NVIDIA × TSMC
How AI demand
becomes silicon.
NVIDIA measures the appetite for AI compute. TSMC measures what that appetite looks like when it reaches the factory floor. Read together over multiple quarters, the connection is clear — but the timing is not mechanical.
NVIDIA is the demand amplifier. TSMC is the manufacturing enabler. AI demand travels upstream into leading-edge manufacturing, but not one-for-one and not on a fixed quarterly lag.
Total revenue, up 106% year over year.
Hyperscale plus AI cloud, industrial and enterprise demand.
Revenue mix from High Performance Computing.
Wafer revenue from 7nm and below.
Demand
NVIDIA says the AI buildout is still accelerating.
NVIDIA added roughly $15 billion of revenue sequentially in Q2 FY27 and reached $96.2 billion for the quarter. Data Center accounted for $89.0 billion: $48.7 billion from hyperscale customers and $40.3 billion from AI clouds, industrial and enterprise customers.
The important signal is not just the size of the quarter. NVIDIA also began production shipments of Vera Rubin in early August while Blackwell demand remained strong. The company is scaling one architecture while beginning the next ramp rather than waiting for demand to normalize.
Manufacturing
TSMC sees the same cycle as a change in mix.
TSMC's 2Q26 revenue reached US$40.2 billion. HPC represented 66% of revenue, up from 61% in the prior quarter, while HPC revenue itself grew 20% sequentially. Smartphone revenue fell 4% quarter over quarter.
The technology mix moved in the same direction. 2nm contributed 3% of wafer revenue, 3nm 30%, 5nm 33% and 7nm 11%. Together, 7nm and below represented 77% of wafer revenue.
That is the upstream version of NVIDIA's story: growth is not spreading evenly across semiconductor manufacturing. It is concentrating where high-performance compute needs the most advanced process technology.
Six-quarter read
The relationship is real. The lag is not simple.
Looking across six quarters changes the interpretation. NVIDIA's Data Center business climbs almost continuously. TSMC's HPC share does not. Seasonal smartphone strength can dilute the percentage even when HPC dollars remain high, while foundry capex moves in large, uneven steps.
From $39.1B in Q1 FY26 to $89.0B in Q2 FY27.
From roughly $15.1B in 1Q25 to $26.5B in 2Q26.
From NT$331B in 1Q25 to NT$496B in 2Q26.
| Signal | 1Q25 / Q1 FY26 | 2Q25 / Q2 FY26 | 3Q25 / Q3 FY26 | 4Q25 / Q4 FY26 | 1Q26 / Q1 FY27 | 2Q26 / Q2 FY27 |
|---|---|---|---|---|---|---|
| NVIDIA Data Center · $B | 39.1 | 41.1 | 51.2 | 62.3 | 75.2 | 89.0 |
| TSMC HPC mix | 59% | 60% | 57% | 55% | 61% | 66% |
| Est. TSMC HPC revenue · $B | 15.1 | 18.0 | 18.9 | 18.6 | 21.9 | 26.5 |
| TSMC advanced nodes | 73% | 74% | 74% | 77% | 74% | 77% |
| TSMC capex · NT$B | 331 | 297 | 287 | 357 | 351 | 496 |
TSMC HPC revenue is an approximation: quarterly U.S.-dollar revenue multiplied by the disclosed, rounded HPC revenue mix. Quarter labels pair TSMC calendar quarters with the closest NVIDIA fiscal-quarter reporting period; they are not identical calendar windows.
HPC share fell while NVIDIA accelerated.
In 3Q25 and 4Q25, NVIDIA Data Center revenue jumped from $51.2B to $62.3B while TSMC's HPC mix moved from 57% to 55%. That does not mean AI demand weakened. Smartphone and other platforms were also moving through a seasonal cycle.
TSMC HPC revenue stayed high, then reaccelerated.
Estimated HPC revenue was roughly $18.9B in 3Q25 and $18.6B in 4Q25, then rose to about $21.9B in 1Q26 and $26.5B in 2Q26. The share percentage alone hides that progression.
Capex does not follow demand quarter by quarter.
TSMC capex fell through much of 2025, then stepped up to NT$357B in 4Q25 and NT$496B in 2Q26. Manufacturing capacity is committed in projects and ramps, not in a smooth line beside NVIDIA revenue.
There is no stable one-quarter lead/lag in this short history. What we can see is a transmission mechanism: AI demand appears first in system revenue, then shows up through foundry dollars, technology mix and large capacity commitments at different speeds.
Cloud, enterprise and sovereign customers commit more capital to compute.
Blackwell scales while Vera Rubin starts its production ramp.
More leading-edge capacity is required to turn system demand into physical product.
HPC dollars, advanced-node mix and capital intensity express the cycle upstream.
Capital
The strongest connection is capacity, not revenue attribution.
NVIDIA explicitly says it is investing in critical supply and capacity while ramping new systems. TSMC spent NT$496 billion on capital expenditure in 2Q26, up from NT$351 billion in the prior quarter.
This is where the two presentations connect most clearly. NVIDIA's product cadence pulls manufacturing requirements forward; TSMC has to commit physical capital before all downstream demand is fully realized.
TSMC's margins show that the burden is not destroying economics so far. Gross margin reached 67.7% in 2Q26, helped by higher capacity utilization and cost improvement, even as overseas fabs diluted margin.
The important difference
TSMC is not simply a proxy for NVIDIA.
TSMC's HPC category includes far more than one customer or one type of accelerator. It can benefit from GPUs, custom AI accelerators, CPUs and networking silicon at the same time. That makes the relationship asymmetric.
If NVIDIA continues gaining AI infrastructure share, both companies can benefit. But if custom accelerators take share from NVIDIA while remaining on TSMC's advanced processes, NVIDIA could weaken without producing the same weakness at TSMC.
These presentations do not support a precise estimate of how much TSMC revenue comes from NVIDIA. The useful comparison is the direction of demand, mix, capacity and outlook — not an invented customer percentage.
Q3 FY27 revenue outlook, ±2%. The guidance assumes no Data Center compute revenue from China.
3Q26 revenue guidance, with gross margin expected between 65% and 67%.
NVIDIA demand growth and TSMC HPC dollars should be read together; percentage mix by itself is not enough.
Watch list
What could break the connection?
- 01
NVIDIA Data Center growth slows materially even as new manufacturing capacity becomes available.
- 02
TSMC's estimated HPC dollars and advanced-node utilization weaken while headline AI infrastructure spending remains high.
- 03
Custom accelerators capture enough share to weaken NVIDIA while keeping TSMC's fabs full.
- 04
Power, financing or data-center construction becomes a tighter constraint than semiconductor supply.
Bottom line
The connection is visible. It is not mechanical.
NVIDIA gives the fastest read on AI compute demand. TSMC shows how that demand spreads into manufacturing dollars, leading-edge process mix and capacity investment. The six-quarter history argues against a simple one-quarter lag — and for reading the two companies as different layers of the same system.