Singapore Just Landed on the US Chip Licence List and I Rewrote 4 Lines of Our 2027 Engineering Plan the Same Afternoon

Advanced AI processor on a circuit board, representing the new United States export licence requirement that now covers shipments of advanced chips to Singapore
Sebastian

Sebastian

Mobile App & Hiring Expert · November 17, 2026 · 9 min read

TL;DR

  • •What changed: a new US rule requires licences to ship advanced chips to Singapore, Malaysia, Thailand, Indonesia and Brazil, and requires a licence for data centres anywhere to serve large volumes of advanced compute to companies headquartered in China, Macau or other embargoed countries.
  • •What Singapore said: on 17 November its digital development and trade ministries said they are studying the measure carefully, engaging affected companies, and exploring how to facilitate trusted access for Singapore companies.
  • •Why it reaches hiring: a licence is a queue, and a queue turns compute from something you buy into something you schedule. That rewards engineers who make existing hardware go further.
  • •The move I would not make: freezing AI requisitions. The constraint is on silicon, not on the people who decide how well you use it.

I had a call booked on Tuesday morning to sign off a 2027 headcount plan for a Singapore product team. We spent the first twenty minutes of it reading a government statement instead, and by the end I had changed four lines of the plan. None of them were the lines anyone expected.

What actually happened on 16 and 17 November

On Tuesday Singapore said it was seeking “trusted access” to advanced AI chips after the United States imposed new curbs aimed at stopping Chinese firms from reaching the technology through overseas data centres. The statement came jointly from the ministries overseeing digital development and energy, trade and industry.

“We are studying the details of this new measure carefully. We are engaging affected companies to understand the potential impact on their businesses.”

The same statement said the city state is talking to the relevant US authorities on the scope and implementation of the measure, as well as exploring how to facilitate trusted access for Singapore companies. That wording is worth reading twice. It is not the language of a government that expects the door to close. It is the language of one that does not yet know how long the queue is.

The rule itself has two moving parts. Under the first, data centres anywhere in the world now need a US licence to provide large volumes of computing power from advanced chips to companies headquartered in China, Macau or other countries under US arms embargoes. That closes a gap that had let Chinese firms rent AI computing power from data centres in Southeast Asia and elsewhere running on hardware they could not legally buy. Under the second, the rule requires licences to ship advanced chips to Singapore, Malaysia, Thailand, Indonesia and Brazil, countries the US commerce department says present a “higher-than-normal risk of diversion”.

A day earlier, speaking at Semafor’s Silicon Valley & The World, Singapore’s minister for digital development and information Josephine Teo said her government is “engaging with our US counterparts as to the scope of implementation”, and added: “I think more importantly, how we can facilitate trusted access to this very important AI infrastructure.” She noted that she had spoken with US Deputy Secretary of State Christopher Landau.

The backdrop is uncomfortable and local. Singapore-based cloud computing firm Megaspeed is suspected of helping Chinese companies get around US export controls, a story that broke on 8 November and that ended with Jensen Huang leaving as Nvidia chief. Indonesia, also named in the rule, said on Tuesday it was reviewing the measure and seeking industry input on its potential impact on operations and investment, including in the data centre sector.

Our expert take

Every employer I spoke to on Tuesday read this as a supply story. It is really a scheduling story. A licence requirement does not usually mean you cannot have the hardware, it means you cannot say when. For a hiring plan, uncertainty about timing is harder to absorb than a known shortage, because you can staff around a shortage you can date. This is why the first thing I changed was not a role, it was the assumption underneath the roles.

Why a chip licence rule lands on your hiring plan, not just your cloud bill

Most 2027 plans I have reviewed in Singapore this quarter share one buried assumption: that accelerated compute is elastic. Need more, buy more, next quarter. Teams size their model work, their data pipelines and sometimes their entire product roadmap on that assumption, and they hire accordingly, mostly for people who build features on top of capacity someone else procures.

A licence step breaks the elasticity without breaking the supply. You still get the hardware. You get it on a timeline set by a process you do not control, which means capacity planning stops being a procurement task and becomes an engineering constraint. The teams that cope are the ones holding people who can answer a different question: not “what can we build if we double the cluster”, but “what can we ship on what we already have”.

Two places the new rule inserts a licence1. Shipping the siliconAdvanced chips sent to Singapore,Malaysia, Thailand, Indonesia, Brazil2. Serving the computeAny data centre, large volumes, to firmsheadquartered in China, Macau, embargoedCapacity becomes a scheduled resourcePlan engineering around what you hold, not what you can order

There is a second, quieter effect. The rule makes the physical location of a workload and the headquarters of its end customer into questions with a regulatory answer. Plenty of Singapore engineering teams serve regional customers from shared infrastructure and have never had to describe, precisely, whose traffic runs where. That description now has to exist, and somebody competent has to own it.

Our expert take

The candidates who will be most in demand six months from now are not the ones with the longest list of model names on their CV. They are the ones who can stand in front of a finance director and explain, without hedging, what a workload costs to serve today and what it would cost on different hardware. That skill used to be a nice-to-have in Singapore because capacity was cheap to add. It is about to become the reason a product ships on time.

The three engineering profiles this rule quietly repriced

Inference efficiency engineers. People who can profile a serving path, quantise or distil a model, batch sensibly and cut cost per request without retraining anything. When compute arrives on a schedule rather than on demand, this is the role that buys you months. In interviews I now ask for a specific instance where the candidate reduced serving cost, with the before and after numbers and what they traded away to get them.

