Tech • AI • Robotics • Game
Apple, Nvidia, AMD: Which machine should you buy to own your own AI?
7/10TL;DR
Choosing local AI hardware now depends less on brand than on two specifications, memory capacity and memory bandwidth, with Apple, Nvidia and AMD each offering different trade-offs against cloud rental.
KEY POINTS
Two numbers now matter most
Local AI performance increasingly hinges on the size of memory and the speed at which a machine can reread it during inference. Capacity determines which models fit on the device, while bandwidth sets the theoretical ceiling for generation speed. Processor branding and industrial design matter less than these two figures for most deployment decisions.
Model size can be reduced without large quality loss
A model such as Qwen 3.8 27B occupies about 17 GB in a 4-bit quantized format, while an 8-bit version is nearly twice as large. Quantization acts like compression, often preserving similar short-task performance while reducing memory needs. That makes local deployment possible on machines that would otherwise be too small.
Memory tiers define practical use cases
With 32 GB of RAM, a user can run a smaller model or a lightweight agent. At 64 GB, a model like Qwen 27B can run smoothly. At 128 GB, there is enough headroom for larger models, long context windows and multiple agents running in parallel, which is why 128 GB has become a key threshold in current buying decisions.
Bandwidth figures differ sharply across vendors
Nvidia’s DGX Spark offers 128 GB of memory with about 273 GB/s of bandwidth. An AMD Strix Halo system has been measured around 215 GB/s. A MacBook Pro M5 Max is cited near 614 GB/s, while the M5 Ultra reaches roughly 1.2 TB/s. On paper, that should translate into much faster token generation on Apple’s higher-end systems.
Real-world speed is more complicated than theory
In practice, tests on Qwen 27B showed similar results of roughly 15 to 30 tokens per second on both a Mac M5 Max class system and a DGX Spark. The gap between theoretical bandwidth and observed output comes from software stacks, workload patterns, concurrency and model architecture. Bandwidth sets a ceiling, not a guarantee.
Dense and mixture-of-experts models behave differently
Dense models activate all parameters for each generated token, making them more strictly memory-bound. Mixture-of-experts designs activate only part of the network, which can yield much higher speeds on the same hardware. One cited example reached about 60 tokens per second on DGX Spark, illustrating why architecture matters as much as raw specifications.
Prompt reading still favors Nvidia
When reading large prompts and context windows, Nvidia retains an advantage through compute and software optimization. A cited measurement reached about 1,950 tokens per second on DGX Spark, versus around 1,000 tokens per second on an earlier Mac Studio generation. For agent systems that repeatedly reread tens of thousands of tokens, that difference can significantly affect wait times.
Each vendor occupies a different niche
Apple offers high memory ceilings, strong bandwidth, quiet operation and familiar ergonomics, but lacks CUDA and uses soldered memory that must be paid upfront. Nvidia sells access to the dominant CUDA ecosystem and better multi-user serving, but hardware is expensive, with a RTX Pro 6000 cited near $16,000. AMD mini-PCs are cheaper, with examples from about $1,700 to $3,449, but they trade away bandwidth and may run hotter under load.
Cloud rental can remain cheaper than ownership
Renting high-end compute for occasional work can cost about $150 per month for limited daily use, while continuous rental of a premium GPU can approach $1,793 monthly, or around €2,000 in France. A local Mac Studio amortized over 36 months with electricity was estimated near $173 per month. Yet for an open model rented cheaply from the cloud, ownership may not break even until roughly 48 million tokens per month, making rental more economical for many users.
Confidentiality and regulation may outweigh price
For legal, accounting or medical firms, keeping client data on-premises can justify local hardware even when cloud pricing looks attractive. Once sensitive files cannot leave the company, the economics become secondary to compliance and risk control. That is one of the strongest arguments for a local AI machine.
Buy now only if core conditions are met
Immediate purchase makes sense if tasks have already been tested on the target model, output quality is acceptable, workload is regular and someone can maintain the machine. Market pressure also matters: JP Morgan is cited as expecting memory pricing to more than quadruple from early 2024 to late 2026, while IDC sees no improvement before early 2028. For some buyers, the only reason to wait is the still-unpublished price of a higher-memory Mac Studio configuration expected later in October.
CONCLUSION
Local AI infrastructure is becoming a choice between memory, bandwidth, software ecosystem and operational constraints rather than a simple brand preference. For many professionals, the most rational strategy is a hybrid one: keep routine and sensitive work on local hardware, and rent frontier models only when their extra capability is necessary.
Ask a question