Post-Midterm AI Sell-off: Data Centre Curbs Are a Headwind, Not the End of the Trade

• Democrats won control of both chambers of Congress, and fears that newly elected Democratic governors will pause data centre construction sparked a sell-off in chip and AI stocks, amplified by leveraged positions.

• We expect the data centre buildout to moderate but not stall, as most 2027 capacity is already being built. Restrictions are likely to vary by state and locality rather than apply nationwide, and hyperscalers can negotiate or build elsewhere.

• Slower construction may delay chip shipments and weigh on near-term earnings for semiconductors, but AI demand remains strong and valuations provide a buffer.

• We continue to prefer hyperscalers over neoclouds, as their scale and pricing power allow them to absorb higher costs, while their existing capacity could become more valuable.

• With much of the fear already priced in, investors should stay invested and avoid panic selling. We see potential upside of 78% for US and 185% for Asian semiconductors even after factoring in greater risks.

Democratic sweep triggers a sharp sell-off in chips and data centre stocks

US stocks fell sharply after Democrats gained control of both the House and the Senate. On 5 November the S&P 500 dropped 4.4% and the Nasdaq Composite 5.5%, their worst day since April 2025, led by a sell-off in semiconductor stocks. Investors feared that newly elected Democratic governors, several of whom campaigned on pausing data centre construction, would slow the AI infrastructure buildout.

Chip stocks bore the brunt in both the US and Asia. Since the close on 3 November, before the election results were clear, the VanEck Semiconductor ETF has fallen 13.7% over the past two days, and the Global X Asia Semiconductor ETF was down 6.1% on 6 November.

Neoclouds fell even harder. CoreWeave dropped 9.9% on 4 November and 9.8% on 5 November, and Nebius 6.7% and 14.9%. Hyperscalers held up better. Microsoft and Amazon rose 6.7% and 6.1% on 4 November, as investors bet that slower construction would make existing capacity more valuable, and fell by less than chipmakers on 5 November.

Fear drove the sell-off, as investors rushed to price in the risk of a slower AI buildout, and leveraged positions amplified the moves. Hedge funds, which had been cutting exposure to crowded AI trades since late October, added to the selling pressure as they reduced positions further. Meanwhile, US leveraged ETFs had taken in about USD15 billion in October, the largest one-month inflow in at least two years, led by retail investors buying leveraged tech funds. As tech selling gathered momentum, those positions were forced to sell, deepening the decline.

Figure 1: A record build-up in leveraged positions amplified the sell-off

Data centre buildout may moderate, but won't stall

The risk to the AI buildout comes mainly from the states. Many of the newly elected Democratic governors campaigned on data centre moratoriums, and the most important example is Texas, which accounts for nearly 30% of the US data centre capacity set to come online in 2028, by far the largest share of any state, according to research firm SemiAnalysis.

Texas governor-elect Gina Hinojosa's campaign platform called for a moratorium on new data centres until legislators establish protections for residents. Her stated concerns included electricity costs, water supplies, public participation and subsidies. Outgoing Republican Governor Greg Abbott had already imposed a permitting and grid-connection freeze during an audit, and Hinojosa's position raises the prospect of a longer pause with additional conditions.

Looking beyond Texas to the US as a whole, much of the data centre capacity expected to come online in 2027 is already under construction, which limits the immediate impact of restrictions on new permits. SemiAnalysis estimates that roughly three-quarters of the 2027 capacity due in the top 20 states is already being built. However, a larger share of the national 2028 pipeline remains in the planning stages and is therefore more vulnerable to permitting delays and regulatory uncertainty.

A sharp slowdown in data centre construction and AI investment, caused by a broad tightening of restrictions across states, could weigh on the entire US economy, but this is not our base case. No single office in a state decides whether a project goes ahead or is halted. Decisions often reflect an interplay between city governments, county officials and state agencies, and local officials who have some say over data centre development may resist sweeping restrictions from the state. Democrats pushing for a moratorium might also face pushback from the construction unions that supported some of their campaigns.

