What would one Sovereign GaiXer-style system cost in the cloud?
Recreating an all-in-one GPU system on three clouds: a Tokyo-region monthly cost estimate
Key takeaway
Recreating the article's single-system workload on one cloud GPU virtual machine costs roughly JPY 400,000–1,200,000 per month when running around the clock. Most of this is GPU cost. The difference comes from the cloud, GPU, and whether you commit for one year. The Japan reference upfront product price used here is JPY 3,806,400 before tax (JPY 4,187,040 including Japanese tax), which the lowest-cost cloud scenario exceeds after about 10 months. This is a workload cost illustration, not a claim of identical hardware or performance.
Renting in the cloud starts at JPY 400,000 per month.
The upfront purchase reference is:
3,806,400 JPYOne-time Japan reference price; no recurring token fee (JPY 4,187,040 including Japanese tax)
The upfront Sovereign GaiXer price in this comparison equals only 10 months of the selected cloud scenario (GCP with a one-year commitment). Azure exceeds the purchase price after a little over four months. Cloud rental payments do not leave you owning the hardware; after the break-even point, purchased hardware has no GPU rental charge. Over three years, the illustrated difference can approach JPY 40 million. The cited Japan offering includes a three-year (36-month) hardware warranty. International availability, warranty coverage, and commercial terms require confirmation.
An all-in-one system: place the workload on one cloud VM too
One Sovereign GaiXer unit contains the application, the LLM, and the database. This cloud scenario likewise puts everything on one GPU virtual machine, plus storage, without separate web, database, or load-balancer services.
The application stack
The web interface (Next.js), API (FastAPI), agent execution, and authentication (Keycloak). This is the lighter part, requiring a reasonable amount of CPU and memory.
AI inference (LLMs)
The estimate keeps three models resident through vLLM, requiring approximately 66 GB of GPU memory. This is the expensive part, accounting for more than 95% of the illustrated cost.
Data
PostgreSQL for history and vector search, plus S3-compatible attachment storage. The scenario assumes a 1 TB disk is sufficient.
Think of moving all your furniture into one studio apartment. A 300-user setup might separate the kitchen (database) and lobby (load balancer), but a single-system setup puts everything in one room, one VM. Since the rent is mostly GPU cost, the GPU you rent determines most of the bill.
Each cloud offers different GPU rooms to rent
The requirements in this scenario are at least 66 GB of GPU memory and a Tokyo region. The qualifying VMs differ across the three providers, and those differences feed directly into the monthly total. Figures are from July 12, 2026, converted at USD 1 = JPY 160.
NC40ads H100 v5
One H100 NVL (94 GB). The source scenario treats this as the closest fit to its inference configuration, with minimal configuration changes. Approximately JPY 1,619/hour, or JPY 1,214/hour with a one-year reservation.
p5.4xlarge
One H100 (80 GB). It meets the 66 GB requirement but needs memory-allocation adjustments. Approximately JPY 1,376/hour. The one-year discounted rate requires a quote because a published rate was not available.
g2-standard-48
Four L4 GPUs (24 GB each), 96 GB total. Lowest cost of the three in this comparison: approximately JPY 822/hour, or JPY 518/hour with a one-year commitment. However, larger models require reconfiguration across four GPUs, with more modest speed than the H100 configurations.
Monthly cost comparison across three clouds
These totals cover a GPU VM, 1 TB disk, and network costs, running 24/7 (730 hours per month). On-demand pricing charges for usage at the listed rate; a one-year commitment provides a discount in exchange for committed usage.
