Portrait

Where Huawei's petaflops are actually landing — the SMIC wafer choke point

Petaflops coming online, Huawei vs. Nvidia

How much AI compute is each side of the US–China divide actually putting into customer data centers per year? One recent datapoint on why the Huawei side of that question is worth pricing: Meituan's LongCat-2.0, a 1.6-trillion-parameter MoE, was pre-trained on a 50,000-card cluster of Huawei Ascend 910Cs — no Nvidia silicon in the loop.Reuters — China's Meituan says new AI model trained on domestic chips The framing chart from the parent piece — total FP16 dense EFLOPS shipped in year, Huawei Ascend vs. Nvidia, on Huawei's disclosed Atlas roadmap cadence — makes the point that the annual-flow gap narrows on paper from ~14× in 2025 to ~5–6× in 2027, but only if Atlas 950 (GA 2026) and Atlas 960 (GA 2027) actually land on their disclosed calendar. The single hardest supply constraint under that "if" — the one this note drills into — is advanced-node wafer capacity at SMIC.

Total AI-accelerator FP16 dense EFLOPS shipped per year, Huawei Ascend vs. Nvidia

The bottleneck: SMIC N+2 wafers

910C fabs on SMIC's N+2 process — 7 nm-class, multi-patterned without EUV. That works for 910C volumes; Atlas 950's 8,192-NPU pod (32× the A3 pod's NPU count) and Atlas 960's denser silicon in 2027 do not obviously fit inside the N+2 capacity SMIC has today. And SMIC does not break out N+2 capacity, N+2 yield, or the wafer share reserved for Ascend anywhere in its filings — the three numbers you would need to size the 950 supply plan.

What management does say, on the Q4 2025 call, is that the delivery clock is now slipping under export-control pressure: 2025 capex printed $8.1B (above the year-start plan) because "the company has procured some key equipment in advance," but "supporting equipment may not be purchased yet," and "the procured equipment may not be able to form production lines this year." The result: 2025 added ~50,000 12-inch wafers/month of capacity; management guides 2026 to add only ~40,000, with 2026 depreciation up ~30% YoY as new fabs exit start-up. Utilization was already 95.7% in Q4 2025 and 12-inch was "nearly fully loaded," so incremental Ascend allocation is a zero-sum fight against a Chinese customer book that grew ≥60% YoY in industrial/automotive. SMIC Q4 2025 Earnings Call, Feb 11, 2026 SMIC Q1 2026 Earnings Call, May 15, 2026

The tool-chain: who supplies the bottleneck

The wafer-capacity number is a function of what tools SMIC can actually take delivery of. Export controls have progressively pushed the advanced-node tool stack toward domestic Chinese vendors — the same three names below carry the load, and their revenue lines are the cleanest public proxy for how fast SMIC's N+2 tool base is actually growing.

Process step Restricted (Entity List / VEU) Domestic substitute picking up the load
Lithography (EUV) ASML — blocked at advanced nodes SMEE (private) — no advanced-node DUV immersion tool shipped in volume yet
Etch Lam Research, Tokyo Electron — restricted at advanced nodes AMEC (688012.CG), NAURA (002371.CS)
Deposition (CVD/ALD/PVD) Applied Materials, Lam, TEL — restricted at advanced nodes NAURA, Piotech (688072.CG)
Wet clean & ECP Lam, TEL, SCREEN ACM Research (ACMR.UQ / 688082.CG)
Metrology / inspection KLA — restricted at advanced nodes RSIC, Skyverse (both private / early revenue)

All three listed China WFE names are on the mutually-confirmed SMIC supplier graph, alongside their restricted foreign counterparts — so the substitution is visible in the observed trade relationships, not just narrated.

SMIC capex vs. domestic WFE supplier revenue, FY2021–FY2027E

The chart pulls the story into one frame. SMIC capex plateaus at ~$8B/yr on VA consensus through 2027 — i.e. no incremental spend to relieve the constraint. But the domestic WFE supplier revenue lines run harder than that: NAURA compounds ~30% p.a. to ~$9.7B by 2027E, AMEC to ~$3.4B, Piotech to ~$1.7B. Two readings sit inside that gap. Bullish: SMIC's capex dollars are moving out of restricted-vendor tools into domestic tools, so a flat capex line still buys accelerating N+2 tool intensity — the "form production lines" problem gradually solves itself as domestic substitutes qualify. Bearish: domestic revenue growth includes CXMT (HBM), YMTC, and mature-node builds, so the share hitting SMIC N+2 specifically is unknown, and every quarter of "supporting equipment may not be purchased yet" defers Ascend wafer starts one-for-one.

Watch: (i) SMIC's next capex re-guide — a step-up would signal management sees the tool-delivery cliff clearing; a flat-to-down line reads as capacity-constrained on the supply side; (ii) NAURA / AMEC / Piotech gross-margin trajectory — a jump signals advanced-node tool mix taking share of a still-flat top line, which is the metric the raw revenue line hides.

Sources

  • S&P global SMIC Q4 2025 Earnings Call, Feb 11, 2026
  • S&P global SMIC Q1 2026 Earnings Call, May 15, 2026
  • Visible Alpha standardized data (revenue, capex) for 981.HK, 002371.CS, 688012.CG, 688072.CG — chart backing series
  • Portrait company graph — SMIC (981.HK) suppliers, mutually confirmed
Where Huawei's petaflops are actually landing — the SMIC wafer choke point
Where Huawei's petaflops are actually landing — the SMIC wafer choke point

← Back to China HPC production