The frustration emerging from the technical community is clear: the Mac M5 Ultra with 256GB of unified memory is on the market, but anyone looking for credible numbers to evaluate Apple's platform against Nvidia GPU-centric workflows finds only influencer videos testing 8-billion-parameter models. This is not a complaint for its own sake: it is the symptom of an information gap that has concrete consequences for those evaluating on-premise LLM deployments.

Apple has built a machine with an amount of shared VRAM that has no direct equivalent in the Nvidia workstation world at the same price. The promise is to run large models locally, without resorting to cloud clusters or multiple GPUs connected via NVLink. But the promise alone is not enough: repeatable measurements are needed on inference, throughput, latency, behavior with different quantization levels, and above all with models exceeding the 10-20 billion parameter threshold. Tests with 8b models, often run in uncontrolled conditions and without configuration reporting, do not allow estimating the real Total Cost of Ownership nor comparing Apple's unified architecture with the CUDA ecosystem, which has years of consolidated benchmarks.

The issue is not only technical but also about incentives. Content creators have an interest in producing quick and spectacular videos, not in running tedious benchmark suites that require weeks of work. Companies that could fund independent tests, on the other hand, tend to focus on hardware already adopted in data centers, where Nvidia dominates. The result is a vicious circle: Apple silicon is not taken seriously because benchmarks are missing, and serious benchmarks do not arrive because the market does not perceive Apple silicon as a platform for enterprise workloads.

This has second-order implications for those thinking in terms of data sovereignty and local control. An organization wanting to keep its LLMs self-hosted on Apple hardware would need reliable data to justify the CapEx investment before the CFO or CISO. Without numbers, the choice almost automatically falls on Nvidia solutions, not because they are objectively superior in every scenario, but because the perceived risk is lower. In other words, the absence of benchmarks penalizes the diversification of hardware offerings for local inference, further consolidating Nvidia's position.

There is also an aspect related to the nature of the Apple ecosystem itself. The Metal framework and libraries like MLX are improving, but they remain far from the maturity of CUDA in terms of tools, profilers, and support for advanced optimizations. This does not mean that the Mac M5 Ultra cannot be an interesting machine for certain workloads; it means that until independent tests emerge on 30-70 billion parameter models with INT8 or FP16 quantization and stable tokens/sec metrics, the debate will remain dominated by impressions rather than evidence.

For those evaluating on-premise deployment, the implicit advice is not to ignore the platform, but to treat the absence of data as a risk signal. Analytical frameworks exist — such as those discussed on AI-RADAR in the section dedicated to cloud vs on-premise trade-offs — that help structure the evaluation without relying on promotional demos. But the responsibility to fill the information gap cannot fall only on buyers: a more serious, independent, and repeatable benchmark culture is needed, capable of treating Apple hardware with the same rigor reserved for Nvidia GPUs.

The Mac M5 Ultra with 256GB has the potential to shift some inference workloads from the cloud to desktops or small local servers, but as long as the numbers are missing, that potential will remain an unverified bet. And in a market where every euro of CapEx must be justified, an unverified bet is often a losing bet.