OpenAI Jalapeno and the Vertical Integration Move

By Nekud October 16, 2026 17 min read

In June 2026, OpenAI confirmed that it had taped out its first custom AI processor, a Broadcom-co-designed inference chip the company had given the internal name Jalapeño. The disclosure was modest in tone, but the engineering achievement behind it was not. OpenAI had, on the company’s own account, designed the chip and brought it through tape-out in nine months, with the chip’s defining features shaped by OpenAI’s internal models rather than by a generic AI workload. The company had manufactured it on a TSMC advanced process, partnered with Broadcom on the design integration and on the network silicon, and planned to deploy the chip at gigawatt scale across OpenAI’s existing data-centre footprint over the next three to four years. By August 2026, OpenAI was reporting that the first-generation Jalapeño chip was already outperforming Nvidia hardware on some workloads. In September 2026, OpenAI confirmed that the company had also signed a chip-development relationship with Samsung, and that the company was, in parallel, working with Broadcom on the second-generation chip. The Jalapeño program was, in substance, the moment at which OpenAI joined the small set of AI labs that own and operate their own custom AI silicon.

The Jalapeño announcement is the most explicit statement yet by OpenAI of a vertical-integration strategy that the company had been working toward for at least two years. OpenAI had, throughout 2024 and 2025, signed some of the largest AI compute-committed deals in history — the multi-billion-dollar commitments to Microsoft Azure, the multi-year deal with Oracle, the substantial capacity commitments with the SoftBank-backed Stargate project, and the substantial relationship with Amazon Web Services — and the company had, alongside these commitments, been building an internal capability to design AI silicon in-house. The Jalapeño chip is the first fruit of that internal capability. The chip is not OpenAI’s only silicon play. The company has, in parallel, committed substantial resources to the long-running Stargate infrastructure joint venture, has signed a Broadcom partnership on next-generation accelerators, has signed a Samsung partnership on foundry and chip-development work, and has called publicly for the United States government to lead an international effort on frontier-AI standards. The cumulative picture is of an AI lab that is determined to control, in detail, the silicon and the standards on which its products run.

This essay traces the Jalapeño program from the mid-2024 internal decision to design custom AI silicon, through the June 2026 tape-out, the August 2026 reporting on workload performance, the September 2026 Samsung partnership, and the strategic implications for the AI infrastructure industry. It focuses on three threads. The first is the engineering story: how OpenAI built a chip in nine months, what the Jalapeño silicon looks like, what it is designed to do, and what its performance-on-workload shows about OpenAI’s perception of its own silicon advantage. The second is the partnership story: the relationship with Broadcom, with TSMC, with Samsung, and with the data-centre partners that will host the Jalapeño silicon. The third is the strategic-implication story: what it means for OpenAI as a company, what it means for Nvidia, what it means for the broader AI infrastructure industry, and what it means for the long-running debate about the right degree of vertical integration in a frontier-AI business. The thesis the essay develops is the following. The Jalapeño program is the first clear evidence that the largest Western AI lab is going to design its own silicon at scale. The implications of that decision are going to unfold over the rest of the decade.

How OpenAI built Jalapeno in nine months

The most striking engineering fact about Jalapeño is the pace. The chip was, on the public record, designed and brought to tape-out in nine months using OpenAI’s internal models. The nine-month cadence is substantially faster than the conventional cadence for a custom AI chip design cycle, which is more typically two to three years for a first-generation tape-out by a team that has previously designed AI silicon, and substantially longer than nine months for a team designing AI silicon for the first time. OpenAI’s claim to have completed the chip in nine months is, on the basis of the public engineering record, plausible. The reason is that the chip’s design was substantially informed by OpenAI’s internal understanding of the inference workload pattern for its own models — an understanding that no merchant chip designer could match, and that allowed OpenAI to focus the silicon design effort on the parts of the workload that mattered most for OpenAI’s own products.

The technical claim behind the nine-month cadence is that OpenAI’s models and OpenAI’s chips are co-designed: the models were designed with attention to the silicon, and the silicon was designed with attention to the models. The information flow between the two allowed OpenAI to bypass substantial parts of the conventional chip-design process that would normally be spent on workload discovery. Instead of discovering what the inference workload looked like, OpenAI already knew the workload from the company’s day-to-day operations. Instead of running generic machine-learning benchmarks to validate the chip, OpenAI could validate the chip against the company’s actual production traffic. The validation cycle was, in effect, the production cycle of the existing AI service business, and the design cycle was, in effect, the next iteration of the AI service business’s architectural decisions.

