OpenAI balances third-party hardware partnerships with new first-party silicon control

OpenAI is charting a dual path forward for its computing infrastructure, balancing its reliance on a growing roster of hardware partners with the newfound control offered by its first custom inference chip.

The company’s strategy, detailed in a recent update, reveals a deliberate effort to manage its supply chain while simultaneously building proprietary technology to optimize performance and cost.

The centerpiece of this first-party push is "Jalapeño," OpenAI's custom inference chip.

Early benchmark results show it delivering higher peak throughput per kilowatt and lower token latency on the InferenceX benchmark using GPT-OSS 120B compared to commercial systems.

These gains were also observed across other model families, including DeepSeek R1 and Kimi K2.

This chip gives the company direct leverage over how its models run, allowing for tighter integration between the model, serving software, memory, and network.

As the company puts it, this creates a "credible first-party path" alongside existing partnerships.

This move toward in-house silicon does not signal a retreat from third-party hardware. Instead, OpenAI frames it as an expansion of its "portfolio" approach.

The company explicitly acknowledges that Microsoft’s compute and NVIDIA’s chips were "foundational" to its growth.

Today, that portfolio has grown to include AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank.

Each partner brings different strengths, from cloud infrastructure to low-latency inference.

The goal, the company states, is to remain on the "Pareto frontier"—constantly seeking the strongest mix of capability, speed, reliability, efficiency, and cost for each specific workload.

Our analysis suggests this is a sophisticated risk management strategy.

By owning a piece of the chip supply, OpenAI insulates itself from potential bottlenecks or pricing pressures in the broader market, particularly as demand for AI compute skyrockets.

At the same time, maintaining a diverse set of external partners preserves optionality.

The company can direct "premium" training workloads to the most capable systems while routing high-volume inference tasks to the most cost-efficient hardware.

This dual sourcing gives OpenAI significant pricing leverage and prevents any single vendor from holding a monopoly over its computational capacity.

The implications for smaller AI companies and developers are notable.

OpenAI’s ability to drive down its own inference costs through custom silicon like Jalapeño could eventually translate to lower API pricing for users.

The company explicitly notes that "better technology creates better economics," and that these cost savings can be passed through to customers.

For a startup building on top of OpenAI’s models, this means the potential for more affordable scaling.

However, it also means that OpenAI’s competitive moat deepens, making it harder for smaller players to match the cost-performance ratio of a fully integrated system.

For the average user, this strategy likely means faster, more reliable, and cheaper AI services over time.

The company points to its GPT-5.6 Sol model, which achieved a new high on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than a leading competitor.

This type of efficiency gain, driven by both software and hardware optimization, translates directly into fewer retries, faster results, and longer, more dependable agent workflows.

What should ordinary companies do in light of this? First, watch for pricing adjustments. As OpenAI’s hardware costs decline, expect potential reductions in API fees for high-volume inference tasks.

Second, consider the strategic value of using a provider that controls its own hardware stack. This integration typically leads to more consistent performance and fewer unexpected latency spikes.

Third, diversify your own AI dependencies. While OpenAI’s custom chips make its platform more efficient, relying on a single vendor's hardware strategy carries its own risks.

Keep an eye on how the performance of Jalapeño compares to offerings from partners like AMD and Broadcom in real-world workloads.

Source: OpenAI