The short answer
The best computer for AI is the one that fits the models you actually plan to run, and that depends on memory far more than on brand.
If you use AI through a browser or an app such as ChatGPT, Copilot, Claude or Gemini, almost any recent laptop will do, because the heavy lifting happens in the cloud. If you want to run models on your own machine for privacy, running cost or offline work, memory becomes the deciding factor, and the right choice depends on how large those models are.
- Everyday AI tools: a recent laptop with 16GB of RAM or more. A Copilot+ PC adds on-device features.
- Local models up to roughly 30B parameters: a desktop with an NVIDIA RTX graphics card carrying 16GB to 32GB of VRAM.
- Bigger local models (70B and up): a unified-memory machine such as a Mac Studio, an AMD Ryzen AI Max+ system or an NVIDIA GB10 mini PC.
This guide covers every major platform and PC maker, with the trade-offs worth knowing before you spend. Prices are US dollars from manufacturer and retailer listings in October 2026, before NZ GST and shipping, and memory prices have been moving, so confirm a current quote before you buy.
What makes a computer good for AI
Memory capacity decides which models you can load, and memory bandwidth decides how fast they run. A model’s weights have to fit in fast memory before anything else matters. ComputingForGeeks puts the weights of a 4-bit (Q4) 8B model at roughly 5GB, a 32B model at about 20GB and a 70B model at about 40GB. Add 15 to 20 percent for the model’s working memory at short context, and much more at long context.
Four other things shape the choice:
- Bandwidth: once a model fits, generation speed is bound by how quickly memory can feed the processor. Desktop graphics cards lead, and unified-memory machines are slower per gigabyte but hold far bigger models.
- NPU: Windows Copilot+ PCs need a 40 TOPS neural processing unit, which runs built-in features such as Live Captions and Studio Effects. It is not what runs a 70B language model.
- Software: NVIDIA’s CUDA is the default target for most AI tools. AMD and Apple work well with the common local runners, but expect more friction with CUDA-first training scripts.
- Soldered memory: unified-memory machines (Macs, AMD Ryzen AI Max+ and NVIDIA GB10 systems) cannot be upgraded later. Buy the capacity you will need in three years, not today.
Here is how the main platforms compare on the two numbers that matter.
| Platform | Memory a model can use | Memory bandwidth | Found in |
|---|---|---|---|
| NVIDIA RTX 5090 graphics card | 32GB GDDR7 | 1,792 GB/s | Desktop PCs |
| NVIDIA GB10 | 128GB unified | 273 GB/s | Lenovo ThinkStation PGX, HP ZGX Nano, Dell Pro Max with GB10 |
| AMD Ryzen AI Max+ 395 | 128GB unified, about 96GB usable by the GPU | about 256 GB/s | HP Z2 Mini G1a, ZBook Ultra G1a, mini PCs |
| Apple M5 Max | Up to 128GB unified | Up to 614 GB/s | Mac Studio |
| Apple M5 Ultra | Up to 512GB unified | 1.2 TB/s | Mac Studio |
The pattern is clear: the fastest memory sits on a graphics card with the least room, and the roomiest memory sits in machines that trade some speed for capacity. Most of this guide is about choosing where on that line you want to be.
Laptop or Desktop?
Choose a desktop if you will run local models most days, and a laptop only if you need to take the work with you or your AI use is mostly cloud tools.
Why a desktop wins for heavy use. A desktop graphics card has its own fast memory, steady cooling and room to be swapped later. The RTX 5090 carries 32GB and draws 575W, and it has a 1,000W power supply for it. A desktop also lets you replace or upgrade individual parts as your needs change, which a sealed laptop cannot.
The catch in late 2026 is price. The same source reports street prices of about US$7,900 to $9,000 for an RTX 5090 (its MSRP was $1,999), about $4,400 for a new RTX 4090 and $1,700 or more for a used RTX 3090. A memory shortage has pushed GPU and RAM prices up across the market. A high-end graphics card now costs more than a complete unified-memory mini PC.
Why a laptop can still make sense. Unified-memory laptops have narrowed the gap. The HP ZBook Ultra G1a offers up to 128GB of memory in a body of about 1.5kg, enough room for models that no 24GB graphics card can hold. The trade-offs are sustained speed, battery life under load and soldered memory that cannot be upgraded.
- Pick a desktop for daily local models, fine-tuning, long sessions and the option to upgrade parts later.
- Pick a laptop for travel, client demos, or cloud-first AI work where portability matters more than raw speed.
