TL;DR: TSMC's third-quarter disclosure puts two-nanometer technology at eight percent of wafer revenue. That establishes a commercial revenue contribution, not eight percent of wafers, chips, capacity, or customer devices. Interpreting the ramp requires keeping revenue mix, wafer volume, die economics, packaging, and company margin in separate accounts. The missing conversions should remain assumptions until evidence supports them.
A node-mix chart can appear to answer more questions than it does. A colored slice expands, and readers begin estimating wafer starts, naming customers, inferring yield, and forecasting finished products. The chart has not supplied those conversions. It has supplied a revenue distribution under the company's stated definition.
On October 15, TSMC reported that two-, three-, five-, and seven-nanometer technologies represented eight, thirty-three, twenty-eight, and ten percent of third-quarter wafer revenue respectively. Together, those categories accounted for seventy-nine percent. The quarter's consolidated revenue was US$46.72 billion and gross margin was 66.4 percent. The discussion below uses the disclosures available by October 28 and labels its own scenarios explicitly. TSMC's earnings release filed with the SEC
Open the denominator ledger
The first ledger entry is the reported numerator: wafer revenue associated with a particular process technology. The second is the reported denominator: total wafer revenue. Their ratio describes revenue mix. It does not directly describe physical output because different products and processes can carry different revenue per wafer.
Consolidated company revenue is another denominator. Before multiplying a node percentage by a company total, an analyst should establish that the scopes match. A figure explicitly labelled wafer revenue should not silently become a percentage of every activity included in consolidated sales. If the necessary reconciliation is unavailable, the result should remain an estimate with the assumption stated.
The time boundary matters too. A quarterly share describes the quarter's recognized revenue distribution. It is not necessarily the share of equipment installed on the last day, the share of wafers beginning production during the period, or a forecast for the next quarter. Manufacturing and accounting stages occur on different calendars.
The ledger sounds elementary because the arithmetic is elementary. Its value is preventing a correct percentage from being applied to the wrong question. Most of the unsupported conclusions in a node-ramp story begin at that boundary rather than in a complicated model.
Eight percent is commercially meaningful on its own
The two-nanometer contribution is evidence that the technology is generating reported wafer revenue. It belongs in the account of how TSMC's business is evolving. A careful interpretation does not need to invent a physical shipment count to make that fact significant.
The next useful questions are about the ramp's economics and scope. How does management describe demand? What costs accompany the early ramp? Which later disclosures show the contribution changing? The earnings call provides management's explanation, while the node percentage remains one observed measure within it. TSMC's October 15 earnings-call transcript
The result should not be reduced to a binary label saying mature or experimental. Commercial production can exist alongside continuing process improvement, changing product mix, and a steep cost ramp. The investor needs evidence about each dimension rather than one word that suggests they move together.
The same discipline applies to a technical comparison. A node name identifies a process family in the company's portfolio. It does not by itself provide all the design, power, performance, area, and cost information needed to compare actual products. A revenue chart is a business observation, not a complete engineering specification.
A price assumption changes the inferred wafer share
Consider a fictional foundry with one hundred revenue units. Its new node contributes eight and its other nodes contribute ninety-two. Suppose, solely for illustration, that the new node earns twice as much revenue per wafer as the combined other-node average. The implied relative wafer quantities are four and ninety-two, giving the new node about 4.17 percent of this simplified wafer total.
Now suppose the new-node revenue per wafer is four times the other-node average. The same eight revenue units correspond to two relative wafers against ninety-two, or about 2.13 percent of the total. The observed revenue share is identical in both examples. The physical interpretation changes with an undisclosed conversion.
Neither ratio estimates TSMC's actual wafer share. Real product pricing, contracts, process variants, and revenue recognition are more complicated than this two-bucket exercise. The example shows why an analyst cannot recover a unique volume result from one revenue percentage alone.
The direction of a change is not always enough either. A higher revenue share can reflect more wafers, higher revenue per wafer, a different mix within the node, or changes elsewhere in the denominator. The model should name which explanation it assumes and identify the evidence that could distinguish it from the alternatives.
