AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: From Local Mac To Cloud API: Finding The Right Place For AI on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

A November 19 report compares cloud APIs, Mac mini and Mac Studio systems, and Nvidia-based PCs for running AI. Its central finding is that the right choice depends on whether the work needs a continuously running agent, large amounts of local model memory, Nvidia-first software, or access to leading closed models.

ThorstenMeyerAI.com published a comparison on November 19 of cloud APIs, Apple Macs and Nvidia graphics-card PCs for running AI, arguing that the best fit depends on the workload rather than one machine being right for everyone. The report draws on published specifications and software documentation, and says its comparisons are not a buying recommendation or a substitute for testing a user’s own tasks.

The report distinguishes an always-on agent from a machine that runs a large model locally. When an agent sends requests to a cloud model, the local computer mainly runs the agent loop, files and tools; it does not need to hold the model’s weights. The source notes Apple lists the base Mac mini at 4 watts when idle, but that figure is an idle specification, not a measure of consumption while doing work.

For local models, the report identifies memory capacity as a key constraint, a central consideration in the Mac-versus-GPU-tower tradeoff. It says the Mac mini tops out at 64 GB of unified memory, while Mac Studio configurations reach 128 GB with M5 Max or 512 GB with M5 Ultra. Nvidia’s GeForce RTX 5090 has 32 GB of GDDR7 memory. A model that exceeds GPU memory can be split between the GPU and system memory, a configuration that may affect performance; the report cites Ollama documentation describing partial GPU and CPU loading.

Software support also shapes the choice. The report says popular Mac options include MLX, LM Studio, Ollama and llama.cpp, and that Apple’s MLX tools support LoRA and QLoRA fine-tuning. It contrasts that with Nvidia’s strong support in tools built around CUDA: vLLM provides precompiled CUDA binaries and directs Apple silicon users to a community-maintained Metal plugin. These are descriptions of software support, not direct speed-test results.

At a glance
reportWhen: Published November 19, 2026; specificat…
The developmentThorstenMeyerAI.com published a comparison of where to run AI workloads, from cloud APIs to Apple Macs and Nvidia GPU PCs.

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 shape the choice

The comparison matters because “running AI locally” can mean very different things. Someone keeping an agent online to organize files or call tools may need a modest computer and a cloud API, while a person who wants private, offline inference needs enough accessible memory for the model and its working context. The report’s rule of thumb estimates that, at 4-bit precision, weights alone require about 0.5 GB per billion parameters: roughly 35 GB for a 70-billion-parameter model and 336 GB for a 671-billion-parameter model. Those estimates exclude overhead and context, so they are not complete memory requirements.

The hardware trade-off has practical consequences. A Mac Studio can offer substantially more unified memory than one RTX 5090, potentially allowing a larger model to fit in a single machine. A GPU PC may suit users whose applications depend on Nvidia CUDA, especially when the model fits in GPU memory. The report says neither Apple nor Nvidia publishes a direct token-speed comparison between the cited systems, so its specifications do not establish which machine is faster across workloads.

Energy and cost also differ, though the cited figures are not directly equivalent. Apple lists the M5 Ultra Mac Studio at a maximum system draw of 385 watts; Nvidia rates the RTX 5090 card at 575 watts and recommends a 1,000-watt power supply for the full system. Apple’s US listed starting prices on November 19 were $2,499 for an M5 Max Studio and $5,499 for an M5 Ultra Studio. The source gives no RTX 5090 price, and a PC’s total cost depends on its other parts.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

From Mac mini to model scale

The report follows an earlier ThorstenMeyerAI.com article about what models can run within the Mac mini’s 16 GB to 64 GB memory range. Its new comparison broadens the question from a single computer to the location and hardware best suited to different tasks: a cloud service, an Apple desktop or a PC with an Nvidia GPU.

