A computer can look powerful on paper and still be the wrong machine for AI work. The usual comparison points—processor generation, screen quality, battery life—only tell part of the story. An AI engineer may also need to run local models, hold large datasets in memory, build inside several containers or keep a GPU busy for hours.
That does not mean every engineer needs an oversized workstation. Plenty of AI products are built through hosted APIs and remote infrastructure. In that setting, spending thousands more on a top-end local GPU may change little. Meanwhile, someone working with computer vision or testing open-source models locally can hit a hard memory limit on an otherwise excellent laptop.
The sensible place to start is therefore not a brand or benchmark. It is the work itself. Once you know what must run on the computer—and what can run elsewhere—the specification becomes easier to get right.
First, work out what the computer will actually do
“AI engineering” now describes a broad set of jobs. One engineer may spend most of the day writing application code around external models. Another may process image datasets, fine-tune models and test inference locally. They should not be shopping from the same checklist.
Think about a normal week rather than the most demanding experiment you might run once a year. Which tools stay open together? Do you use Docker? How large are the datasets? Are models running locally, through an API or on a remote server? Does any part of the stack require CUDA?
| Typical work | What matters most locally |
| Building with hosted AI APIs | Capable CPU, 32 GB RAM, battery life and a good development environment |
| RAG, agents and local databases | More RAM, fast storage and capacity for containers and indexes |
| Local inference or computer vision | Supported GPU with enough VRAM, backed by adequate RAM and cooling |
| Large-scale training | Reliable everyday hardware and access to remote or cloud compute |
This prevents two expensive mistakes: buying power that will sit idle, or choosing a sleek laptop that cannot handle the one task performed every day.
The GPU question is really a memory question
If models will run locally, the graphics processor deserves close attention. But the model name printed beside “GPU” is not enough. Video memory, or VRAM, determines what the card can hold while it works. If the model and its working data do not fit, a faster chip will not make the problem disappear.
An 8 GB GPU is useful for learning and smaller experiments, but it can become restrictive quickly. Moving into the 12–16 GB range gives local inference and computer vision work more room. Engineers who handle larger models may need 24 GB or more. These figures are reference points, not promises: precision, quantisation, batch size, context length and the framework all change actual memory consumption.
Software support matters just as much. NVIDIA remains the straightforward option for work built around CUDA. Other hardware can perform well, but “runs AI workloads” is not the same as “supports this library without awkward workarounds”. Check the tools, drivers and models already in use before committing to a platform.
Frequency matters too. If a high-end GPU would sit idle for most of the month, renting a cloud computer for occasional heavy jobs may cost less and leave you with a lighter, quieter everyday computer.
Do not stop at the GPU
For professional development, 32 GB of RAM is a comfortable starting point. It allows an IDE, browser, notebooks, containers and local services to coexist without constant housekeeping. A machine with 64 GB makes more sense when large datasets, virtual machines, local models or several development environments are part of the routine.
Whether that memory can be upgraded later deserves attention. Many slim laptops use soldered RAM. There is nothing wrong with the design, but the capacity chosen at purchase is the capacity the machine keeps. Saving on memory now may shorten its useful life.
Storage fills up in an unremarkable way. Model weights, package caches, container images, datasets and experiment outputs accumulate until a 512 GB drive becomes a daily nuisance. One terabyte is a reasonable minimum for most engineers buying a primary computer. Two terabytes gives local AI work more breathing room.
If an existing computer still performs well but storage or memory has become the problem, replacement is not necessary. A computer upgrade may extend its useful life for less than a new workstation.
How much CPU power is enough?
The CPU still handles a large amount of work around the model: preparing data, compiling code, running application logic, managing containers and keeping the development environment responsive. A recent multi-core processor is worth having, though paying for the largest available CPU rarely makes sense unless the workload can use it.
Look for a balance of strong single-core performance and enough cores for parallel work. On a laptop, sustained performance is more revealing than a short benchmark. Some thin machines run quickly for a few minutes, heat up and then reduce their clock speed. That may go unnoticed during email and browsing but becomes obvious during a long build or inference session.
We see the physical side of this problem regularly. Dust, ageing thermal paste and worn fans all reduce a cooling system’s ability to move heat. If performance starts well and drops after ten or fifteen minutes, thermal throttling may be involved. Our explanation of why laptops overheat covers the warning signs in more detail.
Laptop or desktop?
For engineers who travel, work from several locations or rely mainly on hosted services, a laptop is the natural choice. Modern machines handle demanding software development, and remote computers can take over when a job exceeds local resources.
The compromise appears when heavy work needs to happen on the device itself. A laptop has less room for cooling, its GPU is constrained by power, and most components cannot be replaced. Mobile and desktop GPUs may share a product name while offering different performance and memory configurations, so check the exact part rather than the family badge.
A desktop costs less for a given level of sustained performance. It can accommodate a larger GPU, more memory and storage, and cooling that does not have to fit beneath a keyboard. It is also easier to upgrade in stages. For someone running local models daily, that flexibility can matter more than portability.
There is a useful middle ground: a modest laptop for everyday work and a desktop accessed remotely for GPU jobs. It avoids carrying a heavy workstation everywhere, although it also means maintaining two devices.
Mac, Windows or Linux?
There is no universal winner. The correct operating system is the one that supports the development stack without adding friction.
