Both halves of the headline are true. AI hardware that fits on a desk now has 128 GB to 512 GB of memory the GPU can use, enough to run models of around 100–120 billion parameters locally. Examples are NVIDIA's GB10 boxes (DGX Spark and partner versions), AMD Ryzen AI Max+ 395 mini PCs, the Mac Studio M5 Ultra and RTX Spark laptops. At the same time, Amazon, Microsoft, Alphabet and Meta are on course to spend roughly US$720–740 billion between them on capital expenditure in 2026, and all four have signalled more for 2027. The two trends don't compete. Small boxes are taking over private, single-user and small-team work, while data centres keep frontier training, large-scale serving and anything that needs hundreds of concurrent users or more memory than a desk can hold.
How small has AI hardware become?
| System | GPU-usable memory | Memory bandwidth | Power | Typical role |
|---|---|---|---|---|
| AMD Ryzen AI Max+ 395 mini PC (128 GB) | About 96–120 GB of 128 GB | About 215 GB/s (measured) | Mini PC class | One user, local inference |
| NVIDIA GB10 (DGX Spark, ASUS Ascent GX10 and others) | 128 GB unified | 273 GB/s | 240 W adapter; about 160–200 W under load | Developer box, small-team inference |
| RTX Spark laptop (e.g. ASUS ProArt P16 128 GB) | Up to 128 GB unified | Not published for the laptop | Laptop | Mobile AI and creative work |
| Apple Mac Studio M5 Ultra | 96–512 GB unified | Up to 1.2 TB/s | 480 W maximum | Large local models, creative studios |
| NVIDIA DGX Station GB300 | 748 GB (252 GB HBM3e + 496 GB LPDDR5X) | 7.1 TB/s (HBM3e) | 1,600 W; needs a dedicated circuit | Department-scale AI at the desk |
| GB300 NVL72 rack | 72 GPUs linked as one system | NVLink fabric | About 132–142 kW per rack (OEM figures) | Data-centre training and serving |
Source: NVIDIA, AMD, Apple and ASUS specifications; measured bandwidth and power from published reviews (ServeTheHome, Tom's Hardware); rack power from Supermicro, GIGABYTE, Schneider Electric and Vertiv figures; compiled 8 Nov 2026. SourceIT has not tested these systems.
What can a desk-sized AI computer actually run?
Memory decides which models fit, and bandwidth decides how fast they answer. A GB10 box with 128 GB runs gpt-oss-120b, a 120-billion-parameter mixture-of-experts model, at about 55–59 tokens per second for a single user with llama.cpp, according to published reviews. That's comfortable reading speed. Commentators in mid-2026 put the practical local ceiling at roughly 100–120 billion parameters for MoE models at 4-bit precision. The largest open models, such as GLM 5.3 at 753 billion parameters, need clusters or a cloud API. NVIDIA showed an eight-node GB10 cluster running GLM 5.3 at 31 tokens per second for one user (NVIDIA blog, 22 Oct 2026). That is a demonstration of what's possible, not a cheap way to serve a team.
So the small box is a real tool, not a toy. It handles private document chat, coding assistants, fine-tuning small models with LoRA, image generation and prototyping. Your data stays in the office, which matters under Singapore's PDPA when the documents contain personal data. We go deeper on that trade-off in can a mini AI supercomputer replace an entire server?
Why aren't data centres going away?
Three things keep the big rooms busy.
Concurrency. One GB10 box serving gpt-oss-120b through vLLM reached 69 to 671 tokens per second in aggregate as the batch grew to 64 requests (StorageReview, 3 Mar 2026). That's fine for a team. It is a long way from serving a whole company or a customer-facing app at peak, which is what H200 and GB300-class servers are built for.
Prompt processing and long context. Small boxes are memory-rich but compute- and bandwidth-limited. Reading a 100-page contract into the model (prefill) is where data-centre GPUs pull far ahead. Clusters of small boxes also scale sub-linearly: NVIDIA's eight-node GB10 result was about 4.4 times one node, not eight.
Training and frontier models. Training a large model, or running one with hundreds of billions of parameters at speed, still needs racks like the GB300 NVL72 drawing 132–142 kW each. In Singapore, Nxera runs GB200 systems at about 200 kW per rack, and NSCC's ASPIRE 2B, launched 8 Jun 2026, has more than 1,500 H200 GPUs. No desk can host that.
