🔍 Read the full analysis: A Field Guide To Choosing Machines For AI Work on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A field guide published by ThorstenMeyerAI.com on Nov. 19 compares cloud APIs, Apple Macs and Nvidia PCs for different AI tasks. Its central finding is that memory capacity, software compatibility and workload matter more than choosing one machine as a universal AI computer.
ThorstenMeyerAI.com published a field guide on Nov. 19 comparing cloud APIs, Apple Macs and Nvidia graphics-card PCs for AI work, arguing that the right choice depends on whether a user is running an agent, loading a local model, fine-tuning or serving models to a team. The guide draws on manufacturer specifications and software documentation, rather than head-to-head performance tests, so its comparisons describe capabilities and constraints rather than proving which machine is faster.
For an always-on agent that calls a cloud model, the guide says a modest, low-power computer can be enough because the device mainly runs the agent loop, files and tools; the model itself runs remotely. Apple lists the base Mac mini’s idle power at 4 watts. That figure is an idle specification, not a measurement of power use while an agent is working.
For local models, the guide puts memory capacity at the center of the decision. It lists Mac Studio configurations up to 512 GB of unified memory with M5 Ultra, compared with 32 GB of graphics memory on Nvidia’s GeForce RTX 5090. Apple’s shared memory pool can accommodate models that exceed a single graphics card’s capacity, while a model too large for the 5090’s onboard memory must be split between the GPU and system memory, according to the guide’s explanation of Ollama documentation.
The guide gives a rough sizing calculation: at 4-bit precision, model weights use about half a gigabyte per billion parameters, before additional memory for overhead and conversation context. On that basis, 70 billion parameters require about 35 GB for weights alone; 671 billion require about 336 GB. Those are arithmetic estimates, not guarantees that a model will run well in a given setup.
Decision guide · local AI hardware
Which machine fits which AI job?
Five common jobs, one after another, with the machine that fits each best on today’s specifications.
First question: does the model run in the cloud, or on your own machine? If it runs in the cloud, the box only keeps the agent going.
The job
- 1An agent that calls cloud models, running around the clock
→ Mac mini - 2A private assistant with a small open model
→ Mac mini - 3Huge open models with hundreds of billions of parameters
→ Mac Studio - 4Tools built for Nvidia’s CUDA, training, serving many users
→ GPU PC - 5The strongest models and the hardest tasks
→ Cloud API
The machine
Mac mini
up to 64 GB · 4 W at idle
Mac Studio
up to 512 GB · 385 W at most
GPU PC
32 GB on the card · 575 W card
Cloud API
strongest models · per-token bills
Sources: Apple, Mac mini and Mac Studio technical specifications and power pages; Nvidia, GeForce RTX 5090 specifications; all retrieved November 19, 2026. Matching of jobs to machines: our reading, not a buying recommendation.
Graphic: Powered by Thorsten Meyer AI
Match Hardware to the AI Workload
The comparison matters because “running AI” covers several different jobs. A computer that keeps a cloud-connected agent available does not need the same hardware as a workstation loading a large model into local memory. Buying for the wrong workload can mean paying for unused capacity or finding that a model or software tool will not run as expected.
For a model that fits inside one GPU’s memory, a PC with an Nvidia card may be a better fit when the chosen software is designed for CUDA. The guide says many AI tools support Nvidia first and notes that vLLM distributes CUDA binaries, while its Apple silicon support relies on a community-maintained plugin. A Mac can still run common local-model tools, including MLX, LM Studio, Ollama and llama.cpp, and Apple’s MLX tools support LoRA and QLoRA fine-tuning.
Power and cost also vary. The guide cites Apple’s maximum system power of 385 watts for the M5 Ultra Mac Studio and Nvidia’s 575-watt rating for the RTX 5090 card alone; Nvidia recommends a 1,000-watt power supply for the system. Apple US list prices cited in the article start at $2,499 for an M5 Max Mac Studio and $5,499 for an M5 Ultra version. Nvidia’s product page does not list an RTX 5090 price, and a complete PC’s price depends on its other components.
Popular right now

Roborock Qrevo S Pro Robot Vacuum and Mop (2026)

Samsung 55-Inch S90H OLED 4K UHD Smart TV Bundle with Full-Motion Mount and Surge Protector

