AI Day Singapore ran on 22 and 23 September 2026 at the Raffles City Convention Centre, and NVIDIA used it to publish a regional announcement rather than a product launch. There was no new chip and no new flagship model. There was a list of fifteen named organisations across five countries, each with something at research, fine-tuning or production stage.
That shape is the point. The announcement NVIDIA posted on 22 September sorts its partners into three themes: moving public sector AI from pilot to production, expanding Nemotron open models for region-specific work, and one enterprise adopting the Vera Rubin platform ahead of everyone else in ASEAN. AI Day Singapore is being used as a checkpoint on whether Southeast Asia has stopped experimenting.
This article counts the partners, sets out the model sizes NVIDIA did not put in the blog post, checks the one user number the announcement does quote, adds the SEA-LION benchmark figures published four days earlier, and separates what is deployed from what is still described as research.
Table of contents
- What AI Day Singapore Actually Announced
- The AI Day Singapore Partner Map, Counted
- Singapore’s Own Agencies Took the AI Day Singapore Stage
- The Nemotron Models Behind AI Day Singapore
- Beyond Singapore: Malaysia, Vietnam, Thailand and Brunei
- Sea Limited and the First ASEAN Vera Rubin Deployment
- SEA-LION v4.8: The Numbers AI Day Singapore Did Not Quote
- Reading the AI Day Singapore Claims Carefully
- What AI Day Singapore Means If You Build in the Region
- Frequently Asked Questions About AI Day Singapore
- References
What AI Day Singapore Actually Announced
Three separate announcements were bundled into the one AI Day Singapore post. Keeping them apart makes the whole thing easier to judge.
Public sector AI from pilot to production
NVIDIA framed this around four areas: enhancing government operations and service delivery, developing accessible citizen services and empowering local businesses, strengthening critical infrastructure and public safety, and supporting startups, developers and researchers to build national AI capability.
Region-specialised models on Nemotron
The second strand covers enterprises, technology providers and research organisations building on NVIDIA Nemotron open models, datasets and libraries. The pitch is ownership: customise, control and own models rather than rent them, and own the big data feeding them rather than shipping it abroad.
One enterprise on Vera Rubin
The third is a single name. Sea Limited became the first enterprise in the ASEAN region to adopt the NVIDIA Vera Rubin platform. That is the only hardware news in the entire AI Day Singapore announcement.
The framing is sovereign AI
NVIDIA’s stated ambition in the post is to “enable all nations to be AI nations”, providing technology, infrastructure, ecosystem and expertise. The tags on the announcement include Sovereign AI and Economic Development, which tells you who the intended reader is: ministries, not developers.
What was not announced
No new NVIDIA model launched at AI Day Singapore. No pricing changed. No sovereign cloud build-out was disclosed with a figure attached. Reading it as a hardware event would be a mistake.
The AI Day Singapore Partner Map, Counted
The AI Day Singapore announcement names organisations quickly and without a tally, so here is one.
| Organisation | Country | What it is building | NVIDIA stack |
|---|---|---|---|
| HTX | Singapore | Public safety research | Nemotron 3 Super, Nano Omni |
| NCS | Singapore | Agentic AI, humanoid robotics | Nemotron, VSS Blueprint |
| ST Engineering | Singapore | AI Studio, Marine MRO | NeMo, cuOpt |
| AI Singapore | Singapore | SEA-LION model family | Nemotron, NeMo |
| Hummingbird Bioscience | Singapore | Toxicity Knowledge Graph | Nemotron 3.5 Lightning, NeMo Retriever |
| Bitdeer AI | Singapore | Open Models AI Codefest | GPU cloud infrastructure |
| Sea Limited | Singapore | Shopee, Monee, Garena AI | Vera Rubin |
| Viettel AI | Vietnam | Legal AI agent harness | Nemotron 3 Super |
| FPT Smart Cloud | Vietnam | Nemotron-Personas-Vietnam | Nemotron datasets |
| iApp Technology | Thailand | OpenThai 2.0 Legal, Thanoy | Nemotron 3 Nano, NeMo |
| Big Data Institute | Thailand | ThaiLLM Collaboration | Nemotron evaluation |
| AS-TECH | Thailand | Airport passenger flow | Cosmos, VSS Blueprint |
| YTL AI Labs | Malaysia | Enterprise and citizen services | Nemotron fine-tuning |
| ITMAX | Malaysia | City traffic operations | Cosmos, VSS Blueprint |
| Antrique | Brunei | Food sector productivity | AI innovation platform |
Singapore takes nearly half
Seven of the fifteen country-attributed organisations named at AI Day Singapore are Singaporean. Thailand contributes three, Vietnam and Malaysia two each, and Brunei one. A sixteenth partner, LynxKite, appears alongside Hummingbird Bioscience without a country attached, so it sits outside the count.
