Nvidia is a victim of the compute marketplace it created

Nvidia shares are down 15% from their May peak even as projected revenue continues to grow, with the stock now trading cheaper than the S&P average on an earnings multiple basis. The article argues the compute “price” is sliding—spot time on an Nvidia H100 peaks around ~$3.20/hour in May and then declines—while DRAM spot pricing has risen dramatically (reported up to ~10x over the past year) as data centers face a memory bottleneck. Memory suppliers like Micron have nearly tripled in value, benefiting from demand outpacing supply and from fewer remaining constraints as the earlier GPU shortage eases.

Analysis

The market is starting to price AI infrastructure less as a pure compute monopoly and more as a bandwidth bottleneck. That shifts incremental bargaining power away from accelerators and toward the parts of the stack that are hardest to substitute quickly, which is why memory names can keep rerating even if GPU demand stays healthy. For hyperscalers like AMZN, GOOGL, and MSFT, cheaper accelerator pricing is offset by structurally higher memory intensity per rack, so capex does not disappear — it becomes more vendor-diverse and more capital hungry.

The near-term risk for NVDA is not an earnings collapse but multiple compression if spot compute keeps drifting lower while consensus still assumes scarcity economics. The key catalyst path is 1-3 months: hyperscaler capex commentary, HBM allocation updates, and any sign that custom silicon is good enough for internal workloads and inference. If that substitution continues, NVDA’s growth remains strong but the market may stop paying peak scarcity multiples.

Contrarianly, the current move may be underestimating how cyclical memory is. The same supply tightness that is supporting MU and peers can reverse quickly once capacity additions land, so the upside is probably strongest over the next 2-4 quarters, not 2-4 years. What would falsify the bearish NVDA view is a re-acceleration in GPU pricing or a new platform cycle that restores pricing power faster than custom ASIC adoption erodes it; what would falsify the memory bullishness is a sharp roll-over in HBM spot pricing or an aggressive supply response from SK hynix/Samsung/Micron.

Key Decisions for Investors

• Trade the relative-value spread: long MU / short NVDA for 1-3 months. Thesis: the market is overpaying for scarce memory economics while underpricing accelerator commoditization; risk/reward is attractive as long as HBM pricing stays firm and NVDA spot compute keeps easing.
• For a cleaner sector expression, go long SMH or SOXX and pair it against an NVDA underweight rather than a naked short. This captures the rotation into semis while limiting single-name idiosyncratic risk if NVDA prints a strong quarter.