In the previous article, I argued that hybrid intelligence makes the device part of enterprise AI architecture. Application teams already distribute work across the endpoint and cloud based on what each step requires. Here, I look at how IT teams can decide which steps run on a Surface device, which stay in the cloud, and when local execution may help reduce the cost of getting the work done.
Surface makes the endpoint side of that choice practical. Across the portfolio, Surface devices support local AI work ranging from NPU-accelerated app features to larger models used by developers, engineers, and creative professionals. Each component suits different work: CPUs handle general processing, GPUs accelerate parallel workloads including AI, NPUs provide power-efficient AI acceleration, and memory determines the size of workloads the device can support. Organizations can match capacity to each role, running device-appropriate steps locally and sending more demanding work to the cloud.
Work tied closely to device data and applications may run well on the endpoint. To compare it with the cloud option, teams can first look at how the work runs on the device, including application responsiveness and resource use. They can then assess the quality of the results and how long employees spend correcting them, along with the cost of falling back to a cloud model when the local model can’t complete a task reliably.
Measure the workflow, not the prompt
An employee may see one AI action while the application handles several steps behind the scenes. Each operation can use a separate model with its own cost and performance profile.
Consider a developer who asks an assistant to fix an error. The application may inspect the code, propose a change, draft a test, and summarize the result before the developer sees an answer. Looking only at the prompt hides much of the work and makes the real cost hard to see.
One practical place to begin is a high-volume feature that already has usage data. For an application built in Microsoft Foundry, the Monitor view and its connected Azure Monitor Application Insights resource can show token use, latency, errors, and success rates in production. Azure Cost Management shows the related spend. Those numbers set a baseline for a local or hybrid test on representative Surface devices, and the test shows whether the change is worth making.
Surface can improve the value of endpoint capacity
Cloud and endpoint costs work differently. Cloud AI may be billed through metered APIs, subscriptions, or reserved capacity. As usage grows, recurring cloud AI costs can rise quickly. For high-volume workloads that perform reliably on the device, shifting repeatable work to the PC has the potential to reduce recurring cloud spend.
When organizations purchase Surface devices for employees, they also get local compute capacity for AI work. Local models can run alongside employees’ applications and data without a new cloud inference charge every time. Those models also share the device’s resources with everything else employees run. A local model only makes sense if it leaves enough capacity for the rest of the employee’s work.
Volume creates a crossover point
Local AI requires an upfront investment. As task volume grows, that investment is spread across more completed work, resulting in lower cost per task. This is the opposite of cloud AI, where usage-based billing typically adds cost with each task.
When comparing cloud and local models, this creates a crossover point where local AI becomes less expensive overall than cloud AI for a given workload. The location of this depends on several factors, such as:
- The relative quality of the output between cloud and local AI and the time employees spend correcting it
- The impact on device performance and whether it affects other work
- The frequency of local runs failing and “falling back” to the cloudLocal AI starts with higher fixed costs but can have a lower incremental cost per completed task. As task volume grows, the cumulative local cost rises more slowly than usage-based cloud cost. The two become equal at the crossover point; beyond it, local execution may cost less overall
If the local output is good enough for the task, doesn’t negatively affect other work, and runs on the device reliably, the crossover point will be reached sooner.
The Surface portfolio gives IT several levels of local AI capacity. Teams can choose the device that fits the role instead of paying for the same level of hardware everywhere. Surface Pro and Surface Laptop support Copilot+ PC experiences and NPU-accelerated applications. 1 Surface Laptop Ultra extends the portfolio to more demanding local AI workloads with its NVIDIA RTX GPU, up to 128 GB of unified memory, and CUDA support.2,3 Its additional memory can accommodate larger models and working datasets, while CUDA supports a broad ecosystem of GPU-accelerated AI tools.
A developer workflow shows the crossover
A frequent developer task is a good place to test the value of local AI. A useful evaluation pairs a specific job—such as resolving a familiar coding error—with baseline data on its current cloud cost, response time, and how often developers retry or correct the output.
The same workflow can then be tested locally or through a hybrid approach on representative hardware. Response time, memory pressure, and cloud fallbacks show whether the local path holds up while the developer runs an integrated development environment, browser, and collaboration tools.
Some steps may work well locally, such as explaining familiar errors or working with proprietary project data. Tasks requiring multi-file reasoning or enterprise data may still call for a cloud model. Developers should continue to review outputs and follow standard testing and security practices.
The same evaluation can apply to other repeatable workflows with clear success criteria, such as answering questions from internal documents or generating images. The results show whether the work can run locally and which Surface device it needs.
Management costs shape endpoint economics
Local AI is not free simply because it runs on hardware the organization owns. At enterprise scale, managing devices requires staff time and budget, and those costs belong in the comparison with cloud inference.
Surface can reduce some of that operational complexity by fitting local AI into existing Windows endpoint tools and processes. Secured-core PC protections, Microsoft Pluton, and BitLocker support device and data policy, while Microsoft Intune, Device Firmware Configuration Interface (DFCI), and Surface firmware management support day-to-day administration.
These capabilities make local AI more manageable, but they do not eliminate ongoing servicing and support costs. A wider rollout also brings decisions about data access, model updates, and cloud connections, all of which add to the cost of local AI.
Bring the endpoint into AI planning
Hybrid intelligence gives cloud and endpoint resources complementary roles. Surface is designed with layered security and management built by Microsoft, so teams can run work close to employees with visibility and control. Cloud models are available when a task calls for greater capability, context, or scale. Applications can use both paths within the same Windows workflow.
The best place to run a model is wherever it gets the job done at the right cost. As AI spreads work across devices and the cloud, workload requirements should inform device selection and refresh planning. The PC is now part of the enterprise AI system, and endpoint strategy should reflect that role.
1Copilot+ PC features available only on select configurations with 16+GB RAM; additional system requirements may apply. See more at Windows 11 Specs and System Requirements | Microsoft Windows.
2 Memory varies by configuration. 128GB total system memory in a unified memory architecture shared between CPU and GPU. Maximum amount addressable by the GPU depends on system configuration and workload and is less than the total.
3CUDA is a registered trademark of NVIDIA Corporation.
Microsoft