Taiwan outlook H2 2026: Taiwan’s irreplaceable AI supply chain and why the re-rating is not over

• AI demand is not a bubble: Taiwan’s exports and export orders continue to grow strongly despite a high base, while leading AI labs are moving toward profitability, proving that compute demand is backed by real cash flow.
• TSMC still has room for re-rating: Advanced-node capacity remains tight, but pricing has been relatively restrained. As customers use prepayments and long-term agreements to secure supply, TSMC’s pricing power should gradually be reflected in earnings.
• Taiwan’s edge lies in its full ecosystem: From advanced nodes, packaging and testing to racks, cooling and chip design, Taiwan’s AI hardware supply chain is difficult to replicate, supporting further re-rating of Taiwan equities.

Six months ago, the market was still worried about an “AI bubble”. Looking back now, although AI-related share prices have seen occasional pullbacks, the bubble has not materialised. On the contrary, Taiwan’s exports have continued to grow on an already high base. More importantly, as AI infrastructure investment continues to expand, the beneficiaries are no longer limited to TSMC alone, but extend across Taiwan’s entire hardware ecosystem — from advanced process technology, packaging and testing, to racks, thermal solutions and chip design — all of which play critical roles in this round of AI supply-chain re-rating.

AI Demand Is Not a Bubble: Exports and AI Monetisation Are Both Confirming the Trend

Taiwan’s exports have not slowed despite a high base

At the beginning of the year, the market was concerned that Taiwan’s export growth would slow sharply due to a high-base effect. However, demand for AI servers, cloud computing, and high-performance computing (HPC) has remained strong. In the first five months of this year, Taiwan’s cumulative exports rose 48.7% year-on-year.

Export orders, a leading indicator, also confirm that AI-related demand remains robust. May export orders reached USD 89.48bn, up 47.2% year-on-year, marking the 16th consecutive month of growth and the second-highest May reading on record. This shows that related demand has not cooled meaningfully despite the high base.

Table 1: Taiwan Export-Related Indicators

Figure 1: Taiwan export order growth remains elevated

AI labs are starting to turn profitable, proving that compute demand is supported by cash flow

Frontier AI labs in the US are also gradually moving toward profitability. Anthropic is a good example. AI companies are no longer relying purely on “cash burn” to drive growth; instead, they are beginning to demonstrate strong monetisation capability. Its annualised revenue has surged from around USD 9bn to more than USD 44bn, while gross margin has exceeded 70%, with profitability expected as early as the second quarter of this year.

More importantly, Anthropic’s main bottleneck at this stage is not weak end demand, but insufficient compute supply. The company still needs to restrict usage for some users. At the same time, in order to secure more compute, Anthropic is paying SpaceX around USD 1.25bn per month to rent data-centre capacity. This shows that the real constraint on further AI revenue expansion is not demand, but the fact that compute supply has not kept up.

In short, AI demand is not only real; it is directly translating into orders for TSMC’s advanced process nodes.

Figure 2: NVIDIA GPU rental prices show that demand for AI hardware remains strong

Of course, even if AI demand is real, it does not automatically mean that Taiwan’s semiconductor supply chain can fully benefit. The key question is whether TSMC and Taiwan’s broader supply chain can convert this demand into their own earnings.

Demand Is Turning Into Earnings: TSMC Still Has Room to Raise Prices

In our article "NVIDIA | Beyond Chips: The Ecosystem Driving Its Next Growth Phase", we noted that hardware suppliers controlling the bottlenecks in AI compute have not chosen to raise prices aggressively, but have deliberately kept pricing relatively restrained.

TSMC has, in fact, adopted a similar strategy. N3 (3nm-class) is now one of the tightest links in the entire semiconductor supply chain. Key AI chips from NVIDIA, Broadcom, MediaTek and AMD all rely heavily on this process, and major chipmakers are competing for limited capacity. Market estimates suggest that AI-related demand will consume close to 60% of TSMC’s leading-edge N3 capacity in 2026, and could rise further to around 86% in 2027. Yet despite such tight supply, TSMC has not sharply increased its average wafer selling price. The average selling price of its N3 wafers rose from about USD 20,000 in 2023 to an estimated USD 23,300 in the first half of 2026, an increase of only around 16% over more than two years — a relatively steady trajectory.

Table 2: TSMC N3 Average Wafer Selling Price

At the same time, although TSMC’s profitability has improved significantly, the increase has not been excessive. The company’s gross margin rose from a trough of around 53% in 2023 to 66.2% in the first quarter of 2026, while operating margin also increased to 58.1%. This shows that AI demand has started to lift profitability, but compared with other core players in the AI infrastructure boom, TSMC has still not fully captured the value of the supply chain.

For example, NVIDIA, as the leading AI chip designer, saw its gross margin rise sharply from around 56% to 60% in 2023 to above 70% in 2026. Micron’s gross margin reached an even higher 84.9%. Although chip design and memory have different business models and margins cannot be compared directly, TSMC is the most critical manufacturing bottleneck for advanced process technology. In this context, its 66.2% gross margin and 58.1% operating margin still do not fully reflect its unique position in the AI supply chain.

