NVDA Blackwell Gross Margin Analysis: Why the Street Is 350bps Too Low

The Core Argument

NVIDIA's Blackwell architecture delivers a 35–40% ASP premium over Hopper at equivalent configurations. TSMC manufacturing cost for B100/B200 wafers is higher than H100 — but not proportionally. The delta flows directly to gross margin. My FY2025 GM estimate is 78%, versus Street consensus at 74.5%. That 350bps gap is worth approximately $3.5B in gross profit on $100B revenue.

H100 vs H200 vs B200: ASP Stack

The ASP progression across generations is the most important data point in the NVDA thesis:

SKUList ASP (est.)GM contribution (est.)H100 SXM5 (80GB)$30,000–$35,000~73%H200 SXM5 (141GB)$40,000–$45,000~76%B200 SXM6$55,000–$65,000~78%

The jump from H100 to H200 is primarily HBM3e memory capacity — the compute die is nearly identical. The jump to B200 involves a new compute architecture AND HBM3e, commanding a larger premium.

TSMC's N4P node costs are higher than N5 (H100), but the scale of NVIDIA's volume keeps the per-die cost increase below the ASP increase. The margin expansion is real.

Supply Chain Evidence

Three observable signals confirm the B200 ramp is ahead of schedule:

1. Lead time compression: H200 lead times moved from 52 weeks (Q1) to 22 weeks (Q3). B200 pre-orders filling faster than original guidance implied.
2. TSMC capacity allocation: NVDA's share of TSMC CoWoS advanced packaging capacity has increased from ~35% to ~45% of total CoWoS capacity per analyst day estimates.
3. Hyperscaler procurement cadence: Two of the four major hyperscalers increased Blackwell orders vs the original Q4 run rate. Channel checks are publicly citable from earnings commentary.

Why Consensus Is Wrong

Consensus margin models lag ASP transitions. In the H100 → H200 transition, consensus was 200bps below actual GM for two consecutive quarters before revising. The same phenomenon is repeating with Blackwell: