I Chose a Laptop for ML Platform Engineering (Without Wasting Money on the Wrong Things)

Photo by Gioele Fazzeri on Unsplash

I recently went down the rabbit hole of choosing a new laptop for machine learning platform engineering work, and I came out the other side with a strong opinion: most people buying for this role buy the wrong machine. They picture machine learning, imagine training enormous models, and reach for the heaviest, most GPU-laden monster they can afford. Then they end up with a brick that’s tethered to a wall socket, runs hot, and is overkill for what the job actually involves. I want to walk through what I learned, because the right machine for platform work looks quite different from what the hype suggests — and it happens to be a far nicer machine to live with.

First, understand the actual workload

The single most important realization is that a platform engineer’s daily work is not the same as a researcher’s. My days aren’t spent training neural networks on my own hardware. They’re spent building and running the infrastructure that lets training happen elsewhere: containers and orchestration, data pipelines, infrastructure-as-code, CI/CD, monitoring dashboards, cloud SDKs, and a lot of local development against services that mimic production.

When heavy model training does happen, it runs in the cloud or on a cluster — on rented GPUs sized for the job — not on my lap. That’s not a compromise; it’s how the role is designed. Training is bursty and enormous, and it makes no economic or practical sense to carry that capacity around in a backpack.

This reframing changes everything about what to prioritize. Instead of obsessing over a giant GPU, I should be optimizing for the things that make container-heavy, multi-service, all-day development smooth. And, pleasantly, those same choices are exactly the ones that give you a light, cool, long-lasting machine.

RAM is the number that matters most

If you take one thing from this, make it this: memory is your bottleneck, not graphics. A platform engineering environment is a swarm of things running at once. I’ll routinely have several containers up, a local lightweight cluster, a data-processing session, my editor and its language servers, and the inevitable wall of browser tabs — all alive simultaneously. Every one of those holds memory, and they add up faster than you’d believe.

I’d treat 32GB as the absolute floor and 64GB as the target. With 32GB you’ll function, but you’ll feel it pressing against the ceiling the moment you spin up a realistic local stack. With 64GB the whole experience stays fluid, and you stop thinking about memory entirely — which is the goal. If a particular machine offers a choice between soldered and upgradeable memory, I lean toward upgradeable, so future-me has an escape hatch.

CPU: efficient cores beat raw wattage

For the processor, I look for a modern, power-efficient chip with at least eight cores, ideally a design that mixes high-performance cores with efficiency cores. The performance cores chew through container builds and data jobs; the efficiency cores quietly handle background work without draining the battery.

The temptation is to chase the highest-wattage, highest-clock desktop-replacement chip available. I’d resist it. Those run hot, throttle under sustained load anyway, and torch your battery. A newer power-efficient architecture gives you most of the real-world speed with a fraction of the heat and a battery that lasts. For a workload that’s bursty rather than constantly pinned at 100%, efficiency wins.

Storage: more than you think, and fast

Container images, datasets, virtual machines, and caches consume disk space silently and relentlessly. I’d start at 1TB of NVMe SSD and seriously consider 2TB. The “NVMe” part matters as much as the capacity — a fast SSD makes builds, image pulls, and data reads dramatically snappier than the older, slower SATA drives. This is cheap performance, so don’t skimp.

The GPU question, answered honestly

Here’s where I diverge most from the conventional buying advice. A discrete GPU is the single biggest enemy of battery life, the main source of heat and weight, and — for platform work — largely unnecessary, because the serious training lives in the cloud.

A small discrete GPU with 8GB or more of memory is genuinely nice for local prototyping and the occasional small experiment. But you pay for it every single day in reduced battery life and a heavier bag, whether or not you’re using it. My honest recommendation for this role is to favor an efficient integrated-graphics machine and rent cloud GPUs on demand when you actually need to train something. You’ll get a cooler, lighter, far longer-lasting laptop, and you’ll pay only for serious compute when you genuinely use it. The only reason to go discrete is if you specifically intend to train or fine-tune models locally and are willing to accept the battery and weight penalty with open eyes.

Battery: design for going all day unplugged

This was a top priority for me, because nothing kills focus like hunting for a power outlet. The good news is that all the choices above — efficient CPU, no power-hungry GPU — naturally lead to excellent battery life. On top of that, I look for two specifics.

First, a large battery, in the region of 90 watt-hours or more. There’s a practical ceiling around 99Wh because that’s the limit for carrying a laptop onto a plane, so the best machines cluster just under it. Pair a battery that size with efficient internals and genuine all-day, unplugged use becomes realistic rather than a marketing claim.

Second — and this is the one people overlook — charging over USB-C Power Delivery. This is the real answer to never hunting for a charger. A USB-C machine charges from a small universal adapter, from a laptop-capable power bank, or from the same charger as your phone. You can top up almost anywhere, from almost anything. Contrast that with a power-hungry machine that demands a big proprietary charging brick you have to remember and lug around, and the convenience gap is enormous.

Designing around a wet, grey climate

I do a lot of my work in a notoriously rainy climate, and I want to set an honest expectation here: no mainstream laptop is meaningfully weatherproof, so the rain question is mostly answered by what’s around the machine, not the machine itself.

The thing that actually protects your laptop from rain is a properly waterproof backpack or a padded waterproof sleeve. That single accessory matters more for wet-weather survival than any chassis specification. I’d buy a good one before I’d pay extra for any “rugged” laptop feature.

That said, a couple of build choices help with the reality of working in damp, grey conditions. A solid metal chassis (aluminium or magnesium) survives the bumps of commuting in bad weather better than flexy plastic, and a spill-resistant keyboard, if offered, is a sensible hedge against the coffee-shop-in-the-rain lifestyle. Two display-related features genuinely earn their keep in perpetually overcast light: a matte, anti-glare screen so you’re not fighting reflections in dim rooms, and a backlit keyboard for the short, dark afternoons that come with that kind of climate. These aren’t about the rain hitting the laptop; they’re about working comfortably in the gloom that comes with it.

The one-line summary

If I compress everything into a single target configuration, it’s this: 64GB of RAM, an efficient eight-core-or-more CPU, 1–2TB of fast NVMe storage, a battery around 90Wh or larger, USB-C Power Delivery charging, a metal build, and a matte display — and skip the discrete GPU unless you truly intend to train locally. That machine handles the real platform-engineering workload effortlessly, lasts all day away from a socket, and shrugs off the practical annoyances of a damp climate.

What I take away from it

The biggest lesson was to buy for the work I actually do, not the work the marketing imagines. Platform engineering is memory-hungry and orchestration-heavy, not GPU-hungry, and once I accepted that, the ideal machine got lighter, cooler, cheaper to run, and dramatically better on battery — all at once. Match the spec to the real workload, protect it properly from the weather, and let the cloud handle the heavy lifting it was built for. That’s the machine I’d buy, and the reasoning I’d trust over any spec-sheet bragging rights.

Learn more at my website: iamstanleygeorge.wordpress.com