The 2026 Semiconductor Rally — Everyone But NVIDIA
 |
2026 semiconductor rally NVIDIA earnings AI chip stocks |
On Wednesday, NVIDIA reported the single largest quarter any semiconductor company has ever produced. Revenue of $81.6 billion, up 85% from a year ago. Data center revenue of $75.2 billion, up 92%. Guidance for the next quarter of $91 billion. Every headline number beat Wall Street's consensus, some by wide margins.
The stock fell.
That sentence is the entire story of the 2026 semiconductor trade compressed into four words. The best result in the history of the industry was met by a market that had already priced it — and then the rally moved past the company that produced it. While NVIDIA shares drifted lower after the print, the broad semiconductor benchmarks had been on a run with few precedents in market history. The iShares Semiconductor ETF closed higher for 17 consecutive sessions into mid-May — the longest uninterrupted winning streak for any U.S. semiconductor benchmark since at least 1993, beating the prior record of 15 straight up-days set in 2014. That 17-day run alone delivered more than 42%. The Philadelphia Semiconductor Index (SOX) is up over 65% for the year. AMD, Broadcom, Micron, the equipment makers, and a cluster of GPU-rental companies most retail investors have never heard of were the ones doing the soaring.
This is the line worth reading more than once: in 2026, NVIDIA now screens as the value stock of the semiconductor sector. It trades at roughly 25 times forward earnings. AMD trades at more than double that — recent estimates run from the high-40s into the high-50s. By this one measure, the company collecting the largest tax on the AI buildout appears priced more conservatively than the companies hoping to take a share of that tax away from it.
Two days ago I wrote about the tax NVIDIA collects on everyone else's AI story. This post is about where that tax flows next — and why a record quarter from the king lifted nearly every chip stock except the king itself.
 |
NVIDIA stock priced for perfection earnings expectations |
The Reframe — A Beat Is Not a Catalyst When the Beat Is Expected
The most common mistake retail investors make around earnings is assuming that good results move a stock up and bad results move it down. The relationship is real over years. Over a single earnings reaction, it frequently inverts, and the inversion has a clean mechanical explanation.
A stock price is not a measure of how a company performed last quarter. It is a measure of every expectation the market holds about every quarter still to come. By the time NVIDIA reported $81.6 billion on Wednesday, the consensus had already moved to expect a number very close to that. Analysts had spent three months revising estimates upward. The options market had priced an implied move. Funds had positioned ahead of the print. When the actual number landed in line with — or even slightly above — those elevated expectations, there was no new information left to push the price higher. The good news was already in the price.
This is what people mean, imprecisely, when they say a stock is "priced for perfection." NVIDIA delivered something close to perfection and the stock did nothing, because perfection was the baseline assumption, not the surprise.
The same mechanism, running in the opposite direction, explains why the rest of the sector rallied. Expectations for AMD, for the memory makers, for the networking suppliers, and for the GPU-rental companies were not as fully saturated. When NVIDIA's report confirmed that demand for AI compute was not merely holding but accelerating — guidance of $91 billion against a just-reported $81.6 billion implies double-digit sequential growth again — the read-through to every other company exposed to the same demand was strongly positive. NVIDIA's print was a weak catalyst for NVIDIA and a strong catalyst for everyone downstream of NVIDIA.
A record quarter can be fully discounted in the leader and still be fresh news for the supply chain. That is much of why the sector outran the stock.
The Loop — Where $725 Billion Actually Lands
The engine underneath all of this is the hyperscaler capital expenditure cycle, and it is worth restating the scale because the scale is the thesis. The combined 2026 infrastructure commitment from Microsoft, Alphabet, Meta, Amazon, Oracle, and the second tier of cloud builders has crossed $725 billion, of which roughly $500 billion is earmarked specifically for AI compute. Microsoft alone guided $190 billion in capital expenditure for its fiscal year.
That capex lands inside a semiconductor industry that is itself reaching a historic peak. Global chip sales are on track for somewhere between $975 billion (Deloitte's estimate) and $1 trillion (the Semiconductor Industry Association's) in 2026, up from $791.7 billion in 2025. The composition of that growth is the part worth pausing on: high-value AI chips now account for roughly half of total semiconductor revenue while representing less than 0.2% of unit volume. Deloitte's framing makes the concentration concrete — of the roughly 1.05 trillion chips sold in 2025, only around 20 million were destined for generative AI. The industry's dollar growth is concentrated almost entirely in a premium tier, which is the structural reason the rally is concentrated in the handful of names exposed to it.
