TSMC and the Process Step You Cannot Skip

That is the screen, not a narrative looking for a print. Semiconductors is the widest internal spread in the coverage universe. SOXX is how the market holds the category — one fund, one factor, one AI-chip tape. The eight names above all sit under that label. They do not sit in one business.

A leading-edge chip still has to be designed in Cadence or Synopsys, printed on an ASML scanner, inspected by KLA, and manufactured at TSMC. NVIDIA can be designed around. Broadcom's custom XPU book is the existence proof. IonQ is not on the path at all. The market's variable is the GICS code. The variable that sorts the scorecard is whether the next node ships without you.

The process-step names cluster on system of record and transaction embedding — the design database, the scanner installed base, the inspection recipe, the PDK. NVIDIA is strong on network effects and embedding (CUDA, the rack, the software) and only intact on system of record, which is the tell: the customer can leave the GPU if they are willing to rewrite the stack. Broadcom is strong on system of record where VMware sits and weaker on the network-effect pillars, which is what a custom-silicon franchise looks like. IonQ is blank on the two pillars that define a process step.

The process-step names reported through July and August. Each line is a customer who could not skip the step.

TSMC's July 16 release is the foundry receipt: USD revenue $40.20 billion, up 33.7% year over year, gross margin 67.7%, and a third-quarter guide of $44.6–45.8 billion on the N2 ramp. That is a customer list — Apple, NVIDIA, AMD, the hyperscaler ASICs — that has already paid the redesign cost of being on TSMC's PDK. The print does not invent a second leading-edge foundry. It invoices the one that exists.

ASML's July 15 release is the machine behind that foundry. Sales €9.3 billion, gross margin 54.0%, full-year guide raised to €43–45 billion. Installed Base Management — net service and field option sales — was €2.8 billion in the quarter and the reason the print beat. Low-NA EUV capacity is being lifted about 30% for 2027. No other company shipped an EUV tool. A competitor would need the decade ASML already spent, and the foundry customers who funded it.

KLA's July 28 release does not publish a share figure in the earnings PR. It does say process control sits on the critical path of the AI build, and that advanced packaging is a new growth line on the same tools. Fiscal 2026 closed at $13.58 billion of revenue and $3.77 billion of free cash flow. Cadence's July 27 release is the design-side twin: Core EDA +18% organic, IP +40%, record backlog $8.1 billion, of which $4.2 billion converts inside twelve months. Synopsys's August 26 release is the other half of that duopoly — $2.477 billion, Design IP $474 million back to year-over-year growth, Ansys a year in. The FY2026 raise assumes export-control rules stay substantially similar. That assumption is a risk on the EDA names. It is not a second toolchain.

NVIDIA's August 26 release is the cycle at full steam: $96.221 billion of revenue, Data Center $89.0 billion, Vera Rubin in full production, third quarter guided at $108.0 billion ±2% with no China data-center compute in the outlook. Broadcom's June 3 release is the design-around sitting inside the same tape: AI semiconductor revenue $10.8 billion, up 143%, guided to $16.0 billion in the next quarter on custom accelerators and AI networking. IonQ's May 6 release is the science project the label also contains — $64.7 million of revenue, an adjusted EBITDA loss of $96.8 million, and $3.1 billion of cash. The GAAP net income of $805.4 million is a warrant revaluation. SOXX holds all three.

• The foundry and the scanner. TSMC's lock is the PDK and the yield-learning. Re-taping a leading-edge chip for another foundry is years and a redesign, not a purchase-order flip. ASML's lock is that there is no second EUV vendor. Both fail if a named customer actually ships the next node somewhere else — not if GPU prices wobble.
• The inspection recipe and the design database. KLA sits inline on the wafer. Cadence and Synopsys sit on the tape-out. Pulling either mid-program risks the node, not a licence seat. The EDA export-control assumption is the live risk on that half, not an open-source 45nm demo.
• The cycle. NVIDIA's lock is CUDA and the rack. It is real, and it is also the lock a hyperscaler spends tens of billions to route around. Broadcom's custom XPU book is that spend landing at a vendor. A digestion quarter hits this group first.
• The science project. IonQ's lock, such as it is, is talent and a fidelity roadmap. It is not a system of record and it is not embedded in a tape-out. A quantum print that looks like a foundry would move it. This one does not.

The SOXX frame is not confused about NVIDIA. It is correct that this cycle's profit pool is the GPU stack, and it is correct at a scale the process-step names do not print. Data Center revenue of $89.0 billion in a quarter is not a rounding error on TSMC's $40.2 billion or ASML's €9.3 billion. It is the customer. Jensen Huang's line in the August 26 release — compute is revenue, Vera Rubin in full production — is the tape the market is paying for. A reader who wants the AI semiconductor trade, this year, is looking at NVIDIA. Pricing NVDA as the AI-chip tape is not a misread. It is the business.

The complaint is not that the tape is wrong about NVIDIA. It is that the tape has bundled TSMC, ASML, KLA and the EDA duopoly into the same trade, and has left a door open for IonQ to sit in the same fund. NVIDIA can miss a quarter because a hyperscaler slipped a rack, because China stayed at zero, because Vera Rubin ramped late. TSMC does not miss that quarter the same way. The foundry invoices the wafer whether the die is a Blackwell or a TPU. That is the skip test from the other side: the customer can change the GPU vendor and still has to pay TSMC, ASML, KLA, and Cadence or Synopsys.

Broadcom is the more useful counter, because it can hide inside the process-step story. Custom XPUs are designed on the same EDA tools, printed on the same TSMC nodes, scanned on the same EUV tools. The June 3 print — $10.8 billion of AI semiconductors, +143%, six hyperscaler customers — is a semiconductor vendor winning work that used to be NVIDIA's. If that book keeps compounding, the market's claim that the profit pool is 'semiconductors' rather than 'one GPU vendor' gets stronger, not weaker. It does not make Broadcom a lithography monopoly. It makes Broadcom the proof that the cycle can move. VMware is a different lock, and it is why Broadcom's system-of-record pillar lights up in a way NVIDIA's does not. Mixing the two locks is how the label keeps winning arguments it has not earned.

The EDA half has its own honest objection. Synopsys is digesting Ansys. Cadence's full-year raise, as the CFO commentary states it, assumes today's export-control rules stay substantially similar. A re-restriction on advanced-node China work would cut a real slice of EDA sales without creating a second toolchain — the duopoly would shrink together. An open-source tape-out at a mature node is not the same event. If either happens, the process-step group is not one quality. It is still not IonQ, and it is still not a GPU cycle.

• TSMC loses the node. Apple, NVIDIA or AMD awarding a majority of an N2- or A14-class node to Intel Foundry or Samsung as the primary source. That is the foundry stop. A packaging or HBM bottleneck moving to another vendor is not.
• ASML gets a peer. A production EUV or High-NA order at TSMC, Intel or Samsung taken by someone else. Capacity adds at ASML are the monopoly extending, not the monopoly ending.
• IonQ becomes a step. Foundry-like commercial wafer revenue — a process other chipmakers have to route through — not another $65 million system-sales quarter on $3.1 billion of cash.
• Not the test. SOXX constituents moving together on an AI-capex headline, or NVIDIA printing another $90 billion quarter. The GPU can keep winning the cycle. Watch whether the customer had a second source for the step.

When a category-wide tape re-prices the foundry, the scanner, the GPU and the quantum name identically, the job is to find the names where the label does not fit the book. On this cohort SOXX was the right read on NVIDIA this cycle and a plausible read on Broadcom. It was the wrong read on the process step, and it was not a read on IonQ at all.