LG OLED evo C4 55-Inch 4K Smart OLED TV (OLED55C4PUA) with 144Hz Refresh Rate, HDMI Cable, 24-Month Protection Plan, and Streaming Bundle
As an affiliate, we earn on qualifying purchases.
From Mac Mini to Workstation Choices
The article follows a prior ThorstenMeyerAI.com piece about what models can run within a Mac mini’s 16 GB to 64 GB memory range. The new guide broadens the question from one computer to a set of choices, including cloud APIs, Mac Studio systems and PCs built around Nvidia graphics cards.
Its comparisons cite Apple and Nvidia specifications and documentation from software projects, checked on Nov. 19, 2026. The guide says Apple lists M5 Ultra memory bandwidth at 1.2 terabytes per second and M5 Max bandwidth at 460 or 614 GB/s, depending on configuration. Nvidia’s cited RTX 5090 specifications give a 512-bit memory interface, but the guide says it found no directly comparable token-speed test from either company. It therefore does not establish a performance winner across platforms.
The guide also cautions that training large models is not a desktop task on any of the machines discussed. Macs can be used for smaller fine-tuning jobs, but software support and precision requirements can limit particular workflows. For example, the cited PyTorch documentation says its macOS GPU backend does not support double-precision float64 computations, which must run on the CPU.
“The CPU and GPU have direct access to the same memory pool.”
— Apple, as quoted in the guide’s summary of MLX documentation
Performance Comparisons Still Missing
The guide does not include controlled speed tests comparing an M5 Mac Studio with an RTX 5090 PC on the same models, software versions and settings. It says neither manufacturer publishes a token-speed comparison between the systems, and the Nvidia specifications it cites do not include a memory-bandwidth figure directly comparable with Apple’s published figures.
Actual costs for a complete GPU PC are also unsettled by the cited material: Nvidia’s product page lists no RTX 5090 price, and the total depends on the rest of the build. Results will also vary with model size, quantization, context length, software support and how much system memory is available. The guide’s memory calculations cover weights alone and do not establish the full memory required by a particular workload.
Its discussion of linking Macs is not a claim that a cluster behaves like one large machine automatically. The article says Apple documents clustering, but setup requires Thunderbolt 5 connections between each pair of Macs and a macOS Recovery setting change. It notes that the base M6 Mac mini has Thunderbolt 4, a distinction that may affect such a setup.
Test the Intended Software First
The next step for a buyer is to identify the specific model, agent or serving tool they intend to use, then check its current hardware and software requirements. Users weighing a Mac against an Nvidia PC should test their own model and workload where possible; the guide’s specifications do not substitute for measurements of speed, power use or reliability under that workload.
For teams considering model serving or fine-tuning, software compatibility is a separate check from whether the machine has enough memory. The guide identifies Nvidia CUDA support as a strength for tools such as vLLM, while Apple users may need a different supported tool or a community plugin. No further benchmark, product announcement or follow-up test is specified in the source material.
Key Questions
Which computer is enough for an always-on AI agent?
If the agent sends its requests to a cloud model, the computer mainly runs the agent loop, files and tools. The guide says many such agents do not need a high-end workstation; the appropriate system depends on the agent’s own software and workload.
What matters most for running an AI model locally?
Memory capacity is a key constraint because model weights, overhead and conversation context need to fit in memory the processor can access. The guide’s 4-bit estimate is about 0.5 GB per billion parameters for weights alone, not a full system requirement.
Is a Mac Studio faster than an RTX 5090 PC?
The source guide does not establish a universal speed winner. It says no direct token-speed comparison is published in the specifications it reviewed. A GPU PC may suit software built for CUDA, while a Mac Studio offers configurations with more memory than the RTX 5090’s 32 GB of onboard memory.
Can Macs fine-tune AI models?
Yes, for some workloads. The guide says Apple’s MLX tools support LoRA and QLoRA, and PyTorch supports GPU training on macOS. Model size, memory needs, precision requirements and software support still determine whether a particular fine-tuning job is practical.
Can a desktop run the strongest closed AI models locally?
Not when those models are not available for download. The guide says leading closed models cannot be downloaded, leaving cloud access as the option for using them; it does not claim that every high-capacity open model requires cloud computing.
Source: ThorstenMeyerAI.com
