「AI関連」という風呂敷に、決算書まで包むな——日経のEPS警戒論と9銘柄の分類を検証
One AI Label Is Not an Earnings Model Nikkei’s November 9 article raises a legitimate warning: a low forward P/E is not proof of value when forecast earnings may fal
One AI Label Is Not an Earnings Model
Nikkei’s November 9 article raises a legitimate warning: a low forward P/E is not proof of value when forecast earnings may fall. But an article questioning forecasts must also make its own evidence testable.
The article does not simply declare the entire industry an AI bubble. Nor does it establish that the newspaper dislikes semiconductors. Those accusations should be excluded. The stronger criticism concerns classification, selection and the link between evidence and conclusions.
AI exposure is real; comparable economics are not
Mitsui Kinzoku supplies ultrathin copper foil used in semiconductor packaging and optical modules. Its official announcements discuss data-center and memory-related demand. Calling it AI-related therefore has a factual basis. Calling its entire business a pure AI earnings stream would require much more evidence.
SoftBank Group also has genuine exposure through Arm and other AI-computing businesses and investments. It should not be excluded merely because it is a holding company. Yet it requires a different valuation framework: asset values, debt, liquidity, investment gains and holding-company discounts. Its CFO explicitly discusses the limitations of accounting profit and the importance of NAV.
The nine companies span NAND and SSDs, semiconductor equipment, copper foil, optical connectivity, power devices, silicon wafers and investment holdings. They do not sell the same thing or recognize profit in the same way. AI exposure is a thematic overlay, not a substitute for a business model.
What each business actually requires
Kioxia requires analysis of NAND prices, shipment volumes, mix, inventories, capital expenditure and depreciation. KOKUSAI ELECTRIC and SCREEN depend on customer fabrication investment, equipment orders and delivery schedules. Mitsui Kinzoku needs product-level volume, pricing and utilization analysis. Furukawa Electric and Fujikura need optical-product orders, production capacity and margins. ROHM needs analysis of its wider automotive and industrial exposure as well as the commercialization of AI-server power products. SUMCO requires wafer demand, pricing, customer inventories and utilization analysis. SBG needs asset-value and financing analysis alongside operating businesses.
These are research questions, not bullish conclusions. A documented connection to AI does not establish the size of incremental earnings.
A target-price gap is not an EPS forecasting error
Recalculating the original article’s target-price-to-market-price ratios produces approximately 4.24 for Kioxia, 2.76 for Furukawa, 2.51 for Mitsui Kinzoku, 2.47 for KOKUSAI ELECTRIC, 2.45 for SBG, 2.38 for SCREEN, 2.21 for Fujikura, 2.17 for ROHM and 2.21 for SUMCO. These broadly match the rounded published ratios.
However, a target price can depend on forecast earnings, valuation multiples, the target horizon, asset valuations and the age of the estimate. A large gap does not isolate which input is wrong. To establish excessive EPS forecasts, show actual EPS revisions, forecast dispersion, company guidance, update dates and the assumptions being challenged.
Selection does not establish a sector-wide relationship
The table covers selected Nikkei 225 constituents, not every AI company or the whole semiconductor industry. Anthropic’s absence is therefore not itself a flaw.
The claim that AI-related stocks have larger gaps needs a defined AI universe, the complete distribution and a non-AI comparison using consistent dates. A list of large gaps cannot establish that stronger AI exposure systematically produces larger gaps. Market prices are observable transactions; target prices are forecasts. Neither label establishes the market’s final correctness.
A reproducibility problem
Nikkei’s official November 9 closing index level is 58,775.50. Using the article’s displayed composite target of 84,000 gives:
84,000 / 58,775.50 − 1 = approximately 42.9%, rather than 44%.
Unrounded inputs or a different calculation timestamp may explain the difference. This does not prove a calculation error, but the displayed inputs do not reproduce the claim. The composite methodology should also explain index adjustments, estimate selection and reference dates. There is no basis to assert that the data provider omitted those adjustments.
Low P/E still needs a real earnings model
At an unchanged share price, a P/E of 2 becomes 4 if EPS halves. That alone cannot prove the stock is expensive. Conversely, low P/E is no guarantee of value when earnings are cyclical or capital requirements are heavy. Normalized earnings, free cash flow after investment and balance-sheet risks matter.
Political risk also requires intermediate evidence: which rule, which government level, what legal status, what implementation date and which customer orders are affected? An election result does not itself enact a nationwide data-center restriction. Restrictions might delay projects or alter their location and efficiency investment; these remain scenarios until supported.
TokyoBrainie’s judgment
Keep the warning against treating low forward P/E as automatic value. Reject the use of target-price gaps as a substitute for testing EPS forecasts. Accept documented AI exposure while separating the different mechanisms through which it affects profit.
An AI label is a useful starting point for research. It is an inadequate ending point. Bulls and bears should both show the calculation connecting the theme to earnings.
Sources and the full numerical tables appear in the Japanese analysis below. The illustration uses fictional characters to satirize business classification.