Infrastructure and capacity engineers. Not platform generalists, but people who have planned across more than one region and more than one provider, and who have had to make a workload portable under pressure. The useful signal is whether they have ever actually moved something, not whether they can describe how they would.

Platform engineers who can answer a compliance question. Where does this run, who is the end customer, how would we prove it. For a long time this sat with legal and procurement in Singapore firms. The rule pushes it into the platform team, and the people who can hold both halves are genuinely scarce. The MAS AI risk guidelines pushed financial institutions in the same direction a month ago, so if you are in that sector you are now being pulled twice.

Need these profiles before your capacity question is answered?

Tell us the workload and the constraint. We will come back with pre-vetted engineers who have shipped under exactly this kind of ceiling, available to interview this month.

Let’s discuss it

What I would not do this week

I would not freeze AI engineering requisitions. This is the reflex, and it is the one that costs the most. The constraint announced this week is on silicon moving across a border. It is not a constraint on the people who determine how much you get out of the silicon already sitting in your racks. Teams that paused hiring during the last infrastructure scare in this market spent the following two quarters paying more for the same candidates, and lost the internal argument about roadmap scope in the meantime.

I would also not redesign an architecture around a rule nobody has finished reading. Singapore’s own ministries said on Tuesday that they are still studying the details and engaging affected companies. If the government is still establishing scope, a startup rewriting its serving stack on Thursday is guessing. Build the option, do not spend it.

And I would not treat the earlier server diversion charges and this rule as the same story. They are connected in the news cycle, but they land on employers differently: one is about who you hire and how you screen them, this one is about what your team can promise to deliver and when.

Nine days, November 20268 NovMegaspeed reportbreaks16 NovMinister: engaging on scope,facilitate trusted access17 NovJoint ministry statement:studying the measureScope and implementation still under discussion with US authorities

Our expert take

The honest answer to “how bad is this for our roadmap” on 17 November is that nobody knows yet, including the ministries. What you can control is whether the answer finds you with a team that can adapt or a team that was sized for abundance. I would rather over-hire one efficiency engineer this quarter than discover in March that my only lever is buying hardware I am queuing for.

The four lines I changed in that 2027 plan

  1. I replaced “scale the cluster in Q2” with a capacity range. The plan now carries a low case and a high case, and the headcount attached to each is explicitly different. This took an afternoon and removed the single largest unexamined assumption in the document.
  2. I moved one inference efficiency hire from Q3 to now. Same budget, earlier start. If capacity arrives on time, that person pays for themselves on the cloud bill. If it does not, they are the reason the roadmap survives.
  3. I added workload placement to the platform role’s scope, in writing. Not a new headcount, a clarified one. The job description now says the holder owns the answer to where workloads run and for whom, which also changes who we shortlist.
  4. I removed the assumption that every AI engineer must sit in Singapore. If the binding constraint for the next two quarters is hardware rather than desks, geography is the wrong thing to be rigid about. Our Singapore hiring view still anchors the senior roles locally, but execution capacity no longer has to be.

None of that is a reaction to a crisis, because this is not yet one. It is the cheap version of a response, taken while there is still time to take it cheaply. By the time the scope of this measure is actually settled, the engineers who can work inside it will be fielding three offers each, and the Singapore market has never been forgiving to the employer who starts late. If you want a sharper read on the compute side specifically, our note on hiring AI infrastructure engineers in Singapore covers the screening detail this article skips.

Frequently Asked Questions

What did the new US chip rule change for Singapore?
Two things. First, data centres anywhere in the world now need a United States licence to provide large volumes of computing power from advanced chips to companies headquartered in China, Macau or other countries under US arms embargoes. Second, the rule requires licences to ship advanced chips to Singapore, Malaysia, Thailand, Indonesia and Brazil, which the US commerce department describes as presenting a higher-than-normal risk of diversion. Singapore said on 17 November 2026 that it is studying the details of the measure and is engaging affected companies to understand the potential impact on their businesses.
Has Singapore said it will lose access to AI chips?
No. The joint statement from the ministries overseeing digital development and energy, trade and industry said Singapore is talking to the relevant US authorities on the scope and implementation of the measure, as well as exploring how to facilitate trusted access for Singapore companies. Digital development minister Josephine Teo said on 16 November that the government is engaging with US counterparts on the scope of implementation, and that the more important question is how to facilitate trusted access to AI infrastructure. A licence requirement is a process, not a ban, but the timeline for that process is the part no employer can yet plan around.
Should Singapore employers pause AI engineering hiring because of the chip rule?
Pausing is the expensive reaction. A licence requirement changes how quickly you can add accelerated compute, not whether you need engineers who can use the compute you already have. The teams that handle a capacity squeeze best are the ones with people who can profile a workload, quantise a model, move inference to cheaper hardware and measure the result. Those skills get more valuable when compute is rationed, not less. Freezing a requisition in November and restarting it in March means paying a premium for the same person three months later.
What engineering skills matter more after the licence rule?
Three groups. Inference efficiency engineers who can cut the cost of serving a model without retraining it. Infrastructure and capacity engineers who can plan across regions and providers rather than assuming one cluster will grow on demand. And platform engineers who understand where workloads physically run and who the end customer is, because the new rule turns that question into a compliance question rather than an architecture preference. None of these are exotic profiles, but screening for them takes a different interview than a generic backend loop.

Plan your 2027 engineering team against the constraint, not the hope

We help Singapore employers staff for capacity they can actually count on, with pre-vetted engineers across inference, infrastructure and platform work.

Discuss your 2027 plan