Four of five are ASEAN founding members
Singapore, Thailand, Malaysia and the Philippines were among ASEAN’s founding members in 1967, and three of those four appear here. Brunei joined in 1984 and Vietnam in 1995, and both are represented. The AI Day Singapore map tracks the older half of the bloc more closely than the newer half.
Who is missing
Indonesia and the Philippines are the two largest ASEAN markets by population and neither has a named partner in this announcement. For an event billed as covering Southeast Asia at large, that is a gap worth noticing rather than glossing over.
Singapore's Own Agencies Took the AI Day Singapore Stage
The host country’s contributions to AI Day Singapore are the clearest signal of how far public sector work has moved.
HTX is still at research stage
Singapore’s Home Team Science and Technology Agency is “embarking on research” using Nemotron 3 Super and Nemotron 3 Nano Omni. NVIDIA’s own wording uses “has the potential to” and “could help” throughout. This is not a deployment announcement, and reading it as one would misrepresent it.
What HTX is actually testing
Super is being assessed for complex reasoning and agentic workflows. Nano Omni’s unified vision, audio and language capabilities are being assessed for multimodal applications grounded in real-world operational data. The stated goal is secure, locally controlled AI across Singapore’s Home Team.
NCS is further along
NCS is described as advancing agentic AI adoption across enterprises and the public sector using Nemotron models and the NVIDIA Blueprint for video search and summarisation, while also advancing physical AI for humanoid robotics. The drivers named are security, responsiveness and data governance.
Data governance is the recurring word
Security, responsiveness and data governance are the three requirements NVIDIA names for NCS, and “secure, locally controlled AI” is the phrase used for HTX. Open weights that can run inside a national boundary are the product being sold at AI Day Singapore, more than any particular benchmark.
ST Engineering has a named production use
ST Engineering is using NVIDIA NeMo tools and cuOpt software to build its AI Studio platform and deploy agentic AI across its businesses, with Marine maintenance, repair and overhaul called out specifically. A named business line is a stronger claim than a research programme.
The Nemotron Models Behind AI Day Singapore
The AI Day Singapore announcement mentions four Nemotron variants and gives the size of none of them. The sizes matter, because they decide what a national agency can actually run.
| Model | Total parameters | Active per token | Named user in the region |
|---|---|---|---|
| Nemotron 3 Nano | 30B | 3B | iApp Technology, Thailand |
| Nemotron 3 Nano Omni | 30B | 3B | HTX, Singapore |
| Nemotron 3 Super | 120B | 12B | Viettel AI, Vietnam |
| Nemotron 3 Ultra | ~500B | up to 50B | None named |
| Nemotron 3.5 Lightning | Not published | Not published | Hummingbird Bioscience |
Sparse activation is the regional story
A 120B model that activates 12B parameters per token costs roughly a tenth of a dense 120B model to serve. Divide 120 by 12 and you get the ratio that makes a national-scale deployment affordable on a modest GPU estate. That arithmetic, not the model quality, is why Nemotron keeps appearing at AI Day Singapore.
Nano Omni replaces a pipeline
Nemotron 3 Nano Omni is a hybrid mixture-of-experts Transformer-Mamba model with Conv3D video layers and efficient video sampling. NVIDIA reports roughly twice the throughput and 2.5 times lower compute for video reasoning against running a computer vision model and a speech model as separate pipelines. For a public safety agency processing camera feeds, that is the whole argument.
Lightning is undocumented in the post
Nemotron 3.5 Lightning powers Hummingbird Bioscience’s toxicity knowledge graph alongside NeMo Retriever, and it is the only model named at AI Day Singapore whose size NVIDIA has not published in the same family table as the others. Treat its capability claims as unverified until it is.
Ultra is absent
No AI Day Singapore partner is named as using Nemotron 3 Ultra. At roughly 500 billion parameters with up to 50 billion active, it is the tier a sovereign programme would reach for last, and nobody in this announcement has.
Beyond Singapore: Malaysia, Vietnam, Thailand and Brunei
The regional partners at AI Day Singapore are, in several cases, further into production than the host country’s agencies.
Thailand has the only user number
iApp Technology adapted Nemotron 3 Nano by fine-tuning OpenThai 2.0 Legal on Thai-language legal data with the NeMo framework. The model is released open source for Thai developers and powers Thanoy, a legal-assistant chatbot already serving approximately 43,000 users. That is the single deployment figure in the entire AI Day Singapore announcement.
Vietnam has the only benchmark claim
Viettel AI fine-tuned Nemotron 3 Super for Vietnamese and agentic applications, and reports the highest ranking on both the VMLU benchmark and its own in-house product benchmark. The model is set to be adopted in Legal AI, an agent harness serving Viettel Group employees and external customers.