There is also room for negotiation, and hyperscalers have the option of building elsewhere. Meta's president, Dina Powell McCormick, said one governor-elect who had made data centres a big part of their platform asked to meet after the inauguration, because they did not want to lose the investment and jobs. She expects much more review of new projects, which will make starting new data centre builds in the US very challenging. However, she said expanding existing sites should be easier, because local communities have seen the benefits over several years. Meta is also looking at other countries where it may find a friendlier environment, and some hyperscalers are already building abroad: Google, for example, is building its first data centre in Malaysia as part of a USD2 billion investment.

A nationwide moratorium on data centres is unlikely, given the US-China technology race, the potential impact on the US economy and the presidential veto it would face. National security concerns may favour a lighter regulatory approach, because policymakers view AI as strategic infrastructure and worry that regulation could slow innovation relative to China. Political incentives point the same way: with the next presidential and congressional elections due in 2028, Democrats will not want to be seen as stalling the US economy.

Congress is also unlikely to pass a federal moratorium, despite the scale of the Democratic victory. Overriding a presidential veto requires a two-thirds majority in both chambers, which Democrats do not have. In the Senate, Republicans can also use the filibuster to prolong debate and delay or block a vote, and ending a filibuster generally requires the support of at least 60 of the 100 senators. Without majorities large enough to clear these hurdles, we expect gridlock in Congress over the next two years rather than sweeping federal restrictions.

Ultimately, the impact will depend on state and local restrictions rather than federal action, and on how long any pauses last. A conditional pause, such as the one proposed by Ohio governor-elect Amy Acton, lifts once projects meet set requirements such as covering their own power costs, so it raises costs but keeps projects moving. A pause that lasts until new legislation is in place, as Hinojosa has proposed in Texas, has no fixed end date and could last considerably longer.

Overall, we expect a moderation in the pace of data centre buildout, but not a collapse in AI investment.

Delayed revenue for semis, higher costs for hyperscalers, but the AI trade holds

AI demand is strong, and we expect it to continue in 2027. GPU rental prices are rising, a sign that demand for computing power still outstrips supply. Token consumption also continues to grow, and the Jevons paradox should keep usage expanding as the cost per token falls, since cheaper AI tends to increase total usage rather than reduce spending.

Hyperscalers’ cash-generative core businesses, improving cloud margins and the monetisation of agentic AI should sustain capital expenditure in 2027, although a rising share is now debt-funded. Accelerating AI spending and data centre expansion supported our positive view on semiconductors, but permitting risks weaken that assumption, making us more cautious. Slower construction could delay hardware orders, shipments and revenue recognition, weighing on earnings. With much of the 2027 capacity already under construction, we expect greater exposure in the 2028 pipeline. Our base case remains a delay rather than a broad loss of demand; lasting losses would become more likely if projects were cancelled and capacity could not be deployed elsewhere. Meanwhile, continued growth in AI usage within a constrained power budget could encourage upgrades to existing infrastructure, supporting demand for newer platforms, more efficient chips and upgraded memory.

Asian chipmakers sit at the centre of this upgrade cycle, supplying much of the memory and advanced manufacturing behind AI servers. For Korea, margin sustainability is the main risk: slower demand alongside expanding supply could weaken pricing. However, Samsung expects shortages to tighten in 2027, while SK hynix sees them persisting beyond 2028. Latest semiconductor exports still show strong shipments, and long term agreements (LTAs) offer some downside cushion, depending on delivery and pricing terms. Inference, longer contexts and agents support demand across HBM, conventional DRAM and NAND. HBM upgrades can improve energy efficiency, while yield gains lower unit costs.

On valuation, Samsung and SK hynix trade at around 2.7x and 2.2x forward earnings, respectively. Even if forecast EPS were cut by 50%, their P/E ratios would rise to only around 5.3x and 4.5x at unchanged share prices, still below their historical medians of 7.5x and 7.2x respectively. Both companies also have substantial net cash positions and sizeable shareholder return plans, including dividends and share buybacks. These balance-sheet and valuation buffers suggest that both companies could withstand a substantial decline in earnings before their valuations become stretched at current share prices.

