Singapore's HTX Deploys Nvidia GB300 Public Safety Robots in Sovereign AI First

Singapore's Home Team Science and Technology Agency has deployed Nvidia's GB300 accelerator, the first organisation in the country to do so, and it has aimed the chip at robots rather than chatbots. HTX disclosed the work at Nvidia AI Day Singapore, held Sept. 22-23 at the Raffles City Convention Centre. The agency says the hardware is the base for training larger multimodal and embodied AI models and for running inference at scale. The Nvidia GB300 public safety robots programme is a procurement event as much as a technical one: a government agency bought the newest tier of AI compute and aimed it at machines that work next to people.

The GB300 is the Blackwell-generation part that reached commercial availability in the first half of 2026. Nvidia positions it as roughly twice as fast as the preceding GB200 on AI deep-thinking workloads, with additional gains on text, image and code tasks. Cloud providers have led adoption, buying racks of the chip to serve models over a network. HTX's installation differs: the compute sits with the agency that owns the robots, and the models it trains stay inside the operator's control.

HTX runs NGINE on the hardware, an AI infrastructure layer that processes text, video and audio in one pipeline and is built to control robots at scale. Ang Chee Wee, HTX's chief AI officer, says the platform lets the agency train larger multimodal and embodied models and run inference at scale while keeping code and data in house. That control matters because the same models read sensor feeds, coordinate machines and handle operational material from Singapore's police, fire and civil defence services.

What the GB300 Public Safety Robots Deployment Signals

Data-centre AI and edge physical AI have been treated as separate markets with separate buyers: cloud capacity in one column, robot fleets in another. Putting GB300-class silicon inside a government robot programme merges them. The chip generation that anchors large training clusters now sits behind hazard-sensing machines, the kind HTX has described working alongside firefighters at a chemical plant blaze. Once one part serves both roles, the buyer list widens, and vendors have to sell to procurement offices as well as to cloud engineering teams.

Sovereignty carries much of the weight in that pitch. A public safety agency cannot route video from an incident scene through a foreign cloud without accepting legal and jurisdictional exposure. HTX is separately researching Nvidia's Nemotron 3 Super and Nemotron 3 Nano Omni models for public safety operations, including agentic workflows and multimodal applications grounded in real operational data. Local compute plus locally controlled models is the bundle being offered to governments, and Nvidia is offering it across the region.

The common objection is that a GB300 is an odd fit for a robot. Edge inference usually trades throughput for power efficiency, and most fielded robots run compact quantised models on modest accelerators, because a flagship Blackwell part draws far more power than a patrol machine can carry. The objection misreads where the compute sits. Training larger multimodal and embodied models is a data-centre-class job, and HTX frames GB300 as the engine for that work, with inference run at fleet scale rather than on each individual robot. The machine is the endpoint; the GB300 is the training and coordination layer.

Whether that split holds decides if the deployment pays off. If training capacity was the binding constraint on HTX's roadmap, the agency has removed it. If inference latency on the robot was the constraint, GB300 changes little, and the test will be published throughput numbers rather than capability descriptions.

Nvidia's account of the Singapore event points to a second shift, from experimentation to production-scale deployment across the region's public and private sectors. That distinction matters for budgets. Pilots are cheap and reversible; production deployments lock in training pipelines, model choices and hardware roadmaps for years. An agency that standardises on GB300 for training and Nemotron for models commits to one vendor at every layer, and switching costs climb with each retrained model.

Southeast Asia has become the proving ground for this approach. Nvidia used the Singapore event to lay out a regional roster spanning public safety, legal services, healthcare and transport.

The roster shows how far the sovereign argument reaches. AI Singapore, Viettel AI, FPT Smart Cloud and Thai partners are adapting Nemotron resources for regional languages and cultural contexts, which extends the case for locally controlled AI from hardware to models. Thailand's OpenThai 2.0 Legal model is set to be open source, and the Nemotron-powered Thanoy legal assistant serves about 43,000 users, one of the few sovereign models with a measurable user base.

Procurement is the mechanism that carried the technology into a government robot programme. HTX, a state agency, bought frontier AI compute and installed it inside its own operations rather than renting equivalent capacity from a cloud provider. That keeps the training pipeline and the operational data under the agency's control, and it makes the agency, not a hyperscaler, the buyer of record for the hardware.

For decision-makers outside Singapore, the lesson is about sequencing. HTX paired a sovereign training stack with research on smaller Nemotron variants, so the agency can test model behaviour before committing production workloads. Vendors selling into public safety will have to answer the question HTX is answering now: which parts of the pipeline must stay on local hardware, and which can sit in someone else's cloud.

HTX going first shows what regional buyers can purchase, as well as what is technically feasible. Malaysia's work on enterprise and citizen services, traffic operations and airport passenger-flow tracking points the same way, and NCS is pairing Nemotron models with Nvidia's video search and summarisation blueprint on public-sector projects while pursuing physical AI for humanoid robotics.

Why this matters

A government agency has bought the top tier of Nvidia's accelerator line and aimed it at robots, which makes flagship AI compute a public-safety purchase rather than a niche for cut-down edge chips. The next wave of AI hardware revenue will be argued in procurement offices as much as in data-centre buildouts, and governments that want control over their models and data will pay a premium for it. HTX's deployment shows the terms of that trade are being set now, one agency at a time.

Sources

At AI Day Singapore, NVIDIA and Partners Showcase AI Advancements Across Southeast Asia

Related Articles

• NVIDIA Strengthens South Korean Ties to Advance Sovereign AI and Robotics Infrastructure
• Sharon AI Asia-Pacific Expansion Secured Through $950 Million Cloud Infrastructure Deal
• Singtel and Mistral AI Collaborate to Scale Sovereign AI Infrastructure in Southeast Asia

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.