A year on from their positive note on Emerging Markets (EM) technology, “DeepSeek impact – an evolution, not a revolution”, the GAM Emerging Market Equity team continues to believe that further upside lies ahead. However, selectivity is already important, and becoming increasingly so.
16 February 2026 | Ygal Sebban
While discussions around artificial intelligence (AI) often centre on the US, some of the most critical technological advances enabling the AI boom are increasingly emerging from Asia. The region has become integral to the global AI ecosystem, supported by world-class capabilities in semiconductors, integrated circuit design and a rapidly expanding software landscape.
In our view, EMs offer a powerful way to access this opportunity set. Asia is home to globally competitive leaders across the AI value chain: foundry (TSMC), memory (SK Hynix and Samsung Electronics), integrated circuit (IC) design (MediaTek) and the broader hardware ecosystem, including packaging, testing, IDM and ODM1. Meanwhile, EM home-grown software and application platforms, particularly in China, continue to scale rapidly and present compelling growth prospects, in our view.
The influence of EM tech is increasingly reflected in market representation. The EM tech sector now accounts for 29% of the MSCI EM Index, up from 21% in mid-2023 and 14% in 2018, even surpassing the tech weighting of the MSCI World (26.5%).2 The key upstream hardware leaders – TSMC, SK Hynix and Samsung Electronics together represent nearly 20% of the EM index.
What is shaping the AI and EM tech era?
- The rapid evolution of generative AI (GenAI) capabilities means the degree of AI’s integration into daily lives is still unknown.
- Although adopting is at an early stage, “The Economist” estimates that 16% of the global workforce uses GenAI tools monthly3, but it is early in understanding its full impact.
- With visibility on end applications still limited, we believe sustained growth in AI-related capital expenditure (capex) and model development are likely to support tech performance through 2026.
- EM Asia is home to many of the world’s leading tech companies with critical roles across both the AI hardware and software ecosystems.
- China’s top-down AI strategy, driven by low-cost, open-source models, positions the country well for broad economic deployment and long-term productivity gains, though selectivity remains key when identifying private-sector companies best aligned with profitability alongside state goals.
Last year’s concerns around “peak capex” were eased by consistent upgrades. The top four cloud service providers (CSPs) are now expected to spend USD 332 billion in 2025 (up 52% year-on-year), with forecasts for 20%+ growth again in 2026.4 Longer-term projections point to AI investment exceeding USD 1 trillion by 20305, with UBS forecasting USD 1.3 trillion, equivalent to 1% of global GDP (IMF basis) 6. Despite the rapid increase, CSP capex still sits comfortably below operating cashflow, supporting stable and positive free cashflow – a reassuring signal amid “bubble” concerns.
However, while capital-intensity remains in acceptable territory overall, the market is becoming increasingly selective in its exposure. Companies where rising investment implies capex moving ahead of operating cashflow have share price punished - the best example being Oracle (ORCL US), whose shares fell 11% over H2 2025 (versus Nasdaq 100 +11%) after updating spending plans projected capex exceeding operating cashflow on analyst forecasts7, despite a strong AI narrative.
Chart 1: AI Capex - Top 4 CSPs capex relative to operating and free cashflow
Source: GAM, JP Morgan, as at November 2025. The views expressed herein are those of the manager and are subject to change. USD bn = USD billion, EBITDA = Earnings Before Interest, Taxes, Depreciation and Amortisation; OCF = Operating Cash Flow; FCF = Free Cash Flow, LHS = left hand side, RHS = right hand side.
Fabrication at the heart of the AI story – TSMC
At its recent Q4 results briefing, Taiwan Semiconductor Manufacturing Company (TSMC) raised five-year AI growth guidance from 40% to 50%, and 2026 capex guidance to USD 52-56 billion (versus market expectations of around USD 48 billion).8 Chairman and CEO Dr Che-Chia Wei emphasised that discussions extended beyond TSMC's direct customers, revealing tangible business benefits and financial returns from AI adoption. TSMC is aligning its long-term capacity strategy accordingly.
From Generic chips to application-specific silicon – MediaTek
AI is accelerating the shift from generic semiconductors (chips) to application-specific integrated circuits (ASIC). Google’s Gemini 3 models, powered solely by in-house TPUs (tensor processing units) rather than Nvidia GPUs (graphic processing units), illustrate this trend. Asia is home to leading ASIC developers, including Taiwanese company MediaTek, where ASIC-related revenue is expected to overtake legacy handset revenue by 2027.9 With 2026 likely a transition year, we see valuations as supported by current levels ahead of expected ASIC-driven growth from 2027 onwards.10
Architecture over transistors – ASE Technology
For decades, semiconductor progress was driven by one main principle: packing more transistors into a chip meant better performance. As chips moved to smaller nodes, measured by nanometres, and became more densely packed, processing power increased.
