OpenAI argues that AI progress accelerates most when improvements are made across the entire stack — data centers and chips, frontier models, developer platforms, consumer and enterprise products, and AI‑native devices. The company released initial measured performance for Jalapeño, its first custom inference chip, and outlined a multi‑vendor infrastructure approach, a data‑center project in Georgia, and internal efficiency metrics such as GPT‑5.6 Sol using 54% fewer output tokens on a cited index.
Jalapeño — OpenAI’s first custom inference chip
OpenAI describes Jalapeño as its first in‑house inference chip. According to the company’s reported measurements:
- On the InferenceX public benchmark using GPT‑OSS 120B, Jalapeño delivered higher peak throughput per kilowatt and lower token latency than the commercial systems included in the comparison.
- The chip also performed strongly on DeepSeek R1 and Kimi K2 benchmarks, suggesting its improvements apply across multiple model families.
OpenAI says Jalapeño gives it greater control over model execution and the economics of serving models. By co‑designing the model, serving software, chip, memory, and network, the company aims to improve throughput, latency, energy efficiency, and cost as a single system. OpenAI adds that it already has working first‑party silicon with measured results and that future generations of the chip are under development.
System approach and following the Pareto frontier
The company emphasizes that different workloads impose different demands: frontier training, high‑volume inference, and always‑on agents have distinct chip, software, network, power, and latency requirements. OpenAI’s stated objective is to remain on the Pareto frontier — continually seeking the best mix of capability, speed, reliability, efficiency, and cost for each workload.
To achieve that, OpenAI maintains a portfolio of partners and hardware. While Microsoft’s compute and NVIDIA’s chips have been foundational, the company lists additional providers in its current portfolio: AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy and SoftBank. These partners contribute varied strengths across cloud infrastructure, accelerated computing, low‑latency inference, data‑center development, and energy delivery.
OpenAI says it actively manages this portfolio for both capability and economics: using premium systems where capability is critical and optimizing for efficiency where scale and cost matter more. Maintaining choice among providers and deployment models, the company argues, lets it direct demand toward the best performance per dollar, preserve pricing discipline as market conditions change, and adapt as stronger technologies appear. The firm states it partners where the ecosystem speeds progress and builds where co‑design yields a meaningful advantage.
Data centers and Project Camellia
OpenAI cites Project Camellia in Georgia as an example of how facility design can be aligned with customer workloads while delivering local jobs, supporting local businesses, covering project infrastructure and energy costs, conserving water via a closed‑loop system, and subjecting commitments to an annual independent public audit.
Operational metrics and economic impact
OpenAI frames the system’s value in terms of how much useful intelligence it produces per unit of compute. It lists concrete internal metrics:
- On the Artificial Analysis Coding Agent Index, GPT‑5.6 Sol with maximum reasoning reached a new high while using 54% fewer output tokens than another leading model.
OpenAI says such improvements translate for customers into faster results, more reliable products, fewer retries, agents that complete longer workflows, and lower total cost for successful work. The company argues that the most favorable economics come from increasing useful intelligence per dollar.
OpenAI also invokes Jevons’ paradox: increasing efficiency makes more applications economically viable, which in turn can expand consumption and create new economic activity through more completed work, better decisions, new products, and additional revenue.
Conclusion
According to OpenAI, more productive compute and a more competitive supplier base enable serving more customers at lower cost and allow efficiency gains to flow through to users. Growth, the company says, funds continued investment in research, infrastructure, and safety — creating a reinforcing cycle where technological improvements improve economics, and improved economics fund the next wave of progress.