Table 3: TSMC’s Gross Margin and Operating Margin Have Improved, But Only Moderately

TSMC has not chosen to maximise near-term profit through aggressive price increases during the AI boom. This is not because it lacks pricing power, but because it is pursuing a longer-term strategy. By keeping pricing relatively restrained, the company maintains long-term relationships with key customers, preserves ecosystem stability, and reduces the risk of customer pushback and regulatory pressure. NVIDIA CEO Jensen Huang has even said that TSMC’s wafers “should be more expensive”. As the most important company in the semiconductor industry, TSMC is seeking to gradually convert the scarcity of advanced process capacity into more stable and sustainable earnings through long-term supply agreements, capacity guarantees and prepayment arrangements.

In fact, TSMC is already gradually raising prices. According to reports, TSMC has begun informing customers of price increases for all advanced process nodes at 7nm and below, with increases ranging from 5% to 10%. Some reports also indicate that 3nm prices could rise by as much as 15% in the second half of 2026, followed by another 5% to 10% increase in 2027. At the same time, customers are showing a willingness to bear higher costs in order to secure capacity. For example, AMD announced in May 2026 that it would invest more than USD 10bn in Taiwan’s ecosystem, deepening its cooperation with TSMC and other partners in advanced packaging and AI infrastructure.

The Moat Is Not Just TSMC, But the Complete Ecosystem

The US experience illustrates the same point from the opposite direction. In the third quarter of 2025, TSMC’s Arizona fab was reportedly forced to halt production temporarily because of unstable gas purity supplied by a vendor, causing losses for the company. This clearly shows that chip production does not depend only on a single fab or the most advanced process technology; it relies on support from the entire ecosystem. If any key link in the supply chain is missing, even an advanced fab may face the risk of disruption.

According to SemiAnalysis, TSMC’s US fabs also faced several early operational challenges, including sulphuric acid costs at around five times the level in Taiwan, some specialty gases still needing to be shipped from Asia, and greater difficulty in managing the local cleanroom environment. More importantly, advanced CoWoS packaging remains largely in Taiwan. In other words, the true moat in semiconductor manufacturing is not just the fab itself, but the entire set of basic supply conditions surrounding it, including water, gases, chemicals, equipment maintenance and packaging capability. These links may seem inconspicuous, but they determine whether an advanced fab can achieve stable volume production.

Talent is another core link that is even harder to replicate. Close to 90% of TSMC’s employees and 88% of its management team are based in Taiwan. The precision manufacturing culture built over decades cannot simply be airlifted overseas. This is also why TSMC sends US engineers back to Taiwan for training. Most of TSMC’s tier-one material suppliers still have not set up facilities in Arizona, and the company continues to rely on Asian supply chains and engineering capability. Taiwan’s law also requires overseas fabs to remain at least one generation behind Taiwan’s most advanced process, while core process R&D remains in Taiwan.

Foreign investors are voting with their action

Even as the US government actively pushes for semiconductor manufacturing to return to the US, many American companies are instead increasing their investment in Taiwan.

Micron, one of the world’s largest memory manufacturers, is also one of Taiwan’s largest foreign investors. Between 2024 and 2025, the company acquired AUO’s Tainan and Houli plants, as well as a Taichung facility owned by Guangyao. It then agreed in early 2026 to acquire Powerchip’s Tongluo P5 fab for USD 1.8bn. Even against the backdrop of tariffs in 2026, Micron has continued to add capacity in Taiwan. The reason is that three decades of investment, talent formation and supply-chain integration mean that a sudden pivot to the US could create capacity risk and cause the company to miss the memory supercycle.

AMD is also increasing its investment in Taiwan, as discussed above. On 21 May 2026, Lisa Su announced that AMD would invest more than USD 10bn in Taiwan’s industrial ecosystem. This is AMD’s largest-ever Taiwan supply-chain investment commitment, covering more than 11 Taiwanese companies across foundry, advanced packaging, ABF substrates and ODM assembly. The investment focuses on advanced packaging, substrates and rack-scale systems, securing capacity ahead of the AI server ramp from the second half of 2026 to 2029. Su described it as a “huge vote of confidence” in Taiwan’s technology, and said that when AMD asks partners to accelerate production, it should also share in the related investment. This “secure-capacity and prepayment” model echoes the dynamic discussed earlier: customers are willing to pay in advance to lock in supply.

Even Intel, the representative company of US manufacturing, cannot completely avoid Taiwan’s supply chain. Foxconn and Intel announced a strategic partnership in June to jointly develop next-generation AI infrastructure, including rack-scale systems based on Intel Xeon processors and AI accelerators, combined with high-speed interconnect, liquid cooling and energy-efficiency designs. The collaboration also extends to edge AI and physical AI, such as AI agents and robotics. The significance is that even Intel, a representative company in the US reshoring push, still needs to rely on Taiwan’s largest system integrator to bring rack-scale AI to market. Intel provides silicon and software, but Foxconn provides manufacturing scale and system integration capability — and this density exists only in Taiwan.