The chip itself, on the basis of the public reporting through mid-2026, is a custom inference accelerator optimized for the OpenAI workload. The chip is a TSMC design on an advanced node, carries substantial on-package memory in the form of HBM4 or near-equivalent, has a dataflow architecture tuned to the decode-heavy nature of OpenAI’s user traffic, and is integrated with Broadcom’s networking silicon to allow scaling across a large rack-scale deployment. The chip is not a training chip, in the way that Nvidia’s GPU is a training chip; Jalapeño is, on the public record, an inference chip. The decision to design for inference rather than for training is consistent with the broader industry shift toward inference-first silicon that the hyperscaler custom-silicon programs are pursuing, and it reflects OpenAI’s realization that the company’s inference footprint is the dominant cost component of its operations, just as it is for Microsoft and for Google.

The August 2026 reporting that Jalapeño was outperforming Nvidia hardware on some workloads was a substantial validation of the design choice. The workloads on which Jalapeño has been reported to outperform Nvidia are inference workloads that are tight loops of prefill and decode on long contexts, which is exactly the workload that OpenAI’s user traffic produces. The inference-first design of the chip, combined with the OpenAI-tuned knowledge of how to layout the silicon for OpenAI’s specific kernels, has produced a chip that, on the workload that matters most to OpenAI, beats the merchant GPU even when the merchant GPU is in the same generation. The result is not necessarily surprising — OpenAI knew what workload it wanted to optimize for, and the chip is, accordingly, optimized for that workload — but the magnitude of the reported performance gap was substantial enough to give OpenAI confidence that the Jalapeño program was the right strategic bet.

The Broadcom, TSMC, and Samsung partnerships

The Jalapeño program is, in the architecture of its production relationships, an example of a partnership model that has become increasingly common in the hyperscaler custom-silicon world. OpenAI does not own a foundry. OpenAI does not own a back-end packaging house. OpenAI does not own a network-silicon design team comparable in scale to Broadcom’s. What OpenAI owns is the architectural specification of the silicon, the workload specification, and the validation infrastructure. The chip itself is fabricated by TSMC. The chip’s network-attach silicon and other auxiliary silicon are designed and supplied by Broadcom. The OpenAI engineering team works with the TSMC foundry team and the Broadcom design team on a continuous-integration basis. The partnership model is, in design, similar to the Google-Broadcom relationship on the TPU program and to the Amazon-Annapurna Labs (now under AWS) relationship on the Trainium program. The model is, in substance, a less vertically integrated version of the Apple’s full-stack silicon approach, with OpenAI capturing the architectural value and the workload-knowledge value, and the foundry and the network-silicon specialists capturing the manufacturing and the network value.

The Broadcom relationship is the most architecturally distinctive of the Jalapeño partnerships. Broadcom is one of the largest custom-silicon houses in the world and has been a substantial supplier of AI infrastructure silicon since the late 2010s. Broadcom’s design team is the principal supplier of the network switches, the network-attached silicon, and the chip-to-chip interconnect silicon that have been used in the largest hyperscaler AI training and inference clusters. The Broadcom relationship with OpenAI is, on the public record, a multi-generation design partnership in which Broadcom supplies both the inference accelerator (in collaboration with OpenAI’s internal team) and a substantial amount of the network-silicon ecosystem that allows the accelerator to scale to the gigawatt scale at which OpenAI’s data-centre footprint requires.

The TSMC relationship is the most production-critical of the Jalapeño partnerships. TSMC is the only foundry in the world with the advanced-node and the high-volume-manufacturing capacity needed to fabricate an AI accelerator at the scale OpenAI requires. The TSMC relationship on Jalapeño is, on the public record, a multi-year capacity commitment that gives OpenAI a contracted allocation of advanced-process and advanced-packaging capacity. The TSMC relationship is also the most-exposed of the Jalapeño partnerships to the geopolitical risk of TSMC’s manufacturing concentration in Taiwan. OpenAI’s September 2026 Samsung partnership announcement should be read, in part, as a hedge on that concentration risk: Samsung is the only foundry outside TSMC with a comparable advanced-node capacity, and the Samsung partnership gives OpenAI a second foundry option for Jalapeño’s next generation.

The Samsung partnership is, on the public record, broader than just Jalapeño. The September 2026 announcement framed the partnership as one on next-generation chip development and on enterprise AI services. The chip-development piece is, in substance, a multi-year foundry and chip-design arrangement in which Samsung’s foundry and Samsung’s chip-design organization work with OpenAI on the second and third generations of Jalapeño. The enterprise AI services piece is a commercial arrangement in which Samsung’s enterprise-customer base becomes a channel for OpenAI’s AI services, with Samsung providing enterprise-deployment infrastructure that OpenAI’s own sales force does not directly cover. The combined Samsung partnership is, in other words, both a manufacturing hedge and a commercial hedge for OpenAI’s broader business, and it is one of the more strategically complex partnership announcements of the late 2020s.