| Item | Azure | AWS | GCP |
|---|---|---|---|
| GPU virtual machine(one VM; 730 hours) | JPY 1,182,000$7,388 | JPY 1,004,000$6,278 | JPY 600,000$3,752 |
| 1 TB disk, network, and other costs | JPY 22,000$140 | JPY 18,000$111 | JPY 21,000$132 |
| Total (on demand) | Approx. JPY 1,200,000$7,528 | Approx. JPY 1,020,000$6,389 | Approx. JPY 620,000$3,883 |
| Total (one-year commitment) | Approx. JPY 910,000$5,681 | Quote requiredReference: approx. JPY 1,140,000* | Approx. JPY 400,000$2,496 |
Azure
- GPU virtual machine(one VM; 730 hours)
- JPY 1,182,000$7,388
- 1 TB disk, network, and other costs
- JPY 22,000$140
- Total (on demand)
- Approx. JPY 1,200,000$7,528
- Total (one-year commitment)
- Approx. JPY 910,000$5,681
AWS
- GPU virtual machine(one VM; 730 hours)
- JPY 1,004,000$6,278
- 1 TB disk, network, and other costs
- JPY 18,000$111
- Total (on demand)
- Approx. JPY 1,020,000$6,389
- Total (one-year commitment)
- Quote requiredReference: approx. JPY 1,140,000*
GCP
- GPU virtual machine(one VM; 730 hours)
- JPY 600,000$3,752
- 1 TB disk, network, and other costs
- JPY 21,000$132
- Total (on demand)
- Approx. JPY 620,000$3,883
- Total (one-year commitment)
- Approx. JPY 400,000$2,496
Like apartment rent, a one-year commitment lowers the illustrated monthly price by roughly 25–35%: Azure from JPY 1.20 million to JPY 910,000, and GCP from JPY 620,000 to JPY 400,000. Short trials suit on-demand pricing; an annual commitment can make sense once continued use is clear. For weekday-only on-demand usage, deallocating resources outside working hours can reduce GPU charges to around a quarter, roughly JPY 150,000/month for GCP. Commitment and resource billing rules still need checking.
Frequently asked questions
Which is the lowest-cost configuration in this comparison?
GCP g2-standard-48 with a one-year commitment is approximately JPY 400,000/month. Models need reconfiguration across four GPUs, and response speed is more modest than the H100 scenarios.
Which configuration most closely fits the source workload?
The source estimate favors Microsoft Azure NC40ads H100 v5 (94 GB) because it requires few configuration changes for its inference workload. A one-year commitment is approximately JPY 910,000/month. This does not establish identical performance to the GB10 product hardware.
Why is the range as wide as JPY 400,000–1,200,000?
More than 95% of the illustrated cost is GPU-related, determined by GPU type (H100 or 4 × L4) and contract type (on demand or annual commitment). Disk and network costs stay around JPY 20,000/month.
Can stopping it when unused reduce the bill?
Yes, for on-demand compute when the resource is properly stopped or deallocated under the provider's billing rules. Weekday daytime use (180 hours/month) cuts GPU hours to about one quarter: approximately JPY 150,000 for GCP, JPY 250,000 for AWS, or JPY 290,000 for Azure. Storage can still incur charges, and committed charges continue under their terms. Choose the contract for your actual operating pattern.
Which offers better value: cloud or purchase?
For continuous 24-hour use, the JPY 3,806,400 purchase price is much lower in this illustration. The selected cloud scenario exceeds it in about 9.5 months, Azure in about 4.2 months. Cloud suits trials lasting weeks or months, situations without installation space, or temporary capacity increases. Cloud for short trials and purchase for ongoing production can be a reasonable approach; compare full operating costs and performance before deciding.
Does cloud hosting lose the benefit of keeping data in-house?
The LLM and database can stay inside a dedicated cloud environment without passing data to an external AI service. The physical location is nevertheless a cloud provider's data center. It is therefore more accurate to describe this as an option when installing hardware is difficult, or as a test environment, rather than as physically on-premises.
Based on the official pricing pages of OpenAI and Google AI, as of June 2026. Prices, model names, and specifications can change; check current official information before deciding. Estimates use USD 1 = JPY 160 and approximately 1.5 tokens per Japanese character. They are illustrative Japan-market references, not quotations or estimates for English-language text. This series does not recommend any particular service. Contact sales to confirm current product pricing, delivery, warranty, and support for your country.
Company, product, and service names mentioned are trademarks or registered trademarks of their respective owners.