Advanced packaging and systems: the concrete expression of a complete ecosystem

Beyond wafer manufacturing, Taiwan’s moat is also reflected in its complete ecosystem across advanced packaging, testing, racks and system assembly. An AI chip is not built by a GPU alone; it requires the efficient integration of the GPU, HBM and other chips, together with power, thermal, rack and system design. Therefore, Taiwan’s advantage is not limited to TSMC itself. It lies in the ability of the entire supply chain to coordinate rapidly around TSMC.

Advanced packaging is one of the most critical bottlenecks. As AI chips become larger and use more HBM, packaging technology directly affects chip performance, power consumption and mass-production speed. TSMC’s CoWoS is one of the most important advanced packaging technologies today. However, to deliver the huge volume of AI orders smoothly, backend packaging, assembly and testing support is still required.

ASE Technology is Taiwan’s representative company in advanced packaging and testing. ASE can take on part of the backend packaging, assembly and testing work when TSMC’s AI orders rise rapidly, allowing more demand to remain within Taiwan’s supply chain. Driven by AI demand, the company’s advanced packaging and testing revenue is expected to double this year, while capital expenditure has been raised from USD 7.0bn to USD 8.5bn and could increase further, reflecting continued tightness in related capacity.

Beyond packaging and testing, AI servers also require complete rack, power and thermal systems. As the power consumption of high-end AI chips rises, liquid cooling, chassis design and system integration become increasingly important. Chenbro (8210), a global leader in server chassis, has entered the liquid-cooled rack market for NVL72 and ASIC platforms, showing that Taiwan’s supply chain has extended beyond wafer manufacturing into the AI server system layer.

From TSMC in wafer manufacturing, to ASE in advanced packaging and testing, and then to Chenbro and peers in racks and thermal solutions, Taiwan covers almost every critical node in the AI server stack. This complete network — spanning advanced process technology, packaging, testing, substrates, racks and cooling — is the real-world expression of Taiwan’s semiconductor moat.

However, one key layer is still missing from this picture: chip design itself. This is where MediaTek comes in.

MediaTek: From Smartphones to AI Infrastructure

MediaTek is transforming from a traditional smartphone chip supplier into a provider of custom cloud AI chip design services.

In the past, MediaTek mainly sold standardised smartphone chips. Today, however, its role is gradually shifting toward helping customers turn product specifications into manufacturable chip solutions. The company provides key chip modules such as input/output chip design and high-speed data transmission technologies, as well as backend system-integration services. It also helps customers coordinate TSMC advanced process and CoWoS capacity, creating tighter collaboration across design, manufacturing, packaging and the system layer. This is no longer a traditional product-sales model, but a higher-value-added design-service model.

Google has become an important customer for MediaTek in cloud AI chips. MediaTek is participating in Google’s next-generation cloud AI accelerator project, responsible for part of the chip design and backend integration. Market data indicates that the project is progressing well, with related revenue expected to contribute around USD 2bn in the fourth quarter of 2026 and expand further in 2027. The company has also raised its 2027 cloud custom chip market-size forecast to USD 70bn to USD 80bn, reflecting demand far above previous expectations.

More importantly, MediaTek’s design capability is also extending from cloud AI into end devices. During Computex Taipei 2026, NVIDIA launched the RTX Spark superchip, positioned as a next-generation Windows PC platform for personal AI agents. The goal is to allow AI agents to run locally on devices rather than relying entirely on the cloud. MediaTek participated in the custom CPU design for RTX Spark, contributing to power efficiency, performance and connectivity. In addition, the platform uses TSMC’s 3nm process, meaning that as AI compute moves from data centres into PCs and end devices, Taiwan’s chip design and advanced process capabilities remain key beneficiaries.

In summary, MediaTek is moving from a consumer electronics chip company into an important part of the cloud AI infrastructure and edge AI device supply chain. Together with TSMC’s advanced process technology, ASE’s packaging and testing, and the system capabilities of Chenbro and peers, it helps form an even more complete Taiwan AI hardware ecosystem. Whether AI compute stays in the cloud or gradually spills over into PCs and end devices, it remains difficult to bypass Taiwan’s supply chain.

Investment Implication| Taiwan’s Equity Re-rating Is Not Over

On valuation, the Taiwan Weighted Index is currently trading at around 20x 2026 forecast earnings. Although this is around one standard deviation above the historical average, it remains relatively reasonable compared with the US semiconductor sector, which trades at more than 30x, and is supported by strong AI-driven earnings growth. Based on 2028 forecast earnings and a fair P/E of 20x, the Taiwan Weighted Index still offers around 37% potential upside from current levels.

Figure 3: Although the Taiwan Weighted Index’s forward P/E of more than 20x is around one standard deviation above its historical average, it remains relatively reasonable compared with the US semiconductor sector, as represented by the MVIS US Listed Semiconductor 25 Index, at more than 31x

Figure 4: TWSE Index and EPS

Table 4:  Projections for the TWSE Index and Upside Potential

Related products: