1.4nm Won't Be Made
in Southeast Asia.
Some of it will be finished here.

TSMC and Intel are converging on 2028. For ASEAN, the node matters less than which company wins the customers — because the customer list decides where the chips go next.

On 22 September, TrendForce put two roadmaps on one chart. Intel's 14A was on the left, TSMC's A14 on the right, and a yellow box in the middle read 1H27.

The chart travelled fast. In Japan, semiconductor analyst Takeshi Hattori summed it up in one line: Intel and TSMC begin 1.4nm risk production in the first quarter of 2027, move to mass production in 2028, and pull away from Samsung and Rapidus.

The direction is right. The detail is softer than the headline.

What was actually reported

TSMC. Commercial Times, citing IC design sources, says A14 pilot production at Fab 20 in Baoshan, Hsinchu, could start as early as 1Q27. Fab 25 in Taichung, the main A14 site, is due for completion in 2Q27, with pilot runs there possibly from 3Q27. Across the two sites, A14 capacity is expected to reach 60,000–70,000 wafers a month. TSMC's own guidance is risk production in 2027 and volume in 2028.

Intel. CEO Lip-Bu Tan reportedly expects 14A risk production in 1Q27. Intel's official timeline is more careful: 2H27 for internal products and a high-volume ramp in 2028. Defect density was reportedly 0.5 by June 2026, against a target of 0.1–0.2 by 1Q27.

Samsung. SF1.4 mass production is now set for 2029, two years later than originally planned. Rapidus isn't in this race. Its target is 2nm mass production in 2027.

Six out of sixty-four

Now the Southeast Asia part. At SEMICON Southeast Asia this year, SEMI put a number on the region's position. Across Asia, 64 new fabs are expected to come online by 2029. Six of them are in Southeast Asia.

None was ever going to be. That sounds like bad news. We don't think it is — or at least not for the reason people assume.

The chip isn't finished when it leaves the fab

An A14 or 14A wafer holds very small, very expensive transistors. It is not yet an AI accelerator.

To become one, the logic die is placed next to stacks of high-bandwidth memory and wired to them through an interposer or a silicon bridge. Then it is mounted, tested, burned in, and tested again. At these nodes, a growing share of a chip's real-world performance depends on how well those pieces talk to each other.

That is the work Southeast Asia already does. Malaysia alone handles close to 13% of the world's chip assembly, packaging and testing.

So the useful question for ASEAN isn't "will we make 1.4nm chips?" It's this: where does a 1.4nm chip go after it leaves the fab?

Same node. Different map.

This is where the TSMC–Intel race stops being a technology story and becomes a geography story.

TSMC's answer is CoWoS. Capacity is reported to go from about 130,000 wafers a month at end-2026 to 260,000 by end-2028. The ramp is centred on AP7 in Taiwan and the Arizona campus. Arizona is there so advanced chips don't have to fly back to Taiwan for packaging. Southeast Asia is not on that map.

Intel's answer is EMIB and Foveros, and one of the places those flows run is Penang. The complex is part of Project Pelican, a RM12bn (about US$2.9bn) facility that was 99% built when Intel revived it. Its first phase, advanced-packaging assembly and test, is due to start later this year. It is designed for die sort, prep, and both EMIB and Foveros production.

Analysts cited by TrendForce estimate Intel's EMIB-T capacity at 15,000–20,000 CoWoS-equivalent wafers a month in 2027, rising to 40,000–45,000 in 2028. That is the same year 14A ramps.

Put the two together and the logic is simple. An AI customer that TSMC wins at A14 will most likely be packaged in Taiwan or Arizona. An AI customer that Intel wins at 14A has a route through Penang.

TrendForce's chart lists Intel's potential 14A customers: Amazon, Apple, AMD, Google, NVIDIA. For Malaysia, whether one or two of those names actually sign is worth more than any fab incentive the government could offer.

Every AI customer Intel signs for 14A is a small vote for Penang. Every customer TSMC signs for A14 keeps the packaging in Taiwan and Arizona.

Country by country

Malaysia: the swing vote. Its upside is tied to Intel Foundry winning outside customers. If that happens, Penang goes from assembling Intel's own CPUs to packaging other companies' AI silicon. That is a different business, with different margins and different engineers.

Singapore: wins either way. Micron's roughly US$7bn HBM packaging plant, the first of its kind in Singapore, starts operations in 2026. Micron says it will contribute meaningfully to HBM supply in 2027, the same year both 1.4nm nodes start their first runs. Every 1.4nm AI accelerator sits beside HBM, whoever made the logic. Micron is also building a US$24bn NAND fab on the same site. Singapore's bet is on memory, and memory goes into whichever accelerator wins.

Vietnam: volume now, frontier later. Amkor's US$1.6bn plant in Bac Ninh, which Amkor calls its most advanced facility, started with system-in-package and memory. Intel's Ho Chi Minh City site, about US$1.5bn invested, is the largest assembly-and-test operation in Intel's network. Samsung plans a US$1.5bn testing plant for November 2027, and Bloomberg has reported a separate US$4bn packaging plan. By July 2026, Bac Ninh province alone had drawn 33 semiconductor-related projects worth US$2.74bn.

But look at what those lines actually do: SiP, memory, test. None of it is CoWoS-class 2.5D integration. The 1.4nm cycle will send Vietnam more volume. On its own, it won't send Vietnam the frontier. Viettel's first domestic fab at Hoa Lac, targeting trial production at the end of 2027, is a separate and much longer bet.

Thailand: adjacent, not central. Thailand's role is backend work and power chips. Infineon sold its Bangkok backend site to Malaysian Pacific Industries and is building a new backend fab in Samut Prakan. Its link to AI runs through power semiconductors and data-centre electrical gear more than through the accelerator itself.

Indonesia: data centres, not chips. Firmus is planning a 360MW NVIDIA-powered AI data centre in Batam — an overflow valve for Singapore's compute, not a packaging hub.

The chips get smaller. Johor gets bigger.

The least obvious channel is electricity.

Both companies are selling 1.4nm on performance per watt. TSMC says A14 uses 25–30% less power than N2 at the same speed. Customers won't bank that saving. They will spend it on more compute per rack, and then on more racks.

In Johor, JLL counts 850MW of data-centre capacity completed, 1,800MW under construction and 2,700MW in the pipeline. Knight Frank puts co-location vacancy at 0.7%, against 4.9% in Singapore and 23.3% in Bangkok. Wood Mackenzie estimates data-centre maximum demand in the state at about 3.8GW, nearly one and a half times Johor's current electricity demand. It expects data centres to reach around 40% of the state's power use by 2035, up from about 24% today.

Wood Mackenzie's conclusion is that the bottleneck is no longer generation. It is transmission and distribution.

A 1.4nm transistor is measured in angstroms. What it ends up needing in Johor is measured in substations.

The case against

We see three reasons to be careful.

Intel has a track record here. As recently as early 2025, the Penang project was on indefinite hold and its engineers had been moved to New Mexico. The 1Q27 date for 14A is a CEO's expectation. The official date is 2H27.

Penang is not Intel's only packaging option. EMIB is already running at Amkor's Songdo K5 plant in Korea. A 14A customer win does not automatically become a Penang win.

Johor's constraints are real. In November 2025, authorities asked investors to delay water-cooled expansions for about 18 months because of drought. A pipeline is not the same as capacity.

What we're watching

Everyone is watching who reaches 1.4nm first. For Southeast Asia, that is the wrong scoreboard.

We are watching who signs the customers. The customer list decides the shipping route, and the shipping route decides whether Penang gets the work or just the headline.

Related analysis