Jalapeno at gigawatt scale

The Jalapeño deployment plan, as it stood in late 2026, called for the chip to be deployed at gigawatt scale across OpenAI’s data-centre footprint over the next three to four years. The phrase “gigawatt scale” is, in the AI infrastructure industry, a precise unit. A gigawatt of AI compute is a substantial fraction of the largest single hyperscaler’s total compute capacity, and it is the order of magnitude at which the AI infrastructure industry is now planning new data-centre campuses. OpenAI’s stated gigawatt-scale deployment plan for Jalapeño is, in substance, a commitment to substantially replace the merchant GPU capacity in OpenAI’s data-centre footprint with Jalapeño over the next several years, with the OpenAI workload running on Jalapeño silicon and the merchant GPU capacity being retained for the workloads that Jalapeño is less well-suited to handle (typically training and certain mixed-training-and-inference workloads). The plan is, in scale, comparable to the Microsoft Maia gigawatt plan, comparable to the Anthropic TPU commitment, and substantially larger than any publicly disclosed custom-silicon deployment commitment from a Western AI lab other than the hyperscalers.

The data-centre partners through which Jalapeño will be deployed are, in aggregate, a substantial part of OpenAI’s existing AI infrastructure footprint. The partnerships include the Microsoft Azure deployments (the multi-year capacity commitments from the early 2020s are still being delivered), the Oracle data-centre deployments (which have, on the public record, become substantially larger through 2026), the Stargate infrastructure joint venture (the substantial SoftBank-backed build-out announced in early 2025), the AWS deployments (which OpenAI has expanded as part of its broader commercial strategy), and a number of additional neocloud and dedicated AI infrastructure deployments. The Jalapeño silicon is, in effect, a chip that will be deployed across all of these data-centre partners, with the chip being shipped to each data centre as it is fabricated and the OpenAI workload being migrated to Jalapeño in waves as the silicon volume becomes available.

The financial implications of the Jalapeño program are substantial. The chip is, by OpenAI’s planning assumption, expected to be a substantial fraction of OpenAI’s total inference silicon cost over the next several years. The cost saving relative to the merchant-GPU alternative is, on the assumed magnitude of the OpenAI inference fleet, comparable to several billion dollars per year in operating cost. The cost saving is, however, conditional on the chip hitting its design performance at volume production. The engineering risk that the chip will fall short of its design performance at volume is, despite the August 2026 reporting, real. The TSMC manufacturing risk that the chip’s production yield will not meet OpenAI’s volume requirements is also real. The financial benefit of the Jalapeño program is, in other words, contingent on the engineering and the manufacturing risks being managed, and the contingency is the largest single execution risk in OpenAI’s current business plan.

The strategic significance of the Jalapeño deployment at gigawatt scale is that OpenAI is, in substance, building a vertically integrated AI lab in the style of the hyperscalers. The vertical integration has three layers. The model layer is the OpenAI capability on which the company has spent the last decade building. The infrastructure layer is, increasingly, OpenAI’s own silicon plus the data-centre capacity that the silicon is deployed in. The deployment layer is, increasingly, the OpenAI customer-facing products on which the revenue is realized. The vertical integration is not, in absolute terms, as complete as Microsoft’s or as Amazon’s vertical integration. OpenAI does not own the foundry. OpenAI does not own the data centres in which the silicon is deployed. OpenAI does not, for the most part, own the customer relationship through which the AI service is sold to enterprise customers. The vertical integration is, however, substantially more complete than OpenAI’s integration was in 2022, when the company depended, for almost all of its infrastructure, on third-party cloud and third-party silicon.

The standards push and the strategic-context story

On September 21, 2026, OpenAI called for the United States government to lead an international effort to develop technical standards for frontier AI, including for systems “capable of recursive self-improvement.” The call was, on the surface, a request for governance. It was, in substance, an attempt to shape the standards environment in which OpenAI and its competitors are operating. The standards environment matters to OpenAI in part because of the export-control regime: the United States export-control regime on advanced AI silicon has, since 2022, been the principal supply constraint on the AI compute base in China and on the AI compute base available to Chinese AI labs. OpenAI’s call for international standards was, in part, a request that the next iteration of the export-control regime be coordinated with allies rather than imposed unilaterally, and a request that the standards be scoped narrowly enough that the export-control burden falls on specific applications rather than on general-purpose AI silicon.

The standards push is also, in substance, a strategic-move that is tightly connected to the Jalapeño program. A coordinated international standards regime on frontier AI is, from OpenAI’s perspective, an opportunity to shape the regulatory environment in which the company’s vertically integrated stack — OpenAI models running on Jalapeño silicon, deployed across OpenAI’s data-centre footprint, sold through OpenAI’s enterprise-product channels — operates. The standards regime would, on OpenAI’s proposal, set the terms of competition between OpenAI and its competitors, would set the technical requirements that frontier-AI systems would have to meet, and would set the export-control burden on the AI compute base that OpenAI’s competitors in China depend on. The proposal is, in substance, an attempt by OpenAI to use the standards regime as a strategic instrument in the competition between OpenAI and its rivals.