Vietnam also contributed a dataset
FPT Smart Cloud co-developed Nemotron-Personas-Vietnam, an open dataset grounded in Vietnamese demographic and cultural data, so local developers can post-train and evaluate localised models. Datasets outlive model releases, which makes this the more durable contribution.
Smart cities are a Cosmos story, not a Nemotron one
Malaysia’s ITMAX uses NVIDIA Cosmos open world models with the video search and summarisation Blueprint for city traffic operations. Thailand’s AS-TECH applies the same stack to airport passenger flow. Both are the same architecture pointed at different footfall problems.
Legal work dominates the use cases
Two of the five countries put their flagship effort into legal assistance: Thailand’s OpenThai 2.0 Legal and Vietnam’s Legal AI harness. Statute is high-volume, jurisdiction-specific, poorly served by English-first models and safe to pilot, which makes it the obvious first target.
Brunei is the smallest entry
Antrique built an AI innovation platform to raise productivity across Brunei’s food sector. It is one organisation, one sector and no published numbers, which is a fair reflection of where the country sits in the regional build-out.
Sea Limited and the First ASEAN Vera Rubin Deployment
The Vera Rubin adoption is the only hardware commitment at AI Day Singapore, and Sea is a serious place to put it.
The scale behind the name
Sea operates Shopee in e-commerce, Monee in digital financial services and Garena in gaming, serving hundreds of millions of users. Its second-quarter 2026 results put numbers on what “hundreds of millions” means in practice.
| Division | Q2 2026 headline figure | Year-on-year | AI use named by NVIDIA |
|---|---|---|---|
| Group revenue | $7.8 billion | +48.1% | – |
| Shopee GMV | $38.3 billion | +28.4% | Listings, discovery, engagement |
| Shopee gross orders | 4.2 billion | +27.5% | Seller tooling |
| Monee loan book | $11.1 billion | +62.5% | Fraud detection, credit risk |
| Garena bookings | $763.5 million | +15.5% | Gaming experiences |
| Garena quarterly paying users | 68.1 million | +10.2% | Safe online spaces |
Why the numbers matter here
Sea reported $458.1 million of net income on $7.8 billion of revenue in the quarter. A company clearing that much can fund frontier hardware from cash flow, which is what makes it a credible first ASEAN adopter rather than a logo on a slide.
The small-business argument
NVIDIA’s framing is that AI on Shopee lets small and medium-sized enterprises with limited resources use enterprise-grade capability previously available only to large corporations. With 4.2 billion gross orders in a quarter, even a small per-listing improvement compounds into something material.
Monee is the sharpest AI case
A loan book that grew 62.5 per cent year on year to $11.1 billion while the 90-day non-performing loan ratio held at 1.0 per cent is the clearest commercial argument in the announcement. Fraud detection and credit risk assessment are where model quality converts directly into retained capital.
What Vera Rubin is for
The stated purpose is developing and deploying models and intelligent agents at greater scale. No cluster size, no GPU count and no spend figure were disclosed, so treat the adoption as directional rather than quantified.
SEA-LION v4.8: The Numbers AI Day Singapore Did Not Quote
AI Singapore’s SEA-LION work gets one sentence in the AI Day Singapore announcement. The detail published four days earlier is far more useful.
Two models, two sizes
On 18 September 2026, AI Singapore announced Nemotron-SEA-LION-v4.8 in two variants: a 30B-A3B built on the Nano tier and a 120B-A12B built on the Super tier. They map exactly onto the Nemotron sizes in the table above.
The benchmark movement is large
On SEA-HELM, the 30B-A3B variant moved from 46.89 to 51.57, a gain of 4.68 points. The 120B-A12B variant moved from 49.30 to 63.44, a gain of 14.14 points. The larger model gained roughly three times as much, which is the opposite of the usual diminishing-returns pattern.
Low-resource languages gained most
At the 120B scale, Burmese rose from 4.98 to 31.35 and Tamil from 21.83 to 56.35 — increases of about 6.3 times and 2.6 times respectively. Those are the languages commercial models serve worst, and they are the reason a regional model family exists at all.
The training budget was modest
The 30B variant used 150 billion tokens of continued pre-training and the 120B variant used 33.5 billion. The smaller model consumed roughly 4.5 times more tokens than the larger one and gained less than a third as much on the benchmark.
Seven languages, not thirteen
The v4.8 release covers seven regional languages: Burmese, Filipino, Indonesian, Malay, Tamil, Thai and Vietnamese. Earlier SEA-LION releases claimed broader coverage, so quote the version number alongside the language count or the figure becomes meaningless.
Reading the AI Day Singapore Claims Carefully
Regional announcements compress a lot of different commitment levels into one voice. Here is how to separate them.