That spending does not land in a single company. It distributes across a stack, and each layer of the stack is a separate investable category with a separate set of public companies attached to it. NVIDIA captures the largest single share — somewhere around half of the AI-specific portion — but "the largest single share" of $500 billion still leaves $250 billion flowing past NVIDIA to everyone else.
This is the part of the AI trade that the single-stock obsession with NVIDIA tends to obscure. The capex of the hyperscalers is not just NVIDIA's revenue line one quarter delayed. It is also the memory makers' revenue line, the networking suppliers' revenue line, the foundry's revenue line, the equipment makers' revenue line, and increasingly the revenue line of a new category of company that exists only to rent out the GPUs themselves.
When NVIDIA confirmed on Wednesday that the demand at the top of this funnel was still accelerating, it confirmed the demand for every layer beneath it at the same time. That is why one company's earnings call functioned as a sector-wide catalyst.
 |
semiconductor sector layers GPU memory networking equipment |
What the Numbers Actually Say
The figures from this week, taken together, describe a rally with unusually broad participation.
NVIDIA (the trigger).
- Q1 FY2027 revenue: $81.6 billion (versus $78.86 billion expected), up 85% YoY, up 20% sequentially
- Data center revenue: $75.2 billion, up 92% YoY, up 21% sequentially
- Non-GAAP EPS: $1.87 (versus $1.76 expected)
- Net income: $42.96 billion, more than double the $18.8 billion of a year earlier
- Q2 FY2027 guidance: $91 billion ±2%
- China data center (Hopper) shipments: $0, down from $4.6 billion a year ago
- Hyperscale customers were approximately 50% of data center revenue; the other 50% came from AI clouds, enterprise, industrial, and sovereign buyers
The China figure deserves a moment. NVIDIA produced a record quarter with zero data center revenue from what was, two years ago, one of its largest markets. The demand elsewhere was strong enough to absorb a complete write-off of a multi-billion-dollar revenue stream and still grow 85%.
AMD (the closest comparable). AMD reported its own most recent quarter a few weeks ahead of NVIDIA, and the figures describe the same demand seen from the number-two position.
- Total revenue: $10.3 billion, up 38% year-over-year
- Data center segment revenue: $5.8 billion, up 57% year-over-year
- Non-GAAP EPS: $1.37
- Q2 guidance: approximately $11.2 billion (±$300 million), ahead of consensus
- At least 20 brokerages raised price targets after that report
The index (the breadth signal).
- The Philadelphia Semiconductor Index (SOX) is up over 65% year-to-date in 2026
- The iShares Semiconductor ETF posted a 17-session winning streak into mid-May — the longest for any U.S. semiconductor benchmark since at least 1993, up roughly 42% over that run
- SOX set an intraday all-time high on May 14
- Its outperformance versus the S&P 500 reached the widest margin in more than a year
The hidden tell — GPU rental prices. On the call, NVIDIA noted that rental prices for H100-class GPUs continued to climb. This may be the most telling data point for anyone trying to separate real demand from speculative enthusiasm. Rental prices tend to rise when the people actually using the compute cannot get enough of it. The read-through sent the "neocloud" companies — Nebius, CoreWeave, IREN, and others whose entire business is buying GPUs and renting them out — sharply higher. When the rental rate on the underlying asset is rising, the businesses built on renting that asset are the most direct beneficiaries.
The Five Layers of the Rally
The cleanest way to hold this in your head is as a stack, top to bottom, with public companies at every level. A retail investor who understands the stack understands the rally.
Layer one — the accelerators. NVIDIA, AMD. The GPUs and AI chips themselves. The largest single dollar capture, the most concentrated competition, the most visible names.
Layer two — the memory. Micron, SK Hynix, Samsung. Every AI accelerator needs high-bandwidth memory (HBM) stacked alongside it, and HBM has been supply-constrained through the entire buildout. Memory is the layer where a commodity historically prone to brutal price cycles has, for now, become a scarce input with pricing power.
Layer three — the networking and custom silicon. Broadcom, Marvell, Arista. Connecting tens of thousands of GPUs into a single training cluster requires enormous networking capacity, and Broadcom in particular also designs the custom AI chips that hyperscalers use to reduce their NVIDIA dependency — a company positioned to win whether the buildout favors merchant GPUs or in-house silicon.
