Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Data governance - retention, ZDR and training
The AI Engine terms commit Sakura in writing not to use inputs, not to give them to third parties (model providers included) and not to train on them. They do not promise zero retention.
Article 10(1) (第10条第1項) of the AI Engine service terms reads, in OneHill's translation, that inputs and outputs are "in principle not stored" (原則として保存されません). They may be stored "for a certain period" (一定期間) in Sakura's facilities for maintenance or fault handling, are then kept encrypted, and are deleted after that period and not stored permanently. The "certain period" is not defined in any of the three terms documents.
Article 10(2) reads: Sakura does not use input information and does not provide it to third parties, "including, but not limited to, model providers". Input information is also not used for AI training by Sakura or the model provider. The sentence covers inputs. Outputs are not named in it.
Translation uncertainty: 恒常的には is read as "not on a standing basis"; 一定期間 could also be read as "a fixed period", which still leaves the length open.
The product page, FAQ and Private Edition page contradict the terms. They say input and response data are "not stored at all", that prompts are "not cached", and that the service is "ZDR (Zero Data Retention)". Article 10(1) permits temporary encrypted storage for an undefined period, so zero retention is a marketing claim and not a contract term.
The manual adds that documents uploaded to the RAG feature are kept until the customer deletes them, a case the terms do not address. For Private Edition, article 9(4) says Sakura does not handle input and output "in principle", with exceptions for fault investigation, support the user asks for, and legal or other justifiable reasons (正当な理由, open-ended).
Data and IP ownership
Nothing in the terms takes a licence over inputs or outputs. Article 10(3) says outputs may be used freely within each model provider's licence. The two closed models narrow that. The PLaMo licence bars entering personal information (Art. 4(7)), fine-tuning, and using outputs to train other language models (Art. 3(2)). The cotomi conditions bar distillation and derived models (Art. 2(8), 2(9)).
There is no mutual confidentiality clause naming customer data. What exists is the express duty in article 10(2) (inputs only) and the statutory secrecy of communications in basic terms article 24(1). Article 24(2) allows disclosure to courts and regulators. Article 24(3) allows disclosure of information within a user's secrecy of communications to third parties where the user commits a prohibited act.
Basic terms article 17(2) says Sakura does not involve itself in user data and is not liable for its leakage or loss, subject to article 29. Under Private Edition, rights in the customer's model stay with the customer (article 8(4)), and Sakura may not use the model to train its own or anyone else's AI (article 9(2)).
Residency and jurisdiction
Japan-only processing is claimed but not contracted. The top-page FAQ says model execution, communication and storage are completed entirely on cloud infrastructure in Japan, and the manual says requests are processed in domestic data centres. None of the three terms documents contains a location clause.
Basic terms article 2(5) (in OneHill's reading) says websites designated in the terms are not part of the terms unless expressly stated. No data centre is named for AI Engine, and there is no region choice in the API.
The contract is under Japanese law (article 32), with the Tokyo District Court as the exclusive first-instance court for the customer's claims. Sakura may choose arbitration in Tokyo before one arbitrator it selects, in Japanese (article 33(2)). Sakura is a Japanese company, so the US CLOUD Act does not apply to it as a US company.
The terms do import US export control: users on the Japanese METI Foreign End User List, the US BIS lists or from embargoed countries are barred (AI Engine terms articles 3 and 12). Applicants must live in a "designated country" (basic terms article 5(3)), and the list is not published.
Compliance and attestations
Sakura holds independently audited certifications, but none of the scopes names AI Engine. ISMS ISO/IEC 27001 (ICMS-SR0063) is a company-level certificate valid to 2027-04-12. The ISMS cloud security certificate (ISO 27017 based), the ISO/IEC 27018 certificate and the ISMAP registration (C21-0030-2) all name Sakura Cloud (Tokyo and Ishikari regions). PCI DSS covers physical requirements 9 and 12 only.
A public SOC 3 report from an independent auditor (security and availability, 2025-06-01 to 2026-05-31, opinion dated 2026-07-31) covers the Ishikari data centre, not AI Engine. SOC 2 Type 2 (data centre) and SOC 2 Type 1 (Sakura Cloud) reports are disclosed on request and are unverified.
No DPA, GDPR addendum or sub-processor list has been identified. The privacy policy concerns personal information Sakura collects about its own customers and staff and says nothing about AI Engine inputs. The check sheet (updated 2026-11-13) says billing, invoicing and customer support are outsourced, without naming contractors. The transparency report omits AI Engine. Compliance is therefore scored on what ties to this service and not on the company's full certificate list.
Security controls
Documented for AI Engine: temporarily kept inputs and outputs are encrypted (article 10(1)); access is by account token; the Sakura Cloud service endpoint gateway gives a private route to the API; service policies control resources per organisation.
Not documented for AI Engine: tenant isolation, a penetration-test statement, a bug-bounty or vulnerability-disclosure policy, and an incident history. Sakura's cloud security policy describes full virtualisation for customer-created servers, which is a different product. Private Edition dedicates the named GPU platform, but article 2(2) says VM, storage, API and network exclusivity is not guaranteed unless agreed.
Pricing and cost model
Pricing is pay-as-you-go in yen, tax included, per 10,000 tokens (gpt-oss-120b: 0.15 yen input, 0.75 yen output), per 60 seconds of audio, per 10,000 mora of speech and per 100 RAG chunks. A free foundation-model plan gives a monthly request allowance (3,000 chat requests).
Payment is by credit card unless Sakura approves otherwise (article 6), billed in yen on the 10th of the next month. RAG storage bills until every document is deleted. Private Edition is a flat monthly fee per GPU on quote with a two-month or one-year initial term. Prices are not compared with other providers.
Reliability posture
AI Engine is outside Sakura's SLA, per the manual ("品質保証(SLA)の適用対象外です"). Private Edition article 11 gives best effort unless an SLA is written into the service site, order form or another document.
The only remedy is basic terms article 29(3). If a continuously provided service is wholly unusable for 24 hours or more for a reason attributable to Sakura, the user may request compensation of actual ordinary damage. It is capped at one thirtieth of one month's basic fee per whole day of outage, or a term extension applies when the sum is under 10,000 yen. How that applies to a usage-billed service is not stated.
Sakura's public status page lists Sakura Cloud components and not AI Engine, so no uptime history is published.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you use it freely and leave without lock-in?
ModerateMostly Apache-2.0 and MIT models behind OpenAI and Anthropic compatible APIs, but no fine-tuning on the shared service, two closed models with no-fine-tune and no-distillation terms, a military-use ban (Art. 12(1)(vi)), and a dedicated option for Japanese corporations only (Private Edition Art. 3).
TransparencyAre the binding terms published, legible and independently checkable?
ModerateThree binding terms documents are published, but they are Japanese only, there is no DPA or sub-processor list, certification scopes do not name AI Engine, and marketing claims (ZDR, Japan-only) are stronger than the terms.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.