← Back to the carnage
REPLICATE

🟑 Multiple models across H100, L40S, and CPU hardware are ghosting prediction and training requests. Turns out 'artificial intelligence' doesn't mean the infrastructure gets a pass on basic competence. Engineers are investigating why their expensive GPUs suddenly developed amnesia. (Auto-resolved: no updates in 24h)

⏰ Jan 20, 2026 Fixed (for now) spotted via status_rss

// The official story

Predictions and training unavailable for multiple models

// Status page says

Status: Investigating We are seeing indications that multiple models are not receiving prediction or training requests correctly. Affected components H100 Hardware (Degraded performance) L40S Hardware (Degraded performance) CPU Hardware (Degraded performance)

πŸ”— Official status page (for the PR-approved version)

https://www.replicatestatus.com//incidents/01KFESSJ2K93ZTF97QM1BH077D

Replicate's Greatest Hits

🟒 REPLICATE Aug 26, 2026
🟠 Replicate's A100 hardware is ghosting everyoneβ€”the infrastructure apparently got tired of working and self-quarantined itself. Kubernetes said 'nope' and cordoned off the machines like they're contagious. Your training jobs are now stuck in purgatory. (Auto-resolved: no updates in 24h)
🟒 REPLICATE Jul 31, 2026
🟠 Replicate discovered their caching layer was playing gatekeeper with HuggingFace models, choking out H100/L40S/A100 setups. The fix? Nuking the cache and hoping nobody asks questions while they play detective. GPU performance is bruised but stabilizing. (Auto-resolved: no updates in 24h)
🟒 REPLICATE Jul 16, 2026
🟑 Replicate's model downloads are getting 504'd because HuggingFace's CDN decided to ghost them. The usual "we're monitoring" dance has begun while engineers refresh the status page like it's a slot machine. (Auto-resolved: no updates in 24h)
🟒 REPLICATE Jun 03, 2026
🟑 Your L40S and H200 clusters are playing hot potato with model setup times. High contention means everyone's waiting in line while the hardware pretends to be busy. Engineers are investigating why expensive silicon doesn't want to cooperate. (Auto-resolved: no updates in 24h)