Semiconductor Stocks Rose 86% in 2026, Nvidia 28%: Where the Money Went

The PHLX Semiconductor index has gained 86% in 2026, while Nvidia, its largest member, is up 28%. That gap opened in a year when Nvidia reported $96.2 billion of revenue in a single quarter. Semiconductor stocks did not rally on AI in general. They rallied on the companies whose AI earnings surprised, and in 2026 the surprise sat in memory.

Short answer. Nvidia stock underperformed the chip sector in 2026 because its growth was already expected. Investors paid instead for earnings that changed shape: memory makers whose margins doubled, plus equipment and turnaround names. Nvidia grew revenue 106% in its latest quarter, yet trades near 17 times forward earnings, close to a 10-year low.

Semiconductor stocks in 2026: the scoreboard

Start with what the index actually did, because the year-to-date number hides the path. The SOX hit an all-time high of 14,655 on 22 June 2026, after gaining more than 100% since January. It then fell as much as 29% by late July, according to Disruption Banking's account of the index.

By 7 October the index closed at 13,066.15, still about 11% below that June peak. On 8 October it was trading at 12,624.75, down 3.38% on the day, per Nasdaq's index page. A single session can move the year-to-date figure by several points.

Inside that, the leaders are not the names most people associate with AI. Bloomberg credited the gains to Micron, Marvell and Intel, whose shares are each up more than 200% this year. Nvidia, the stock that defines the AI trade, added 28%.

Nvidia itself was down 11% for the year on 30 March. The index posted a 16-day winning streak in April, the longest in its 32-year history, per Deutsche Bank figures reported by Yahoo Finance. Then came the June peak and the July slide.

That matters for how you read any "up X% this year" claim. Pick a different start date and the ranking changes. I use 6 October because that is the most recent date for which a Tier 2 source published both numbers side by side.

Several outlets quoted lower figures in September, including one that put the index "over 60%" for the year. That figure is weeks older than Bloomberg's and comes from a less established outlet. When two numbers differ this much, take the newer one from the better source and state the date. That is what this post does throughout.

Why Nvidia stock underperformed a record quarter

On 26 August 2026, Nvidia reported second-quarter fiscal 2027 revenue of $96.2 billion, up 106% from a year earlier, per the company's results release. Data Center revenue alone was $89.0 billion, up 117%. GAAP and non-GAAP gross margins were both 75.0%.

By any operating measure that is an extraordinary quarter. The stock response over the full year was modest anyway. The reason is not that investors doubt Nvidia. It is that they had already paid for this outcome.

A share price moves on the gap between results and what the market already assumed. Nvidia's demand has been forecast quarter by quarter, in public, by its own guidance. Beating that guidance by a few billion dollars is a small surprise relative to a market value that Bloomberg put just under $5.8 trillion.

Bloomberg reported Nvidia trading at about 17 times expected earnings over the next 12 months, close to its lowest level in 10 years and below the S&P 500 at 19 times. Larry Tentarelli of Blue Chip Daily put it plainly: Nvidia is "almost a victim of its own success," and "it just takes so much to move the needle," he said.

I think that is the right reading, and it is not a bearish one. A company can grow earnings faster than its share price for a long time when the multiple is compressing. That is what happened to Nvidia in 2026. The business outran the stock.

The margin line pointed the wrong way

There is a second, more specific signal. Nvidia guided third-quarter gross margin to 74.0%, plus or minus 50 basis points, against the 75.0% it had just reported. That is a one-point step down at a company whose pricing power is the core of the bull case.

One explanation circulating in coverage is memory cost. If high-bandwidth memory, the stacked DRAM packaged next to each accelerator, gets more expensive, some of that cost lands in Nvidia's cost of goods. I cannot prove that link from the filings alone, and Nvidia's release does not state it. It is consistent, though, with where the margin went next.

