Meta’s new local model forces enterprises to recalculate AI costs and ROI

Meta has introduced a new AI model called Muse Glimmer, which is optimized to run on a PC or Mac using a single GPU for always-on local workflows. However, its hardware requirements, such as a GPU with at least 24GB of VRAM, may make it challenging to justify for deployment at scale. While enterprises show interest in running models locally, determining if switching from cloud to local makes financial sense is complex due to uncertain RAM and cloud pricing costs over the next 12-18 months.

Analysts and consultants agree that the economics and architecture of AI are increasingly moving towards the edge, but calculating ROI for local AI deployment is not a simple apples-to-apples comparison. Quantized models like Muse Glimmer, compressed to roughly 4-bit precision, are cheaper than full-precision cloud models, but the comparison may not be entirely fair because cloud APIs typically serve full-precision models.

Moreover, local deployment introduces additional costs like hardware, power, support, and hardware refresh cycles, which cloud providers typically handle.

Enterprise CISO Mike Wilkes believes that Meta's move gives enterprises more options and that the financial comparison now involves capital expenditure that can be amortized over several years versus an effectively perpetual cloud operating expense. This could make local AI deployment more predictable and cost-effective, providing enterprises with control over model versions and offline availability.

Independent cybersecurity and risk advisor Steven Eric Fisher, however, notes that the RAM requirements are context-dependent and that practical memory needs may exceed the 32 GB available on an Nvidia RTX 5090 GPU.

Justin Greis, CEO of consulting firm Acceligence, argues that moving inference from the cloud to the endpoint does not automatically result in lower total cost of ownership. While Muse Glimmer is technically viable, the enterprise ROI still depends on factors such as variable cloud costs, deployment, endpoint management, support, security, model updates, and accelerated hardware refresh cycles for local devices.

Greis maintains that local AI deployment is a milestone, but determining enterprise ROI requires a more comprehensive evaluation.