Nvidia · one lens
Hamilton Helmer on Nvidia
OpenAI’s Jalapeño benchmarks attack Nvidia’s performance edge, but they do not yet show that a customer has copied Nvidia’s scale economics or paid the full cost of leaving its platform.
Nvidia still clears the Benefit-plus-Barrier test from 7 Powers most convincingly in Scale Economies and Switching Costs, but OpenAI now supplies a harder field test. OpenAI said Jalapeño delivered 1.5x–1.9x as much AI work per watt as Nvidia chips and 1.7x–3.6x lower latency than Nvidia chips across the named benchmark set (TEXXR 1239748). SemiAnalysis reported that OpenAI and Broadcom developed the ASIC over 16 months from project start to the reported chip (TEXXR 1239755). OpenAI therefore attacks Nvidia’s Superior and Significant performance Benefit, but OpenAI has not established a Sustainable Barrier of its own.
Nvidia’s Scale Economies status remains held. Nvidia’s prior record grounds the Benefit in spreading chip-design and CUDA fixed costs over the industry’s largest unit volume, while Helmer’s Barrier lies in the unattractive economics of matching that volume. OpenAI’s benchmark reports do not disclose unit cost, shipped volume, pricing, or challenger margin. OpenAI has shown a performance route around Nvidia, not equivalent scale economics.
Nvidia also retains Switching Costs, now under clearer pressure. OpenAI and Broadcom built a custom accelerator rather than buying an interchangeable substitute. OpenAI’s build supports the Barrier because a customer funded a bespoke escape route. OpenAI’s benchmark does not show workload migration, production deployment, software retraining, or total transition cost, so the record cannot yet show that the separate cost of leaving has disappeared.
Nvidia said its Groq inference accelerator entered full production with Nebius as launch customer (TEXXR 1239690). Nvidia’s launch shows operational response, but Helmer separates operational excellence from Process Power. The record does not show embedded routines built over decades that rivals cannot copy.
Nvidia’s Cornered Resource case remains contested because the prior record describes leading-edge TSMC capacity as shared rather than exclusive. The fresh reports do not establish preferential access that rivals cannot buy. The record still does not argue Network Economies, Counter-Positioning, or Branding.
The TEXXR coverage record logged 162 Nvidia articles in completed 2026Q2, 49 above the trailing eight-quarter average of 113. The DeadRisk market layer recorded an August 26 close of $211.75 per share, compared with the prior close of $213.05. The supplied record provides an August 26 earnings marker but no filed earnings result, so Nvidia’s Power call cannot rest on that event yet.
The seven Powers, one by one
| Power | Status | Why |
|---|---|---|
| Scale Economies | held | the prior record supports fixed-cost advantages across chip design and CUDA, while Jalapeño benchmarks show performance rather than equivalent unit economics or volume |
| Network Economies | absent | not argued in the record |
| Counter Positioning | absent | not argued in the record |
| Switching Costs | held | OpenAI funded a custom accelerator route, but the record does not show workload migration, retraining cost, or broad production displacement |
| Branding | absent | not argued in the record |
| Cornered Resource | contested | leading-edge TSMC capacity remains shared rather than exclusive, and the fresh reports do not establish preferential access unavailable to rivals |
| Process Power | absent | not argued in the record |
An editorial application of Hamilton Helmer's published framework — not their actual view, affiliation, or advice.
Where this lens comes from
Hamilton Helmer's framework is set out at the lens page, drawn from 7 Powers. This page applies it to Nvidia and nothing else — the company's full record is in the dossier.