TSMC's Q3 Revenue Rose 50% to $46.7B. What It Says About AI Demand in 2027

The 50% in TSMC's Q3 headline is a New Taiwan dollar figure. In US dollars, revenue grew 41.2% to $46.72 billion, which is still the fastest the foundry has grown at this scale. TSMC's Q3 2026 earnings matter less for that number than for three others: the share of revenue now coming from high-performance computing, an unchanged capex budget, and the 2027 figures management declined to give.

What TSMC reported in its Q3 2026 earnings

The short answer

TSMC reported Q3 2026 revenue of NT$1,494.24 billion, or $46.72 billion, on 15 October 2026. Net income was NT$763.06 billion, up 68.7%. Gross margin was 66.4%. High-performance computing reached 68% of revenue, and TSMC guided Q4 revenue to $52.3 billion to $54.3 billion.

Those numbers come from the earnings release TSMC filed with the SEC. Operating margin was 58.7% and net margin 51.1%. Diluted earnings per share were NT$29.43, against NT$17.44 a year earlier. The release notes that Q3 figures had not yet been approved by the board, which is routine but worth stating.

A foundry this size does not usually grow like a startup. TSMC added about $13.6 billion of quarterly revenue in a year, on our arithmetic from its own growth rates. That increase alone is larger than most listed chip companies' entire quarterly sales.

The 50% is partly a currency story

TSMC reports in NT dollars. Revenue grew 50.9% in that currency and 41.2% in US dollars. The gap of nearly 10 points reflects a weaker NT dollar against a year earlier. TSMC's average rate in Q3 2026 was 31.98 per US dollar, per its earnings presentation.

For a reader tracking AI chip demand, the US dollar figure is the one to use. TSMC's customers pay in dollars, and the hyperscalers whose budgets drive those orders report in dollars. The NT dollar number is accurate. It also flatters the demand signal by roughly a fifth.

We think every headline that led with 50% got the story slightly wrong. The real growth rate is remarkable enough without the help.

Margins slipped, and 2-nanometer is the reason

Gross margin fell from 67.7% in Q2 to 66.4% in Q3. CFO Wendell Huang attributed the decline mainly to the early ramp of 2-nanometer production. That node rose from 3% of wafer revenue in Q2 to 8% in Q3, per Focus Taiwan's report of the call.

New nodes start at lower margins because yields and volumes are still climbing. A margin dip caused by the most advanced product ramping is the opposite of a demand warning. It tells you customers are moving to the newest, most expensive wafers as fast as TSMC can make them.

HPC at 68%: the AI chip demand signal in the platform mix

TSMC splits revenue into six platforms. High-performance computing, or HPC, covers data centre processors, AI accelerators and custom chips designed by cloud companies. It is not a pure AI line, but it is the closest TSMC comes to publishing one every quarter.

HPC was 68% of Q3 2026 revenue, up from 66% in Q2 and 57% in Q3 2025. HPC revenue grew 22% on the quarter. Smartphone revenue grew 10% and was 21% of the total. Automotive grew 23% but is only 4% of sales.

What the share means in dollars

TSMC does not publish HPC revenue in dollars, so the next figures are ours. Applying 68% to $46.72 billion gives roughly $31.8 billion of HPC revenue in Q3 2026. Applying 57% to the implied Q3 2025 total of about $33.1 billion gives roughly $18.9 billion.

That is growth of about 68% in a year, against 41% for the company. Because the shares are rounded to whole percentages, treat the result as accurate to a few points, not to the decimal. The direction is not in doubt.

North America is 79% of revenue

Customers headquartered in North America produced 79% of Q3 revenue, up from 76% a year earlier. That is where Nvidia, AMD, Broadcom, Apple and the cloud companies designing their own chips are based. China-headquartered customers were 5%.

The geographic figure reinforces the platform one. TSMC's growth is concentrated in a small number of US chip designers, and their own revenue depends on a smaller number of US cloud buyers. We cover that second layer of concentration in the breakdown of which customers carry Nvidia's guide.

