◇Custom Inference Silicon Displaces Merchant GPUs

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Now 40 · -6 since Aug 27 · ranged 34 to 40

Why the conviction moved

• Aug 28
  Weakened

  Amazon, the operator furthest along on in-house accelerators with Trainium, committed to two million additional Nvidia GPUs for 2027 and 2028 and put Annapurna Labs to work as Nvidia's first NVHBM memory partner rather than purely on displacement silicon. The largest custom-silicon builder deepening both its purchase volume and its engineering entanglement with the merchant vendor weakens the claim that steady-state inference is migrating off merchant GPUs.

• Aug 27
  Strengthened +5

  OpenAI's Jalapeño claims 1.5–1.9x throughput per kilowatt versus Nvidia's GB300 on a suite OpenAI chose and ran itself. If the perf-per-watt gap holds under independent testing, the highest-volume steady-state inference workload at one of the largest buyers leaves the merchant GPU tier.

• Aug 27
  Strengthened +6

  Nvidia is opening NVHBM — a custom memory design moving the controller into the HBM stack for a claimed 30 percent more bandwidth and 15 percent lower memory power than HBM4E — to rival accelerators via NVLink Fusion, with Amazon's Annapurna Labs as the first partner. Nvidia selling its memory advantage to the in-house silicon team most likely to displace its GPUs is an incumbent conceding that the inference tier will be contested and choosing to be paid inside it.

Source trail

• Supporting · August 27, 2026

  Nvidia Opens Its Custom Memory Design to Rival Accelerators Through NVLink Fusion

  Nvidia is opening NVHBM — a custom memory design moving the controller into the HBM stack for a claimed 30 percent more bandwidth and 15 percent lower memory power than HBM4E — to rival accelerators via NVLink Fusion, with Amazon's Annapurna Labs as the first partner. Nvidia selling its memory advantage to the in-house silicon team most likely to displace its GPUs is an incumbent conceding that the inference tier will be contested and choosing to be paid inside it.

  NVIDIA Blog

• Supporting · August 27, 2026

  OpenAI Publishes First Benchmarks for Its Jalapeño Inference Chip Against Nvidia Systems

  OpenAI's Jalapeño claims 1.5–1.9x throughput per kilowatt versus Nvidia's GB300 on a suite OpenAI chose and ran itself. If the perf-per-watt gap holds under independent testing, the highest-volume steady-state inference workload at one of the largest buyers leaves the merchant GPU tier.

  AI Post (Telegram)