- Pick a compact unified-memory box (Mac Studio, Ryzen AI Max+ or GB10 mini PC) when you want big models on a desk without a tower, a 1,000W supply or the noise.
NVIDIA AI Computer
NVIDIA is the safe default for AI because almost every AI tool is written for its CUDA software, and it now covers everything from a gaming graphics card to a palm-sized AI mini PC.
RTX graphics cards (desktop). The RTX 5090 is the fastest single consumer card. In a testing, it generated about 250 tokens per second on an 8B model and 71 on a 32B model, but a 70B model does not fit in its 32GB. Cheaper cards still work: a 16GB RTX 5060 Ti (about US$780 to $830) suits models up to roughly 13B, and a 12GB RTX 3060 (about $465) is the cheapest card the same source would buy for local models. Street prices for the top cards are inflated right now, as covered under Laptop or Desktop?.
RTX workstation cards. For a 70B model on one card, the RTX 6000 Ada has 48GB (about US$8,000 to $8,450). The RTX PRO 6000 Blackwell has 96GB but sells near US$20,000.
DGX Spark and GB10 mini PCs. The GB10 chip combines a 20-core Arm CPU and a Blackwell GPU with 128GB of unified memory at 273 GB/s, in a case about 150mm square. It can run models of up to 200B parameters, or 405B with two units linked. NVIDIA sells its own DGX Spark, and Dell, HP, Lenovo and ASUS sell versions of the same machine. Dell lists its 2TB configuration at US$5,903.
RTX Spark laptops and desktops (new this month). RTX Spark is NVIDIA’s Windows on Arm chip, a laptop and compact-desktop sibling of the GB10. TechSpot reports that laptops built on it began shipping on 16 October with up to 128GB of unified memory, including the Microsoft Surface Laptop Ultra (from US$2,600), Dell XPS 16, HP OmniBook Ultra 16, Lenovo Yoga Pro 9n, MSI Prestige N16 Flip AI+ and Asus ProArt P16. In a lab test, a 128GB Surface let the GPU use 64GB and ran 20B to 35B models at 58 to 89 tokens per second, while a laptop with an RTX 5090 was about 2.3 times faster on anything that fit in its 24GB. The same review saw two Surface units cut their own power under one extreme sustained workload, so wait for more testing if you plan heavy daily use.
The honest trade-offs. A GB10 machine runs NVIDIA’s Linux-based DGX OS, not Windows, so it is a developer’s tool rather than a general-purpose PC. Its 273 GB/s bandwidth is a fraction of the RTX 5090’s 1,792 GB/s, so on models that fit both, the graphics card generates faster. The Spark’s strength is capacity and a full CUDA stack in a small, quiet box.
Best for: developers and anyone whose tools assume CUDA. Skip it if: you want a Windows PC that also runs AI, or you only use cloud AI services.
AMD AI Computer
AMD’s Ryzen AI Max+ 395 is the cheapest way to get 128GB of model memory in a Windows or Linux machine, and it is the chip behind some of the most interesting AI laptops and mini PCs.
The chip combines 16 Zen 5 cores, a Radeon 8060S graphics unit and a 50 TOPS NPU, all sharing 128GB of LPDDR5X memory. About 96GB of that pool can be used by the GPU on Linux, which lets one small box hold a model that no consumer graphics card can fit. The same source quotes cited speeds of about 52 to 56 tokens per second on the 120B mixture-of-experts model gpt-oss-120b, but only 4 to 6 on a dense 70B model, because its roughly 256 GB/s of bandwidth is the limit.
Where you find it. HP builds it into the ZBook Ultra G1a laptop and Z2 Mini G1a workstation, ASUS into the 1.2kg ROG Flow Z13 tablet-laptop, and several brands into mini PCs including the Framework Desktop, GMKtec EVO-X2, Bosgame M5 and Minisforum MS-S1 Max.
Price. The 128GB mini PCs cost between US$3,449 and US$4,349 in October, roughly double their launch prices of around US$2,000, because soldered memory has become expensive. AMD’s newer Ryzen AI Max+ PRO 495 offers up to 192GB, with about 160GB usable by the GPU, from about US$6,500 to US$7,400.
AMD graphics cards. For a desktop, the 24GB Radeon RX 7900 XTX (about US$1,130 to $1,300) runs Ollama and llama.cpp well, and on the Vulkan backend it generates faster than an RTX 3090 on the community benchmark ComputingForGeeks cites. Prompt processing is slower, and CUDA-first tools such as many training and fine-tuning scripts are harder to run on it.
