Sector Rotation in the AI Rally: Three Years In, Where Are We, and Who's Next?

This stage was tightly concentrated in chips and their close cousins. In its three-year retrospective at the end of 5/2026, Morningstar gave the panoramic figure: since 5/2023, the U.S. semiconductor index has returned a cumulative 423.9%, against 85.6% for the broad U.S. market over the same span. The size of the whole sector was rewritten with it, with total semiconductor market value swelling from 2.2 trillion dollars to 9.4 trillion (as of 4/2026, Morningstar).

Stage Two (2024–2025): The Bottleneck Spills Over, Power and the "Pick-and-Shovel" Names Take the Baton

A bottleneck never stays put. As GPU supply gradually ramped, the market started pricing the "supporting shortage" around compute, and the clearest case was electricity.

Power stocks broke out ahead of most people's awareness: nuclear operator Vistra rose 261.3% in 2024, the second-best name in the S&P 500 that year, behind only Palantir.

The logic kept getting confirmed over the next two years. The hardest evidence is in GE Vernova's order book, where CEO Scott Strazik laid out a set of figures on the 2026-04-22 earnings call (Utility Dive).

The gas-turbine backlog (including reserved capacity) reached 100 gigawatts by the end of the first quarter of 2026, up from 83 gigawatts at the end of 2025. The delivery window is tighter still: 2029 and 2030 combined leave only about 10 gigawatts of capacity to sell, and the company has begun taking orders for 2031 and beyond. The scarcity is written straight into the price, with new quotes running 10%–20% above the backlog pricing of the fourth quarter of 2025.

A five-year wait to buy a turbine is what a "hard bottleneck" means.

The demand side is just as blunt: GE Vernova's electrification business booked roughly 2.4 billion dollars of data-center orders in the first quarter of 2026, more than its entire 2025 total (company results).

Nuclear power thus became a buying target for the tech giants. On 2026-01-09, Vistra signed a 20-year power-purchase agreement with Meta (announcement). Under it, three of its nuclear plants will supply Meta a combined 2,609 megawatts of zero-carbon power, of which 433 megawatts come from uprate upgrades. Meta's head of energy put it plainly: nuclear is "critical to advancing our AI ambitions."

In 2020 these three plants were still on a retirement track. Our "Energy · Power Supply Chain" topic lays out, layer by layer, the full path by which electricity went from "utility" to "AI strategic resource."

In the same stage, the optical modules that connect GPUs were first to convert into a major rally, and they did it on the A-share market: by Sina Finance's year-end tally, Eoptolink rose 424% in 2025 and Innolight rose 396%, the top two in the A-share AI compute group (roundup). Optical modules are the networking gear that wires a GPU cluster "into one big machine," the top global suppliers are mostly Chinese, and the worldwide demand windfall for this link landed directly on A shares.

Stage Three (2025–2026): The Memory Super-Cycle, and the Mirror Rally in the China Chain

The next baton, again, showed up first on the leaderboard. The top two U.S. gainers of 2025 were both memory companies, SanDisk (+559%) and Western Digital (+360%), and that already telegraphed the shift in the main line. Into 2026, memory became the hottest link on the whole chain:

• The price moves are nearly vertical. TrendForce data show that in the first quarter of 2026, global DRAM (memory-chip) industry revenue reached 97 billion dollars, up 81% quarter over quarter, with commodity DRAM contract prices up about 93%–98% in the single quarter; it forecasts a further 58%–63% rise in the second quarter, with NAND flash contract prices up 70%–75% (TrendForce month-6 report, month-3 forecast). The mechanism: HBM (high-bandwidth memory bolted next to the AI chip) is crowding out commodity memory capacity, while the cloud vendors build out commodity servers for AI inference, so both ends are grabbing for supply at once.
• Three giants joined the trillion-dollar club in the same month. In 5/2026, Samsung (5/6), Micron (5/26, up 19% in a single day) and SK Hynix (5/27) each crossed a 1 trillion-dollar market value in turn; SK Hynix was up more than 200% on the year (Bloomberg, CNBC, US News).
• The price hikes are starting to bite the upstream buyers. Microsoft attributed roughly 25 billion dollars of its 2026 capex to higher memory and component prices (per CFO Amy Hood, Tom's Hardware); Meta's main reason for raising its 2026 spending guidance was likewise memory. Pricing power at the bottleneck eventually shows up on someone else's cost line.

