Autoresearch: As of 19 Sep 2026, does TSMC disclose that the inference-ASIC wave is material to utilization/mix? Is Jalapeño/Etched/Cerebras on a node/HBM generation Nvidia does not consume? Any independent Etched benchmark?

Generated by /autoresearch on 2026-09-19. Sibling pass 2026-09-19-autoresearch-tsm-agentic-cpu-split-dram already found TSMC will not size CPU vs GPU vs XPU; this pass is the ASIC-additivity / Jalapeño supply-chain test. See Provenance. Context: vault/projects/stock-market Research-only. No buy/sell/size.

Summary

TSMC’s Q2 2026 print still does not size custom ASIC / XPU versus GPU. Net revenue NT$1,270.38 billion; HPC 66% (from 61% in Q1); 3nm 30%, 5nm 33%, 7nm 11%, 2nm 3%; advanced (≤7nm) 77% (TSMC 2Q26 earnings release; 2Q26 management report; SEC HTML). OpenAI’s Jalapeño is widely described as Broadcom-co-designed, TSMC-fabbed, leading-edge packaging, eight HBM stacks — i.e. the same constrained CoWoS/HBM/leading-edge pool as Rubin, not Etched’s “4nm / different HBM / more-is-more” supply chain. Cerebras still claims the escape hatch (N5, no HBM, no CoWoS) in the same week’s commentary. No independent Etched/Sohu benchmark or 2026 rack-shipment primary was retrieved. Step 4 stays open. Do not re-date TSM/NVDA.

Findings

TSMC will not give the chain its number

2Q26: “Our business in the second quarter was supported by strong demand for our leading-edge process technologies” (Huang). Q3 guide US$44.6–$45.8 billion. Platform mix: HPC 66% / Smartphone 22% / IoT 5% / Auto 4%. HPC +20% sequential. Geography: North America 78%. No “ASIC,” “custom,” “XPU,” “Etched,” “Jalapeño,” or “Cerebras” token in the release or management-report revenue tables (earnings release; management report). The 5nm bucket (33%) can contain N4/N5 ASIC wafers; TSMC does not say so. That is the same wall the agentic-CPU pass hit.

Jalapeño is a TSMC capture — and a Step 3 problem

Matterfact’s 29 Jun newsletter (The Information TITV + Bloomberg Tech week): Jalapeño “has eight stacks of HBM… leading edge TSMC capacity and they’re doing co-loss and advanced packaging. So this is not a bottleneck optimized chip. This is a performance optimized chip.” Translation they draw: OpenAI is in the same CoWoS/HBM/leading-edge queue as Nvidia (Matterfact 29 Jun). Tom’s Hardware: first-gen Jalapeño defined by OpenAI, co-designed with Broadcom, made by TSMC; next-gen could add Samsung Foundry as a second source — which would be a Step 4 non-TSMC risk if it happens, and is not a 2026 wafer start (Tom's Hardware).

If Jalapeño is the largest inference buyer’s ASIC, additivity (Step 3) fails for this name: it consumes the flagship packaging/HBM stack. Etched’s 4nm / other-HBM claim can still be true for Etched and false for the wave.

Cerebras remains the additive anecdote; Etched remains unbenchmarked

Same Matterfact week: Feldman — “we don’t use HBM… we don’t use [CoWoS]… we’re at the five nanometer node”; record revenue $191 million, +92% YoY, in that recap (Cerebras’s already-ingested Q2 core $209.9M is the better first-party number on the sibling wafer-scale page). Cerebras stock was punished on the Jalapeño headline — the tape read OpenAI in-house as a threat to a named OpenAI order, which is the opposite of “TSMC wins regardless / Nvidia is the only risk leg.”

No Sohu rack volume, no third-party concurrency-per-MW bench, no Etched 10-Q. Step 2 stays founder-and-investor-sourced.

Contradictions and open questions

• Step 4 “TSMC captures the wave regardless of vendor” is directionally plausible (Jalapeño is at TSMC today) and unquantified (no mix %).
• Step 3 “deliberately not Rubin’s chain” is contradicted for OpenAI and untested for Etched.
• Samsung second-source talk is a live Step 4 falsifier, future-dated.
• TSMC Q3 print will still not size ASIC unless Wei changes disclosure policy.

Provenance

Rounds run: 3 of 3. X sources: attempted; spend-cap 403. Not used. Grokipedia: not used for 2026 claims.

Web sources:

• TSMC 2Q26 earnings release — official — NT$1,270.38B; node mix; no ASIC split.
• TSMC 2Q26 management report — official — HPC 66%; 5nm 33%.
• TSMC SEC HTML — official reprint.
• Matterfact 29 Jun — secondary — Jalapeño 8×HBM / leading-edge TSMC; Cerebras N5 claim.
• Tom's Hardware — trade — TSMC today; Samsung maybe next-gen.

Generated: 2026-09-19