Rule of thumb: if the work is one person or one small team with private data, it can probably live on a desk. If it needs many concurrent users, high uptime or a model bigger than about 128–512 GB, it belongs in a data centre or the cloud.
Is the hyperscaler spending a bubble?
It might cool, but the direction is clear. Company guidance for 2026, as reported, is about US$220 billion for Amazon, about US$170 billion for Microsoft (calendar year, lowered because of third-party capacity delays), US$195–205 billion for Alphabet and US$135–145 billion for Meta. That adds up to roughly US$720–740 billion (CALC; the companies define capex slightly differently, so treat it as approximate). Alphabet says 2027 spending will "increase significantly", Amazon expects "meaningfully more", and Meta says it will "continue investing aggressively".
Microsoft trimming its own figure is a reminder that capex can't rise forever. But nothing in the guidance suggests small AI boxes are pulling demand away from data centres. Both are growing at once, for different jobs.
What does this mean for a Singapore business?
Treat it as a placement decision, workload by workload:
- On a desk: developers, analysts and creative teams who need private, always-available AI. A GB10 box running flat out around the clock at about 160 W uses about 1,402 kWh a year, roughly S$437 at SP Group's 31.16¢/kWh tariff including GST (CALC, illustrative; business tariffs differ).
- Department scale: a DGX Station GB300 brings 748 GB to the office, but at 1,600 W flat out it uses about 14,016 kWh a year, roughly S$4,370 at the same tariff (CALC), and it needs an electrician-checked dedicated circuit. Our comparison of DGX Station and GPU servers covers when that beats a rack server.
- Data centre or cloud: customer-facing services, training runs and anything that must stay up when one box fails. Singapore's data centres run at only about 2–4.8% vacancy, so colocation space is tight. Plan early. For the path from one box upwards, see how NVIDIA scales from your desk to the data centre.
- Neither: for many SMEs, a SaaS tool or an API costs less than any hardware. Before buying, read whether businesses are buying AI servers before they have a strategy.
Choosing a platform also means choosing a software ecosystem. Whether that dependence on NVIDIA is becoming a business risk is the subject of NVIDIA vs AMD for enterprise AI.
Buying in Singapore
In the GB10 class, the ASUS Ascent GX10 64 GB / 1 TB is S$6,670, the ASUS Ascent GX10 128 GB / 1 TB S$10,350 and the NVIDIA DGX Spark 4 TB S$9,300 (all listed 8 Nov 2026). Compare more brands in the GB10 collection. For RTX Spark in a laptop, the ASUS ProArt P16 128 GB is S$10,773.20 (listed 8 Nov 2026). The ASUS ProArt GR1X, Apple Mac Studio and DGX Station GB300 systems such as the MSI XpertStation WS300 are price on request. For budgeting, public US pricing of roughly US$93,000–125,000 for most OEMs works out at about S$119,000–160,000 before GST at USD/SGD 1.279 (6 Nov 2026); the final figure depends on OEM, storage and options, so ask us for a quotation. That is an assumption, not a quotation. Prices change; ask for a current quotation. For data-centre GPU servers and colocation-ready builds, ask us for a sizing discussion. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can a small AI computer replace a data centre? For one person or a small team running models up to about 100–120 billion parameters, often yes. For many concurrent users, training or very large models, no.
How big a model can a 128 GB GB10 box run? Reviews show gpt-oss-120b running at about 55–59 tokens per second for a single user. Larger open models need clusters or a cloud service.
Will hyperscalers keep building data centres? Their 2026 guidance adds up to roughly US$720–740 billion across Amazon, Microsoft, Alphabet and Meta, and all four have signalled more for 2027, though Microsoft has trimmed its 2026 figure.
How much electricity does a desk AI box use? A GB10 box at about 160 W around the clock uses about 1,402 kWh a year, roughly S$437 at SP Group's 31.16¢/kWh including GST. A 1,600 W DGX Station uses about ten times that.
Do I need either for my business? Not always. If your AI use is occasional or uses public data, a SaaS tool or API may cost less than buying hardware.
