Semiconductor 4Q26: Longer runway, cheaper entry; Upgrade to 3.5 Stars Attractive

To recap, our previous outlook carried a Neutral call on the industry — not owing to any deterioration in the investment thesis, but because valuations already appeared priced for perfection. While we saw limited room to add exposure to the broader sector, we identified foundry and memory as our highest-conviction plays within the space.

That said, as we conclude 3Q26, throughout the latest earnings season, we see a substantial improvement in the industry, pointing to still rosy, or to some extent, an even greater AI development ahead. Fabless players like Nvidia and Broadcom are guiding substantial growth into FY28, while stronger memory and server CPU demand is widening the earnings contribution across the supply chain. With that, we believe investors now have a longer earnings runway at a lower valuation after the July correction, having a greater entry opportunity to add semiconductor exposure into investors’ portfolios heading into 4Q26.

Figure 1: SMH YTD performance and our previous call.

Guidance extends the AI runway

The more meaningful takeaway from last quarter was not the scale of the earnings beats, but how far management guidance extended beyond our expectations. In 2QFY27, Nvidia guided 70% revenue growth (vs consensus at 45%) in FY28 despite supply constraints; this is the first time the company has provided full-year guidance during the fiscal year. Similarly, Broadcom, the biggest custom ASIC player, raised its FY27 AI revenue guidance to USD115 billion and introduced FY28 guidance of USD230 billion, two consecutive years of doubling in revenue, as OpenAI and Meta deployments add to its existing Google TPU business.

Beyond accelerators, Micron’s FQ4 guidance of USD50 billion in revenue at 86% gross margin points to the strength of memory pricing and demand, with management expecting supply to remain tight beyond 2027. Meanwhile, as agentic AI adds tasks such as code execution, data retrieval and workflow coordination, the need for greater CPU for inference is rising fast alongside GPUs. Throughout the quarter, Intel and AMD have both highlighted stronger-than-expected server CPU demand, with AMD expecting server CPU revenue growth above 80% in 2H26 and above 70% in 2027.

The AI buildout hasn’t slowed, and the pie is expanding. Following a strong quarter, we now expect SMH’s underlying earnings to grow 87% in 2026 and 40% in 2027, up from 69% and 30% in our previous review.

Figure 2: Earnings surprises

Source: Bloomberg Finance L.P., iFAST compilations. Data as of 27 August 2026.

Figure 3: Server CPU Unit TAM Expansion by Workload

AI capex supercycle expected to remain robust through 2027.

As highlighted in our various previous write-ups, we see the AI capex as a double-edged sword, an indicator we monitor closely where the spending sustainability could fundamentally alter the industry outlook as AI capex spending is underwriting every layer of the industry. 

The question of capex spending sustainability thus lies in two things, with both showing healthy signs

1.      Has the demand for AI computing slowed?

Based on a proxy for token consumption from OpenRouter, token usage has been reaching its weekly highs, suggesting that the demand for AI computing is still strong.

2.      Has the return on investment (ROI) improved?

Hyperscalers stand to benefit from increasing demand for AI computing. As more hyperscalers, including Meta (personal AI agent - Muse), Google (Gemini), and Microsoft (recently revamping Copilot with code generation, agentic AI tools), expand their own AI offerings, we see stronger commercial reasons for them to sustain capex spending.

At the same time, hyperscalers are also increasingly focused on measures to improve cost efficiency, such as workforce rationalization and greater vertical integration through the development of proprietary AI silicon. On top of that, improving cloud revenue to capex ratio further strengthens our view that ROI is stabilising, with our estimates suggesting that the combined capex at Amazon, Alphabet, Meta, Microsoft and Oracle will reach approximately USD830 billion in 2026 and USD1.16 trillion in 2027.

As such, we see both growing AI usage and greater ROI are pointing to greater hyperscalers capex sustainability, thereby supporting demand and earnings across the semiconductor sector.