The China chain ran a "same structure, own tempo" rally over the same window, and the catalyst was the DeepSeek moment of 2025-01-27: this domestic model app topped the free charts in both Apple's U.S. and China stores that day.

The U.S. market's reaction was violent: NVIDIA fell about 17% in a single day and shed roughly 589 billion dollars of market value, the largest single-day market-value loss in U.S. history (CNBC). A "cheap, good model" made the market reprice two things at once: how fragile the U.S. compute narrative was, and how real domestic compute demand had become.

Over the following year, domestic compute began rewriting the price ranking on A shares: on 2025-08-28, Cambricon overtook Kweichow Moutai at a stock price of 1,587.91 yuan, briefly becoming "the highest-priced A share."

The support was delivered earnings. First-three-quarter revenue was 4.607 billion yuan, up roughly 2,386% year over year, and net profit of 1.605 billion yuan swung from loss to profit (21st Century Business Herald, Securities Times).

Policy is the variable unique to the China chain, and 2025–2026 ran a full swing of it:

• In 1/2025, the Biden administration issued the "AI diffusion rule," tightening exports.
• On 2025-05-13, the Trump administration rescinded that rule (TechCrunch).
• In 1/2026, it approved NVIDIA H200 exports to China, on terms that include the U.S. government taking a 25% cut (CNBC).
• On 2026-05-31, the Commerce Department's BIS issued fresh guidance to block the channel by which Chinese firms procure advanced chips such as Blackwell and Rubin through third-country subsidiaries in Singapore, Malaysia and elsewhere (VOA).

The upshot of the swing: the H200 that was cleared met a cool reception in China (White House AI head Sacks claims the Chinese side is declining to buy, which is his own one-sided account), while domestic substitution actually accelerated, with SMIC's capacity utilization at 93.1% in the first quarter of 2026 (report).

The three-year, 380 billion-yuan AI infrastructure plan that Alibaba announced in 2/2025 was, by that year's month-11 earnings call, called "possibly on the conservative side" by management; on the same call CEO Eddie Wu left the remark that "an AI bubble doesn't really exist within three years," his argument being that all GPUs old and new are running at full load (Sina Finance, management's own judgment).

In the first half of 2026, the leading links of the A-share AI chain mirrored the U.S. market. The two optical-module champions rose another roughly sixty percent on the year, AI memory ran the hottest, and Cambricon was roughly flat after its ex-rights adjustment, with money again "concentrating toward the bottleneck."

Where We Are Now: 6/2026, How to Read Three Dashboards

Money has run all the way from chips to memory and power, and now it is time to answer the question from the top: where is the rally? This piece reads three sets of figures.

Set one: capex is still accelerating, but where the money comes from has changed. The four hyperscalers spent roughly 410 billion dollars of capex combined in 2025, already a record.

The 2026 plans add another 77% on top of that, roughly 725 billion dollars combined (Tom's Hardware), broken out below.

Expectations have been chasing company guidance the whole way: in early 1/2026 the market's combined expectation for the four was still around 440 billion dollars (per Fortune's reporting at the time), and 4 months later the companies' own guidance pushed the number up by sixty percent.

Institutions look further out: after the late month-4 earnings season, Evercore and BofA estimated that big-tech AI spending would top 1 trillion dollars in 2027 (CNBC, analysts' forecast).

The cost is just as real. The four generated roughly 200 billion dollars of free cash flow combined in 2025, below the roughly 237 billion of 2024. The gap is starting to be plugged with debt. Alphabet issued 25 billion dollars of debt in 11/2025, and its long-term debt roughly quadrupled on the year to 46.5 billion dollars.

Oracle goes further: free cash flow was -23.7 billion dollars in fiscal 2026, it has already issued 43 billion dollars of debt this year, and it plans to raise another roughly 40 billion in the next fiscal year (including equity issuance) (Oracle results).

The infrastructure build-out is shifting from "funding itself out of cash flow" to "adding leverage," and that is a classic signal that the texture of the cycle is changing.

