Daniel Nishball
· SemiAnalysis
· May 01, 2026 at 02:30
· ⏱ 34 min read
| Read on Substack ↗
Summary
The article argues that agentic AI has permanently increased token value and inference margins, shifting value capture to AI labs like Anthropic. Despite surging demand and tight supply, Nvidia and TSMC have not repriced their products to reflect scarcity, leaving significant room for future price increases that would benefit their margins and GPU rental economics.
•Anthropic's ARR exploded from $9B to over $44B in 2026, with inference gross margins rising from <40% to >70%.
•Agentic workloads (e.g., Claude Code) have increased token value; blended price per million tokens for Opus 4.7 on agentic tasks is ~$0.99 vs $5/$25 sticker due to high cache hit rates.
•GB300 NVL72 throughput is ~17x higher than H100 with FP8, and ~32x with FP4, while TCO per GPU is only ~70% higher.
•Nvidia has not raised system pricing despite demand far exceeding supply; the article sees room for a ~40% increase in VR NVL72 server pricing while still offering below-trend cost/FLOP improvements.
•TSMC's N3 capacity is at >100% utilization but wafer pricing has not increased; the author argues TSMC is 'leaving value on the table' and could raise prices or demand larger prepayments.
•DRAM/LPDDR5X pricing has surged 6x over the past year; SOCAMM memory in Rubin systems allows Nvidia to explicitly price memory at ~60% gross margin, a new pricing lever.
•Cost-based GPU rental pricing for VR NVL72 implies a floor of ~$4.92/hr/GPU, while value-based pricing (via $/FLOP parity) implies a ceiling of ~$12.25/hr/GPU, indicating substantial upside for Neoclouds and Nvidia.
•The 'One Chart to Rule Them All' framework shows that at current Nvidia system pricing, Neoclouds can earn 15% IRR while offering a 60% drop in cost/FLOP vs GB300, suggesting Nvidia has significant pricing power to capture more value.
Article notes DRAM pricing has surged 6x in the past year and is likely to remain elevated due to memory supply tightness (90%+ fab utilization). SOCAMM contract pricing is projected to exceed $13/GB
Article notes DRAM pricing has surged 6x in the past year and is likely to remain elevated due to memory supply tightness (90%+ fab utilization). SOCAMM contract pricing is projected to exceed $13/GB by end of 2026. Memory is the primary cost driver in Rubin systems, benefiting memory vendors like Micron.
Risk: Memory demand is cyclical; if AI capex slows or alternative memory technologies (HBM next-gen) appear, pricing could peak.
Article highlights TSMC's N3 capacity is the tightest bottleneck (>100% utilization) but pricing has not increased. The author calls this a 'strategic error' and says TSMC could raise prices and custo
Article highlights TSMC's N3 capacity is the tightest bottleneck (>100% utilization) but pricing has not increased. The author calls this a 'strategic error' and says TSMC could raise prices and customers would accept it. Any eventual repricing would directly boost TSMC's revenue and margins.
Risk: TSMC's long-term relationship-focused strategy may keep it from fully pricing to scarcity; geopolitical risks could disrupt wafer supply.
Article explicitly states Nvidia has not raised system pricing despite structural demand/supply imbalance and that there is room for a ~40% increase in VR NVL72 server pricing while still offering bel
Article explicitly states Nvidia has not raised system pricing despite structural demand/supply imbalance and that there is room for a ~40% increase in VR NVL72 server pricing while still offering below-trend cost/FLOP improvements. Nvidia also has a new pricing lever via SOCAMM memory margin (~60%). Rising GPU rental floors and value-based ceilings support margin expansion.
Risk: Regulatory scrutiny (antitrust) or customer pushback could delay repricing; alternative compute (TPU, Trainium) may gain share if Nvidia prices too aggressively.
This newsletter, published May 01, 2026,
features Daniel Nishball
discussing MU, TSM, NVDA.
3 trade ideas extracted by AI with direction and confidence scoring.