TSMC's Q4 guidance and the 44% year

In July, TSMC guided Q3 revenue to $44.6 billion to $45.8 billion, per its investor relations guidance table. It delivered $46.72 billion. That is about $0.9 billion above the top of the range and about 3.4% above the midpoint.

Beats of that size are not unusual for TSMC, which guides conservatively. What changed is the full-year frame. TSMC now expects 2026 revenue growth of about 44% in US dollars. In July it said slightly more than 40%, according to Focus Taiwan.

Q4 implies another record quarter

The Q4 guide is $52.3 billion to $54.3 billion, assuming 31.8 NT dollars per US dollar. Gross margin is guided to 65% to 67% and operating margin to 56.5% to 58.5%. The CFO tied the margin guide to the continued 2-nanometer ramp and to dilution from overseas fabs.

At the midpoint, Q4 alone would be about $53.3 billion. That is a 14% sequential increase from a company that already grew 16% in Q3. Two consecutive double-digit sequential gains at this revenue level do not come from inventory restocking. They come from customers taking every wafer available.

TSMC's own capex says more than its revenue

Revenue tells you what customers bought this quarter. Capex tells you what TSMC thinks they will buy in two to three years, because a fab started today produces wafers well after that. That makes TSMC's capex budget the more forward-looking of the two signals.

The budget is $60 billion to $64 billion for 2026. In January it was $52 billion to $56 billion, per MarketBeat's summary of that call. It was raised once, by about 15% at the midpoint, and in October it was held.

Capex held while revenue guidance rose

This is the detail we would weight most heavily. TSMC raised its revenue outlook and did not raise its spending. A supplier that believed the boom was fragile would cut. A supplier chasing it without discipline would raise. TSMC did neither.

Management framed the next stretch instead. Spending from 2026 to 2028 will exceed the previous three years, and Huang said that if TSMC does its job, revenue growth will outpace capex growth. He also acknowledged that tool price inflation is real, which means part of any future capex rise buys less capacity than the headline suggests.

TSMC's Q3 capex was NT$443.21 billion, about $13.8 billion. Operating cash flow was NT$919.77 billion, which left NT$476.56 billion of free cash flow, per the Q3 earnings presentation. Capex was about 30% of revenue in the quarter.

That contrast matters. Several of TSMC's largest end customers now fund data centres partly with debt and equity, as covered in our comparison of hyperscaler capex with operating cash flow. The foundry at the top of the chain is the one participant still paying for its build in cash, with plenty to spare.

CoWoS and the packaging bottleneck

CoWoS stands for Chip-on-Wafer-on-Substrate. It is TSMC's advanced packaging process, which places a large processor die next to stacks of high-bandwidth memory on one substrate. Almost every leading AI accelerator uses it, and packaging capacity has repeatedly been tighter than wafer capacity.

On the call, Huang said CoWoS demand remained strong and that TSMC would work with customers to ensure sufficient capacity. C.C. Wei, the chief executive, said CoWoS was still customers' top choice and cited satisfactory yields, per Focus Taiwan.

TSMC gave no CoWoS capacity figure and no packaging revenue figure. That is consistent with past practice. It also means every CoWoS wafer estimate you read is a broker's channel check, not a disclosure.

On demand overall, Wei said demand remains higher than supply and the gap is still large, according to a third-party transcript of the call that we could not match line by line against TSMC's own PDF. In January, TSMC said AI accelerators were a high-teens share of 2025 revenue. It raised its forecast for AI accelerator revenue growth to a mid to high 50s compound annual rate from 2024 to 2029.

Our position: a company that declines to publish packaging capacity while saying demand exceeds supply is telling you allocation, not volume, is the constraint. Allocation favours the largest and earliest committed buyers, which is why the downstream order book matters so much.

