lifecycle

The SHI Spend Optimization Delivery Framework – creating value across the technology lifecycle

For CFOs, the real opportunity is not just to reduce costs at renewal. It is to make sure technology investments keep earning their place from sourcing and provisioning through day-to-day use, optimization, renewal, and retirement. When organizations focus only on the purchase point or the contract deadline, they miss too many chances to improve value in between.

That is true across cloud, SaaS, and software licensing. Better intake decisions reduce duplication before it starts. Clear ownership and usage data helps teams make smarter calls while spend is active. Stronger optimization and renewal discipline protect budget before contracts roll forward. In practice, that starts to look different across each part of the technology estate.

How to improve cloud, SaaS, and software licensing spend

In cloud, cost needs to be visible early enough to influence behavior. Rightsizing, automated shutdowns, cleanup of unattached resources, and stronger retention policies all help, but only when they are part of normal operations.

In SaaS, visibility alone will not stop sprawl. Organizations also need approved catalogs, clear ownership, regular usage reviews, and renewal playbooks that begin well before the deadline.

In software licensing, better usage data, benchmarking, harvesting, tier optimization, and portfolio rationalization improve both savings and negotiating leverage. When vendors are met with evidence instead of urgency, the conversation changes.

Which technology spend metrics matter most

The most useful metrics are not just the ones that show whether the bill went down. They show whether the organization is getting better at making intentional decisions. That includes total spend and spend growth rate, but also forecast accuracy, the share of renewals with optimization action, the amount of underused spend being removed, and how clearly ownership is assigned across teams.

The right question is not whether spend increased. It is whether the increase was intentional, visible, and tied to value. That gives CFOs a stronger basis for decision-making than cost reduction alone ever will.

Why technology cost optimization matters now

Technology optimization is not just a cost exercise. Done well, it creates investment capacity. Money recovered from idle cloud resources, underused licenses, duplicate SaaS applications, and poorly prepared renewals can be redirected toward AI, modernization, resilience, and growth.

In short, optimizing technology spend now means organizations can fund what comes next.

How SHI helps improve technology spend management

If your organization is trying to get more value from cloud, SaaS, and software licensing, the right place to begin is with an assessment. It gives you a clear view of your current position, your appetite for improvement, and the gaps between where you are and where you want to go.

It looks across people, process, technology, and commercial arrangements to help chart a practical path forward. For CFOs, that means a stronger foundation for reducing waste, improving control, and funding what comes next.

NEXT STEPS

Start with an assessment to see where value is being lost today, and what it will take to recover it. Connect with an SHI expert today.

NVIDIA has delivered decades of technological innovation, and 2026 has been no exception. Alongside releases like the NVIDIA NemoClaw blueprint and continued investment in AI infrastructure, the company has also introduced something CIOs will find equally important: a simple metaphor that defines the full AI stack.

In what CEO Jensen Huang calls a “five-layer cake”, NVIDIA outlines the layers required to sustain AI at scale. Starting from the foundational layers that power intelligence, the metaphor works its way up to the applications users interact with every day.

Speaking at the 2026 World Economic Forum in Davos, Switzerland, Huang emphasized that while enterprise value is realized at the application layer, AI outcomes require a full-stack outlook. As he put it, “This application layer could be in financial services, it could be in healthcare, could be in manufacturing. But you can’t build that top layer without everything underneath it.”

For CIOs and business leaders, this structure can land less as a metaphor and more as a reality check. An organizational appetite to capture AI’s benefits — often driven by the latest tools and announcements — can pull focus to the top of the stack and unknowingly obscure the critical foundational work required across every layer. Over-indexing on the application layer can lead organizations to pay a premium and miss chances to achieve the same outcome more efficiently by making smarter decisions lower in the stack.

How can CIOs widen their view and operationalize the full AI stack?

This is where the SHI and NVIDIA partnership moves the five-layer structure from theory to execution. Within it, SHI acts as a unifying force across the AI stack, supporting organizations to make informed decisions at any layer while maintaining a full-stack view.

Through a well-defined, three-phase process called Imagine. Experiment. Adopt., SHI works alongside organizations’ business and technical stakeholders to optimize each layer of the AI stack through a consultative, lab-enabled approach.

Each phase brings the right layers of the stack into focus at the right time, before early decisions become too expensive to unwind. Backed by NVIDIA’s technology foundation, SHI enables AI layers to connect end-to-end through vendor partnerships and comprehensive wraparound services.

Now, let’s take a closer look at how SHI’s Imagine. Experiment. Adopt. approach enables organizations to create cohesive, scalable, and secure enterprise AI environments.

Imagine.

