• Serves models on self-hosted vLLMvLLMAn open-source engine for serving large language models fast. Self-hosting it means running models on your own GPUs.confirmed
  confirmed: engineering blog · may 12

  →Serving cost and latency are their team's problem directly. That is the budget line you are selling into.

• Two open roles for inferenceinferenceRunning a trained AI model to get answers. This, not training, is where most of the GPU bill goes. optimizationconfirmed
  confirmed: careers page · may 18

  →They are trying to hire their way out of this. A pilot that lands before those hires ramp is cheaper and faster, and they know it.

• GPUGPUThe chips AI models run on. Renting them is one of the biggest line items on an AI company's bill. spend flagged in Q2 planningconfirmed
  confirmed: discovery call transcript · may 20

  →The pain has a named line in a planning doc and exec attention this quarter. Deals move when that is true.

• Evaluating managed endpointsmanaged endpointsSomeone else hosts the AI model and you call an API. The opposite of running it yourself.inferred
  inferred: founder forum thread · may 15

  →They are already shopping. Assume at least one competitor call has happened; speed matters more than polish now.

This deal is about serving cost at scale, and the person who feels it is the infra lead, not procurement. They have exec attention on the GPUGPUThe chips AI models run on. Renting them is one of the biggest line items on an AI company's bill. bill and hiring underway, which means they will fix this with or without you inside two quarters. The way in is a benchmark on their own traffic that beats what their new hires could build by summer.

• 1
  Priya NairHead of Infrastructurestart here

  Owns the GPUGPUThe chips AI models run on. Renting them is one of the biggest line items on an AI company's bill. bill and wrote the blog post you will be quoting. Technical, direct, allergic to marketing language. Open with the idle-replica number from her own call.

• 2
  Dan OkaforVP Engineeringnext

  Final sign-off. He does not care what it costs; he cares that p95p95The response time the 95th-slowest request gets. The number engineers watch to know users are not waiting. never regresses. Bring the latency chart, not the pricing page.

• 3
  Procurementnot yet engagedhold

  Correctly out of the loop for now. Bringing them in before the pilot is scoped adds four weeks for nothing.

• “We're burning half our GPUGPUThe chips AI models run on. Renting them is one of the biggest line items on an AI company's bill. budget on idle replicasidle replicasModel servers kept running with no traffic, just in case. Paid by the hour, used never..”

  Priya Nair · discovery call · 14:22

• “If p95p95The response time the 95th-slowest request gets. The number engineers watch to know users are not waiting. goes over 400ms, product will notice before we do.”

  Dan Okafor · discovery call · 21:05

• they say

  “We can just build this ourselves.”

  you say

  Agree that they could. Then price the engineering months against the pilot cost and offer the benchmark on their own traffic. The comparison does the arguing.

• they say

  “Managed endpointsManaged endpointsSomeone else hosts the AI model and you call an API. The opposite of running it yourself. lock us in.”

  you say

  Deploy in their VPCVPCA private, walled-off slice of a public cloud. Where careful companies keep their data.. Weights and routing configs stay theirs, and the export path is shown up front, not promised.

• they say

  “Not this quarter.”

  you say

  The GPUGPUThe chips AI models run on. Renting them is one of the biggest line items on an AI company's bill. bill is already flagged for Q2 planning. A two-week pilot lands before planning closes; waiting is the expensive option.

• Run the demo against a sample of their traffic, not a toy model
• Send Priya the pilot scope with the warm-pool sizing
• Ask for the latency SLOSLOService level objective: the response-time promise an engineering team commits to keeping. on the demo call; it is still unknown