For years, AI competition was described mainly as a model race: better benchmark scores, larger context windows, and faster release cycles. That framing leaves out the physical system required to deliver a model to millions of users. Inference consumes chips, memory, networking, electricity, cooling, data-centre capacity, and serving software every time a person asks a question.
OpenAI's August 25, 2026 infrastructure essay makes that hidden stack explicit. The company disclosed measured results from Jalapeño, its first custom inference chip, and described a strategy that combines first-party silicon with a wide portfolio of external compute and energy partners. The announcement is a company account rather than independent testing, but it provides concrete evidence of why custom hardware has become a product and negotiating strategy.
Jalapeño is an inference chip, not a replacement for every accelerator
Jalapeño is designed for inference: running trained models in production rather than performing every kind of AI workload. OpenAI reported results on InferenceX using GPT-OSS 120B and said the chip delivered higher peak throughput per kilowatt and lower token latency than the commercial systems included in its comparison. It also reported gains on DeepSeek R1 and Kimi K2, an attempt to show that the hardware is not useful only for one OpenAI model family.
At the comparison's previous-best time-between-tokens, OpenAI's published chart reports large mixed-token-throughput-per-kilowatt advantages for Jalapeño: 53.7 times for GPT-OSS, 104.3 times for DeepSeek R1, and 56.1 times for Kimi K2. Those are striking company-reported benchmark results, but they should not be read as a universal speed ranking. Hardware performance changes with model, batch size, latency target, software stack, power measurement, and workload mix.
That distinction is important because an inference fleet optimises several things at once. A system can maximise total tokens while making each user's response slower, or deliver extremely low latency at poor utilisation. OpenAI frames Jalapeño around the ability to improve throughput, latency, energy efficiency, and cost together, which is the practical value of controlling more of the stack.
OpenAI reported working first-party silicon, but its strategy is not to replace every partner. It is to create credible choice across chips, clouds, data centres, and energy.
Co-design turns engineering metrics into product economics
A company that develops the model, serving software, chip, memory, and network as one system can make trade-offs that are difficult when every layer is purchased separately. A model's operators can tune scheduling and memory movement for known workloads; chip designers can prioritise the operations that dominate those workloads; software teams can route requests based on real capacity and latency.
The financial effect appears at enormous scale. Better throughput means more paid or useful work from the same power and infrastructure. Lower latency makes interactive products more usable. Greater energy efficiency can reduce both operating cost and the capacity needed to support demand. These gains do not make data centres cheap, but they can change the cost per successful task.
OpenAI offers a model-side example: GPT-5.6 Sol with maximum reasoning reached what the company calls a new high on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than another leading model. The comparison is separate from Jalapeño, but it supports the full-stack argument: efficiency can come from model behaviour as well as hardware, and the strongest economics emerge when both improve.
The strategy is diversification as much as ownership
OpenAI is not describing a future in which it abandons outside suppliers. Microsoft compute and NVIDIA chips remain foundational in its account. Its named portfolio also includes AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank, spanning cloud capacity, accelerators, low-latency systems, data-centre development, and energy delivery.
Jalapeño creates a first-party option inside that portfolio. The strategic advantage is choice: premium systems can serve workloads where frontier capability matters most, lower-cost systems can handle work where efficiency dominates, and demand can shift as prices, capacity, and technology change. A credible internal path also gives a buyer more leverage in supplier negotiations even when external partners continue to carry most workloads.
Data-centre design completes the picture. OpenAI points to Project Camellia in Georgia, where it says facilities will use closed-loop water systems, cover project infrastructure and energy costs, support local employment, and undergo annual independent public audits. Those commitments need scrutiny over time, but their inclusion shows how chip strategy now reaches into power, water, local economics, and public legitimacy.
Why only a few companies can make this bet
Custom silicon requires specialised teams, fabrication partners, software investment, long lead times, and enough predictable demand to justify the programme. A smaller AI company usually gains more by choosing among available clouds and accelerators than by designing a chip. Even a large company can lose years if its workload assumptions change before hardware arrives.
The broader lesson is therefore not that every AI startup needs silicon. It is that hardware decisions now shape model availability, product latency, margins, resilience, and bargaining power. Teams building on top of foundation models should understand those dependencies because provider pricing or capacity can alter their own product economics overnight.
Jalapeño is best read as evidence that the largest AI companies are becoming infrastructure portfolios. The model remains visible to users, but the durable advantage may come from coordinating everything underneath it more efficiently than a competitor can.
AI companies are building chips because model capability is only valuable when it can be served quickly, reliably, and at a sustainable cost. Jalapeño gives OpenAI a measured first-party inference path alongside Microsoft, NVIDIA, and a growing supplier portfolio. Whether the published advantages hold across independent workloads remains an open question, but the strategic direction is clear: competition is moving from the model alone to the entire system that makes the model usable.
Back to all writing