| name | gpu-customer-usage-conversion |
| description | Analyze whether GPU resources in a DCE environment have converted into real customer usage and billable consumption. Use when a user asks "当前环境的 GPU 资源有没有真正转化成客户用量", "哪些 GPU 只是有资源没客户用", "GPU 算力有没有变现/产生账单", "GPU 资源利用和客户 Token 用量能不能闭合", "GPU customer usage conversion", "GPU monetization", or similar questions about connecting GPU supply to workspaces/customers, LLM/API usage, runtime activity, and Billing Center revenue evidence.
|
GPU Customer Usage Conversion
Judge whether deployed GPU capacity has become real customer usage by joining
DCE evidence across GPU supply, workspace/customer carrier, LLM/API activity,
runtime workload activity, and Billing Center billing.
REQUIRED SUB-SKILL: Use dce for command discovery, auth checks, module
availability checks, and read-only command execution.
SCOPE: Read-only analysis and recommendations. Do not create, update,
delete, scale, reprice, recalculate bills, change quota, or change routing.
Core Rule
Do not treat "GPU exists" as customer usage. A conversion conclusion requires
evidence that connects:
GPU resource -> workspace/customer carrier -> actual usage -> fee/revenue/chargeback
If a link is missing, say "基于当前可获取数据" and classify the result as partial
conversion or insufficient evidence. Treat voucher/credit consumption separately
from cash revenue.
Quick Start
Run the bundled collector from this skill directory:
sh scripts/collect_gpu_customer_usage.sh \
--hostname https://<dce-host> \
--start 2026-06-01 \
--end 2026-06-30 \
--cluster <gpu-cluster> \
--workspace <workspace-id>
If no cluster or workspace is supplied, run the collector anyway, inspect the
generated clusters.json and workspaces.json, then rerun with the GPU
clusters and relevant workspace ids. The collector writes raw JSON evidence and
trace.tsv; after collection, read the JSON files and calculate the final
tables from real data.
Workflow
- Establish scope.
- Default window: last 30 days.
- Use
YYYY-MM-DD for report and billing commands.
- Use RFC3339 timestamps for LLM Studio usage commands when needed.
- Use Unix seconds only for Billing Center commands that require Unix time.
- Check access and module availability.
- Run
dce auth status --hostname <host> when a host is provided or needed.
- Use
dce global-management about list-g-product-versions -o json if a
module route returns 404.
- Inspect unfamiliar commands before execution.
- Use
dce commands show <path...> --json.
- Prefer
-o json.
- Use only read-only commands.
- Collect GPU supply evidence.
- Clusters, GPU devices, GPU modes, node/device inventory, utilization,
allocated/requested GPU, memory total/used/allocated when available.
- Collect carrier evidence.
- Workspaces, namespaces, shared resources, quotas, queues, model services,
API keys, and owner/customer mapping when present.
- Do not call a workspace a customer unless DCE data or the user confirms
the mapping.
- Collect usage evidence.
- Token counts, request counts, active users, active API keys, model instance
usage, pod/runtime activity, queue activity, and last-used time.
- Do not treat model deployment or API key existence as usage.
- Collect billing evidence.
- Billing Center
amountDue, productName, voucherPayment, chargeback,
and workspace/account aggregation.
- Report voucher/credit consumption separately from cash revenue.
- Classify conversion and report in the required output format.
Conversion Levels
已转化: GPU supply can be connected to workspace/customer, sustained usage,
and billing/revenue or chargeback evidence in the analysis window.
部分转化: GPU has a carrier and usage, but billing is missing/low, usage is
unstable, utilization is weak, or evidence is concentrated in one
workspace/customer.
未转化: GPU exists but no proven customer/workspace usage or billable
consumption is visible.
证据不足: DCE data cannot connect GPU, carrier, usage, and money strongly
enough to classify.