Its examples rely on specifications and documentation checked on November 19, 2026, including Apple and Nvidia product information and documentation for MLX, Ollama, PyTorch and vLLM. The source also says Apple’s M5 Ultra has 1.2 terabytes per second of listed memory bandwidth, while M5 Max configurations list 460 or 614 GB/s. Nvidia’s cited RTX 5090 page gives a 512-bit memory interface but, according to the report, no bandwidth figure. These figures do not by themselves provide a like-for-like performance test.

There are limits beyond memory. The report says PyTorch’s Mac backend supports GPU training but does not support torch.float64; computations that genuinely require double precision must run on the CPU. It also cautions that training large models is not a desktop task on any of the systems discussed. The leading closed models remain available only through their providers, according to the source, rather than as downloadable weights for local use.

“The decision matrix is our reading of the specifications and documentation below. It is not a buying recommendation, and it does not replace a test with your own work.”

— Thorsten Meyer, ThorstenMeyerAI.com

Performance comparisons remain open

The report does not provide head-to-head benchmark tests for the Mac Studio and RTX 5090, or a measured comparison of their token speeds, power use under the same workload or total purchase costs. Published memory bandwidth and power ratings are useful specifications, but do not settle performance for a particular model or application.

Actual results will depend on model size and quantization, context length, software versions, memory use, and whether the task is inference, fine-tuning or serving several users. The source also does not test the community-maintained vLLM-Metal plugin against CUDA deployments. Readers should treat its decision matrix as a guide to constraints and compatibility, not a guarantee that a system will meet a specific performance target.

Test against the intended workload

The report’s practical next step is to match the machine to the task: decide whether the model must run locally, check that its weights and working memory fit, and confirm that the required tools support the chosen hardware. For a cloud-backed agent, the local machine may only need to run the agent and its supporting tools; for local inference, memory capacity and software compatibility become central.

Anyone choosing between a Mac Studio and GPU PC still needs to test their own model and workflow. Future software support, hardware configurations and independent benchmarks could change the comparison. The source does not announce a subsequent test or a new product; its current conclusion is a workload-based framework rather than a single recommended machine.

Key Questions

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

Not necessarily. The report says that if an agent calls a cloud model, the local computer mainly runs the agent loop, files and tools rather than holding the model itself.

Which system has the most memory in the comparison?

According to the report’s November 19 specifications, Mac Studio configurations reach 512 GB of unified memory with M5 Ultra. The cited RTX 5090 has 32 GB of graphics memory; the report does not compare complete PC system memory configurations.

Is a Mac faster than an RTX 5090 for local AI?

The report does not establish that. It cites specifications but says neither company publishes a direct token-speed comparison between the systems. Performance depends on the model, memory fit, software and task.

Can a Mac fine-tune AI models?

The report says Apple’s MLX tools support LoRA and QLoRA fine-tuning, and PyTorch’s Mac backend supports GPU training. It also notes that larger models need substantial memory and that desktop systems are not suited to training large models.

When does a cloud API remain the better fit?

A cloud API is relevant when users need access to leading closed models that cannot be downloaded, or when they do not want to run model weights locally. The report does not compare API prices, privacy terms or service performance.

Source: ThorstenMeyerAI.com

You May Also Like

2026’S Top AI Smartwatches: 9 Devices For Tech And Fitness Lovers

Discover the nine leading AI-powered smartwatches of 2026, highlighting features, ecosystem compatibility, battery life, and who each device suits best.

10 Best Noise Cancelling Headphones For 2027, Compared

A 10-product roundup names Bose QuietComfort its all-around over-ear pick, with Sony, Soundcore, Apple and Beats among the alternatives.

2026’S Most Innovative AI-Powered NAS Solutions

Discover the most advanced AI-driven NAS solutions of 2026, highlighting key features, industry impact, and what to expect next.

How AI Note Takers Can Help With Meeting Records: 15 Picks For 2027

A 2027 buyer’s guide compares 15 AI meeting recorders, including Plaud and Comu models, while flagging gaps in pricing, privacy and performance data.