Nvidia stock has fallen 15% since its peak in May 2026, and by one measure the company now trades cheaper than the S&P average — investors are paying less per dollar of projected Nvidia profit than they do for the typical large American company. Revenue forecasts are still climbing. What has broken is the assumption that GPU scarcity would keep compressing margins upward forever. The GPU shortage that dominated 2025 headlines has eased, and the bottleneck moved down the bill of materials.
The new bottleneck is memory. Micron, one of the world's largest makers of DRAM, has nearly tripled in value over the same window Nvidia has declined. DRAM spot prices have risen roughly 10x over the past year, according to Datatrack data going back to 2023, and high-bandwidth memory used in AI accelerators has become the piece of the stack buyers cannot get enough of. Money is still pouring into AI infrastructure equities — it is just routing around Nvidia and into the memory suppliers.
The compute side of the ledger tells the mirror-image story. An hour of time on an Nvidia H100 GPU peaked around $3.20 in May on the compute marketplace Ornn, and has fallen steadily since. Nvidia's stock chart and the H100 spot-price chart trace nearly the same arc. When compute gets cheaper, the company whose margins depend on selling compute gets cheaper too.
Key facts
- 01Nvidia stock has fallen 15% since its May 2026 peak even as projected revenue keeps growing.
- 02H100 spot compute peaked near $3.20 per hour in May and has declined steadily since.
- 03DRAM spot prices have risen roughly 10x over the past year, making memory the new AI bottleneck.
- 04Micron has nearly tripled in value over the same period Nvidia declined.
- 05Google, Amazon, Microsoft, and OpenAI have all launched custom processors to reduce Nvidia dependence.
The technical asymmetry is worth naming. Nvidia's rise rests on genuinely hard engineering — CUDA as the default programming layer for AI research, and a GPU roadmap running at a cadence few competitors have matched. The H100 and its successors are among the most complex commercial chips ever fabricated. DRAM, by contrast, has been getting incrementally better for 20 years, without dramatic architectural shifts. The memory companies did not invent anything new in the summer of 2025. The industry simply underestimated how much memory the data-center buildout would consume, and prices adjusted.
Nvidia's problem is that its own success created the competition on the compute side. Google, Amazon, Microsoft, and OpenAI have all launched custom silicon to reduce their dependence on Nvidia. None of those chips need to beat the latest H-series or B-series part outright — they only need to be good enough to bleed off marginal demand and cap the price of GPU-hours. That is exactly what has happened.
“More GPU and accelerator players are entering the market. Everyone wants to make their own silicon, but no one is making their own DRAM.”— Wayne Nelms, Ornn co-founder and CTO
Ornn co-founder and CTO Wayne Nelms framed the split as straightforward supply and demand. Compute has a growing roster of alternate suppliers; memory does not. Every major cloud is building an accelerator. None of them are building DRAM.
The memory side has no equivalent pressure valve. HBM production is concentrated in a small number of fabs, capacity additions take years, and there is no hyperscaler-funded internal HBM program equivalent to what Google, Amazon, and Microsoft have built for compute. The result is that memory suppliers get to keep raising prices while GPU vendors have to defend theirs.
Nelms expects the pattern to hold until the underlying supply-and-demand picture changes. Nothing on the public roadmap suggests that shift is imminent.
For Nvidia, the near-term math is uncomfortable rather than dire. Revenue is still growing, gross margins on the highest-end parts are still healthy, and the CUDA moat has not eroded — most of the custom silicon coming out of the hyperscalers still runs alongside Nvidia inside the same data centers, not instead of it. What has changed is the ceiling. When compute is scarce, the company selling compute captures the scarcity premium. When compute is abundant and memory is scarce, that premium moves upstream.
The broader signal for the AI market is that the trade is rotating from the picks-and-shovels layer everyone recognized in 2023 and 2024 into a less glamorous tier — the DRAM and HBM suppliers, the packaging specialists, the cooling and interconnect vendors. That rotation says something about where the marginal dollar of AI capex is actually going. It is not going into buying more GPUs at any price. It is going into feeding the GPUs already installed, and the companies making the feed are the ones getting rich.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
