What’s The Right Machine For Your AI Workload?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What’s The Right Machine For Your AI Workload? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

A November 19 guide from ThorstenMeyerAI.com compares cloud services, Apple Macs and Nvidia GPU PCs for AI agents, local model use, fine-tuning and serving. Its central finding is that the right machine depends on the workload: memory capacity shapes which models fit, while Nvidia software support and cloud access remain important for other jobs.

ThorstenMeyerAI.com published a guide on November 19 comparing cloud APIs, Apple Macs and Nvidia graphics-card PCs for different AI workloads, arguing that there is no single machine suited to every job. Its comparison focuses on memory capacity, software compatibility, power use and workload, from running an always-on agent to serving large models or fine-tuning them.

The guide says many always-on agents do not require a powerful local computer if they send model requests to a cloud service. In that setup, the machine mainly runs the agent’s workflow, files and tools. Apple lists the base Mac mini’s idle power at 4 watts, according to the article. That figure describes idle use, not the computer’s power draw while handling an active workload.

For models run locally, the guide puts memory ahead of headline processor specifications. It estimates that weights at 4-bit precision take about half a gigabyte per billion parameters, before adding overhead and conversation context. On that estimate, a 70-billion-parameter model needs about 35 GB for weights alone, while a 671-billion-parameter model needs about 336 GB. These are calculations, not a guarantee that a model will run well with that amount of memory.

The article cites Apple configurations ranging from 64 GB maximum on Mac mini to 512 GB on a Mac Studio with M5 Ultra. Nvidia’s GeForce RTX 5090 has 32 GB of graphics memory. A model that exceeds the card’s capacity may have to use system memory as well; the guide cites Ollama documentation describing a model split between CPU and GPU memory. It says Nvidia PCs often have an advantage for tools designed around CUDA, while Apple’s MLX, LM Studio, Ollama and llama.cpp offer options for running models on Mac.

At a glance
reportWhen: Published November 19, 2026; specificat…
The developmentThorstenMeyerAI.com published a workload-based comparison of cloud APIs, Apple Macs and Nvidia GPU PCs for running and developing AI systems.

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

  1. 1An agent that calls cloud models, running around the clock
    → Mac mini
  2. 2A private assistant with a small open model
    → Mac mini
  3. 3Huge open models with hundreds of billions of parameters
    → Mac Studio
  4. 4Tools built for Nvidia’s CUDA, training, serving many users
    → GPU PC
  5. 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

Memory and Software Set the Limits

The comparison matters because “running AI” can mean very different workloads. An agent that calls a cloud model may need little local compute, while a user who wants a large open model entirely on-device must find a system with enough memory. Buying for the wrong task can mean paying for capacity that goes unused or discovering that a model does not fit the available hardware.

Memory capacity is not the only constraint. Software support can determine whether a machine is practical for a particular job. The guide says many AI tools support Nvidia CUDA directly, while some tools reach Apple silicon through separate projects or community-maintained plugins. That distinction can affect setup, compatibility and the options available for serving a model to multiple users.

Power and cost also vary. The article cites Apple’s maximum whole-system power figure of 385 watts for an M5 Ultra Mac Studio, compared with a 575-watt rating for the RTX 5090 card alone. Nvidia recommends a 1,000-watt system power supply for a PC using the card. Those figures are not directly equivalent: one describes a whole computer’s maximum power and the other a graphics card’s rating. Apple’s listed U.S. starting prices were $2,499 for an M5 Max Mac Studio and $5,499 for an M5 Ultra model on November 19; the cited Nvidia product page did not list a 5090 price.

Bose QuietComfort Ultra Wireless Noise-Cancelling Over-Ear Headphones with Spatial Audio, White Smoke

Bose QuietComfort Ultra Wireless Noise-Cancelling Over-Ear Headphones with Spatial Audio, White Smoke

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

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

As an affiliate, we earn on qualifying purchases.

A Mac Mini Question, Expanded

The guide follows an earlier ThorstenMeyerAI.com article about what models a Mac mini with 16 GB to 64 GB of memory can run. The new comparison widens the question from one small desktop to a range of systems, including cloud services, Mac Studio configurations and PCs built around Nvidia graphics cards. It also refers to Apple’s August 25, 2026, description of its new Mac mini as “the leading desktop for always-on agentic computing.” That wording is Apple’s product characterization, not an independent test result.

