Both frontier AI labs now have people in Singapore. That changes what you can ask for.
Anthropic announced on 16 September that its Singapore office opens in October — its first in Southeast Asia and fifth in Asia-Pacific. OpenAI has had its Asia headquarters here since late 2024 and is in talks to add 100,000 sq ft. Anthropic's stated reason is straightforward demand: Singapore ranks second out of 121 countries for Claude usage per head, behind only Australia.
Two years ago buying frontier AI from Singapore meant a web form and a support queue in another timezone. That is changing, and it has practical consequences well before it has strategic ones.
If you're mid-negotiation, say so. Two vendors with local staff chasing the same reference customer in the same city is leverage on price, on support response times, and on contract terms that simply wasn't available a month ago. Ask what local support actually means — named contact, hours, escalation path — and get it written down rather than described.
If you've been waiting for local presence before committing, that reason has mostly expired. Waiting further costs you the months, not the terms.
If you're hiring, move sooner rather than later. Both labs are recruiting in Singapore for exactly the profile you want — someone who can run an AI programme, not someone who can build a model.
Does this affect you? The week's announcements, translated
You see a dozen AI launches a week. Most change nothing about how your business runs. Here are the three from this week that might, and what each is actually worth to you.
An "open source" model just stopped being free to use commercially
What happened. Alibaba shipped a new version of its Qwen image model on 18 September under a research-only licence rather than the open Apache terms that covered earlier versions. The text is blunt: use is permitted "FOR NON-COMMERCIAL PURPOSES ONLY," and anything commercial requires a separately negotiated licence.
Matters to you if anything you sell or operate has open-weight models inside it — whether you know it or not. The trap is that open-source models are licensed per version, not per family. A team that picked Qwen a year ago because it was freely usable will not necessarily get a warning when they pull the next release, and "we host it ourselves so licensing isn't an issue" is precisely the assumption that fails here.
You can ignore it if everything you run is a commercial API from a named vendor with a contract.
The thirty-minute version. Ask whoever builds or maintains your systems for a list of AI models in production, with the version and licence of each. If that list doesn't exist, that's the finding.
Anthropic hired its own safety auditor. Note how the claim will be worded.
What happened. On 18 September Anthropic and Accenture announced that Accenture staff will embed inside Anthropic to stress-test its models and safety practices, with both sides expecting to spend at least $1 billion each over five years. Anthropic says the arrangement makes its accountability "more verifiable." Critics, who expected independent non-profits in the role, called it self-policing with extra steps.
Matters to you if vendor safety claims appear anywhere in your procurement. Third-party AI assurance is becoming a product, which means "independently evaluated" is about to start appearing in proposals on your desk. It is not a meaningless claim — but it isn't self-evidently a strong one either.
You can ignore it if no one has yet put an AI-related assurance claim in front of you.
The one question. Who chose the assessor, and who pays them? In this week's flagship example the answer to both is the vendor. That may still be worth something; it just isn't independence, and knowing the difference is the whole value of asking.
Several new models shipped and got cheaper. Nothing you run changes.
Chinese labs released a run of models this week, some undercutting Western pricing substantially. If you buy inference in volume, your procurement team should be watching that price curve. If you don't — and most businesses buy finished software, not tokens — this is industry news, not your news. The test to apply every week is whether an announcement changes a process you own, a cost you carry, or a risk you're accountable for. Most don't.
What going first exposed: everyone has deployed it, almost nobody has connected it
Research presented on 14 September found 98% of companies have deployed AI somewhere in the customer journey — and only 15% have combined it with orchestration across departments. Asked what was blocking them, respondents named compliance (50%), security (48%) and disconnected systems (45%).
Read the source with appropriate suspicion: the study comes from Talkdesk, which sells the orchestration layer it has diagnosed a shortage of, and its finding that mature adopters see far better satisfaction scores is doing commercial work. The barrier numbers are the honest part, and they describe an integration problem rather than an AI one. Nearly half of companies are held up because their systems don't talk to each other — a constraint that predates AI by a decade and that no model release addresses.
A chatbot bolted onto a support inbox is the easy 98%. The value sits in the part where a customer request actually reaches the system that can resolve it, and that work is unglamorous integration between things you already own. One useful check: when your AI can't answer something, what happens next? If the answer is "it tells the customer to email someone," you are in the 98%, not the 15%. Worth noting too that on Talkdesk's own numbers the satisfaction gains were large while the cost-per-contact gap was narrow — if your business case rests on headcount savings, test that assumption early.
Government initiatives
What governments in the region are funding and building, as distinct from what they're regulating.
PSG, EDG and MRA close on 29 September — six days from publication — and nothing further has been published. The EDGE grant page still shows only the headline terms: up to 70% for SMEs, up to 50% for non-SMEs, and a S$100,000 annual cap across all activities. It states that support levels differ by activity without yet publishing the activity-level detail — which is the part that decides what a system project actually recovers. If you are an SME with a scoped project, this is the last week. If you are not, there is still nothing to act on; set a reminder for when the activity pages appear.
The Philippines published a $34.4 billion AI build-out plan on 9 September. It runs to 2033, splits roughly $21bn private and $13.5bn public, and puts about $14.6bn into data centres — targeting growth from roughly 50MW of capacity today to 1.5GW by 2033, anchored on the Clark-Bataan corridor, with a stated goal of 500,000 AI-related jobs. Treat the numbers as indicative rather than committed: roughly $16.3bn of it is still to be found, and the responsible secretary's own framing is that serious investors may appear this year for projects starting in 2027. Relevant mainly if you are weighing where regional data should live.
Hiring signals
What's happening to AI-related roles, in Singapore where the data exists.
Singapore's AI hiring is overwhelmingly for users, not builders. PwC's Global AI Jobs Barometer, Singapore edition (published 15 June, covering about 1.6 million Singapore postings across 2025) puts AI at 5.3% of job postings, up from 3.3% the year before — around 84,000 postings. About 82% of them are for people who use AI in their work; only 18% are for people who build it. If you have been assuming an AI hire means a technical specialist, the market disagrees.
The wage premium is highest where you'd least expect. The same study puts the AI wage premium at 107% in government and the public sector, and 96% in consumer markets. If you are losing people and can't see where to, those are the two places to look.
You are now competing with both frontier labs for the same profile. Anthropic's October opening and OpenAI's expansion are recruiting in Singapore for programme-level AI people — the same person a mid-sized company needs. Expect salary expectations to move before headcount does.