Cloud Run GPUs vs Replicate Deployments

Cloud Run GPUs 78

OAuth 2.0 tokens from service accounts under IAM. Google's APIs accept access_token as a query parameter, the 10-point deduction other Google API listings take (20). IAM roles separate viewing, deploying and invoking, and services can require IAM authentication for every call (16). The platform runs your own container and returns its output (10). Cloud Audit Logs record Admin API calls, and request logs go to Cloud Logging (15). Google's vulnerability reward programme, security.txt and certifications (17). 20+16+10+15+17 = 78.

Replicate Deployments 40

Bearer tokens starting r8_, several per account, each can be disabled; no scopes or expiry, re-read on 2 November (20). No read-only or per-model token; every token can create, update and delete deployments (0). Returns your own model's output (10). No audit log or per-token usage view found; predictions are listed per account (5). Replicate scans public GitHub for leaked tokens, disables them and emails the owner. replicate.com/.well-known/security.txt answers 404, and no bug bounty or certification was found in the docs we read. OSV lists no advisory for the replicate packages, replicate-mcp or Cog (5). 20 + 0 + 10 + 5 + 5 = 40.

Cloud Run GPUs 20

No x402, MPP or L402 (0). Per-second GPU, CPU and memory prices published without a login (20). The Cloud Run free tier covers CPU and memory, not GPUs, and the $300 trial credit needs a payment method (0). A person creates the project and billing account in a browser (0). 0+20+0+0 = 20.

Replicate Deployments 30

No machine payment protocol (0). Per-second prices for each hardware SKU published without a login, re-read on 2 November, now with a small CPU tier at $0.09 an hour and H200 SKUs on committed-spend contracts (20). Accounts without a card can run predictions at up to 6 a minute; private deployments still bill set-up and idle time (10 of 20). Signup is a browser flow (0). 0 + 20 + 10 + 0 = 30.

Cloud Run GPUs 66

Platform reading. One gcloud run deploy with --gpu and --gpu-type adds a GPU to a container, and responses carry the service object with fieldMask on updates (15). List calls page with pageSize and pageToken (15). Google's status codes and messages, with quota errors that name the limit (12). Revisions are immutable, traffic can be moved back, and etags guard updates, but no idempotency key on create (12). Client libraries in many languages and gcloud, though a GPU deploy needs instance-based billing, minimum CPU and memory and, past 3 GPUs, a quota request (12). 15+15+12+12+12 = 66.

Replicate Deployments 73

The MCP server exposes one tool per HTTP operation and has an experimental code mode that collapses them into two tools (search the SDK docs, run TypeScript) (20). List responses page with next and previous cursor URLs, the predictions list filters by created_after, created_before and source, and the models list sorts, but there is no page size or field selection (15 of 20; was 10, pagination not verified on 1 October). Coded errors with suggested fixes, detail messages on HTTP errors (15). No idempotency keys; Prefer: wait, a Cancel-After header, webhooks and cancel cut polling, and a failed run still bills its active time (8). Sensible defaults (min_instances 0 allowed) and official Python and JavaScript clients plus Cog (15). 20 + 15 + 15 + 8 + 15 = 73.

Cloud Run GPUs 70

Hosted reading. Google Cloud's status page with a JSON incident feed (20). In the 90 days from 8 August 2026 the feed names Cloud Run in three multi-product incidents, a service outage on 20 August (15:40 to 19:20 UTC, 27 products, global and us-west1), a network degradation in us-central1-b on 1 September and a europe-west2 disruption on 20 October. One major (10). Resource limits and GPU quota are published with numbers (3 GPUs a region by default), but no request rate for the Admin API was found on the quotas page (10). Google's client libraries retry with backoff and the Admin API takes etags, while no GPU-specific retry guidance was found (10). An SLA for GPU services, 99.95% with zonal redundancy and 99.5% without (10). GPU services generally available since 7 April 2025 (10). 20+10+10+10+10+10 = 70.

Replicate Deployments 75

Replicate incidents post under a Replicate component on cloudflarestatus.com, and replicatestatus.com redirects to that page filtered to Replicate (20). The component's 90-day bars, read on 2 November, show 5 days with incidents, all in September 2026, and none since 28 September. Some third-party models couldn't scale out for 15 hours 41 minutes on 14 and 15 September, a Pruna-specific issue ran 20 hours on 17 September, backend services returned intermittent 500s for 1 hour 54 minutes on 24 September, and hot-swapped Flux models stuck for 17 minutes on 28 September. Each names a subset of models or intermittent errors, so each reads as partial, minor only (20). 600 prediction creates a minute, 3,000 a minute on other endpoints, and 6 a minute without a card, re-read on 2 November (15). A 429 body says when the limit resets ('resets in ~30s') and the error-code page gives retry advice per code, but there's no Retry-After header or idempotency guidance (10 of 15). No SLA found (0). Deployments are GA (10). 20 + 20 + 15 + 10 + 0 + 10 = 75.

Cloud Run GPUs 84

Closed service under the Google Cloud terms, with Apache-2.0 client libraries (15). The Cloud Data Processing Addendum and the privacy notice cover customer data, per the Document AI listing's 1 November reading; no Cloud Run statement on GPU instance data after shutdown was found (22). The Cloud terms promise 12 months' notice before a backwards-incompatible API change, and driver versions are announced in the release notes (16). Regions per GPU type listed, with Google's subprocessor list and data-location terms (16). 15+22+16+16 = 69.

Replicate Deployments 79

Closed service; Cog is Apache-2.0 (20). API prediction inputs, outputs, files and logs are deleted after one hour by default, web predictions are kept until deleted, and a subprocessor page is published, per the 1 October check, not re-read on 2 November (22). Dated deprecations in the changelog (streaming default in July 2024, spend limits in July 2025) but no written policy (12). Subprocessors listed; data locations not stated in what we read (15). 20 + 22 + 12 + 15 = 69.