Set two: the demand order book is enormous, but the increments are decelerating. Oracle's RPO (remaining performance obligations, which you can read as "signed but not yet delivered order book") is a thermometer for AI compute demand: in the fourth quarter of fiscal 2026 (released 2026-06-10) it reached 638 billion dollars, far above Wall Street's expectation (Oracle).

But break the same fiscal year into single-quarter additions: the first quarter added roughly 317 billion (tied to the roughly 300 billion-dollar OpenAI mega-deal booking), the second quarter added roughly 68 billion, and the third quarter added only roughly 29 billion. The total is still hitting records, but the momentum has clearly slowed.

On the night of the earnings beat the stock fell instead of rising, as the market began pricing infrastructure on "ability to deliver," with the sheer size of the order book no longer enough to move it.

Set three: application-layer revenue is, for the first time, "worthy of" the scale of the infrastructure. This is the biggest difference from a year ago:

• Palantir posted first-quarter 2026 revenue of 1.633 billion dollars, up 85% year over year, its fastest growth since its 2020 IPO, with U.S. commercial revenue up 133% year over year, and it raised full-year guidance to roughly 7.65 billion dollars (SEC filing). It was also the top S&P gainer of 2024 (+340.5%) and rose another roughly 135% in 2025. The strongest name in the application layer has delivered three years running.
• Model-layer revenue is a steep curve: Anthropic's annualized revenue went from roughly 9 billion dollars at the end of 2025 to past 30 billion by early 4/2026, with more than 1,000 enterprise customers paying over a million dollars a year (TechCrunch); research group Epoch AI puts its growth at roughly 10x a year, far faster than OpenAI's roughly 3.4x, and extrapolates that the two annualized revenues might cross around 8/2026 at roughly 43 billion dollars (Epoch AI, an extrapolated forecast that comes with its own wide confidence band).
• The usage evidence is more direct still: Google disclosed that its monthly token processing (a token is the unit of text a model is billed on) grew from 9.7 trillion in mid-2024 to roughly 3,200 trillion in 6/2026, about 330x; Google Cloud revenue was 20 billion dollars in the first quarter, up 63% year over year, with an order book of roughly 462 billion dollars (Alphabet investor materials). At the month-3 GTC conference, NVIDIA CEO Jensen Huang declared that "the inference inflection point has arrived" and raised the cumulative order outlook for the Blackwell and Rubin platforms across 2025–2027 from roughly 500 billion dollars to at least 1 trillion (CNBC, company figure).

Mind one contrast, though: the revenue surge is concentrated in the model companies and a few top software names, while the "pure-play application" group in the public market is still waiting as a whole. In the same Morningstar three-year retrospective, the software-applications industry (the Salesforce and Adobe type) rose just 17.6% over three years, more than twenty times less than semiconductors' 423.9%. Put another way, the rotation has reached the door of the application layer and has not yet pushed through it.

Sentiment, for its part, just went through a stress test. The bubble debate boiled over in 11/2025, with 45% of managers in that month's BofA fund-manager survey naming an "AI bubble" the biggest tail risk and 54% saying AI stocks were already in a bubble (Fortune).

The most representative bull-bear clash was between Burry and NVIDIA. Short-seller Michael Burry publicly likened NVIDIA to Cisco in 2000 ("I'm not saying NVIDIA is Enron, it's obviously Cisco"), citing the nearly 3 trillion dollars in spending commitments from cloud vendors over the next three years and the precedent of the late-1990s fiber glut.

NVIDIA took the rare step of circulating a seven-page memo to sell-side analysts rebutting each point, and its defense of depreciating GPUs over 4–6 years became the focal point of the bull-bear fight (CNBC).

More telling was the market's reaction to good news. When NVIDIA delivered revenue of 57 billion dollars that month, up 62% year over year, the stock fell rather than rose, and the 2025 full-year gain settled at roughly 39%. The market has refused to keep paying up for "merely good results."

Across 4–5/2026, the tech and semiconductor ETFs each rose about twenty percent in a single month, a sharp melt-up. Then came the hard brake. On 6/4, the Nasdaq fell 4% in a single day, and the semiconductor-sector ETF fell about ten percent over two days, with the trigger nothing more than Broadcom's results failing to raise full-year AI guidance (CNBC). That was the Nasdaq's worst single day since 4/2025. A single supplier's hesitation on guidance was enough to punch through the entire chain, a direct read of how crowded the trade is.