At CES (Consumer Electronics Show) 2026 in Las Vegas, Nvidia CEO Jensen Huang made clear that the AI bottleneck is shifting from compute to contextual memory and storage, “Moore’s Law11 has largely slowed. The number of transistors we can get year after year can’t possibly keep up with the 10-times-larger models,” he noted, emphasising the strain imposed by long-context inference workloads. Traditional network storage is inefficient for these tasks, prompting Nvidia to redesign its architecture around new “inference context memory” systems.12 As a result, system-level architecture now matters more than transistor density. Performance increasingly depends on advanced memory integration and heterogeneous packaging.
This evolution elevates the importance of companies specialising in advanced packaging and testing. ASE Technology, a Taiwanese global leader in outsourced assembly and test (OSAT) and electronic manufacturing services (EMS), looks well-positioned to benefit from growing industry demand for high-performance heterogeneous integration, alongside tight advanced-packaging capacity at TSMC.
Memory “super cycle” narrative strengthening –SK Hynix and Samsung Electronics
We have been positive on memory since mid-2023 (see Emerging markets: on the path to the industrial revolution 5.0), but the outlook has upgraded significantly since late 2025. Market conviction in a memory “super cycle” has strengthened, particularly for leaders such as SK Hynix13 and Samsung Electronics, as their strategic importance in the AI supply chain continues to expand. Two key developments are reinforcing this narrative.
- Broad‑based memory demand
Throughout the second half of 2025, it became increasingly evident that proliferation of inference-based AI models is driving a broad uplift in memory demand. This includes not only leading-edge high bandwidth memory (HBM) but also legacy DRAM (Dynamic Random Access Memory) nodes, which remain essential for cost-effective AI deployment at scale. Tight supply across both advanced and legacy nodes has pushed spot prices sharply higher. Several analysts now forecast DRAM shortages extending well into 2027, with the potential to run into early 2028.14
Chart 2: Spot price and contract price of DRAM (DDR4 8GB)
Source: GAM, Bloomberg, inSpectrum, as at 8 January 2026.
- Memory as the new bottleneck
The shift in AI workloads towards longer-context inferences has reinforced the fact that memory, not compute, is increasingly becoming the limiting factor in system performance. As highlighted at CES 2026, Jensen Huang has stressed that architectural constraints, rather than pure processing power, are now shaping the pace of AI advancement.15 This transition elevates not only HBM, but also traditionally lower-margin NAND16, as memory moves from a passive cost input to a core performance driver in AI systems.
Despite surging demand, supply has been slow to adjust. Constraints stem from supply discipline within the DRAM oligopoly (three dominant producers), persistent demand for legacy nodes (which historically would have been migrated to newer ones), and most importantly, a shortage of clean-room capacity across semiconductor equipment suppliers – meaning insufficient specialised facilities and equipment to meet the backlog of orders. As a result, industry backlogs are unlikely to be cleared before late 2027 or even 202817.
Recent industry research indicates that leading Asian memory stocks, including SK Hynix and Samsung Electronics, are being supported by a “stronger-for-longer” industry thesis and upgraded street forecasts for earnings per share and book value18. Should confidence in the memory “super cycle” continue to build, we believe a valuation re-rating could unlock further upside beyond historical trading ranges. For second-tier memory companies, we note that valuations have risen sharply due to retail-driven interest, resulting in less favourable risk-reward characteristics compared with leading names.
Application AI and physical AI – Hesai
A year ago, we highlighted how investment focus in tech had shifted from the very upstream segments towards mid-stream hardware. Early-2026 performance now suggests this momentum is moving even further down the stack, towards software-driven application AI and the rapidly emerging domain of physical-world AI.
At CES, a number of technology majors unveiled new suites of robotics. While some robotics-focused manufacturing companies are trading at high valuations19, attractive opportunities are emerging beyond the obvious thematic names, particularly in the enabling tech that forms the backbone of physical intelligence systems.
One such example is Hesai, a Chinese tech company listed in both the US and, more recently, Hong Kong. As a global leader in LiDAR sensors (light detection and ranging), critical for autonomous driving, robotics and a broad spectrum of physical-AI applications, Hesai represents a classic “pick-and-shovel”20 play, providing essential infrastructure to a rapidly expanding ecosystem.
The risks
There are plenty of concerns about the returns profile and potential bubble characteristics of GenAI. An MIT report from July 2025 estimated that around 95% of GenAI projects are failing21, while valuations across parts of the tech ecosystem continue to look demanding. At the same time, memory shortages are putting pressure on industrial and consumer-electronic producers – especially PC and handset manufacturers – through constrained supply and elevated component pricing negatively effecting both volume and margin.
Stay in the theme, but stay selective
Despite the noise, we believe it is far too early to fade the GenAI theme, while selectivity is essential. In our view, the winners will be businesses with clear leadership or defensive niche positions within the AI value chain, backed by disciplined capital allocation, reasonable valuations and attractive long-term earnings trajectories. We believe established industry leaders within the AI value chain are well placed to participate as adoption accelerates and the technology’s economic impact becomes increasingly evident.