For local development, the source points to Apple’s MLX tools for LoRA and QLoRA fine-tuning and PyTorch’s macOS GPU backend. It also notes a limitation in PyTorch’s Mac backend: it does not support float64 computations on the GPU, so calculations that genuinely require double precision must run on the CPU. For model serving, it contrasts vLLM’s precompiled CUDA binaries with vLLM-Metal, which the article describes as a community-maintained plugin for Apple silicon.

The source says clustering Macs is possible but requires setup, including Thunderbolt 5 connections between machines and a setting changed in macOS recovery. It identifies a hardware difference relevant to that arrangement: the base M6 Mac mini has Thunderbolt 4. The guide presents clustering as a project rather than a simple substitute for a single larger system.

“The CPU and GPU have direct access to the same memory pool.”

— ThorstenMeyerAI.com, summarizing Apple’s MLX documentation

Performance Depends on the Model

The comparison is based on published specifications and software documentation, not a controlled benchmark of the listed machines. The article says Apple and Nvidia do not publish a token-speed comparison between the M5 Ultra and RTX 5090. It therefore does not establish which system will be faster for a particular model, prompt length or software configuration.

Actual performance can depend on whether model weights and conversation context fit in fast-access memory, how a tool divides work between CPU and GPU, and whether it supports the chosen hardware well. The guide’s 4-bit memory estimates cover weights alone; overhead and context require additional capacity. Nor does the article provide a single, directly comparable total price for an RTX 5090 PC, whose cost depends on the rest of the system.

The source does not test power use under matching workloads, detail the performance of the community-maintained vLLM-Metal plugin, or establish how well any particular Mac cluster performs. Its hardware specifications and software references were checked on November 19, 2026, so later product or software changes are outside the comparison.

Match the Machine to the Job

The guide’s practical next step is to identify the workload before choosing hardware: whether the system will call a cloud model, run an open model locally, fine-tune one, or serve requests to a team. Readers considering local use should estimate memory needs for weights, overhead and context, then check that their chosen software supports the machine.

For a specific purchase, the comparison leaves room for hands-on tests using the reader’s model and tools. A useful follow-up would compare model speed, power draw, setup and total system cost under the same workload. The source does not announce such a benchmark or a later update; its figures describe the specifications and documentation available on November 19, 2026.

Key Questions

Does an always-on AI agent need a powerful computer?

Not necessarily. According to the guide, if an agent sends requests to a cloud model, its local computer mainly runs the workflow, files and tools. Requirements will differ if the model itself runs locally.

How much memory does a local model need?

The article’s rough estimate is about 0.5 GB per billion parameters for weights at 4-bit precision, before overhead and conversation context. It estimates 35 GB for the weights of a 70-billion-parameter model; that is not a full-system memory recommendation.

Should I choose a Mac Studio or an Nvidia PC?

The guide says a Mac Studio offers configurations with more unified memory, while an Nvidia PC may suit tools built first for CUDA. The right choice depends on the model, available memory, software and workload; the source does not report a direct speed benchmark.

Can Macs fine-tune AI models?

Some fine-tuning is supported. The source cites Apple’s MLX tools for LoRA and QLoRA and PyTorch’s GPU training backend for macOS. It does not claim that every model or training workload will fit or perform well on a Mac.

Can a desktop run the largest AI models?

It depends on the model and hardware. The guide says the strongest closed models are not downloadable, making cloud access necessary to use those models. It also estimates that a 671-billion-parameter model’s 4-bit weights alone take about 336 GB, before additional memory needs.

Source: ThorstenMeyerAI.com

You May Also Like

Inside The AI Restrictions Nobody Was Paying Attention To In China

An analysis of China’s lesser-known restrictions on Chinese optical transceivers and their implications for AI infrastructure and global supply chains.

The United States: The High-Variance Bet

The US is pursuing a minimal regulation, market-driven strategy for AI and social safety nets, with significant local experimentation amid federal inaction.

Refusing Certain AI Safety Measures? Here’s Why That Doesn’t Mean Discarding AI As A Whole

A new paper argues safety should target harmful subsets within topics, not entire topics, highlighting nuanced approaches to AI safety and deployment.

The Hidden Labor Of AI: What Reviewing Its Work Means For Employees

Research finds AI can speed some work but add review time and errors, leaving employees responsible for checking its confident output.