Memory stocks and HBM: where the earnings surprise landed

Micron's fiscal fourth quarter, which ended on 3 September 2026, is the clearest single data point of the year. Revenue was $54.23 billion, against $11.32 billion a year earlier, according to its results filed with the SEC. That is growth of about 379% in one year.

The margin change is the bigger story. GAAP gross margin was 86.8%, up from 44.7% in the same quarter of 2025. A memory company now earns a higher gross margin than the accelerator vendor it supplies.

Micron reports by business unit. Its Core Data Center unit brought in $18.0 billion in the fourth quarter against $1.58 billion a year earlier, roughly 11.4 times as much. The Cloud Memory unit grew from $4.54 billion to $16.28 billion. Together those two units were about 63% of the quarter's revenue.

The release does not break out HBM as its own line, so any precise "HBM revenue" figure you see quoted for Micron did not come from this filing. What the filing does show is the data centre mix, and that mix moved by an order of magnitude.

Why memory re-rated and Nvidia did not

Memory is historically a cyclical, commodity business, priced for the bottom of the cycle. When that kind of business doubles its margin, the market has to throw out its old model of the company. That is a re-rating in the literal sense. Nvidia's model did not need throwing out. It needed a spreadsheet update.

My position: the 2026 chip trade was a bet on who captures the next marginal dollar of AI spending, not on who earns the most of it today. On this year's evidence, memory captured more of the marginal dollar than anyone expected, and that is what the share prices paid for.

AI chip stocks in custom silicon and networking: a split verdict

A common story this year is that custom accelerators and networking carried the index. The evidence supports half of that.

Broadcom grew fast and was not rewarded

Broadcom reported fiscal third-quarter 2026 revenue of $29.6 billion, with AI semiconductor revenue of $16.7 billion, up 221% year on year. Those figures come from Pulse 2.0's coverage of the results. CEO Hock Tan said demand for custom AI accelerators and networking "continues to be very strong."

That growth rate is roughly double Nvidia's. Yet Broadcom is absent from Bloomberg's list of leaders. Its earlier guidance for the third quarter disappointed, and weaker Broadcom AI guidance was one of the triggers Disruption Banking cited for the July sell-off. The market punished a shortfall against expectations, even inside a 221% growth year.

Marvell re-rated on modest growth

Marvell is the mirror image. Its most recent data centre revenue figure reported by Disruption Banking was a record $2.17 billion, up 46% year on year. Its shares are up more than 200% this year, per Bloomberg.

A 46% grower outrunning a 221% grower by a wide margin tells you the market was not paying for the growth rate. It was paying for change in the outlook. Marvell started the year priced as a laggard in custom silicon and finished it priced as a beneficiary.

Intel is the move filings cannot explain

Intel is the third name Bloomberg lists as up more than 200%. Intel does not report an AI revenue line comparable to Micron's data centre units or Broadcom's AI semiconductor figure. Whatever drove its re-rating, it cannot be traced to an AI revenue number in a filing, and I will not invent one.

How the SOX index structure shaped the 2026 result

Part of Nvidia's lag is arithmetic, not sentiment. Since April 2024 the SOX caps its three largest members at 12%, 10% and 8% at each quarterly rebalance, and every other member at 4%. The index has 30 constituents.

In the Nasdaq weighting file dated 30 June 2026, Nvidia carried 10.34% of the index. In a market-cap-weighted index of the same 30 names, a company worth nearly $5.8 trillion would dominate. Here it is roughly one tenth.

When weight is capped, a 200% move in a 5% member contributes about as much to the index as a 100% move in a 10% member. The SOX return is therefore closer to "how did the typical big chip stock do" than "how did the chip industry's market value change".

That is a design choice, and a sensible one for an index meant to represent a sector. It does mean that "Nvidia trailed the SOX" partly compares one stock against a basket built to limit that stock's influence. Both facts are true at once.