Three tiers of certainty
“Is embarking on research using” (HTX) is the weakest. “Is fine-tuning” or “is adapting” (YTL AI Labs, iApp Technology) is middling. “Already serves approximately 43,000 users” (Thanoy) is the only claim in the announcement with a measured outcome attached.
Benchmark rankings need a date
Viettel AI’s claim to top the VMLU leaderboard is a point-in-time statement about a public leaderboard that changes. It is a real result, but it is not a durable property of the model.
“First in ASEAN” is a procurement fact
Sea being the first ASEAN enterprise on Vera Rubin says something about purchasing order, not about capability delivered. Nothing has been published about what is running on it yet.
Potential is not deployment
Counting the hedges is a fast sanity check. The public sector section of the AI Day Singapore post uses “has the potential to”, “could help” and “could strengthen” for its host-country flagship, and reserves the plain past tense for partners outside Singapore.
Open source is doing real work here
The genuinely checkable claims are the open ones: OpenThai 2.0 Legal is released for Thai developers, Nemotron-Personas-Vietnam is an open dataset, and the SEA-LION benchmark numbers are published with before-and-after scores. Those can be verified without NVIDIA’s help.
What AI Day Singapore Means If You Build in the Region
Practical reading of AI Day Singapore for anyone with a deployment in or near these markets.
Start at the Nano tier
Every deployment named at AI Day Singapore whose model size NVIDIA has published sits on the 30B or 120B tier, and none on Ultra. A 30B model with 3B active parameters is the current regional default for a reason, and it is where a first localisation attempt should start.
Language coverage is a version number
If Burmese or Tamil matters to your product, the SEA-LION v4.8 gains are the most concrete evidence available that a regional fine-tune beats a general model. Check the specific release, because the language list changes between them.
Video workloads have a stack now
Cosmos plus the video search and summarisation Blueprint appears twice at AI Day Singapore, once for traffic and once for airport passenger flow. If you are doing anything with camera feeds in the region, that pairing is the pattern others have already validated.
Robotics is arriving alongside it
NCS is explicitly advancing physical AI for humanoid robotics under the same programme. NVIDIA published its Isaac ROS 5.0 agentic robotics release on the same day as this announcement, which is the developer-facing half of the same push.
Own the dataset, not just the model
FPT Smart Cloud’s persona dataset and iApp’s open legal model are the two contributions most likely to still matter in three years. Model weights are replaced; curated regional data is not. Our AI models and tools hub tracks where those releases land.
Frequently Asked Questions About AI Day Singapore
When and where was AI Day Singapore held?
It ran on 22 and 23 September 2026 at the Raffles City Convention Centre in Singapore, with hands-on training and expert-led sessions alongside the partner announcements.
How many partners did NVIDIA name?
Sixteen organisations are named in total. Fifteen have a country attached: seven in Singapore, three in Thailand, two each in Vietnam and Malaysia, and one in Brunei.
What is the biggest deployment announced?
By measured users, iApp Technology’s Thanoy legal-assistant chatbot in Thailand, serving approximately 43,000 users on a fine-tuned OpenThai 2.0 Legal model built from Nemotron 3 Nano.
Did NVIDIA announce new hardware at AI Day Singapore?
No. The only hardware item was Sea Limited adopting the existing Vera Rubin platform as the first enterprise in ASEAN. No cluster size or spend figure was published.
Which Nemotron models are Southeast Asian partners using?
Nemotron 3 Nano and Nano Omni at 30B total and 3B active parameters, Nemotron 3 Super at 120B total and 12B active, and Nemotron 3.5 Lightning at Hummingbird Bioscience. No partner is named on Nemotron 3 Ultra.
Is SEA-LION part of this?
Yes. AI Singapore is expanding SEA-LION to include Nemotron open models and NeMo tools, and published Nemotron-SEA-LION-v4.8 on 18 September 2026 in 30B-A3B and 120B-A12B variants.
References
At AI Day Singapore, NVIDIA and Partners Showcase AI Advancements Across Southeast Asia
Announcing Nemotron SEA-LION v4.8 in collaboration with NVIDIA
NVIDIA Launches Nemotron 3 Nano Omni, Unifying Vision, Audio and Language
Nemotron 3 Nano Omni Powers Multimodal Agent Reasoning in a Single Efficient Open Model
iApp Technology: OpenThai 2.0 Legal launch
ITMAX Advances AI for Malaysia Smart Cities With NVIDIA Cosmos
Antrique launches NVIDIA-powered AI innovation platform in Brunei
Sea Limited Q2 2026 Earnings Call Transcript
Sea Q2 2026 slides: revenue surges 48%, Shopee eyes $1B EBITDA
SEA-LION: Empowering Open Multilingual AI for Southeast Asia
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