Map conversion levels to risk labels:
| Conversion level | Risk label |
|---|
| 已转化 | 正常 |
| 部分转化 | 关注 |
| 未转化 | 风险 |
| 证据不足 | 异常 |
Data Sources
Use catalog discovery over hardcoded assumptions. Useful searches:
dce search "gpu devices" --json
dce search "workspace resources" --json
dce search "model serving" --json
dce search "llm studio token usage" --json
dce search "api key usage statistics" --json
dce search "billing aggregation" --json
dce search "workspace report" --json
Common read-only commands may include:
dce global-management workspace list-workspaces --page 1 --page-size 200 -o json
dce global-management workspace list-shared-resources-by-workspace --workspace-id <id> -o json
dce container-management cluster list-clusters --page 1 --page-size 200 -o json
dce container-management devices list-gpu-devices --cluster <cluster> -o json
dce llm-studio modelservingmanagement list-model-serving --page.page-size -1 -o json
dce llm-studio apikeymanagement get-api-key-usage-statistics2 --start-time <rfc3339> --end-time <rfc3339> --period TIME_PERIOD_DAY -o json
dce llm-studio wsdashboardmanagement get-ws-dashboard-summary --workspace <id> --start-time <rfc3339> --end-time <rfc3339> -o json
dce llm-studio wsdashboardmanagement list-ws-user-token-usage --workspace <id> --start-time <rfc3339> --end-time <rfc3339> --page.page-size -1 -o json
dce llm-studio wsdashboardmanagement list-ws-instance-token-usage --workspace <id> --start-time <rfc3339> --end-time <rfc3339> --page.page-size -1 -o json
dce billing-center bill get-account-bill-aggregation --workspace-id <id> --start-time <unix-seconds> --end-time <unix-seconds> -o json
If a command is absent or fails, inspect the command catalog and module
availability before concluding the data does not exist.
Output Format
Answer in Chinese when the user asks in Chinese. Follow this Markdown structure
exactly unless the user asks for another format. Do not output a tool-call
waterfall, skill-loading details, retry chatter, or JSON processing internals.
# 结论
基于当前可获取数据,<GPU 是否已转化为客户用量的判断>。当前风险等级为:<正常/关注/风险/异常>。
## 关键指标
| 指标 | 当前值 | 状态 |
|---|---:|---|
| GPU 总量/可用量 | ... | 正常/关注/异常 |
| GPU 平均利用率/分配率 | ... | 正常/关注/异常 |
| 活跃 workspace/customer | ... | 正常/关注/异常 |
| Token/request 用量 | ... | 正常/关注/异常 |
| 账单金额/抵扣金额 | ... | 正常/关注/异常 |
| 证据闭合度 | ... | 正常/关注/异常 |
## 主要发现
1. **<发现 1>**
<说明影响。>
2. **<发现 2>**
<说明影响。>
3. **<发现 3,可选>**
<说明影响。>
## 原因分析
### 原因 1:<原因>
证据:<真实数据证据。>
影响:<对转化判断的影响。>
### 原因 2:<原因>
证据:<真实数据证据。>
影响:<对转化判断的影响。>
## 建议动作
### 立即处理
1. <具体动作>
2. <具体动作>
### 持续观察
1. <具体动作>
2. <具体动作>
### 后续优化
1. <具体动作>
2. <具体动作>
## 后续可以继续追问
- 帮我查看未转化 GPU 的详细原因
- 帮我生成 GPU 客户用量转化提升方案
- 帮我导出一份给交付 / 老板看的报告
Keep 关键指标 to 3-6 rows. Include at least GPU capacity/utilization, active
workspace/customer, token/request usage, billing amount, or an evidence gap. Put
data gaps in the metric table or reason analysis; do not bury them in a command
trace.
Evidence Rules
- Every number must come from pulled DCE output, a user-provided file, or an
explicitly labeled assumption.
- Use
基于当前可获取数据 when Billing Center, LLM Studio, GPU metrics, or
workspace/customer joins are incomplete.
- Do not infer revenue from GPU utilization or deployment inventory.
- Do not infer usage from API key existence, model service existence, or quota.
- Do not merge voucher/credit consumption with cash revenue.
- Do not report precise revenue, utilization, or conversion rate when only
directional evidence exists.
- Show concise data-source coverage when useful, but avoid operational
walkthroughs unless the user explicitly asks.