Layer four — the equipment and materials. ASML, Applied Materials, Lam Research, Corning. The companies that sell the machines that make the chips, and the materials that go into the data centers. ASML's monopoly on EUV lithography means it sits underneath the entire stack regardless of which chip company wins. This layer also contains some of the most extreme individual performers of 2026 — companies whose niche position in the supply chain has produced returns exceeding even NVIDIA's.
Layer five — the renters. Nebius, CoreWeave, IREN. The newest layer. These companies buy GPUs at the top of the stack and rent compute to AI labs that cannot or will not build their own data centers. Their economics rise and fall directly with GPU rental rates — which, as of this week, are still rising.
The investment lesson embedded in the stack is the one this blog returns to repeatedly: the question is rarely which single stock, it is which layer, at what valuation, with what risk. Every layer is exposed to the same demand. Each carries a different competitive structure and a different multiple.
 |
AI chip stock bubble valuation risk 2026 correction |
What the Numbers Do Not Prove
A rally this broad and this fast invites a particular kind of self-congratulation. The honest analysis has to hold the strong demand and the genuine risks in the same frame.
One — the rally has already had a correction inside it. The same SOX index up 65% on the year fell roughly 6.8% from its peak in mid-to-late May, driven not by anything about chips but by rising U.S. Treasury yields. This matters because it shows the sector is now sensitive to the cost of money, not only to AI demand. High-multiple growth stocks fall when yields rise, regardless of how strong the underlying business is. A genuine demand story and a rate-driven drawdown can coexist, and did, within the same month.
Two — some serious observers are drawing the 1999 comparison. Analysts pointing to the dot-com parallel are not cranks. The pattern they are describing is specific: a real technological transformation, real demand, real revenue — wrapped in valuations that priced a decade of perfect execution into eighteen months of stock performance. The internet was real and changed everything. Cisco, the great tax-collector of that buildout, still lost roughly 80% of its value when the cycle turned and has never reclaimed its 2000 high. Both things were true. The technology was real and the stock was a catastrophic buy at the peak.
Three — the NVDA reaction itself is a signal. A market that does not reward a record-beating quarter is, arguably, a market that has run ahead of its fundamentals at the leader. That is not yet a top — saturated expectations can persist for a long time — but it tends to be a condition that precedes one. When a stock stops responding to good news, it often means the marginal buyer has already bought.
Four — concentration risk runs through the entire stack. NVIDIA's top two customers are roughly 36-39% of its revenue. The neocloud companies are even more concentrated, often depending on a handful of AI labs. If a single large buyer — a hyperscaler, a frontier lab — slows its spending, the effect propagates down every layer of the stack at once. Broad participation on the way up means broad participation on the way down.
None of these four observations argues that the demand is fake — the demand itself is not seriously in dispute. They point to something else: price and demand are different variables, and the gap between them is where retail investors have, on the historical record, most often been hurt.
 |
semiconductor index ETF diversification core satellite portfolio |
How a Retail Investor Might Hold This
Three observations, in order of importance.
One — the index is the honest expression of "I believe in the buildout." An investor who is convinced the AI infrastructure cycle is real, but who cannot confidently name which layer or which company wins, is describing a thesis that a semiconductor index fund (the SOX-tracking ETFs, SOXX and SMH being the common ones) expresses more accurately than any single stock. The buildout has many beneficiaries. Owning the basket captures the thesis without requiring a correct guess about which name leads. That is not a recommendation of any specific fund — it is an observation that the structure of a broad thesis matches the structure of a broad instrument.
Two — chasing the extreme individual performers is the most dangerous version of this trade. The companies that have returned 300%+ this year are the ones most likely to appear in a "stocks that obliterated NVIDIA" headline, and most likely to be bought by retail investors at the exact moment the easy gains are behind them. A stock that has tripled has, by definition, already rewarded the people who held it through the uncertainty. Buying after the triple is buying the certainty at the price of the upside. The position sizing rule from the math of avoiding ruin applies with full force: a single chip name, however compelling, belongs in the Satellite, sized so that a 50% drawdown changes nothing about the household.
