SGLang cold start study shows GPU billing starts long before first user request

A developer ran a performance test on October 5, 2026, using a T4 GPU on Google Colab with the SGLang 0.5.21 framework. They found that while the server was marked ready after approximately 387 seconds, the first user request wasn't completed until 563 seconds after billing began. This means 88% of the total billed time elapsed before any user request was served. The test highlighted a significant gap between a server being 'ready' and actually being capable of handling user requests efficiently. For comparison, a similar test using the vLLM framework showed a much shorter delay between readiness and request completion.