Weight by revenue — which means letting Nvidia, TSMC, and the memory makers dominate the math — and the industry’s trailing-twelve-month gross margin is running around 59%, with net margin near 42%, according to CSIMarket’s Q2 2026 industry data. Take the median public semiconductor company instead, and gross margin drops to roughly 43%, with operating margin around 19–20%. Same industry, two very different answers, and most write-ups on this topic only give you one of them.
| Measure | Gross margin | Operating margin | Net margin |
|---|---|---|---|
| Revenue-weighted industry average (TTM, Q2 2026, CSIMarket) | ~59% | ~46% | ~42% |
| Median public semiconductor company (multi-year) | ~43% | ~19–20% | — |
Revenue-weighted figures skew heavily toward a handful of large, AI-exposed companies. The median tells you what a typical chipmaker actually earns.
Margin by segment: this is where the real variation lives
“Semiconductors” covers four genuinely different businesses. Lumping them into one number erases the part that actually matters.
| Segment | Example companies | Typical gross margin, 2026 | What drives it |
|---|---|---|---|
| Fabless logic/AI | Nvidia, AMD | 72–75% | No fab to depreciate; owns pricing on scarce, high-demand chips |
| Foundry | TSMC | 56–60% | Owns the fabs, sets prices for everyone else, but carries the depreciation |
| Memory | SK Hynix, Micron, Samsung | 56–76% (2026 cyclical peak) | Currently near all-time highs on AI-driven “memflation”; can go negative in a downturn |
| Analog / legacy / industrial | ON Semiconductor, Texas Instruments | 40–47% | Slower-moving, less AI exposure, feels demand softness first |
Nvidia’s own gross margin illustrates the fabless case well — it ran around 64–65% back in fiscal 2022 and has climbed to 72–75% through 2026, per the company’s quarterly filings. TSMC, carrying the fab depreciation Nvidia doesn’t, has moved from 52% in 2021 to the high 50s by 2023 and stayed there since, per its reported financials. Memory is the outlier worth dwelling on, because it shows how fast this industry’s economics can flip.
The memory swing: from -67% to +76% in three years
Nothing in this industry illustrates the cyclicality better than SK Hynix’s operating margin over the last three years, as reported in the company’s own quarterly disclosures.
| Period | SK Hynix operating margin | Context |
|---|---|---|
| Q1 2023 | -67% | Memory glut, prices collapsed, oversupply from pandemic-era overbuilding |
| Q4 2025 | ~58% | AI-driven memory demand recovery underway |
| Q1 2026 | 72% | HBM allocated to AI servers tightens general DRAM/NAND supply |
| Q2 2026 | 76% | Revenue up 257% year-over-year; DRAM and NAND prices still climbing |
Micron and Samsung’s memory division followed a similar arc. Micron’s own fiscal Q1 2026 results put its gross margin at 56%, with company guidance calling for 67% the following quarter, and industry press covering Samsung’s Q4 2025 results reported the memory division’s gross margin in the 63–67% range — briefly overtaking TSMC for the first time since 2018. That’s not incremental improvement, that’s a business going from underwater to software-company-level profitability in twelve months, purely on pricing.
What actually moves the margin number (beyond “which company”)
The original framing of this question — average margin, one number — misses most of what determines whether a specific chipmaker is profitable. Five factors matter more than the industry label does.
| Factor | Why it moves margin | Real-world example |
|---|---|---|
| Business model | Fabless companies skip fab depreciation; IDMs and foundries carry it | Nvidia (fabless, ~73%) vs. Intel’s foundry arm (IDM, thin/negative margins) |
| Product niche | Leading-edge AI/logic and scarce memory command premium pricing; legacy/analog chips compete on cost | AI GPUs at 70%+ gross margin vs. legacy microcontrollers priced for cost-sensitive industrial buyers |
| Where the fab is built | Boston Consulting Group’s 2023 fab-economics analysis found the 10-year total cost of owning and operating a fab is roughly 30% higher in the US than in Taiwan, South Korea, or Singapore, and 37–50% higher than in China — with the Semiconductor Industry Association noting that 40–70% of that gap is tied directly to the size of government subsidies each country offers | US fabs need CHIPS Act-scale incentives just to close the gap with Asian competitors |
| Cycle position / utilization | Fabs and memory lines carry enormous fixed costs; running below full utilization crushes margin fast, running near capacity during a supply squeeze does the opposite | Memory went from a 2023 glut to a 2026 shortage without adding much new capacity — pricing did all the work |
| Customer concentration | Selling mostly to a handful of hyperscalers gives them leverage to demand price cuts on custom silicon, compressing margin even for otherwise strong designs | Custom AI ASIC deals for large cloud providers typically carry tighter margins than merchant GPU sales |
Geography deserves the extra sentence, since it’s the factor most other coverage skips. The US isn’t building fabs because it’s the cheapest place to build one — by BCG and SIA’s own numbers, it’s the most expensive major option outside a handful of niche cases. The entire premise of the CHIPS Act was to subsidize away a structural cost disadvantage, not remove one that didn’t exist. Any margin discussion that treats “built in America” as a cost advantage has the direction backwards.
History: three violent swings in five years
Margins in this industry don’t drift, they lurch.
| Period | What happened |
|---|---|
| 2020–2021 | Pandemic chip shortage. Everyone from automakers to console makers begged for allocation; chipmakers had pricing power they hadn’t seen in years. |
| 2022–2023 | The hangover. Consumer electronics demand cratered, memory prices collapsed, oversupply piled up. McKinsey’s analysis of roughly 300–400 semiconductor companies found that outside the very top performers, total economic profit fell from $33.5 billion in 2022 to $10.3 billion in 2023, and average economic profit per company dropped 88% between 2021 and 2024. |
| 2024–2026 | The AI buildout reshaped the map, unevenly. Fabless AI leaders held margins near 75%. Memory makers went from fire-sale pricing to “memflation” — Gartner’s April and August 2026 forecasts put the 2026 rise in DRAM revenue at roughly 247% and NAND revenue at roughly 372%, as AI data centers absorbed supply that used to go to PCs and phones. |
Whenever someone quotes a single “semiconductor margin” figure with no date attached, ask which year they mean — 2023 and 2026 aren’t in the same universe.
Where margins go in 2027–2028
| Year | Outlook | Key driver / risk |
|---|---|---|
| 2027 | Growth continues; widely called the “peak constraint” year | Gartner projects roughly $1.9 trillion in global semiconductor revenue, memory alone crossing $1 trillion; demand for advanced logic and HBM still outrunning new capacity |
| 2028 | Growth likely continues, but margin dispersion narrows | New fab and memory capacity funded in 2025–2026 starts reaching mature yield, loosening supply; Deloitte’s 2026 outlook flags that AI data centers will need roughly 92 additional gigawatts of power by 2027, a constraint that could delay chip demand rather than eliminate it; ROI questions on AI infrastructure spend could push out some chip orders if monetization lags expectations |
None of this points to a crash. It points to the unusually wide gap between AI/memory-exposed companies and everyone else narrowing somewhat as new supply comes online, while the industry as a whole keeps growing in dollar terms.