Best for: running big models cheaply in a small machine, and Windows or Linux users who want a single box. Skip it if: you need the fastest generation on small models, where a graphics card wins, or your software requires CUDA.
Apple AI Computer
Apple’s Mac Studio is the quietest way to run very large models at home, and the new M5 Ultra version holds more memory than any other single desktop here.
Apple announced the Mac Studio with M5 Max and M5 Ultra in August 2026, and by Apple’s own figures the two chips offer:
- M5 Max: up to 128GB of unified memory at up to 614 GB/s, from US$2,499 in the base 36GB configuration.
- M5 Ultra: up to 512GB at 1.2TB/s, from US$5,499. The 512GB configuration is due in late October.
Both put Neural Accelerators in every GPU core, and Apple claims up to 4.3x the peak AI compute of the M3 Ultra. Several Mac Studios can also be linked over Thunderbolt 5 to pool memory, and Apple says four units run AI inference up to three times faster than one. Developers get Apple’s MLX framework and a new Core AI framework. These performance numbers come from Apple’s own July 2026 testing, so treat them as claims until independent results appear.
Why it suits AI. Memory is the limit for local models, and a 128GB Mac holds a 70B model with room to spare while staying nearly silent on a desk. That is something a single graphics card cannot do.
The trade-offs. On models that fit in an NVIDIA card’s memory, the card is faster, and the same source notes that prompt processing is slower on Apple silicon. The CUDA ecosystem is not available, so you run llama.cpp and MLX rather than NVIDIA-specific servers. Memory is soldered, so choose the capacity you need when you buy. Pricing here is in US dollars; check Apple’s New Zealand store for NZ pricing.
Best for: quiet, low-power big-model work, and anyone already in the Apple ecosystem. Skip it if: your tools need CUDA or you need Windows.
Microsoft AI Computer
Microsoft does not sell one AI computer. It sets the Copilot+ PC standard that most Windows laptops follow, and sells Surface models that meet it.
What Copilot+ means. A Copilot+ PC needs a neural processing unit (NPU) of at least 40 TOPS, 16GB of RAM, a 256GB SSD and Windows 11 24H2 or later. Qualifying chips come from Qualcomm (Snapdragon X), Intel (Core Ultra Series 2) and AMD (Ryzen AI 300 and 400). The NPU runs features such as Live Captions with translation, Windows Studio Effects, Cocreator in Paint and Click to Do on the device, with little battery cost. Recall, which searches snapshots of your screen, is opt-in, and the same source suggests leaving it off on shared or work-managed machines.
What it does not mean. A Copilot+ badge does not mean you can run big language models. Those need the memory covered in the earlier sections, not an NPU.
Surface models. TechSpot reports that the Snapdragon X2 Surface Pro starts at US$1,499 and the Surface Laptop at US$1,599. Both use chips with an 80 TOPS NPU and offer 16GB to 64GB of memory. Prices rose $100 over the last generation, and up to $500 to $600 over 2024’s entry models.
Surface Laptop Ultra. Microsoft’s AI-focused flagship uses NVIDIA’s RTX Spark chip (covered under NVIDIA). It starts at US$2,600 with 24GB, and the 128GB model costs about US$5,900. Microsoft also takes pre-orders for an RTX Spark Dev Box from US$6,000.
Windows on Arm trade-offs. Most mainstream apps now run natively on Arm, but Evetech notes gaps for anti-cheat games and some niche business software. Check your daily applications before buying a Snapdragon or RTX Spark machine.
Best for: everyday AI features, battery life and a Windows laptop that stays cool. Skip it if: you need CUDA-only software today or rely on a niche x86 application.
HP AI Computer
HP covers the widest range of AI computers in one catalogue: an NVIDIA GB10 mini PC, AMD Ryzen AI Max+ workstations, NVIDIA RTX PRO workstations and a new RTX Spark laptop.
HP ZGX Nano AI Station. HP’s version of the GB10 machine has 128GB of unified memory, measures 150 x 150 x 51mm and comes with 1TB or 4TB of storage. HP says it runs models of up to 200B parameters, or up to 405B with two units linked by a QSFP cable sold separately. It ships with NVIDIA’s DGX OS and the free HP ZGX Toolkit, which adds Ollama and MLflow, and you drive it from a Windows 11 or Ubuntu computer with Visual Studio Code.
HP Z2 Mini G1a and ZBook Ultra G1a. Both use AMD’s Ryzen AI Max+ PRO 395 with up to 128GB of memory. The ZBook Ultra G1a is a 14-inch mobile workstation of about 1.5kg with a 74.5Wh battery, and the Z2 Mini G1a is a desktop about the size of a paperback that takes up to 8TB of storage. They run Windows 11 Pro, which makes them easier to fit into an office than the Linux-based GB10 box.
