Hybrid AI: the PC with agents curbs the token bill

From everything in the cloud to splitting the work between PC and cloud

Over the last two years the conversation about AI in business has revolved almost exclusively around the cloud: large models, APIs and pay-per-use bills. The news coming from Computerworld points to a shift in approach: providers are pushing what Microsoft calls hybrid intelligence, that is, running on the computer itself the AI tasks that do not need a cutting-edge model and reserving the cloud for what truly requires muscle.

The reason is economic. AI is getting more expensive and companies are looking to reduce token costs. The proposal is not to replace the cloud, but to move preliminary tasks to the PC and send to the cloud only those that current equipment cannot handle. Bob O'Donnell, president of TECHnalysis Research Insights, summed it up this way: we will always need cutting-edge models in the cloud for certain tasks, but there is no need to use them for everything.

Pavan Davuluri, head of Windows and Devices at Microsoft, put it at an event on October 7 with an idea worth retaining: the future of AI does not consist of having a single model, but of having the right model at the right time. Cutting-edge capability when needed and smaller, faster models when speed comes first. Satya Nadella added that this hybrid intelligence will be everywhere and will blur the line between AI in the cloud and the PC.

What an “agentic” AI PC is

Microsoft acknowledges that high cloud usage bills make companies hesitate when expanding their use of AI. Its response involves offloading to AI PCs the tasks that do not require cloud models. This is where agentic AI equipment comes in: computers capable of running agents, generating video, writing code and conversing with advanced generative AI models. If the equipment cannot handle the task, it is sent to the cloud.

The technical leap is not minor. A year ago, according to Davuluri, these capabilities were exclusive to cutting-edge cloud models; now they can be on the computer. These devices incorporate powerful GPUs, distant relatives of those in data centers. Nvidia's RTX Spark superchip, with Blackwell GPU, powers many of them, including Microsoft's new Surface Laptop Ultra. Gerardo Delgado, senior director of product management at Nvidia, explains that, unlike older AI PCs, the new ones can run full models and have enough power to do so quickly.

Analyst Jack Gold, of J. Gold Associates, provides a useful forecast for planning: over the next three or four years, between 10% and 15% of AI workloads will run on personal computers, and between 40% and 45% in the cloud. He adds that companies will need equipment with accelerated AI and that we will likely see renewed interest in buying new machines.

The confidentiality and cost argument

HP has presented the ZBook Ultra G3a mobile workstation, with an AMD GPU that, according to the company, generates free tokens and reduces cloud bills. Brian Allen, head of Global Z workstation products at HP, points to a double advantage: it can be installed on-premises, there is no need to worry about confidential information going to the cloud or about tracking tokens or receiving high bills every week.

The device incorporates Perplexity's Portable Computer service, with agents capable of linking a local model with cutting-edge cloud models. The example HP gives is clear: a PC with agents can do basic design tasks in Autodesk's Revit 3D modeling software without an Internet connection; for more advanced work, Perplexity sends the task to one of the 19 cutting-edge cloud models depending on which fits best. Allen argues that designers will probably delegate general tasks to the cloud and keep proprietary work local.

A real case: debugging code without paying for tokens

Microsoft showed how a local model can save the token costs that are usually generated in the cloud. A team could not reproduce a software bug, so GitHub Copilot selected GPT-6 Luna in the cloud to analyze the code. That model delegated creation and testing to local subagents running MAI Code 1.1 Flash on a PC to reproduce the bug. The local work was free: a subagent fed 1.6 million tokens of code and context into the local model, which generated 10,500 tokens in response, according to a GitHub product manager during the demonstration.

What this means for an SME

The quick reading would be “we need to buy new laptops.” The sensible reading is different: start thinking about your AI as a portfolio of tasks with different costs. Not everything deserves a cutting-edge model, and many repetitive operations or operations with sensitive data can stay on the user's device. This connects with what we already discussed when talking about the six or eight AI models that Jassy predicts: the key will not be having the best model, but knowing which one to use at each moment.

It is also worth not rushing to measure performance by the shine of the hardware. As we explained in GPU at 100% does not mean your AI is doing well, a saturated GPU can hide a bottleneck that has nothing to do with computing power. Before renewing your fleet, measure which tasks you consume and how much they cost you.

Practical recommendations from ForgeNEX

1. Inventory of AI tasks. Classify what your team does: quick queries, summaries, code generation, analysis of sensitive documents. Mark which ones need a cutting-edge model and which do not.
2. Cost per token versus equipment cost. Compare the monthly cloud bill with the amortization of equipment with accelerated AI. Do not buy out of fashion; buy when savings or confidentiality justify it.
3. Data policy. If local work avoids sending out confidential information, document what stays on the PC and what is sent to the cloud. Hybrid AI does not eliminate the need for data governance.
4. Test with a pilot device. Choose a specific profile (design, development, analysis) and validate for a quarter whether the local model covers its usual tasks.
5. Do not rule out the cloud. The providers' message is hybrid, not local. The cloud will continue to carry the most demanding workloads.

In short: hybrid and agentic AI is not a promise of a distant future, but a way of managing spending and data sensitivity that is already in manufacturers' catalogs. For an SME, the opportunity lies in choosing well what stays at home and what goes out to the cloud, and in not confusing renewing hardware with having an AI strategy.

Source: Computerworld. Analysis and adaptation: ForgeNEX.