Today (6/12), SpaceX listed on the Nasdaq in the largest IPO in history, raising roughly 75 billion dollars, and the liquidity it siphons from AI positions is part of the backdrop to this bout of turbulence (CNN). Even so, the several buy- and sell-side people CNN interviewed cast this pullback as consolidation after a strong run, with the major indexes still holding high-single-digit to double-digit gains on the year.

Taking the three dashboards together, this piece's stage read is the late-middle of the infrastructure rally overlapping with the early innings of applications taking the baton.

On valuation, this cycle is fundamentally different from Cisco in 2000: NVIDIA's price-to-earnings ratio is currently around 31x, and it was also below 50x at the start of 2026, whereas Cisco's was around 200x at its peak (Macrotrends, Fortune).

That NVIDIA multiple corresponds to the early month-6 point: a stock price of roughly 205–208 dollars, down about 12% from the mid-month-5 high.

The historical lesson on tempo is worth keeping on the desk. Cisco briefly overtook Microsoft as the world's most valuable company in 3/2000, after which the Nasdaq lost more than three-quarters over two and a half years. Cisco itself did not reclaim its old price peak until 2025-12-10 (CNBC).

That round trip took 25 years and 8 months. Amazon and Google, the application kings of the internet, were precisely the ones that grew on the rubble of the infrastructure bubble.

Infrastructure's "narrative top" has historically come before the "delivery top" in applications, and that is the single most important asymmetry in the rotation framework.

Four Candidates for the Next Baton, and the Gate on Each

That answers the first question. For the second, push forward on the "bottleneck drift plus delivery order" framework: the next stage has four candidate links. What follows is this piece's research-grade reasoning, and each candidate comes with the signal that would have to appear for it to hold.

Candidate one: pure-play applications and AI agent software. History stands on its side (the internet and mobile internet both ended with the application layer cashing the largest market value), and the leading evidence on the revenue side has already shown up (Palantir, Anthropic, the 330x in token usage). The gate is whether enterprise AI budgets can move from "pilot" to "standard issue," and whether falling token costs can make application gross margins hold up. Watching the relative strength of the software-applications index against the semiconductor index on Morningstar's basis is the simplest confirming signal.

Candidate two: edge devices and robotics. Compute sinking from the cloud down to phones, cars and robots is the natural extension of the "inference era," and it is the direction laid out in our "Embodied Intelligence · Robotics Supply Chain" topic. The gate: the appearance of a high-volume edge-AI hit product, or humanoid robots winning verifiable mass-production orders. Until then it is closer to thematic investing.

Candidate three: the "first year of volume" for China's domestic compute. With export controls tightening again at the end of 5/2026, the policy premise for domestic substitution is, if anything, more firmly in place; SMIC running at full capacity, Alibaba's bias toward raising capex, and volume ramps for domestic chip clusters (super-nodes) are the mainstream expectations on the brokerage side (Guotai Junan and CITIC Securities, late-2025 views, an institutional judgment). The gate: the quarterly delivery of domestic AI chip shipments and Chinese cloud capex, plus actual H200 sales to China as the reverse indicator.

Candidate four: the long-cycle continuation of power and "physical assets." What makes this link special is that its bottleneck is measured in years. The gas-turbine order book already runs to 2031, and a nuclear power-purchase agreement is signed for 20 years. It may not have another breakout slope like Vistra's in 2024, but its order visibility is the longest on the whole chain. The gate: whether new gas-turbine quotes and PPA (long-term power-purchase agreement) prices keep rising.

At the same time the rotation framework's failure case has to be written down: if cloud-vendor capex guidance turns (any single one cutting it would be amplified by the market), or if AI revenue growth fails to keep pace with depreciation and interest, the rotation turns into a synchronized drawdown across the whole chain. The semiconductor group falling ten percent over two days on 6/4 already showed a mini version of that pattern.

Two more pieces of evidence point the same way: Morningstar analysts expect demand for AI infrastructure products to peak around 2028 and warn that some hardware companies' price-to-sales ratios are already near internet-bubble levels (their institutional judgment); and Oracle's RPO single-quarter increment sliding from 317 billion to 29 billion also signals that the order-driven narrative is switching toward a delivery-driven one.