At this phase, SHI focuses on helping organizations determine where to start. This begins with identifying the critical business objective and where well-designed AI solutions can address execution challenges. Through consultative discovery and readiness assessment with SHI’s AI experts, CIOs define what success must deliver, then identify the capabilities required across the full AI stack. This approach helps avoid diving straight into tool-led decisions that can lead to rework and unnecessary expense later.

What we work through during the Imagine phase:

• The top business objectives the organization is trying to achieve
• Which workflows or functions are most burdened by manual effort
• Whether the focus is solving a specific problem or enabling a broader AI capability
• What success looks like in measurable business terms

By engaging in the Imagine phase, SHI helps organizations establish the architectural and operational clarity needed to move forward with a full-stack roadmap that lays the foundation for experimentation and adoption.

Experiment.

Experiment is where organizations test whether the outcomes defined in the Imagine phase can hold up under real conditions. Through controlled pressure-testing and proof-of-concept validation in SHI’s AI & Cyber Labs, teams assess value, performance, risk, system design, and cost before moving AI initiatives into their own production environments.

What we work through during the Experiment phase:

• Whether any existing services break or degrade when usage scales beyond a pilot
• What operational dependencies or handoffs AI introduces
• How sensitive results are to model choice, prompting, or data quality

In this phase, NVIDIA’s five-layer structure serves as a diagnostic lens for working with SHI. Using controlled testing to surface layer gaps early, organizations know what holds up in practice, what needs adjustment, and whether an initiative is ready to move to adoption.

Adopt.

Adopt is where AI moves from a tested initiative into a production-ready operating reality. At this phase, the challenge is less about technology choices and more about running AI reliably across people, processes, and platforms inside the organization.

What we work through during the Adopt phase:

• How users will be trained, supported, and enabled to use AI confidently in their roles
• How ongoing value and ROI will be measured beyond the initial rollout
• How infrastructure, cost models, and support teams will need to adapt as adoption grows

By the end of the Adopt phase, organizations move beyond isolated deployments and establish AI as a durable operating capability. AI is embedded into day-to-day workflows, supported by trained teams, governed responsibly, and operated at scale.

How the AI stack comes into focus across SHI’s three-phase approach.

Taken together, these phases help leaders slow down, sequence decisions deliberately, and avoid the costly downstream trade-offs that occur when AI initiatives move too quickly to the top of the stack.

Consider a common enterprise AI use case like modernizing an external customer service capability to increase issue resolution speed and improve customer experience. While the idea begins at the application level, delivering it in practice quickly pulls on every layer of the AI stack.

Imagine – outcome-first thinking focused on readiness and prioritization

• Applications layer: Define which customer interactions AI should handle, what success looks like, and where escalation or human interaction is required.
• Models layer: Establish build-versus-buy direction, data sensitivity boundaries, accuracy expectations, and governance intent.
• Infrastructure, chips, and energy layers: Set early guardrails for security, integration, capacity, and cost before tooling decisions are locked in. Define power and cooling requirements needed to support the anticipated workload. These directional decisions shape everything that follows and inform which technologies and partners are a good fit.

Experiment – controlled testing to validate and deliver an MVP prototype

• Applications layer: Test whether prototypes fit into real customer workflows without breaking handoffs or introducing friction.
• Models layer: Pressure-test reliability, grounding quality, and hallucination risk under realistic conditions, including proof-of-concept scenarios. This is typically where short-listed models, platforms, and components from the broader vendor ecosystem are validated before longer-term commitments are made.
• Infrastructure, chips, and energy layer: Validate how latency, security, chip performance, power draw, and cost behave as traffic increases.

Adopt – production-ready execution grounded in security, enablement, operations, and ROI

• Applications layer: Embed AI into frontline workflows as a production-ready operating capability, defining ownership, and training teams so the system is trusted and used consistently.
• Models layer: Shift from selecting to stewarding, while monitoring guardrails and drift management over time.
• Infrastructure, chips, and energy layers: Ensure familiarity with system operations, streamline management of performance and capacity, and optimize cost savings as part of ongoing operations, tying consumption directly back to business value.

SHI turns the NVIDIA AI stack into an enterprise operating reality

NVIDIA’s five-layer AI structure makes it clear that AI outcomes will only hold when every layer beneath it is intentionally built, integrated, and governed. Huang recently said, “AI is no longer a single breakthrough or application — it is essential infrastructure. Every company will use it. Every nation will build it.”

The promise of AI is not just in the technology itself, but in how well it is operationalized. Through Imagine. Experiment. Adopt. SHI can help organizations turn AI ambition into real enterprise transformation and build the foundation to execute and adopt at scale.

NEXT STEPS:

Ready to discuss how your organization can get AI initiatives off the ground?

Connect with one of SHI’s AI experts now and learn more about our NVIDIA partnership.