Read more: Big Tech earnings 2Q26: Demand confirmed; Higher Burden of Proof

Figure 4: AI token consumption has been trending upward exponentially.

Source: OpenRouter, iFAST compilations. Data as of 29 September 2026.

Figure 5: Hyperscalers’ operating margin is improving

Figure 6: AI capex underwrites every semiconductor value chain.

Figure 7: We continue to see robust capex spending from hyperscalers.

Leading-edge node manufacturing and AI packaging capacity expected to remain in constraint till 2028.

Among the semiconductor value chain, the manufacturing bottleneck that we have been highlighting remains in constraint, with TSMC’s wafer capacity this year reportedly already sold.

Although TSMC has raised 2026 capex to USD60–64 billion and lifted revenue growth guidance to slightly over 40%, the expansion is still chasing demand. Meanwhile, although there are competitive pressures coming from Intel and Samsung, we believe they are unlikely to materially challenge TSMC before 2028 as tight 3nm and 2nm supply continues to support pricing power.

At the same time, as demand is broadening beyond mobile due to increasing AI accelerators and custom chips are migrating to 3nm and 2nm, the reallocation of 5nm capacity to newer nodes keeps advanced logic TSMC's main revenue and margin lever through 2027, alongside faster packaging growth.

In short, bottlenecks remain in Asia, and we remain positive on TSMC given its dominant share of the global foundry market. Tight supply should supports pricing, while discipline capacity expansion allows TSMC to meet more AI demand and increases spending on equipment, thus, reinforcing our preference for foundry and WFE.

Read more: TSMC 2Q26 results tell you why we are so bullish on Asian semiconductors

Figure 8: TSMC 2nm is ramping up.

Source: Bloomberg Finance L.P., iFAST compilations. Data as of 3 July 2026.

Memory contracts strengthen earnings visibility

We remain positive on memory. After months to about a year of rallying, the question that bulls and bears have been debating is whether pricing power can be sustained, especially as we are seeing increasing LTAs locking in demand for the memory chips.

While LTAs may cap incremental ASP upside just as the market tightens, we see the contracts as signalling rising demand and reducing oversupply risk by anchoring demand and enabling more prudent capex. At the same time, while bears may argue customers may not fulfil the contracts during a downcycle, several developments reinforced our positive view on the memory:

1.      Rising hyperscaler capex supports memory demand, with AI infrastructure requiring more HBM and server DRAM.

2.      Take-or-pay provisions, as part of LTAs’ agreement structures should strengthen volume commitments, where buyers must purchase an agreed minimum volume or pay a specified fee for the shortfall, making it costly to walk away.

3.      Supply is expected to remain tight into 2028, and the current prepayments signal memory players still have bargaining power, which should continue to sustain the current ASP.

Taken together, these factors support our view that memory is becoming structurally less cyclical. That said, we remain mindful of the sector's inherent volatility and will continue to monitor developments closely.

Similarly, while China supply additions (DRAM and NAND) from CXMT may increase the risk of  shortening the current upcycle, leading edge server DRAM/HBM remains insulated as effective output is constrained by technology gaps, with yields and the restriction of advanced equipment access remaining the major bottlenecks.

Read more: Micron Q326: The Night Is Still Young

Table 1: LTA details announced across major memory makers in 2QCY26 earnings conference.

Source: JPMorgan, iFAST compilations. Data as of 22 September 2026.

Pacing AI models doesn’t change our view

Recently, Anthropic CEO Dario Amodei has called to pace frontier AI development as increasingly ⁠autonomous AI models can behave in unexpected and potentially harmful ways, including sabotaging code, assisting fraud and manipulating information in controlled tests. While this has raised concerns that tighter regulation could slow infrastructure investment, we see it differently.

