The Cloud Just Lost Its AI Monopoly

For most of the current AI cycle, companies have built products around a simple assumption: the intelligence lives somewhere else. An application sends a request to OpenAI, Anthropic, Google, or another provider, the model processes it in the cloud, and the result comes back through an API. That architecture made powerful AI accessible without requiring companies to operate the underlying models themselves, but it also tied AI products to cloud inference, external infrastructure, and usage-based costs.

Meta is now pushing another approach. This week, the company released Muse Glimmer, an open model designed to be lightweight enough to run locally on personal computers while supporting agentic tasks. Meta is also giving developers access to the more powerful Muse Spark 1.2 and says it plans to release the model's weights later.

Muse Glimmer may be one model release, but the broader signal is more interesting. Capable agents are beginning to get another place to run: directly on the devices where the work happens.

Cloud infrastructure has been fundamental to the AI boom because the most capable models require enormous amounts of compute. Companies could access that intelligence through an API instead of building the infrastructure themselves, which dramatically lowered the barrier to adding AI to products.

Local models create another option. An agent running on an employee's laptop or company-controlled hardware can potentially work with files, applications, and other local resources without sending every interaction to an external model provider. For workloads that do not require frontier-level intelligence, that could change how companies think about privacy, latency, connectivity, and inference costs.

The result will probably be a more distributed AI stack. Some work will continue going to the most capable cloud models, while other tasks could happen closer to the user.

Lightweight and local models are not new. What makes this development worth watching is the combination of local execution and agentic behavior.

An AI model that generates text locally has limited operational reach. An agent that can reason about a task and interact with files or applications on the same device has a much larger role inside the software environment.

That opens possibilities for products that operate closer to where data is created. A local agent could work with sensitive company documents, interact with desktop applications, or complete routine tasks without making a cloud API call for every step. Companies could also design systems that route tasks dynamically, using local models for routine work and sending harder problems to more capable cloud models.

For the past few years, model selection has dominated AI architecture discussions. Engineering teams compared providers based on intelligence, latency, context windows, reliability, and price. Local agentic models add deployment location to that decision.

A company building an AI product may eventually decide where each type of work should happen based on the sensitivity of the data, the complexity of the task, available hardware, latency requirements, and inference economics. A privacy-sensitive workflow could remain on company-controlled hardware, while a difficult reasoning task gets routed to a frontier model in the cloud.

This creates a more complicated architecture, but it also gives companies greater control over how they consume AI.

The strongest opportunity may come from combining local and cloud models rather than choosing one deployment model for everything.

Routine agent activity can generate enormous numbers of model calls as systems inspect information, reason through intermediate steps, use tools, and verify their own work. Sending every step to a premium cloud model can become expensive, especially as agents run for longer periods.

Hybrid systems could route simpler work to smaller local models and reserve expensive cloud inference for tasks that genuinely require it. The same architecture could help companies keep sensitive data closer to the device while still giving users access to frontier models when necessary.

Economics becomes particularly interesting at scale. Companies gain another lever for controlling how much AI costs without necessarily reducing how often employees or customers can use it.

This issue of Redeployed is brought to you by Tecla Labs: As AI models gain the ability to run locally, companies have more choices about where intelligence lives across their products. Building hybrid systems that combine local models with cloud AI requires deep experience with model routing, infrastructure, security, performance, and how those pieces work together in production. Tecla Labs designs and builds AI systems for companies that need this expertise now, so teams can ship flexible AI products without locking themselves into a single deployment model.

Moving inference onto devices also moves responsibility onto the companies deploying it. Teams need to account for hardware differences, model updates, performance, security, and the operational complexity of supporting AI across potentially thousands of machines.

Local execution also does not guarantee security. An agent with access to files and applications can create serious risks if its permissions are poorly designed, regardless of whether the model runs on a laptop or in a data center.

Capability remains another constraint. Smaller local models may be sufficient for narrow or repetitive tasks while struggling with work that frontier cloud models handle reliably. Cheap cloud inference could also remain economically superior for many workloads once the full cost of maintaining local infrastructure is considered.

A hybrid AI stack requires engineering teams to think beyond model integration. They need to understand how workloads should be routed, where data can travel, what devices can support, and how model behavior stays consistent across different execution environments.

That makes infrastructure and systems architecture increasingly important to AI development. Teams may need routing layers that select models based on task complexity, privacy requirements, cost, or latency. They will also need observability across the entire system so they can understand where work is happening and how each deployment option performs.

The engineering challenge becomes deciding how to distribute intelligence across the stack.

Muse Glimmer does not prove that enterprise AI is about to leave the cloud. Frontier models still have significant capability advantages, and the economics of local execution will vary widely by workload.

But Meta's release points toward an AI market with more deployment choices. Companies may increasingly combine frontier APIs, open models, local execution, and company-controlled infrastructure depending on what each workflow requires.

That would make AI architecture considerably more flexible and potentially more defensible. Companies would have greater control over their data, costs, performance, and dependence on individual model providers.

Cloud AI made intelligence available everywhere through an API. The next phase may make intelligence part of the software itself.

Connect With Other Technology Leaders

If you want to connect with other technology leaders having real conversations about AI and how it is changing business, check out GILD Curated Circuit.

More to come…

✔️ How Apple’s on-device AI works, and what it changes for your users — Courier

✔️ What smart people are saying about Meta’s open-weight model Muse Glimmer and Mark Zuckerberg’s AI manifesto — Business Insider

✔️ The AI pricing paradox: AI PCs might be the answer to spiralling cloud costs, but cloud isn't going anywhere — TechRadar