Home/Research/AI Compute: From Leader Premium to Second-Line Elasticity
中文
AI Compute: From Leader Premium to Second-Line Elasticity
May 27, 2026 阅读中文版

AI Compute: From Leader Premium to Second-Line Elasticity

Executive Summary

The AI compute trade is no longer a single-stock story around NVDA. From January 2, 2025 to May 26, 2026, local market data show LITE up 965.5%, GLW up 325.6%, AMD up 313.5%, and COHR up 280.7%. NVDA rose 54.7% over the same window. NVDA still anchors the demand story, but the strongest price elasticity has moved into optical links, glass, alternative GPUs, ASICs, networking, and the foundry layer.

That does not mean the second-line fundamentals have been documented with the same clarity. A cleaner description is that the market is capitalizing the next potential bottleneck before every income statement has caught up. The price rotation is clear. The earnings proof is more uneven.

Core Findings

First, AI compute pricing has moved from leader confirmation to supply-chain diffusion. LITE, GLW, AMD, and COHR have all outperformed NVDA since the start of 2025.

Second, the expectation layer still supports the AI capex chain. Analyst consensus estimates show NVDA revenue rising about 65% in FY2026 and 81% in FY2027. AMD is expected to grow about 46% and 53% over the same two fiscal years. TSM is expected to grow about 38% and 27%, using the original reporting currency captured in the local database.

Third, valuation pressure is more visible in the second-line names. NVDA’s trailing PE is 38.7, at the 54.5th percentile of its local five-year history. ANET is at 56.8 times trailing earnings and the 95.8th percentile. LITE is at 506.7 times trailing earnings and the 97.8th percentile, though its history is shorter and therefore lower confidence.

Fourth, news and search data support a hot semiconductor backdrop, but they should not be the main anchor. GDELT and Google Trends still show warm AI and semiconductor language. Aggregated news headlines contain too much source noise to carry the thesis alone.

From Leader Confirmation to Bottleneck Diffusion

If the market is viewed only through NVDA, the 2026 AI compute move looks strong but not extreme. NVDA moved from $138.31 on January 2, 2025 to $213.90 on May 26, 2026, a 54.7% gain. That is a powerful move, but it was not the strongest part of the chain.

The extreme moves came from the side channels. LITE moved from $85.60 to $912.11, up 965.5%. GLW moved from $46.71 to $198.78, up 325.6%. AMD moved from $120.63 to $498.84, up 313.5%. COHR moved from $100.59 to $382.94, up 280.7%.

The market is no longer asking only who sells the GPU. It is asking where the next layer of supply may break. That layer can be optical modules, optical materials, glass substrates, switches, ASIC supply, advanced packaging, or foundry allocation.

This diffusion has two sides. On the constructive side, the AI capex opportunity set has widened. On the risk side, the evidence quality becomes uneven. NVDA has the cleanest revenue and earnings proof. Many side-channel companies do not yet disclose AI revenue with the same granularity.

Expectations Are Still Moving Up

Consensus estimates have not rejected the AI demand story. Analyst consensus shows NVDA’s average revenue estimate moving from about $129.4 billion in FY2025 to $213.7 billion in FY2026 and $387.3 billion in FY2027. That implies roughly 65% growth followed by 81% growth. This is why NVDA remains the anchor: the growth is not only narrative, it is embedded in analyst models.

AMD looks more like second-supplier elasticity. Its average revenue estimate rises from about $34.1 billion in FY2025 to $49.9 billion in FY2026 and $76.3 billion in FY2027, or about 46% and 53%. The market reaction to AMD is not a statement that it has displaced NVDA. It reflects the way a second supplier can gain option value when the primary supply chain is tight.

TSM sits closer to infrastructure. The local database records the estimates in the original reporting currency and shows about 38% revenue growth in FY2026 and 27% in FY2027. Its price elasticity is lower than AMD’s, but the signal is important: AI demand is reaching the manufacturing layer, not only chip designers.

The Crowding Point Is Not NVDA Itself

Valuation percentiles give a counterintuitive result. NVDA is not the most crowded part of the chain. As of May 26, 2026, NVDA’s trailing PE was 38.7, at the 54.5th percentile of its five-year local history, based on 1,254 history days with high confidence.

The crowding is more visible in the side channels. ANET trades at 56.8 times trailing earnings, at the 95.8th percentile, with high confidence. LITE trades at 506.7 times trailing earnings, at the 97.8th percentile, with medium confidence because the history is shorter. CSCO and GLW also sit in high-percentile zones.

That is the risk embedded in diffusion. The names with the highest price elasticity do not necessarily have the cleanest fundamental disclosure. The market is paying for possible next-layer scarcity. If demand momentum slows or customer orders do not convert into revenue, the side channels can be more sensitive than the anchor.

Sentiment Is Backdrop, Not Proof

GDELT and Google Trends support the idea that semiconductor attention remains warm. In local news-tone data (GDELT), the semiconductor query averaged 0.50 over the last month and 0.63 over the last seven days. Artificial intelligence averaged 0.27 over the last month and 0.36 over the last seven days. In Google Trends, semiconductor averaged 60.0 over the past 12 months and 72.4 over the latest 30 days. AI averaged 65.6 over the past 12 months and 76.2 over the latest 30 days.

But headline sentiment is not the evidence anchor. The news-sentiment feed contains many NVDA and AMD references, but a large share comes from Yahoo aggregation or intraday ranking articles. Ticker linkage can be weak and sentiment scores are often absent. The data are useful for monitoring narrative temperature, not for replacing revenue, order, capex, and valuation work.

What We Do Not Know

Local EDGAR fundamentals do not fully cover LITE, GLW, COHR, and other optical or materials names. This report can check price diffusion and valuation percentile behavior. It cannot pin down the exact AI revenue contribution for every company in the chain.

The local database does not contain a full semiconductor sector valuation panel or full-industry forward PE history. The conclusion should not be extended into a blanket claim that the whole semiconductor sector is cheap or expensive.

The headline-sentiment view contains source noise. A cleaner ticker-level sentiment indicator would require URL-level cleaning and related-field validation.

Conclusion

The AI compute chain has entered a second phase. NVDA remains the demand anchor, but price elasticity has diffused into the side channels that look closest to the next bottleneck. That diffusion is real. The discipline is to avoid treating price diffusion as identical to earnings conversion.

A cleaner framework separates the chain into three layers: NVDA as the documented demand anchor; AMD and TSM as measurable supply diffusion; and LITE, GLW, COHR, and ANET as higher-elasticity bottleneck pricing. All three belong to the AI compute chain. They do not carry the same evidence quality or valuation risk.

Data and Sources

This report is an independent KSINQ market observation for informational purposes only. It is not investment advice. Data snapshot: May 27, 2026.