Matching each 2026 move to an AI revenue line in the filings

Here is the scorecard that ties each move back to the most recent reported number. Where no AI line exists in the filing, the table says so rather than filling the cell.

Share moves are from Bloomberg, 6 October 2026, which gives one group figure for Micron, Marvell and Intel. Revenue lines are from each company's latest results: Nvidia and Micron from their own filings, Broadcom and Marvell from secondary coverage of their results. Fiscal quarters do not align across companies.

No single variable sorts these rows. Revenue size does not, growth rate does not, and AI exposure does not. The best fit is the change in what investors expected at the start of the year. Micron and Marvell began 2026 priced for ordinary outcomes and delivered unusual ones. Nvidia began the year priced for an unusual outcome and delivered it.

That is a weaker claim than "memory is the new AI trade," and I think it is the more useful one. It predicts that the next re-rating goes to whichever supplier's margins are about to change, not to whoever is currently largest. The indicators that would signal an AI chip sell-off are the other half of that picture.

Where this argument is weakest

Three problems, stated plainly.

Year-to-date is a fragile unit

Every number in this post depends on 1 January as the starting line. The index gained more than 100% to June, gave back up to 29% by late July, and fell 3.38% in a single session on 8 October. Nvidia was down 11% at the end of March. A post written in late June would have told a far more extreme story about the same companies. Treat the gap as dated to early October and expect it to move.

The sources disagree, and some are secondary

Reported index gains for this year range from "over 60%" in a September piece to 86% from Bloomberg two weeks later. The Broadcom and Marvell figures here come from secondary coverage of their results rather than the companies' own releases, which I could not open during research. Bloomberg gives one group figure for three different stocks, not a per-stock number. A reader with a terminal could tighten every one of these, and should.

The memory case may be the cycle talking

The strongest objection to this post is that memory has done this before. Margins spike when supply is short, capacity gets built, and prices fall. Micron's own filing shows net capital expenditure of $27.37 billion for fiscal 2026. Supply responses on that scale are historically how memory upcycles end. If that happens, the 2026 re-rating will look like a peak multiple on peak earnings, and Nvidia's patient 28% will look like the sober call. I do not think that is the most likely outcome within a year, but it is a genuine possibility and not a footnote.

What the AI chip stock rotation means for your compute costs

If you run a software or services business, you do not own these stocks for a living. You do rent their output, and the 2026 margin shift has a direct line to your invoices.

When memory margins go from roughly 45% to 87%, someone downstream is paying that difference. Accelerator vendors absorb part of it, cloud providers absorb part, and the rest arrives in the price of compute. This is one reason that falling per-token prices can coexist with rising AI bills. The unit gets cheaper while the bundle you consume grows and its hardware inputs get dearer.

It also reframes the capex debate. The hyperscalers' $725 billion build with no published payback math is, from the supplier side, revenue. In 2026 a larger share of that revenue seems to have landed with memory makers than anyone modelled in January. Where the money goes inside the build matters as much as how big the build is.

Finally, it is a reminder about reported earnings. Chips depreciate, and the useful life a buyer assumes for them decides how much of this year's spend shows up as cost. The mechanics are covered in how GPU depreciation schedules shape AI earnings. If memory is now a larger share of each server's cost, the depreciation question gets more expensive, not less.

I would not read any of this as a reason to change vendors. I would read it as a reason to stop assuming your compute costs will fall on their own. The suppliers with pricing power in 2026 were not the ones most buyers were watching.

Frequently asked questions

Where to start

Pull up the last two quarters of invoices from your largest AI or cloud vendor. Note whether the price per unit you consume has fallen, held or risen, and write that number down with the date. That is your own small index, and it will tell you whether the 2026 supplier shift has reached you yet.

Then, when Micron and Nvidia next report, read the gross margin line before the revenue line. If Micron's margin starts falling while its capital spending keeps rising, the memory cycle is turning, and compute prices are likely to follow.

Related reading