成本分析
在巴西拥有自己的AI要多少钱？
云基础设施（Azure、AWS、GCP）、本地GPU和商业API之间的完整比较。巴西境内运营的最新价格，符合LGPD。
云基础设施 — 巴西区域
| 提供商 | 区域 | 实例 | GPU | 月费 |
|---|---|---|---|---|
| AWS | São Paulo | ml.g5.12xlarge | 4x A10G (96GB) | $5,176 |
| AWS | São Paulo | ml.g5.48xlarge | 8x A10G (192GB) | $14,863 |
| AWS | N. Virginia (referência) | ml.p5.48xlarge | 8x H100 (640GB) | $46,209 |
| Azure | São Paulo | NC24ads A100 v4 | 1x A100 (80GB) | $5,804 |
| Azure | São Paulo | ND96asr H100 v5 | 8x H100 (640GB) | $47,998 |
| GCP | São Paulo | a2-highgpu-8g | 8x A100 (640GB) | $51,976 |
| On-Premises | Infraestrutura própria | 8x H100 SXM | 8x H100 (640GB) | $8,500 |
按模型的GPU需求
| 模型 | 参数 | 最小GPU内存 | 推荐 | 预估月费 |
|---|---|---|---|---|
| DeepSeek V4 Pro | 671B MoE (37B ativos) | 640GB | 8x H100 80GB | $48,000 |
| DeepSeek V4 Flash | Quantizado (INT8) | 80GB | 1x A100 80GB | $5,200 |
| MedGemma 27B | 27B | 80GB | 1x A100 80GB | $5,200 |
| Qwen3 235B | 235B MoE | 640GB | 8x A100 80GB | $52,000 |
| Qwen-AgentWorld-35B-A3B | 35B MoE (3B ativos) | 80GB | 1x A100 80GB | $5,200 |
| ContechLLM (Qwen3.6-35B-A3B) | 35B MoE (3B ativos) | 80GB | 1x A100 80GB | $5,200 |
| FinBERT | 110M | 16GB | 1x T4 16GB | $800 |
| Claude Opus 4.6 (API) | Serverless | 0GB | API Anthropic | Variável |
自有AI vs. 商业API
| 提供商 | 模型 | 输入/1M tokens | 输出/1M tokens | 月费@1M请求 | 数据隐私 | LGPD |
|---|---|---|---|---|---|---|
| OpenAI | GPT-4o | $3 | $10 | $12,500 | 无 | ✗ |
| Gemini 2.5 Pro | $1 | $5 | $6,250 | 部分 | ✗ | |
| Anthropic | Claude Opus 4.6 | $15 | $75 | $90,000 | 部分 | ✗ |
| Infraestrutura própria | DeepSeek V4 Pro | — | — | $8,500 | 完全 | ✓ |