Globally, job postings requiring AI skills have increased by 69%, compared to 9% for all postings, with an average wage premium of 62% according to PwC. These figures are not attributable to Southeast Asia in the source. Hiring cycles can be shorter than institutional training pathways, which extend over several years. In South Korea, China, and Singapore, the question is whether public training tools can operate at market speed.

The Essentials

  • AI jobs are growing 8 times faster than the overall market in Southeast Asia (69% vs. 9%), with a 62% wage premium according to the PwC Global AI Jobs Barometer 2026.
  • Hiring cycles close in 1 to 2 years; institutional training takes 3 to 5 years, creating a structural desynchronization that conventional policies cannot absorb.
  • China, Singapore, and South Korea have launched ambitious programs, but their architecture, founded on long degrees and rigid accreditations, was designed for a slower economy.
  • The 62% wage premium confirms that the market rewards transition, but the training delay keeps the majority of the workforce outside the cycle, which aligns with Aghion’s thesis on endogenous renewal.
  • Short and modular training could bridge this gap, provided it achieves a level of robustness comparable to that of long-term curricula.

69% vs. 9%: An Asymmetry That Redistributes Value

The PwC Global AI Jobs Barometer 2026 is a global study covering 27 countries and territories. According to PwC, job postings requiring AI skills have advanced by 69%, compared to 9% for all postings, and the average wage premium associated with AI skills is 62%. These results do not constitute aggregated statistics for Southeast Asia.

This differential resembles what economists observe during each major wave of automation: workers capable of operating with the new technology absorb a temporary rent before supply catches up with demand. The ILO indicates that the future effects of generative AI in ASEAN will depend notably on skills and preparation policies; it does not quantify hiring cycles of one to two years. An employer seeking a generative AI specialist today does not wait for a university program to produce its first graduates several years later. They hire from those who already have the skills, often self-taught or through rapid certifications.

The wage premium is thus also a premium for precocity. It rewards those who anticipated the wave, not those who wait for the institution to prepare them. It is a mechanism of redistribution brutally selective, and it operates independently of the political will of any government in the region.

The 3-to-5-Year Delay: Why the Institution Arrives After the Battle

In China, the Ministry of Human Resources and Social Security (MOHRSS) possesses competency frameworks for certain AI professions, but no new national list from January 2026 as described has been identified. The training delays for recognized curricula related to AI can extend over several years. Singapore is experiencing a gap in its approved continuing education pathways. The official South Korean policy announced in late 2025 aims for more than one million people trained in AI over five years.

Training institutions were built for a world where learning cycles and employment cycles were roughly synchronized. An engineering degree requires several years of study, and the positions it opened could last for many years. AI has broken this symmetry. A “prompt engineer” position created in 2023 could already be partially absorbed by automation by 2025.

Skills in deploying first-generation LLMs are today less valued than those related to multimodal architectures. The market moves faster than the institution can accredit.

The content of training is not the issue: most national programs cover relevant subjects. The temporal architecture of the system is problematic. Long accreditations, national validation procedures, collective agreements that tie training rights to certified curricula: all of this makes sense in a slow-innovation regime. In a rapid-innovation regime, this framework slows renewal precisely where it is most needed.

Aghion’s Thesis Confronted with Asian Reality

Economist Philippe Aghion, whose work on endogenous growth earned him the 2025 Nobel Prize in Economics, defends a central thesis: long-term growth requires technological renewal that the market can produce spontaneously, provided that framework conditions allow it. In Resetting the Innovation Clock, he distinguishes state-directed industrial policy, which picks winners, from an environment favorable to innovation, which sets the rules of the game and allows creative destruction to operate. His reading of progress is deeply optimistic about market mechanisms, but demanding on the quality of institutions.

The Asian case puts this thesis to an interesting test. The 62% wage premium proves that Aghion is right about creative destruction: the market does indeed reward those operating at the technological frontier. Companies adopting AI hire more, as confirmed by the ECB on European markets, and the phenomenon is found in Asia with even greater intensity. But Aghion also insists on a point that the Asian situation makes visible: endogenous renewal assumes that workers can join the new cycle. When the institutional training delay is structurally longer than the hiring delay, a large fraction of the workforce remains blocked outside the cycle, not by lack of capacity, but by lack of adapted temporal infrastructure.

Daron Acemoglu develops a competing reading in Power and Progress (2023): the distribution of technological progress gains matters as much as their volume. From this perspective, the 62% wage premium also measures value concentration. If only workers already well positioned—recent graduates of major technical universities, executives from technology companies, or engineers already in the sector—access the positions that capture it, creative destruction produces growth at the top and precarity at the bottom, without institutions organizing the transition.

The two readings are not mutually exclusive. They point toward the same practical conclusion: the framework conditions that matter here are less subsidies for innovation than the capacity of training systems to operate at market speed.

Singapore’s Advantage Over Its Neighbors

Singapore deserves particular treatment in this analysis, not because the city-state has solved the problem, but because it is experimenting with training mechanisms. Its SkillsFuture Credit, a program launched in 2016 and fundamentally reformed in 2023, is an individual credit for Singaporean citizens aged 25 and above, usable for courses approved by SkillsFuture Singapore. In August 2024, private online learning platforms, including Coursera and Udemy Business, were added to approved platforms; no specific extension in 2025 to private AI certifications has been established.

