Amazon's Low-Cost AI Chips Attract Enterprise Clients, Priced ~80% Below NVIDIA H100
N.R. Finch
Amazon's custom AI chips run inference workloads at roughly 80% less than Nvidia's H100 — drawing enterprise buyers toward multi-vendor strategies, though software maturity and on-premises readiness remain key bottlenecks.
80% cheaper — where does the saving come from?
Per The Information, Amazon's Inferentia2 and second-generation Trainium cut inference costs — the work of running a trained AI model to answer queries or generate content — by roughly 80% versus Nvidia's H100.
This means → running the same AI customer-service bot or risk model could cost a business one-fifth of the Nvidia price tag on Amazon silicon.
In plain terms = training AI is like building a car; inference is like driving it. Amazon is offering the daily fuel bill at an 80% discount.
Who is already testing these chips?
Co Driver Labs consultant Karol Piatek said an insurance client — required by regulators to run AI in its own data center — tested Amazon chips after using Nvidia and got positive results.
The insurer then asked AWS whether it could deploy Inferentia via Outposts — Amazon's product for placing cloud hardware inside a customer's facility — but was told the chip is not yet ready for Outposts testing.
Piatek: "Even if the cost were double the cloud price, they'd still use AWS chips on-premises." This reflects a compliance-driven demand for local deployment that outweighs price sensitivity.
Is Amazon pushing chips beyond the cloud?
AWS sales staff have held exploratory talks with AT&T and other clients about leasing AI chips through Outposts; Amazon has also discussed selling chips outright for customer-owned data centers.
After CEO Andy Jassy floated the idea of selling chips outside AWS in his April shareholder letter, client interest rose noticeably.
The "AWS AI Factories" product launched last December already lets customers co-deploy Trainium alongside Nvidia chips on-premises; Jassy disclosed Trainium's backlog stands at roughly $225 billion.
Why aren't companies sticking with Nvidia alone?
Cognizant's head of cloud and infrastructure, Sriram Kumaresan, said: "AI demand is growing faster than available compute, driving strong interest in alternatives to a single-vendor GPU strategy."
He stressed that Trainium and Inferentia are seen as "trusted, production-grade options" — not stop-gaps for GPU shortages.
This means → the motivation is not just cost. Companies fear they cannot secure enough compute — and that supply crunch is cracking Nvidia's monopoly grip.
What is still missing?
Rhythmic Technologies CEO Cris Daniluk said his clients have discussed alternatives but remain at a very early exploratory stage.
The core concern: Amazon chips require a new software stack, and enterprises worry about migration risk and ecosystem maturity.
In plain terms = the hardware is 80% cheaper, but the supporting software hasn't caught up — like buying a discounted printer when you're not sure the ink cartridges are easy to find.
Content is for reference only, not financial advice.