Why are stocks plunging despite higher earnings expectations?
- Bernstein acknowledges its earlier short-term bullish call on AI infrastructure was wrong: from the close on 28 September to 6 November, its covered stocks fell 12%–42%. During that period, results and guidance from five companies that had reported were above sell-side consensus, but performance against buy-side expectations was mixed.
- Overall earnings feedback was neutral to positive, but share prices fell alongside valuation-multiple compression across the hardware and semiconductor sectors; the SOX was down 38% from its June peak. Bernstein also believes factors such as liquidity needs ahead of IPOs and rising interest rates intensified the selloff.
Do the returns on AI infrastructure add up?
- Weak recurring revenue at AI labs in September raised market concerns about cloud providers’ returns on investment and ability to secure external financing. Bernstein says lab revenue strengthened in October, making the weakness in September look more like pricing volatility; token usage continues to grow, and price cuts are helping expand adoption.
- Narrower margins at the model layer do not necessarily hurt upstream companies: with end demand unchanged, higher token demand means more infrastructure is needed. Bernstein’s rough estimate is that hyperscalers will have about 1.2 trillion dollars of AI infrastructure coming online by 2027; at a required return on investment of 15% and a net margin of 24%, that would imply about 175 billion dollars in net profit and 729 billion dollars in spending. The firm considers this level of spending feasible, but not certain.
Which year’s growth will power and land constraints affect first?
- Bernstein judges that power and land constraints will have a limited impact in 2027: capacity delivered that year is already under construction, while credible Texas capacity targeted to come online in 2027 that has not yet broken ground is only 0.3 GW; given current timelines, those projects may already be delayed.
- Risks are higher in 2028. The firm believes local policies are more likely to require data centers to provide greater economic benefits to states and communities, and projects may relocate or adapt to new rules through negotiation; however, its confidence in its assessment of power and land issues is lower than in its demand outlook.
What do the target-price cuts and rebound outlook each tell us?
- Bernstein left earnings forecasts unchanged but cut its target prices for Lumentum, Celestica and Corning, with the changes mainly reflecting valuation assumptions. Celestica’s target price is $ 340, equivalent to 14 times its 2027 fiscal-year EPS forecast of $ 24; Corning’s target price is $ 120, equivalent to 21 times its 2028 EPS forecast of $ 5.65. Its valuation multiple was cut from 25 times to its five-year average, and end markets for its non-AI businesses are weak.
- Lumentum’s target price was cut to $ 635, roughly matching an acquisition offer calculated on an exchange basis of 4.5 Marvell shares. Bernstein estimates the probability of the deal closing at 67% and considers the valuation high; shareholder opposition and opposition from Chinese regulators are constraints. The firm believes a technical rebound may occur in the short term, but a full recovery will take longer; catalysts are limited ahead of hyperscalers’ capital-spending updates in January next year and Nvidia GTC in March.