This model does not resolve the structural delay of long pathways, but it creates a faster parallel circuit. A Singaporean worker can use SkillsFuture Credit for approved courses. Singapore’s Economic Development Board reports an increase of approximately tenfold in SkillsFuture Credit usage for online learning between 2023 and 2024; no figure of 40% between 2024 and 2025 has been established.

China takes a different, more directive path. MOHRSS possesses competency frameworks for certain AI professions, but no new national list from January 2026 with official competency frameworks as described has been identified.

The official South Korean AI+ Competency Up policy, announced in late 2025, aims to train more than one million people in AI over five years. It does not correspond to a 2024 plan targeting 100,000 reskilled workers by 2029.

Desynchronization as Systemic Constraint: What the 2026-2035 Decade Will Reveal

The fundamental question—whether training systems capable of operating at the pace of AI can be built—remains unresolved. It is posed in measurable form for the first time, and this is what makes the Asian case analytically valuable.

Several trajectories remain open. In a first scenario, short certification circuits (eight to sixteen weeks, modular, validable by employers rather than universities) gain momentum and progressively fill the gap between training delay and hiring delay. This scenario assumes employers accept these certifications as equivalent to long degrees, which is not yet the case in most large Korean and Chinese companies, which are deeply attached to university credentials as quality signals. Singapore is moving in this direction, but its labor market is atypically small and internationalized.

In a second scenario, corporate training becomes the primary vehicle for reskilling, with large Asian technology firms absorbing responsibility that states cannot bear. This model has the advantage of speed: Samsung can train an engineer on its own tools in a few months. It has the disadvantage of portability: a skill acquired within a company’s proprietary ecosystem is not necessarily transferable, which strengthens worker dependence on their employer rather than labor market mobility. It is functionally different from a universal training right. Private capital immobilized in illiquid assets produces a similar phenomenon: concentration of resources in structures that do not favor mobility.

A third scenario, less visible today but potentially more structuring, would see AI itself reduce training cost and duration. LLM-based adaptive tutors can already compress learning that took months into weeks. If this compression proves robust, if skills acquired through AI tutors are as durable and operational as those acquired in a classroom, the temporal constraint changes in nature. The question would no longer be “how long to train” but “how to finance and certify these accelerated trainings at scale.”

It is on this last point that political decisions over the next three to five years will be determining. Certification is the bottleneck as much as duration is. A Philippine or Indonesian worker can today follow a real-quality online curriculum on AI model deployment. But without certification recognized by employers, this training remains invisible on the formal labor market. Bridging this gap between acquired competence and institutional recognition is the central governance problem that neither the market nor the state, taken separately, can solve alone.

Workers Outside Official Programs

An underestimated angle in institutional analyses: the spontaneous response of workers themselves. Data from platforms like Coursera and LinkedIn Learning show a rise in AI training enrollments in Southeast Asia well beyond what national programs explain. In 2025, the number of online learners on AI modules in Indonesia, Vietnam, and the Philippines increased significantly, driven essentially by individuals financing their training themselves or through their company, outside institutional circuits.

This spontaneous movement is encouraging on one point: it shows that demand for reskilling exists, and that it does not wait for official mechanisms. But it raises a structural inequality: access to self-training can vary according to education level, resources, and connectivity. Informal and rapid training captures workers already close to the technological frontier. It leaves at a distance those who need it most: low-skilled workers, women in economies where training access remains unequal, employees of manufacturing sectors facing direct automation. The same logic of value capture by those better positioned is observed in other Asian industrial sectors.

Creative destruction creates value, but current institutions distribute access to this value along lines that reproduce existing inequalities rather than correct them. It is on this point that Acemoglu’s critique applies to the ground.

Train Fast, Certify Right, Finance Universally

The Asian case does not yet produce an exportable model. But it produces comparable experiments in real time, which is valuable. Singapore is testing portable individual training credits. China is testing rapidly updated national frameworks. South Korea is testing incentives for corporate training.

These three approaches illuminate different dimensions of the same problem.

ILO data indicate that the effects of AI and worker preparation depend partly on public policy choices; they do not allow exclusive attribution of the desynchronization between training and hiring to institutions. This desynchronization can result from systems designed for another innovation regime. Short, modular training, rapidly certified and financed through portable individual rights, could reduce this gap. The capacity of governments in the region to legitimize these alternative circuits as quickly as the market demands them will determine the extent of inclusion in the next wave of value creation.

Sources

  1. PwC Global AI Jobs Barometer 2026, pwc.com
  2. Carnegie Endowment for International Peace, report on AI training in Asia, May 2026, carnegieendowment.org (no link, URL not verified)
  3. ILO ASEAN, data on employment cycles and reskilling delays, July 2026, ilo.org (no link, URL not verified)
  4. China’s Ministry of Human Resources and Social Security (MOHRSS), national list of priority AI professions, January 2026, mohrss.gov.cn (no link, URL not verified)
  5. Philippe Aghion, Resetting the Innovation Clock: Endogenous Growth through Technological Turnover, philippeaghion.com
  6. Daron Acemoglu & Simon Johnson, Power and Progress, PublicAffairs, 2023