The question this framework always answers is "where the money is most likely to go next," and the premise for every candidate above to hold is that capex, the master tap, stays open.

How to Track It: Six Public Signals

The rotation is not finished. Every judgment above maps to a public signal you can watch yourself:

The Edges of the Evidence

This piece's factual base comes from one multi-source deep search and an independent cross-check process: 6 search angles, 30 sources, 149 extracted factual claims; of those, 47 load-bearing claims (the price moves, capex, orders, valuations and every key number that made it into the body) were cross-checked across 6 independent live searches, with about half confirmed as written and the rest adopted after correcting the basis or the timing (for example, the Nasdaq's 4% single-day drop happened on Thursday, 6/4, SK Hynix's year-to-date gain is stated as "more than 200%," and the BofA fund-manager survey ran in 11/2025), while another 5 claims that could not be verified or conflicted with primary data (such as "more than 12,000 articles on the AI bubble in month-11," the rumored figure of Alibaba raising capex to 480 billion yuan, and an order-of-magnitude error in Meta's single-quarter free cash flow) were cut out entirely.

Inherent limits: all market figures are point-in-time snapshots; annual gains carry a basis difference between "price return" and "total return including dividends" (flagged where possible in the text); and the forecasts and qualitative calls from TrendForce, Epoch AI, Morningstar, the various brokerages, and Burry, Huang, Wu and others are each their own, which this piece relays attributively. Market conditions swung sharply in early 6/2026, so the shelf life of the "current" figures here is measured in weeks. The stage read and the next-baton reasoning are this piece's research judgment, and it does not constitute investment advice.

Main Sources

• Motley Fool: top two S&P gainers of 2023 (2024-01-02) · CNBC: top five S&P gainers of 2024 (2025-01-13)
• Morningstar: three-year AI-rally chart retrospective (2026-05)
• TrendForce: 1Q26 DRAM industry revenue and contract prices (2026-06-01) · 2Q26 memory price forecast (2026-03-31)
• Bloomberg: Samsung crosses a trillion in value (2026-05-06) · CNBC: Micron crosses a trillion (2026-05-26) · US News: SK Hynix crosses a trillion (2026-05-27)
• Tom's Hardware: the four hyperscalers' 725 billion-dollar 2026 spending plans (2026-04-30) · Microsoft attributes 25 billion dollars to memory price hikes · CNBC: 2027 AI spending may top 1 trillion (2026-04-30)
• Utility Dive: GE Vernova gas-turbine backlog 100 GW (2026-04-23) · GE Vernova 2026 Q1 results · Vistra x Meta nuclear PPA announcement (2026-01-09)
• Oracle FY2026 Q4 results: RPO 638 billion dollars (2026-06-10) · Oracle FY2026 Q3 results: RPO 553 billion
• Palantir 2026 Q1 results (SEC, 2026-05-04) · TechCrunch: Anthropic annualized revenue 30 billion and 3.5 GW TPU expansion (2026-04-07) · Epoch AI: Anthropic vs OpenAI revenue-growth estimate (2026-02-19) · Alphabet investor materials: token volume and Cloud orders (2026-06-03)
• CNBC: the DeepSeek moment, NVIDIA's largest single-day market-value loss in history (2025-01-27) · 21st Century Business Herald: Cambricon overtakes Moutai (2025-08-27) · Sina Finance: 2025 A-share AI compute roundup (2025-12-31)
• CNBC: H200 cleared for China (2026-01-14) · VOA: BIS blocks the third-country subsidiary channel (2026-06-01) · TechCrunch: AI diffusion rule rescinded (2025-05-13)
• Fortune: BofA fund-manager survey and the bubble debate (2025-11-18) · CNBC: NVIDIA's seven-page memo rebutting Burry (2025-11-25) · Fortune: valuation comparison and concentration (2026-01-04)
• CNBC: Cisco reclaims its 2000 peak after 25 years (2025-12-10) · CNBC: the 6/4 Nasdaq drop (2026-06-04) · CNN: the 6/9 AI sell-off and SpaceX IPO backdrop (2026-06-09) · CNBC: GTC 2026 Jensen Huang keynote (2026-03-16)