If the public evidence supports only revenue mix, reporting only revenue mix is a precise result. It is more useful than a detailed wafer estimate whose apparent accuracy comes from hidden assumptions rather than disclosed data.
A wafer is not a fixed number of chips
Moving from wafers to chips requires another conversion. A wafer can contain different numbers of dies depending on die size, layout, and other design factors. Accepted die output also depends on manufacturing results and product requirements. A single wafer count cannot answer those questions without the relevant assumptions.
Use a deliberately simplified example. One product yields five hundred gross dies per wafer and eighty percent meet the defined acceptance condition, giving four hundred accepted dies. Another yields six hundred gross dies with a seventy-five-percent acceptance rate, giving four hundred fifty. The product with the lower acceptance percentage produces more accepted dies in this toy comparison.
The example does not provide a manufacturing-yield formula or claim that either figure describes a real process. It separates two terms that headlines often collapse: the number of opportunities per wafer and the fraction accepted. A percentage without its denominator can mislead even when it is accurately quoted.
An investor also needs to know what is being counted as a chip. A package can contain several dies and other components. The product sold to a customer may therefore not correspond one-to-one with an accepted die at a particular node. The analysis should preserve the unit at every conversion rather than change labels midway through the calculation.
The node chart cannot identify yield uniquely
A strong revenue contribution can coexist with different yield conditions because volume, price, mix, and cost can change together. The chart does not contain enough independent observations to solve for every variable. Assigning a precise yield merely because a ramp is growing is an identification error.
Management may provide qualitative or quantitative information about process progress. That information deserves separate attribution and should retain its scope: a development milestone, a product-specific result, or a broader manufacturing statement. It should not be converted into a universal value for every die made on the process.
For a research note, the useful approach is a sensitivity model. It can show how several plausible assumptions would affect a hypothetical product's accepted output or cost. The note should avoid presenting the central assumption as if it were measured. A range is informative only when the reader understands what varies and why.
This is not an argument against modeling. It is an argument for models whose unknowns remain visible. A transparent conditional result can guide the next evidence search. A falsely precise result can make later news look surprising simply because the earlier model hid its own uncertainty.
Packaging is a separate conversion
Suppose a fictional product has four hundred thousand accepted dies available from a given production stage. If an explicitly assumed later assembly acceptance rate is ninety percent, three hundred sixty thousand remain after that step. If the final product requires two such dies, the simple material ceiling is one hundred eighty thousand packages before other constraints.
Those values are invented to demonstrate units. A real package may use different combinations of dies, memory, substrates, and assembly steps. It may also include rework and other process paths that the simple multiplication omits. The appropriate model depends on the product's actual architecture and manufacturing route.
The important distinction is that a node ramp and a finished-package ramp need not have identical timing. Additional wafers can wait for another required component or an assembly qualification. Conversely, previously accumulated material can support package shipments before the current wafer-production rate changes.
A semiconductor investor should therefore avoid translating a node percentage directly into a system-shipment forecast. The missing links are not minor adjustments to be added later. They define what the physical output unit means and when it can become available to the next customer.
High-performance computing is not a synonym for one AI workload
TSMC's call reports high-performance computing as a platform category and describes several sources of AI-related silicon demand, including accelerators, CPUs, and networking. That is broader than one type of model training or one customer's GPU. The platform and process classifications answer different questions. TSMC management discussion
A node can serve products across several end markets, and an end market can use several nodes. The analyst should not treat the two charts as interchangeable partitions that can be matched row for row. Without an appropriate cross-tabulation, the exact intersection remains undisclosed.
For example, multiplying a new-node share by an HPC share would assume a particular relationship between the categories. Independence would be one possible mathematical assumption, but it would not become a business fact merely because the calculation is easy. The same warning applies to assigning the resulting number to a named customer.
The useful conclusion can remain broader: leading-edge demand has several contributors, and a specific new node is gaining revenue contribution. That is enough to frame questions about diversification and investment without inventing a customer allocation table.
Aggregate gross margin does not reveal a node's margin
TSMC's earnings discussion attributes part of the quarter's margin pressure to the two-nanometer ramp. This is management's causal explanation of an aggregate result. It does not disclose a standalone gross margin for every node. October earnings-call profitability discussion
A separate toy company demonstrates the ambiguity. Suppose a new product is ten percent of revenue and the remaining business is ninety percent. If the old business has a sixty-five-percent margin and the new product a fifteen-percent margin, the combined margin is sixty percent: 0.90 times 65 plus 0.10 times 15.