The strategic-context story behind Jalapeño is, in other words, broader than just the silicon bet. The vertical integration is, on OpenAI’s planning, the foundation of a vertically integrated AI business in which OpenAI controls the model, the silicon, the data centre, and the customer relationship. The standards push is, in turn, the strategic instrument by which OpenAI is trying to make the international environment favorable to a vertically integrated AI business, and to constrain the strategic options available to OpenAI’s competitors. The combined Jalapeño-plus-standards push is, in the late 2020s, the most ambitious vertical-integration strategy in the AI industry outside the hyperscalers, and it is the strategy on which OpenAI’s long-term competitive position is going to substantially depend.

The strategic risk of the Jalapeño-plus-standards strategy is that it depends on substantial success in both the silicon program and the standards push. The silicon program has, on the August 2026 reporting, hit its early performance milestones, but the deployment at gigawatt scale is years away and the long-term economics are uncertain. The standards push has, on the public record, not yet produced the coordinated international regime that OpenAI has called for, and the export-control burden on Chinese competitors is, on the basis of the 2025–2026 reporting, likely to remain in place regardless of the standards regime. The cumulative strategic risk is that the Jalapeño program will succeed as an engineering program but that the strategic context will turn out to be less favorable than OpenAI’s leaders had assumed when they committed to the program. The risk is real and it is large, but it is the cost of betting on a vertically integrated AI business in a strategically competitive industry.

What Jalapeno means for the AI infrastructure industry

The Jalapeño program is the third largest Western custom-silicon program after Google’s TPU and Amazon’s Trainium, and it is the most strategically significant of the programs because it is the first such program owned by an AI lab rather than a hyperscaler. The implication is that the custom-silicon wave is no longer constrained to the hyperscalers; it has, with the Jalapeño program, extended to the AI lab itself. The AI lab is, in OpenAI’s case, the largest single customer of AI compute in the world, and the lab’s decision to design its own silicon is, in effect, a decision to internalize the silicon cost of its own AI service business.

The implication for Nvidia is substantial. Nvidia’s merchant GPU business was, for the second quarter of fiscal 2027 (the period most recently reported through October 2026), running at a 89-billion-dollar data-centre revenue figure on its own, and the data-centre business was approximately 92 percent of Nvidia’s total revenue. The Jalapeño program does not, by itself, displace a substantial fraction of that revenue. The program does, however, indicate that the largest single AI compute consumer is, like Microsoft’s Maia program and like Google’s TPU program, on a trajectory to substantially reduce its dependence on merchant GPU silicon over the next several years. The cumulative effect of the OpenAI, Microsoft, Google, Amazon, Meta, and Anthropic custom-silicon programs on Nvidia’s merchant GPU revenue is, by the late 2020s, expected to be the largest single demand-side risk to Nvidia’s growth trajectory.

The implication for the broader AI infrastructure industry is that vertically integrated AI businesses are likely to become a more significant part of the industry structure than they were in the early 2020s. OpenAI is, on its published investor disclosures, no longer simply an AI model company; it is increasingly an AI infrastructure company that uses its own models to design its own silicon, partners with foundries to fabricate the silicon, and partners with hyperscalers and other data-centre operators to host the silicon. The vertical integration has, by the late 2020s, become a feature of the AI industry that is present in at least three of the largest Western AI companies and is increasingly present in the Chinese AI industry as well. The implication for the AI service market is that the service market is going to be shaped, increasingly, by the cost structure of the AI infrastructure on which the services run, and the cost structure is, in turn, going to be shaped by the custom-silicon decisions that the largest AI companies have made.

The Jalapeño program is, in summary, the vertical-integration move by which OpenAI has chosen to internalize a substantial fraction of the silicon cost of its AI service business. The program is, on the public record, engineering on track: the first-generation chip has taped out, the workload performance has been reported as competitive with Nvidia, and the foundry and network-silicon partnerships are in place. The program is, on the public record, strategically significant: the gigawatt-scale deployment plan is comparable in scale to the Microsoft Maia gigawatt plan, and the OpenAI workload is, in turn, one of the largest single workloads in the AI industry. The program is, on the public record, financially significant: the cost saving relative to the merchant GPU is, on OpenAI’s planning assumption, several billion dollars per year. The cumulative judgment is that the Jalapeño program is one of the most important corporate-strategy moves in the AI industry in late 2026, and that its long-term implications are going to unfold over the rest of the decade.