Other HP Z workstations. HP’s AI workstation range also includes the Z2 SFF, Z1 Tower and Z2 Tower G1i desktops and the ZBook Fury G1i and ZBook X G1i laptops. These are the models to ask about if you want an NVIDIA RTX PRO graphics card with ISV-certified software support.
HP OmniBook Ultra 16. This RTX Spark laptop is listed from US$3,200.
Best for: businesses that want a known brand, workstation-class support and a choice between NVIDIA and AMD. Skip it if: you want the lowest price for 128GB, where a Ryzen mini PC from a smaller maker costs less.
Dell AI Computer
Dell’s main AI computer is a GB10 mini PC with strong business support, and its XPS 16 brings NVIDIA’s new RTX Spark chip to a conventional laptop.
Dell Pro Max with GB10. Dell’s US store lists it as the Dell Pro Precision with GB10. It has 128GB of LPDDR5X memory, ships with NVIDIA DGX OS 7 and measures 150mm square at about 1.3kg. Two 200G QSFP ports let you link a second unit. The 2TB configuration is listed at US$5,903, and a 4TB drive adds US$2,450. Dell also sells ProSupport Plus, which adds next-business-day onsite repair and accidental damage cover. That service is the main reason to buy this version over a cheaper brand.
Dell XPS 16 with RTX Spark. TechSpot lists it from US$3,800, shipping from 16 October. This is the most expensive RTX Spark laptop announced so far, so compare it against the Surface Laptop Ultra and HP OmniBook Ultra 16.
Dell Pro Max and Pro Precision laptops. In StorageReview’s test, the Dell Pro Max 14 Premium had an 8GB RTX PRO 2000 Blackwell graphics chip and could not load models larger than that memory allows. The Dell Pro Precision 5 16S used integrated Intel graphics and managed larger models slowly. Treat these as business laptops with AI features, not local-model machines, unless you configure them with a high-memory graphics option.
Best for: organisations that want one supplier, on-site support and standard purchasing. Skip it if: you want the lowest price per gigabyte of model memory.
Lenovo AI Computer
Lenovo offers a GB10 mini PC for developers, a long-established range of ThinkStation and ThinkPad workstations, and an RTX Spark laptop.
ThinkStation PGX. Lenovo’s first GB10 workstation has a 20-core Arm CPU, 128GB of soldered unified memory at 273 GB/s, one M.2 2242 SSD slot, a 10GbE port and two 200Gbps QSFP ports for linking two or three units. It runs a custom DGX OS based on Ubuntu, weighs 1.2kg and uses a 240W power supply, and Lenovo says it supports models of up to 200B parameters. Lenovo promotes it as a companion to an existing ThinkPad or ThinkStation, so you offload AI jobs to it rather than replace your main computer.
ThinkStation and ThinkPad P-series. Lenovo’s specification site lists ThinkStation P2, P3, P4, P5, P7, P8 and PX desktops and ThinkPad P1, P14s and P16 laptops. These are the models to look at if you want a conventional tower or laptop with a dedicated NVIDIA graphics card, ISV-certified software and Lenovo’s business warranty options.
Lenovo Yoga Pro 9n. This RTX Spark laptop is listed at about US$3,000, the lowest launch price among the third-party RTX Spark models. StorageReview’s headline on its review says the 18-core version keeps pace with 20-core chips and runs cooler, which makes it worth a look if you plan to run it hard.
Best for: teams that already buy ThinkPads and ThinkStations and want an AI box that fits the same support and management. Skip it if: you want Windows on the AI machine itself, since the PGX runs Linux.
Asus AI Computer
Asus sells the widest mix of form factors: a GB10 mini PC, a tablet-style laptop with AMD’s big-memory chip, and a creator laptop with NVIDIA’s new RTX Spark.
ASUS Ascent GX10. This is ASUS’s GB10 machine, with the same chip and 128GB of memory as NVIDIA’s DGX Spark. Reviewers found its inference speed in line with the other GB10 boxes and praised its quiet cooling, front power button and magnetic stacking feet. It has four USB-C ports but no USB-A, one 10GbE port and two 200Gbps QSFP ports for linking a second unit. Its M.2 2242 slot is wired for PCIe 5.0, but the 1TB and 2TB models ship with slower PCIe 4.0 drives. Price is the issue: ASUS launched the 1TB model at US$2,999 in October 2025, and its store now lists the 1TB, 2TB and 4TB versions at US$5,999, US$6,999 and US$7,999, above NVIDIA’s own DGX Spark.