For Taiwan, TSMC’s advanced process and packaging technologies help customers improve computing performance per watt. Hardware upgrades could continue to support TSMC and the wider Taiwanese supply chain even if new data centre campuses are delayed. Similar to the Korean market, valuations also provide support. TSMC and the TAIEX trade at 15.4x and 13.7x forward earnings, below their five-year averages of 16.8x and 15.4x respectively.

Our conviction in Asian semiconductors has moderated, but we do not yet see sufficient evidence to justify turning negative at current valuations. We remain constructive and view the selloff as overdone.

For hyperscalers, we expect more restrictions and higher costs. Fewer tax incentives, responsibility for their own grid upgrades, power and water costs, and fewer available sites will pressure returns, making margins more important for investors. However, hyperscalers can partly offset this by favouring states with fewer constraints and greater local acceptance, or by building overseas. They should also be able to pass much of the extra cost on to customers, as demand already exceeds supply: all three major cloud providers said earlier this year that revenue would have been higher with more capacity. If demand stays strong while new competing capacity arrives more slowly, existing computing capacity could become more valuable. We therefore continue to prefer hyperscalers over neoclouds, as their scale, diversified footprint and pricing power leave them better placed to absorb these costs.

An unintentional slowdown in data centre construction may even be beneficial for hyperscalers' share prices. Investors have grown wary of the cash hyperscalers are consuming on infrastructure spending before returns arrive, and they may see a more measured pace of growth as better, because it allows revenue growth to catch up.

Related article: Lessons from the CLEC Bust: Why we prefer Hyperscalers over Neoclouds

Stay invested and avoid panic selling

Even after factoring in greater risks to the buildout, our updated valuations point to potential upside of 78% for US semiconductors and 185% for Asian semiconductors, as represented by the VanEck Semiconductor ETF (NASDAQ: SMH) and the Global X Asia Semiconductor ETF (HKEX: 3119) respectively.

With much of the fear already reflected in share prices, we think investors should avoid panic selling into the weakness, as sharp moves driven by policy headlines can reverse once the details become clearer. Overall, we remain positive on global semiconductors and continue to favour the broader digital economy, with a preference for hyperscalers over neoclouds.

If the recent volatility has left you uncertain and you would like someone to speak to, please reach out to the FSM Advisory team at advisory@fsmglobal.sg

Table 1: MVIS US Listed Semiconductor 25 Index valuation

MVSMHTR Index	2025	2026E	2027E	2028E
Earnings Per Share (EPS)	446.8	887.2	1147.9	1357.4
Earnings Growth YoY	-	98.6%	29.4%	18.3%
PE Ratio (X)	-	20.6	15.9	13.5
ETF Target price (based on a fair PE of 24x)				797
Upside Potential				78.0%
Source: Bloomberg Finance L.P., iFAST estimates. Data as of 5 November 2026

Table 2: Factset Asia Semiconductor Index valuation

Factset Asia Semiconductor index	2025	2026E	2027E	2028E
Earnings Per Share (EPS)	19.3	67.6	88.1	97.4
Earnings Growth YoY	-	251.0%	30.3%	10.5%
PE Ratio (X)	-	11.3	8.7	7.9
ETF Target price (based on a fair PE of 20x)				405
Upside Potential				185.1%
Source: Bloomberg Finance L.P., iFAST estimates. Data as of 5 November 2026

Table 3: Recommended products

	Recommended Products
Digital Economy	•       Fidelity Global Technology A-ACC-USD •       Eastspring Investments Unit Trusts - Global Technology SGD •       Invesco NASDAQ Internet ETF (NASDAQ: PNQI) •       VanEck Semiconductor ETF (NASDAQ: SMH) •       Global X Asia Semiconductor ETF (HKEX: 3119)

Declaration:

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.

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.