Three — rate sensitivity is now part of the position. This sector falls when Treasury yields rise, as it did this month. An investor holding semiconductors is, whether they intend to be or not, also holding a position on interest rates. That does not argue against the holding. It argues for understanding that the next drawdown may have nothing to do with AI demand and everything to do with the bond market — and for not panic-selling a demand thesis on a rate-driven move.
 |
long term investing discipline semiconductor cycle direction |
Lines to Watch From Here
- GPU rental rates. Perhaps the cleanest real-demand signal available. As long as H100-class rental prices keep rising, the demand at the top of the funnel looks real. A flattening or decline would likely be the first crack — visible before it shows up in any earnings report.
- The next NVIDIA print versus the $91 billion guide. Guidance of $91 billion sets the new baseline expectation. Meeting it does nothing; the only outcomes that move the stock are a clear beat or a miss.
- AMD's data center trajectory. AMD crossing into the low-teens in quarterly data center revenue would mark the first real dent in NVIDIA's accelerator monopoly.
- Memory pricing. HBM scarcity is what gives Micron and the memory makers their current pricing power. The cycle has always turned eventually. Watch for the first quarter of softening HBM pricing.
- Treasury yields. The sector's near-term price is now partly a bond-market story. Rising yields compress the multiple regardless of demand.
- Neocloud customer disclosures. The renters are the newest and least-tested layer. Watch how concentrated their revenue is among a few AI labs — that concentration is the layer's central risk.
A Closing Observation
The math, as always, gets the larger room. The 2026 semiconductor rally is built on a real foundation — demand for AI compute that is, by the most reliable signal available, still accelerating. That is not in serious dispute. What is in dispute, and what no quarterly print can resolve, is whether the prices the market is paying for that demand leave room for a satisfactory return from here.
A record quarter that the leader's stock could not rise on is a precise illustration of the gap. The business performed. The price had already performed for it. Those are two different events, and the discipline of holding them apart — celebrating the demand without confusing it for a guarantee about the price — is most of what separates the investor who survives a cycle from the one who learns its lesson at the peak.
NVIDIA lit the match this week. The whole sector caught. The question worth holding, as the rally broadens, is the oldest one: not whether the fire is real, but how much you are paying to stand near it.
Reading the rally for what it is — real demand at uncertain prices — is the work this week.
Reference figures (verified week of May 20, 2026): NVIDIA Q1 FY2027 revenue $81.6 billion vs $78.86 billion expected (+85% YoY, +20% sequential); data center revenue $75.2 billion (+92% YoY, +21% sequential); non-GAAP EPS $1.87 vs $1.76 expected; net income $42.96 billion (more than double the $18.8 billion prior year); Q2 FY2027 guidance $91.0 billion ±2%; China data center shipments $0 vs $4.6 billion prior year; hyperscale ~50% of data center revenue. AMD most recent quarter (reported ahead of NVIDIA): total revenue $10.3 billion (+38% YoY), data center segment $5.8 billion (+57% YoY), non-GAAP EPS $1.37, Q2 guidance ~$11.2 billion (±$300 million). iShares Semiconductor ETF 17-session winning streak into mid-May — longest for any U.S. semiconductor benchmark since at least 1993 (prior record 15 days, 2014), ~42% over the run. PHLX Semiconductor Index (SOX) +65% YTD 2026, intraday record May 14, ~6.8% correction from peak on rising Treasury yields. NVIDIA forward P/E ~25x vs AMD high-40s to high-50s (recent estimates vary by source). Hyperscaler 2026 capex ~$725 billion (~$500 billion AI-specific). Global semiconductor sales on track for $975 billion (Deloitte) to ~$1 trillion (SIA) in 2026, up from $791.7 billion in 2025; AI chips ~half of revenue at <0.2% of unit volume (of ~1.05 trillion chips sold in 2025, ~20 million for generative AI per Deloitte). GPU rental (H100-class) prices reported rising; neocloud names (Nebius, CoreWeave, IREN) rallied on the read-through. Sources: NVIDIA Q1 FY2027 earnings release, CNBC, Kiplinger, StockTitan, TradingView, Yahoo Finance, Investing.com SOX data, Deloitte 2026 Semiconductor Outlook, Semiconductor Industry Association, Opening Bell Daily. This post is observation, not investment advice.
Related Posts:
NVIDIA's $78 Billion Quarter — The Tax on Everyone Else's AI Story
Microsoft's $190 Billion Capex — The Bet Nobody Is Sizing
The Year Your Portfolio Is Down 30% — A Behavior Guide
