Babbage Index · Research · 002

US vs China hyperscaler CPU pricing

Intel versus AMD, cents per vCPU-hour, on a one-year commit basis for 4-vCPU / 8 GB compute instances. The rebased view strips out on-demand noise and puts US hyperscalers, Oracle, and domestic Chinese CSPs on the same procurement-style axis.

Babbage Research · 002·Published 2026-06-21·1-year commit basis
Chip Wins
AMD 5/7
Intel is cheaper in 2 provider rows
China Avg
1.20¢
60% below the US/Oracle average on this basis
Intel Avg
2.29¢
Across all seven provider rows
AMD Avg
2.17¢
Across all seven provider rows
ProviderIntel
¢/vCPU-hr · instance · chip
AMD
¢/vCPU-hr · instance · chip
CheaperGap
AWS
3.74¢
c8i
Granite Rapids
4.32¢*
c8a
Turin
Intel13.5%
GCP
2.68¢
C4
Granite Rapids
2.50¢
C4D
Turin
AMD6.6%
Azure
3.69¢
Dlsv7
Granite Rapids
2.54¢
Dalsv7
Turin
AMD31.1%
Oracle
2.30¢
Std · X12
Granite Rapids
2.28¢
Std · E6
Turin
AMD1.0%
Volcano
1.04¢
ecs.c4i
Granite Rapids
1.09¢
ecs.c4al
Turin
Intel4.6%
Tencent
1.08¢
S9e
Granite Rapids
0.98¢
SA9e
Turin-Classic
AMD9.0%
Alibaba
1.52¢
c9i
Granite Rapids
1.47¢
c9a
Turin
AMD3.3%

Methodology

Prices are normalized to cents per vCPU-hour for 4-vCPU / 8 GB compute instances on a one-year commit basis. AWS uses a 20% Compute Savings Plan adjustment, GCP and Azure use a 37% committed-use / reserved adjustment, Oracle uses a 20% annual universal-credits adjustment, and the Chinese CSP rows use committed quotes captured on 21 June 2026.

On that basis, the China rows cluster around 1.0-1.5¢/vCPU-hour. Tencent SA9e is the low row at 0.98¢, Volcano's Intel row is 1.04¢, and Alibaba c9i/c9a land at 1.52¢ and 1.47¢. The US and Oracle rows sit higher, from Oracle's 2.28-2.30¢ up to AWS AMD at 4.32¢. The Chinese rows are a static research snapshot, not a daily live feed.

Intel versus AMD

AMD is cheaper in five of seven rows. The largest spread is Azure, where Dalsv7 Turin is 31.1% below Dlsv7 Granite Rapids. AWS is the main exception: c8i is 13.5% below c8a on this price basis. Volcano is the other Intel win, with ecs.c4i 4.6% below ecs.c4al.

The AWS row needs a footnote. AWS c8a runs AMD Turin with SMT off, so a vCPU maps to a full physical core. Its per-vCPU rate is not directly comparable to the SMT-on Intel and AMD rows. Price alone makes AWS Intel look cheaper; performance-normalized work may tell a different story.

Spot overlay · AWS/GCP/Azure · current cohort

The committed rows above answer what a buyer pays after a one-year commitment. Spot pricing answers a different question: how much of a discount appears when schedulable capacity is bought from live, interruptible pools. The overlay keeps those two pricing channels separate rather than mixing spot rates into the committed comparison.

Live BI-C overlay
BI-C · Full
—
on-demand · May 10 = 1000
BI-C · Spot
—
spot signal · May 10 = 1000
Spot discount
—
raw basket vs on-demand
BI-C Spot replaces the AWS, GCP, and Azure portions of the on-demand basket with live spot or preemptible quotes; Oracle remains on-demand. The on-demand and spot signals are independently normalized to 1000 on 10 May, so their index-level gap is not the discount. The displayed discount compares the unnormalized spot and on-demand baskets directly. Regional spread is (max/min - 1) × 100 across m8i.xlarge spot in five AWS regions. This live overlay is separate from the static Chinese quote snapshot used above.

The useful signal

The cleanest cross-border signal is not a single "US versus China" average. It is the gap between procurement channels. Committed Chinese capacity is materially below the rebased US hyperscaler and Oracle rows, while the within-provider Intel/AMD spread is usually smaller than the cross-market spread.

For capacity planning, that means the first-order decision is access: can the workload actually use the domestic Chinese committed channel? If yes, chip selection is a second-order optimization. If no, the apparent discount is not a usable price.

Basis: 1-year commit, 4 vCPU / 8 GB. Gaps use unrounded rates.
BI-CPrevious note