Tying TSMC's numbers to hyperscaler capex

The target reader of this page wants one chain: hyperscaler budgets become chip orders, chip orders become TSMC wafers. The chain is real, but each link adds a lag and a markup, and the disclosures do not line up neatly.

The first two rows are one bank's estimates, not company guidance. Hyperscalers report Q3 results after TSMC, so this table will be revised within weeks.

Bank of America sees calendar 2026 hyperscaler capex above $860 billion, up roughly 80%, with a path toward $1.2 trillion in 2027, according to investingLive's report of the note. TSMC's capex midpoint of $62 billion is about 7% of that 2026 figure.

That ratio is illustrative, not a disclosed relationship. It shows how small the foundry's own spending is against the downstream build it enables. A modest change in hyperscaler intentions is a large change for TSMC's order book, in either direction.

The cash funding that downstream build is examined in our analysis of AI capex and its missing payback math. The memory side of the same chain, where prices rather than volumes moved first, is in the DRAM shortage piece.

What analysts project for 2027, kept separate

This section exists because most coverage of TSMC's Q3 2026 earnings blended management statements with broker forecasts. The table below separates them. Only the left column is TSMC.

Broker figures as compiled by TrendForce, 16 October 2026. They are projections from channel checks and models, and they disagree with each other by more than $20 billion on 2028 capex.

Why the broker spread is the real information

UBS's $90 billion for 2027 would be roughly 45% above the 2026 budget. Morgan Stanley's $83 billion would be about a third higher. The spread between them is a measure of how uncertain 2027 still is, even among analysts with access to supply chain checks.

Goldman's CoWoS projection, a 114% rise in 2027 output, is the most aggressive number in the set. If you only remember one figure from this table, remember that TSMC itself did not give any of them. Wei said he could not say whether 2027 would be tighter, because customers were still raising their requests.

Where reading TSMC as an AI demand gauge fails

We have argued that TSMC is the earliest hard read on AI chip demand. Here is where that argument is weakest.

Orders are not end demand

TSMC sees what chip designers order, not what end users consume. Chip designers order against hyperscaler forecasts, and hyperscalers build against their own forecasts of AI usage. Every layer has an incentive to over-order in a shortage, because being short costs more than holding excess. A foundry reading in a shortage measures fear of missing allocation as well as real demand.

The same pattern in 2000 inflated telecom equipment orders well above the traffic that eventually arrived. The financing side of that comparison is in our piece on Lucent, Nortel and vendor financing. We do not think TSMC's position resembles Lucent's, because TSMC does not finance its customers. The order-inflation risk is still the same shape.

HPC is not an AI line

HPC includes conventional server processors and other data centre silicon, not only AI accelerators. TSMC's only explicit AI figure is the January statement that accelerators were a high-teens share of 2025 revenue. Our dollar arithmetic on HPC overstates AI revenue by an amount nobody outside TSMC can measure.

The counter-case also deserves a fair hearing. Price is part of the growth, since 2-nanometer and 3-nanometer wafers cost far more than older ones. Some of the 41% is mix and pricing rather than more chips. The analyst average selling price projections compiled by TrendForce imply exactly that. That does not weaken the demand signal, but it does mean revenue growth overstates unit growth.

Frequently asked questions

Where to start if your costs depend on AI compute

If you run a business whose margins depend on AI compute, the foundry numbers reach you with a delay of several quarters. Use that delay.

This week, open your two largest AI or cloud contracts and find the clause that governs price changes at renewal. If supply stays tight into 2027, as every broker in the table above assumes, the vendors upstream of you hold the pricing power. You want to know now whether your contract lets them pass it through.

Then set a reminder for the four hyperscaler earnings reports that follow TSMC's this month. Compare each company's capex guide with the TSMC figures here. If cloud budgets rise again while TSMC's capex stays flat, the shortage is getting longer, and your renewal conversation should start earlier. The indicators that would signal the opposite are listed in our guide to AI chip selloff indicators.

Related reading