ROG Flow Z13. This 1.2kg tablet-style laptop uses the Ryzen AI Max+ 395, has a 13.4-inch touchscreen and a 70Wh battery. ASUS says it can run 70B-parameter models. That claim needs a high-memory configuration, and the 32GB version we found is better suited to smaller models, so check the memory size before you buy.
ProArt P16 with RTX Spark. TechSpot lists this creator laptop from US$3,400, shipping from 16 October.
Best for: developers who want a quiet GB10 box with good I/O, and travellers who want a very light laptop that can load larger models. Skip it if: you are price-sensitive on the GX10, where NVIDIA’s own version or an HP, Dell or Lenovo model may cost less.
Samsung AI Computer
Samsung’s Galaxy Book6 laptops are good everyday AI computers, but they are not built for running large models locally.
Galaxy Book6 Pro. The 14-inch model is a Copilot+ PC with an Intel Core Ultra 5 chip, a 47 TOPS NPU, 16GB of soldered memory, a 512GB SSD, a 14-inch AMOLED touchscreen and a weight of 1.24kg. Samsung’s list price is US$1,799.99, with a reseller listing at US$1,924. It adds Samsung’s Galaxy AI features to the Windows ones, which suits note-taking, translation, video calls and photo tools.
Galaxy Book6 Ultra. The 16-inch model uses an Intel Core Ultra 7 chip with a 50 TOPS NPU, an NVIDIA GeForce RTX 5060 with 8GB of graphics memory, 32GB of soldered RAM and a 1TB SSD. It is listed at US$3,699.99 and claims up to 30 hours of video playback.
The limits for AI work. An 8GB graphics card restricts you to small models; ComputingForGeeks says cards under 12GB are limited to 8B models with short context. Both Galaxy Books have soldered memory, so what you buy is what you keep. We did not find a Samsung computer aimed at large local models.
Best for: people who want a polished Windows laptop with on-device AI features and a bright screen. Skip it if: running 30B or larger models locally is the goal. Look at the unified-memory options earlier in this guide instead.
Quick picks: the best computer for AI by use
Start with the model size you want to run, then pick the platform that holds it.
| If you want to… | Look at | Why |
|---|---|---|
| Use cloud AI tools and on-device Windows features | A Copilot+ laptop such as the Samsung Galaxy Book6 Pro or Surface Laptop | 40 TOPS or more NPU, 16GB RAM, long battery life |
| Run models up to about 32B quickly on a desk | A desktop with an RTX graphics card of 24GB to 32GB | Highest memory bandwidth, CUDA support, though prices are inflated |
| Run 70B models quietly at a desk | Mac Studio with M5 Max and 128GB | Large, fast unified memory, near-silent |
| Get 128GB at the lowest price | A Ryzen AI Max+ 395 mini PC or HP Z2 Mini G1a | Cheapest 128GB, Windows or Linux |
| Develop with CUDA and link two machines | A GB10 mini PC (HP ZGX Nano, Dell, Lenovo PGX, ASUS GX10) | Full NVIDIA stack, 128GB, QSFP linking |
| Run big models in a laptop | Surface Laptop Ultra or another RTX Spark laptop, or the HP ZBook Ultra G1a | Up to 128GB unified memory, but newer platforms with early reliability reports |
| Buy for a business with support | Dell, HP or Lenovo workstations | Onsite support and fleet management |
A few rules cover most buying decisions:
- Buy memory first. It is soldered on almost every machine in this guide, so you cannot add it later.
- Match the machine to the model. Dense models need bandwidth, while mixture-of-experts models run well on big-memory machines.
- Check availability and price before you commit. Memory costs have roughly doubled the price of Ryzen AI Max+ mini PCs since June, and the 512GB Mac Studio is not due until late October.
Which AI computer should you buy?
The right AI computer is the one with enough memory for your models, from a platform your software supports, at a price you can justify this month.
For most people, a recent Copilot+ laptop covers everyday AI. For serious local work, a Mac Studio, a Ryzen AI Max+ machine or a GB10 mini PC gives you the capacity that graphics cards cannot match at a sensible price, while an RTX desktop remains the fastest choice for smaller models. Because most of these machines have soldered memory, a wrong guess on capacity is expensive to fix.
Because most of these machines have soldered memory, work out the model size you need before you buy, and check current prices and availability, since both are moving quickly.