Apple silicon Macs are quiet, efficient and pleasant for general development. Unified memory makes them interesting for certain local model workloads too. They suit engineers building with APIs or using software that supports Apple hardware well. They are not native CUDA machines, though, and that matters when the workflow relies on NVIDIA-specific tooling.
Windows offers the widest hardware choice, including systems fitted with NVIDIA GPUs. Windows Subsystem for Linux makes a Linux-style workflow practical for many developers, although driver and container compatibility should still be checked against the team’s setup.
Linux is often the cleanest option for a dedicated AI workstation, particularly when the same tools will later run on Linux servers. The trade-off is that certain commercial applications and laptop features may require more work.
Do not choose an operating system because it is supposedly the professional option. Choose it because the code, libraries and deployment environment work properly on it.
Cloud compute does not make local specifications irrelevant
If training happens in the cloud, it is tempting to assume almost any laptop will do. Technically, it might. In practice, an under-specced computer still slows down coding, local tests, data inspection and containers.
Buy enough local performance for the full development environment, then rent the exceptional computer. Large training runs, multi-GPU experiments and short bursts on expensive accelerators usually belong on infrastructure that can be turned off when the work finishes.
Cloud costs need managing. Idle instances, persistent storage and forgotten experiments can erase the savings quickly. Set spending alerts, automate shutdowns and decide which datasets genuinely need to remain attached to costly infrastructure.
Reliability belongs in the specification
When a work computer fails, its benchmark result becomes irrelevant. The real cost is the time spent rebuilding an environment, recovering data and waiting for parts. For freelancers and small technical teams, a few days without the main machine can disrupt an entire project.
Before buying, look at the warranty, local parts availability and whether ordinary failure points can be replaced independently. Can the battery be changed? Is the SSD accessible? Does a damaged charging port require a new motherboard? We considered those practical questions in our review of the most repairable laptops available in New Zealand.
A repairable laptop will not be the thinnest one in the shop. It may nevertheless be the better professional purchase. Extra storage, more memory or a replacement fan can keep a sound machine working for years.
Maintenance matters on high-performance computers as well. Long AI workloads generate heat and keep the cooling system active. Unusual fan noise, repeated shutdowns or a sudden fall in performance should be investigated rather than treated as normal. Advanced Computers provides laptop repairs in Auckland when hardware faults or cooling issues begin interrupting work.
Protect the work, not only the machine
No amount of careful purchasing prevents every failure. Code should live in version control, important files need a separate backup, and environment setup should be documented well enough to reproduce. If the computer disappeared this afternoon, how long would it take to resume essential work elsewhere?
Local model files and public datasets may be replaceable. Private datasets, research notes, credentials and unpublished work are not. Keep at least one current backup separate from the computer and test that it can actually be restored. If a drive has already failed, avoid repeated restart attempts when the missing data matters; further use can make recovery harder. A professional data recovery assessment is safer than experimenting on the only copy.
Full-disk encryption, a password manager, multi-factor authentication and controlled access to company data are also basic requirements for a portable engineering machine.
What we would check before buying
Start with the heaviest task that happens regularly—not the biggest model mentioned in an online discussion. Note its real memory, storage and GPU requirements, then leave reasonable room for growth.
Check the frameworks and drivers required by current projects, actual VRAM rather than just the GPU name, cooling during sustained tests and which components can be upgraded. Add local warranty support, parts availability and the price of a backup, dock, monitors and cloud usage to the calculation.
For teams, standardising a small number of configurations can simplify support and replacement. It should not mean giving every employee the same machine regardless of role. Someone hiring engineers for AI projects should establish the real workload with the technical lead before ordering equipment; the job title alone is not a hardware specification.
Final thoughts
The best computer for an AI engineer is rarely the most expensive one available. It is the machine that handles ordinary work without friction, supports the required software and leaves enough headroom for the next few years.
API-focused developers can put their budget into memory, storage, battery life and a good display. Engineers running models locally should begin with GPU support and VRAM, then consider cooling and system memory. When workloads become unusually large or unpredictable, cloud computing is often more rational than buying a laptop around the worst-case scenario.
Finally, consider how the computer will age. Upgradeable storage, available parts and competent local support do not appear in performance charts, yet they can decide whether a machine lasts two years or six.
Frequently asked questions
Does every AI engineer need a dedicated GPU?
No. A dedicated GPU is unnecessary for many engineers who build products with hosted APIs or send heavy jobs to remote infrastructure. It becomes important for local inference, computer vision, fine-tuning and GPU-dependent experimentation.
Is 32 GB of RAM enough for AI engineering?
For API-based development and moderate local work, 32 GB is usually comfortable. Sixty-four gigabytes is worth considering when the computer regularly runs local models, several containers, virtual machines or large data-processing tasks.
Can a MacBook be used for AI development?
Yes. Apple silicon MacBooks are capable development machines and can run supported models locally. They do not support CUDA natively, so engineers whose work depends on NVIDIA tooling should choose another local platform or use remote GPU infrastructure.
Should an AI engineer buy a laptop or desktop?
A laptop suits mobile and cloud-first work. A desktop provides better sustained performance, cooling and upgradeability for local GPU workloads. Some engineers use both: a lightweight laptop for daily work and a desktop for remote computation.
How much storage does an AI workstation need?
One terabyte is a practical starting point. Two terabytes or more is preferable when models, datasets, container images and experiment outputs will remain on the computer. Storage requirements tend to grow faster than expected, so upgradeability is valuable.