CoreWeave (@CoreWeave) on X (original) (raw)

The Essential Cloud for AI

• Pinned
  
  Registration for #FullyConnected26 is open: 🗓️ September 29 – Oct 1 📍San Francisco 3 days and 3 tracks to accelerate the superintelligence loop. 2,000+ AI Leaders in the room. This is where you need to be. utm.io/upDdn
• 
  We're pulling back the curtain. From a CoreWeave data center:
  @nvidia
  Vera Rubin NVL72, hosted by
  @SiliconANGLE@theCUBE
  on June 30, 9:30 am PT. Save the date.
• 
  Every training job hits the same 2 bottlenecks: Idle GPUs, which hurts efficiency, and researchers doing IT, which hurts velocity. Stop wasting budget and talent. Watch
  @deok_filho
  break down how to solve these bottlenecks in the clip above and see the full on-demand version
• 
  Prototypes are predictable. Production inference under real-world demand is not. If your platform hides runtime dynamics, the latency shifts are costing you time and money. Stop guessing and start scaling predictably:
• 
  If you're building AI at scale, this is the room you want to be in.#FullyConnected26September 29 – October 1 San Francisco, CA Let us know if you're coming.
• 
  Join us on June 30th! Tune in to
  @SiliconANGLE@theCUBE
  as we discuss
  @nvidia
  Vera Rubin Platform innovations, with our exec team and special guests. Save the date.
• 
  It's almost here: the first episode of Training Tuesday💪 Join our 30-min session on June 23. 👉utm.io/uqfrp
• 
  15 months post-IPO, CoreWeave is joining the
  @Nasdaq
  -100. This milestone is a reflection of our team's hard work and the trust of our customers, partners, and shareholders. We're just getting started. Full release: hubs.la/Q04lczyG0
• 
  Replying to @CoreWeave
  If your inference stack is accumulating drift, the first question isn't "how do we monitor better?" It's: was this infrastructure built for inference, or was inference deployed on infrastructure designed for something else? Full breakdown below.
• 
  Replying to @CoreWeave
  The teams that avoid chronic drift share one thing: They treat reliability as an architectural property, not a monitoring problem. Monitoring tells you drift happened. Architecture determines whether it accumulates. Better alerting won't fix a structural gap.
• 
  Replying to @CoreWeave
  Latency drift is actually 3 problems compounding: 1⃣GPU contention: scheduling overhead negligible at 50 rps. Material at 5,000 2⃣Batching misconfiguration: a large prefill stalls decode for every concurrent request. 3⃣Autoscaling lag: 90s to spin up a pod during a 60s