Firstly, the implementation of stricter regulations to pace the AI model companies could raise the concern of the national AI race, which we see as less incentive for the government to enforce it. This presents a prisoner’s dilemma, where neither countries nor companies want to slow first and allow competitors to gain ground. AI leadership is a strategic priority for the US, while frontier developers have a narrow technological lead to defend. In fact, Anthropic rolled out its new Opus 5.5 model to counter OpenAI’s momentum since its launch of GPT-6 Astra, after the company’s CEO, Dario Amodei, had called for the global AI community to slow the pace of releasing new ​capabilities to address those concerns.

 

On the other hand, based on the AI Gateway data, open-weights models have outpaced closed-weights models, yet we haven’t seen any slowdown in computing demand, with total model consumption continuing to reach new weekly highs. Hence, essentially, no matter whether open-weights or closed-weights models gain stronger market share due to cost efficiencies or model scoring, both would increase the need for higher computing demand, as no one can afford to lose the AI race.

Figure 9: Open-weights model token volume is catching up with closed-weights models. 

Source: AI Gateway, iFAST compilations. Data as of 28 September 2026.

Earnings visibility is getting stronger and valuation is getting more attractive; Upgrade to 3.5 stars Attractive

We see greater upside for the semiconductor sector as stronger guidance extends a longer earnings runway, with sustained capex spending and persistent supply bottlenecks have strengthened our view that semiconductors remain central to AI development, and that the outlook remains firmly constructive.

Our preferred exposure remains foundry, memory and WFE, alongside selective opportunities in fabless. Within equipment, we favour ASML and Lam Research as foundry and memory expansion could benefit equipment orders, while more complex chip structures and advanced packaging increase equipment requirements.

Within fabless, we favour Nvidia and Broadcom, where stronger guidance has lifted our earnings estimates and valuations have become more attractive. On the other hand, while we remain positive on AMD’s server growth, elevated expectations leave less upside in the stock. Meanwhile, Intel’s recovery has yet to justify its valuation given continued foundry losses and heavy reinvestment needs. That said, while continuous rollout of AI agents could sustain the near-term “CPU trade”, we nevertheless remain selective, favouring companies where earnings upgrades offer sufficient support for the valuations being paid.

For SMH, stronger earnings alongside lower valuations now warrant a more positive view of the sector. At a fair P/E of 24x, we derived a target price of USD844, which represents 40% upside potential. The entry point has improved and thus we upgrade the semiconductors from 2.5 stars Neutral to 3.5 stars Attractive.

Table 2: MVIS US Listed Semiconductor 25 IndexValuation.

	FY25	FY26E	FY27E	FY28E
EPS	446.8	836.3	1173.4	1441.1
EPS growth y/y		87.2%	40.3%	22.8%
P/E	35.8	29.42	20.97	17.08
Dividend yield	0.7%	0.6%	0.7%	0.9%
Fair P/E (x)				24
Target price (USD)				844
Potential upside				40%
Source: Bloomberg Finance L.P., iFAST compilations. Data as of 24 September 2026.

Key risks:

1.      Private AI lab funding risk

OpenAI and Anthropic’s growing compute commitments depend on securing further financing, while recent Anthropic confidential filing suggested that the company operation remain loss making. Higher financing costs or delayed fundraising could slow deployments and chip orders.

2.      Data centre execution risk

Power, cooling and construction bottlenecks could push back chip deliveries and deployments.

3.      Mid-term elections policies risk

A Democratic takeover of either congress or senate could potentially intensify congressional scrutiny of AI companies through investigations and hearings, increasing pressure for tighter regulation. 

Declaration:

For specific disclosure, at the time of publication of this report, IFPL (via its connected and associated entities) and the analyst who produced this report hold a NIL position in the abovementioned securities.

This research report was prepared with the assistance of artificial intelligence (AI) tools. iFAST Financial Pte Ltd does not rely exclusively on AI for content generation; the content of this report — including all investment theses, ratings, price targets and conclusions — has been independently reviewed and verified by the research analyst(s) to ensure accuracy and professional integrity.