Deployment & demand · United States of America
Amazon Web Services
AMZN USvia Amazon Work there, own it, buy from it ↓
Hyperscale compute buyer and custom silicon. Holds none of the How this is computedCount of rows in the bottleneck corpus: every material plus every non-material chokepoint.Method →Code → bottlenecks tracked here; listed as context for its step of the chain.
Why graded so · Buys at a scale that moves supply, and designs around it where it can.
Role cited from aws.amazon.com ↗. The grade is this project’s judgement. Report an error
01What it rests on, and what rests on it
Depends on
- High-bandwidth memory stacking yield
“144 GB HBM3e per chip, 4.9 TB/s bandwidth - 1.7x higher than Trainium2.”
aws.amazon.com · primary source · AWS’s own Trainium3 chip. Names HBM3e, not its supplier.
- Leading-edge foundry capacity
“Amazon EC2 Trn3 UltraServers powered by AWS's first 3nm AI chip help organizations of all sizes run their most ambitious AI training and inference workloads.”
aboutamazon.com · primary source · Names a 3 nm chip, which only leading-edge foundries make; does not name the foundry.
A recorded dependency is a sentence, not a share, a volume or a contract. Walk it on a portfolio →
02SEC filings
- SEC 8-K14 SepOther events ↗
- SEC 8-K9 SepDirector or officer change ↗
- SEC 10-Q30 JulQuarterly report ↗
- SEC 8-K30 JulResults of operations ↗
- SEC 8-K9 JulOther events ↗
The company’s own filings, from SEC EDGAR, fetched hourly. Substrata has not judged their effect on any bottleneck.
03Others at the same step of the chain
- MicrosoftHyperscale compute buyerConcentrated
- GoogleHyperscale compute buyer and custom siliconConcentrated
- MetaHyperscale compute buyerConcentrated
- OpenAIFrontier model developerConcentrated
- AnthropicFrontier model developerConcentrated
- xAIFrontier model developerConcentrated
- CoreWeaveSpecialist compute providerCompetitive
04Ways to take part
Not investment advice. This says how the public can reach a company — a job, its shares, its products — not whether anyone should. It takes no account of anyone’s circumstances, and a company holding a chokepoint is not a reason to own it. What this is not
- Work there
- It publishes its roles on its own careers page.Official careers page ↗
- Own part of it
- Not listed on its own: it is part of Amazon, whose shares trade publicly as AMZN US. Owning those is owning the whole group, of which this is one part.AMZN US listing record ↗Its SEC filings
- Buy from it
- What it does in the chain: Hyperscale compute buyer and custom silicon. At this depth of the chain most sales are to other businesses, under contract. Its own page describes what it offers.Its own page ↗
05Hiring
Its roles are published on its own site, which Substrata links rather than reads. Official careers page ↗
06What this holds up
- Software scaffoldone turn: Minutes
depends on High-bandwidth memory stacking yield, which gates it.
Worst judgement on this loop: How this is computedThe highest severity score among the bottlenecks recorded as gates on the loop.Method →Code →, judged 2026-09-15 · this join is primary source
- Frontier training runone turn: Weeks to months
depends on High-bandwidth memory stacking yield, which gates it.
Worst judgement on this loop: How this is computedThe highest severity score among the bottlenecks recorded as gates on the loop.Method →Code →, judged 2026-09-15 · this join is primary source
- Software scaffoldone turn: Minutes
depends on Leading-edge foundry capacity, which gates it.
Worst judgement on this loop: How this is computedThe highest severity score among the bottlenecks recorded as gates on the loop.Method →Code →, judged 2026-09-15 · this join is primary source
- Firmware and programmable logicone turn: Minutes to hours
depends on Leading-edge foundry capacity, which gates it.
Worst judgement on this loop: How this is computedThe highest severity score among the bottlenecks recorded as gates on the loop.Method →Code →, judged 2026-09-15 · this join is primary source
- Frontier training runone turn: Weeks to months
depends on Leading-edge foundry capacity, which gates it.
Worst judgement on this loop: How this is computedThe highest severity score among the bottlenecks recorded as gates on the loop.Method →Code →, judged 2026-09-15 · this join is primary source
A loop is one design, build and measure turn. Severity is a dated judgement on a 0–12 scale, not a measurement, and the judgements are not added together — the figure shown is the worst single one.
07Sources
- aws.amazon.com ↗ — its role in the chain: Hyperscale compute buyer and custom silicon
Discussion
Comments use threadkit (permission is participation). Sign in with OrangeCat to write. Fact-check is the same assistant as Ask, grounded in the corpus.
- No comments yet.