Defined as >50% of tokens generated

Some things that would count as non-GPU: cerebras hardware, a CPU, etc.

Resolution criteria

This market resolves to YES if, as of December 31, 2027, more than 50% of the total tokens generated by OpenAI's flagship frontier model are served on non-GPU hardware. It resolves to NO otherwise.

Key Definitions:

• Flagship Model: OpenAI's most advanced, primary general-purpose or reasoning frontier model available to public and API users as of December 31, 2027 (e.g., the successor to GPT-4, GPT-5, o1, etc.).

• Primarily Serve: Defined as >50% of the total production tokens generated (across ChatGPT, OpenAI API, and enterprise services) for the flagship model.

• Non-GPU Hardware: Processors that are not Graphics Processing Units (GPUs) such as Nvidia's Hopper, Blackwell, or Rubin architectures, or AMD's Instinct line. Non-GPU hardware includes:

  • Custom ASICs / XPUs (e.g., OpenAI's Broadcom-co-developed Jalapeño inference chip, Google TPUs, or other custom AI accelerators).

  • Wafer-scale engines (such as Cerebras CS-3 or CS-4 systems).

  • Central Processing Units (CPUs).

  • Language Processing Units (LPUs).

Source of Truth: This market will resolve based on official press releases or technical documentation from the OpenAI Newsroom, public statements from OpenAI executives (such as CEO Sam Altman or VP of Hardware Richard Ho), or highly credible reporting from specialized semiconductor and tech analysis publications like SemiAnalysis or The Next Platform published on or before February 1, 2028.

If there is a lack of definitive public disclosure regarding the exact token breakdown, the market will resolve based on the clear consensus of credible reports. If no credible information is available to confirm >50% token share on non-GPU hardware, this market will resolve to NO.

Background

To bypass Nvidia's high-margin supply constraints, OpenAI has aggressively pursued a vertical-integration strategy for its inference workloads.

In January 2026, OpenAI announced a multi-year partnership with Cerebras to deploy up to 750 megawatts of Cerebras wafer-scale systems, aiming to bring ultra-low-latency, non-GPU inference to ChatGPT. Following this, in June 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first proprietary custom inference ASIC built on TSMC's 3nm process. Benchmarks presented at Hot Chips 2026 claimed that the Jalapeño chip outperforms Nvidia's Blackwell GPUs in throughput per watt and latency for transformer workloads.

This market tracks whether these custom ASICs and wafer-scale architectures successfully transition from early deployment to handling the majority of OpenAI's flagship token traffic by the end of 2027.

This description was generated by AI. Review and verify everything here yourself. You can edit, replace, or delete any part of this description, including the resolution criteria. You do not need to trust the AI output.