Gross margins just peaked at 76.7% and they explicitly guided them down while ramping opex 30%.
Thesis: The 'Capacity Ceiling' is actually a cycle extender. By physically limiting shipment volume via CoWoS constraints, TSMC is preventing NVIDIA from stuffing the channel. Demand overhang remains massive. The margin compression (76.7% to mid-70s) is mix-driven, not pricing-driven. You don't short a monopoly with a backlog bigger than its capacity.
Verdict: LONG — Conviction: HIGH
Catalyst: H200 ramp in Q2. As yields improve, volume offsets the mix-shift margin headwind.
Key Risk: The 40% inference claim. If this is just reclassified recommender systems and not true Generative AI inference, the 'training cliff' thesis remains valid.
The Tell: Jensen volunteering the 40% inference number unprompted. He knows the 'training peak' narrative is the biggest risk to the multiple, so he's seeding the 'inference is recurring revenue' defense early. 'The estimate is probably understated' is classic deflection masquerading as humility.
Friction Level: MODERATE_FRICTION — Bears see margin compression and a capacity ceiling. Bulls see a supply-constrained supercycle that extends duration.
Report not found
The report data is no longer available. Please return to the archive.