The same sixty-percent aggregate results if the old business margin is sixty-two percent and the new-product margin is forty-two percent. The new product's margin differs substantially between the two scenarios. An aggregate margin and revenue share alone do not identify it.
This example is unrelated to TSMC's actual node economics and deliberately uses a different share and margin. It shows why back-solving one undisclosed business margin from a company total requires additional assumptions. Product mix, cost changes, currency, and other factors can affect the remainder of the business too.
The analyst can still assess management's explanation and track subsequent results. The limitation is narrower: do not present an assumed decomposition as a reported fact. A credible interpretation preserves both the observed margin and the information needed to explain it.
A ramp has a time profile inside the quarter
A quarterly revenue share averages activity recognized over the reporting period. It does not reveal a complete monthly ramp. The same quarterly contribution can arise from a steady rate, a late acceleration, or a changing mix whose revenue effects offset one another.
Consider an invented three-month sequence with new-node revenue of two, three, and five units. Another sequence of three, three, and four produces the same total of ten. Their end-of-quarter rates differ. A forecast that treats the quarterly average as the next month's starting rate would miss that distinction.
The appropriate response is to use later management guidance and other relevant disclosures, while retaining uncertainty about the intra-quarter profile. A chart with one quarterly point should not be interpolated into a detailed monthly production curve without a stated assumption.
The timing also affects costs. Equipment, staffing, qualification, and depreciation can precede the associated revenue at different intervals. A node can gain share while its economic contribution is still changing. The investor should evaluate the ramp as a sequence of investments and outputs rather than a single milestone labelled launched.
The customer's design decision has another denominator
A chip designer does not buy a node percentage. It chooses a process and architecture that can support a product under performance, power, area, cost, schedule, and supply constraints. The best choice depends on the actual design and expected volume, not on a general rule that the smallest named node is always the economical choice.
A product can benefit from a newer process while another remains better suited to an established one. Different functions within one system can also have different requirements. The relevant unit is the customer's product outcome, not a technology label considered in isolation.
This helps explain why several nodes can remain commercially important at the same time. A rising new-node contribution does not require every older node to disappear. The node-mix chart describes the resulting business distribution; it does not dictate the customer's engineering objective.
For the outside analyst, the challenge is to avoid substituting an imagined customer roadmap for evidence. Public announcements can establish certain product relationships. Other allocations remain confidential or unconfirmed. A precise revenue share does not remove that boundary.
A wafer price is not its manufacturing cost
The conversion from revenue to physical volume uses revenue per wafer, not the foundry's cost of producing it. Those quantities should not be substituted for one another. A higher customer price can coexist with higher manufacturing cost and with either a higher or lower margin, depending on how the two change.
For a simple invented example, a wafer sold for twenty thousand currency units and costing sixteen thousand has a twenty-percent gross margin. Another sold for thirty thousand and costing twenty-one thousand has a thirty-percent margin. The second costs more to manufacture and still has a higher percentage margin. These values describe no real node; they demonstrate why price and cost require separate assumptions.
The customer's cost per accepted product introduces the additional die and package conversions already described. A more expensive wafer can be economical for a particular design if its complete product economics improve, or uneconomical if they do not. The node-mix chart cannot decide that comparison by itself. Keeping price, foundry cost, and customer product cost in separate fields makes the financial model much easier to interpret when later evidence changes one of them.
Fixed design costs change with volume
A fictional product requires one hundred million currency units of nonrecurring design and preparation costs. Spread over two million accepted products, that is fifty units per product before recurring production costs. Spread over twenty million, it is five. The underlying fixed cost is the same; the volume assumption changes its unit allocation.
The example is not an estimate of a real mask set, tape-out, or product program. It illustrates why a customer's process decision can depend strongly on expected volume and the probability of reaching it. A technically superior design may have different economics at a smaller scale.
Recurring costs need their own model: wafers, accepted dies, packaging, testing, and other required work. Combining fixed and variable costs into one unexplained cost-per-chip number makes it difficult to understand which assumption drives the result. The investor should preserve the distinction when interpreting adoption.
A delayed launch can change both sides. It can shorten the period in which fixed costs are recovered and alter the available manufacturing or product mix. A schedule assumption is therefore part of the economics, not merely a date appended to an otherwise timeless unit-cost calculation.
Currency can change the growth comparison
TSMC reports consolidated results in New Taiwan dollars and also provides US-dollar figures. The release's growth rates differ by currency. An analyst should compare like with like and use the disclosed basis, rather than place a dollar revenue total next to a local-currency growth rate without explanation.
This distinction matters when attributing growth to node mix. Revenue can move because of physical activity, pricing, mix, or exchange rates. A reported percentage in one currency does not isolate those causes. The model should identify which variables are held constant in any derived comparison.
Guidance introduces another exchange-rate assumption. It describes an expected range under the company's stated outlook and assumptions, not a guaranteed outcome at every future rate. The investor can examine sensitivity without pretending that management promised the midpoint in all circumstances.
A good data sheet therefore records currency, period, consolidation scope, and whether the figure is actual or guided. Those labels take little space and prevent several common errors when numbers travel from an earnings release into charts and social posts.
Capital spending is not immediate production
A capital budget supports future capabilities, but spending, equipment installation, process qualification, available capacity, and recognized revenue are separate milestones. A larger budget can be consistent with a strong demand outlook while leaving the timing of useful output uncertain.
The analyst should follow the relevant company's actual milestone language. A planned fab is not an operating fab; installed equipment is not necessarily qualified for every product; qualified capacity is not necessarily fully utilized. Each transition narrows a different uncertainty.
The same logic prevents overinterpreting a node's current revenue share as proof of the entire future investment return. The share demonstrates one part of commercialization. Returns also depend on future demand, costs, execution, and the price paid for the business's claims on those cash flows.
That last valuation step should remain explicit. An excellent manufacturing result can already be reflected in a market price, while a weaker result can be less damaging than a price assumed. A technology milestone alone cannot settle an investment decision without the financial expectations attached to it.
Keep an evidence table that resists accidental upgrades
A compact table can preserve the current result and the next unanswered question.
| Observation | Supported interpretation | Conversion still needed |
|---|---|---|
| Node share of wafer revenue | Revenue contribution under that classification | Revenue per wafer and scope reconciliation |
| Wafer quantity, if disclosed | Physical activity in the stated unit | Die layout and acceptance conditions |
| Accepted die output | Components available at that stage | Package requirements and later acceptance |
| Company gross margin | Aggregate profitability | A disclosed or explicitly assumed decomposition |
| Management guidance | A dated forward-looking range | Realized demand, currency, cost, and execution |
The table is an original analytical aid, not an additional TSMC disclosure. It makes it harder for a note to change the meaning of a number while retaining its apparent authority. A model input can move from unknown to estimated, and later to observed, without hiding that progression.
The researcher also saves the source date and original wording. If a later quarter changes the definition or presentation, the comparison can be adjusted deliberately. A smooth chart assembled from inconsistent categories can be less informative than a discontinuity clearly explained.
The next quarter should answer the next question
When new figures arrive, compare the node distribution on a consistent basis and read management's explanation of the ramp. Check whether the changes support the earlier scenario or require different assumptions. Do not update only the headline number while leaving every inferred wafer and margin input untouched.
A sensitivity model becomes useful at this point because its assumptions were named. If pricing or product mix appears different, the analyst can revise the appropriate conversion. If public information still does not identify yield or customer allocation, those cells remain conditional.
The goal is not to produce an estimate for every blank. It is to build a coherent account of which observations have changed the investment or operating question. Some news strengthens a revenue-growth thesis; other news clarifies cost, timing, or concentration. Their relevance depends on the decision the analysis is intended to support.
TSMC's two-nanometer revenue contribution is meaningful evidence of an evolving foundry business. Its value is clearest when the chart is allowed to say exactly what it measures. Keeping the conversion from revenue to wafers to chips explicit preserves both the achievement and the uncertainty that the next disclosure must resolve.