Summary
According to a post on X by @thdxr, his team's new GPU rack running their own DeepSeek inference setup handled 6.77 trillion tokens in the last 24 hours. He reported a median time to first token of 1.16 seconds and a median of 143 tokens per second. He said the 95th-percentile numbers are worse because one tray is flaky, and he had earlier claimed the setup beats every inference vendor the team has found on price and performance. These are self-reported figures from a single post, and the snippet does not include the methodology.
AI-assisted summary based on the listed source.
What happened
@thdxr reported that his team's new GPU rack running their own DeepSeek inference setup handled 6.77 trillion tokens in the last 24 hours, with a median time to first token of 1.16 seconds and a median of 143 tokens per second. He said the 95th-percentile numbers are worse because one tray is flaky, and earlier...
Why it matters
Teams comparing self-hosted DeepSeek inference with hosted vendors get a concrete set of throughput and latency numbers to consider. The reported tail latency, tied to one flaky tray, shows that hardware reliability can shape real-world performance as much as median speed. Watch for more detail on the setup or independent benchmarks, since the claim of beating every vendor on price and performance is not supported by data in the post.
What this means for you
Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.
Signal Intelligence
Signal Strength 88%
Technical label WATCH
Public Interest 30
Category ROBOTS & HARDWARE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 0
Novelty Interest Score 94
Consequence Score 46
Curiosity Score 0
Shareability Score 45
Why this is here
VQV surfaced this signal because it is recent, relevant to AI Chips, connected to Scobleizer News Pulse (X).
