OpenAI Samsung Chips Plan Signals a Foundry Split With TSMC

OpenAI says it is working with Samsung on next-generation processors, despite relying on TSMC to manufacture its first custom AI accelerator. The OpenAI Samsung chips disclosure introduces a possible second foundry path for a program already targeting deployment at enormous scale.

Harrison Kim, general manager of OpenAI Korea, described progress in joint processor production and research during a September 9 press conference in Seoul. His remarks, reported by next-generation chip coverage, did not identify a product, manufacturing process, production date, or order volume.

That missing detail is central to the story. OpenAI and Broadcom have already unveiled Jalapeño, an inference accelerator expected to enter OpenAI data centers before the end of 2026. TSMC manufactures that first-generation processor.

Samsung might manufacture a later Jalapeño generation, a separate OpenAI processor, or a jointly developed design. Each possibility carries different technical and commercial consequences.

The most important signal is therefore not a confirmed Samsung production award. It is OpenAI’s willingness to build more than one manufacturing route while expanding its custom silicon program.

That strategy puts pressure on TSMC, Samsung, Nvidia, and OpenAI itself. TSMC must defend a valuable advanced-node customer, while Samsung must prove it can deliver consistent yields and packaging capacity.

Nvidia faces a customer that wants greater control over inference economics. OpenAI must turn an ambitious supplier network into reliable hardware without fragmenting its software and deployment systems.

OpenAI Has Expanded Samsung From Memory Into Processor Work

The relationship now reaches beyond memory procurement, but OpenAI has not defined Samsung’s exact manufacturing role.

Kim said OpenAI had made substantial progress with Samsung in joint production and research for processors under development. He offered no technical specifications or commercial terms during the Seoul briefing.

The wording matters because semiconductor “production” can describe several activities. Samsung could fabricate the processor die, contribute design services, provide packaging, supply memory, or combine several roles.

OpenAI also has not said whether the discussions concern Jalapeño. The first processor in that family was developed with Broadcom and manufactured by TSMC.

The original chip report noted several unresolved possibilities. Samsung could become a second source, manufacture another OpenAI design, or participate in an undisclosed future processor.

These options should not be treated as interchangeable. Porting one advanced chip between foundries requires far more than sending the same design files to another factory.

Each foundry uses its own process design kit, transistor libraries, physical rules, and manufacturing optimizations. Engineers often must revise major parts of a design for another production process.

A true second-source version would therefore require substantial engineering work. It would also need separate validation, packaging qualification, performance tuning, and software testing.

A future generation designed for Samsung from the beginning presents a different path. OpenAI could divide its roadmap between foundries instead of reproducing one processor at both suppliers.

For example, one foundry might handle a high-volume inference processor. Another might manufacture a specialized part with different performance, power, or packaging requirements.

OpenAI has not confirmed such a division. Still, its language suggests that Samsung’s involvement extends beyond the previously announced memory partnership.

That earlier relationship already covered several Samsung companies. Samsung Electronics agreed to supply advanced memory, while other affiliates explored data center, maritime, and infrastructure projects.

OpenAI projected demand reaching up to 900,000 DRAM wafer starts each month for Stargate. A wafer start represents a silicon wafer entering the manufacturing process, before it becomes individual memory chips.

The Samsung infrastructure agreement also included SK hynix, another leading memory producer. The companies described accelerated capacity expansion rather than a fixed near-term purchase schedule.

Adding processor collaboration changes the strategic profile of that partnership. Memory makes Samsung an important component supplier, but processor fabrication could place it inside OpenAI’s core compute roadmap.

The distinction remains important. OpenAI has disclosed collaboration, not a finalized high-volume foundry contract.

Readers should treat the announcement as a directional signal. It reveals how OpenAI is organizing future supply, even though the immediate production details remain private.

The Scale Behind the OpenAI Samsung Chips Strategy

A second foundry becomes rational when expected demand, supply risk, or roadmap breadth outweighs the cost of supporting two manufacturing paths.

OpenAI’s custom accelerator program is not framed as a limited experiment. The company and Broadcom previously announced plans to deploy 10 gigawatts of OpenAI-designed accelerators and networking systems.

Deployment was scheduled to begin during the second half of 2026. The companies said the rollout would continue through the end of 2029.

Ten gigawatts describes power capacity, not chip quantity. The number of deployed processors will depend on rack design, utilization, supporting equipment, and each accelerator’s power envelope.

Even with those qualifications, the target indicates industrial-scale infrastructure. It also explains why OpenAI might value manufacturing redundancy before later processor generations reach production.

The company’s custom accelerator program connects processor design, networking, racks, and model serving. That integration aims to optimize the complete inference system rather than one isolated chip.

Inference is the computation used when a deployed model answers prompts, generates media, or performs agent tasks. Its economics depend on throughput, latency, energy consumption, memory, networking, and utilization.

A processor optimized around OpenAI’s own workloads can improve those variables. However, specialized hardware only creates savings when enough workloads can use it consistently.

That requirement favors volume. OpenAI operates consumer products, business services, developer APIs, coding systems, and research workloads with different demand patterns.

A custom processor must support enough of those workloads to justify its development and deployment costs. It must also keep working as OpenAI’s models and serving methods change.

The company’s infrastructure ambitions extend beyond its processor program. OpenAI introduced Stargate with an intention to invest $500 billion across four years in United States AI infrastructure.

The Stargate infrastructure plan began with OpenAI, SoftBank, Oracle, and MGX as initial participants. Microsoft, Nvidia, Arm, and Oracle were named as technology partners.

OpenAI later expanded the initiative internationally. Its South Korean agreements added memory production, data center planning, and related infrastructure work.

The 900,000 monthly DRAM wafer projection offers another scale indicator. It is not a forecast for processor wafers, and the two figures should not be combined.

However, memory demand and accelerator demand rise together inside AI systems. Advanced accelerators need high-bandwidth memory, which places memory stacks close to the computing die.

Packaging those components creates another potential bottleneck. Manufacturing the processor does not solve capacity constraints if memory or advanced packaging remains unavailable.

A Samsung foundry route could give OpenAI additional options across that chain. Samsung manufactures logic chips, memory, and packaging technologies within one corporate group.

Vertical breadth does not automatically produce better results. Coordination, yield, thermals, cost, and production schedules still determine whether integration creates an advantage.

The likely objective is flexibility rather than a clean replacement for TSMC. OpenAI can negotiate from a stronger position if multiple suppliers can support different parts of its roadmap.

It can also reduce exposure to disruptions affecting one production region or process. That benefit grows when infrastructure commitments extend across several years.

Double-sourcing carries a price. OpenAI and Broadcom would need engineers, validation systems, software support, and operational processes for each manufacturing path.

The company appears willing to explore that trade because processor availability has become a strategic constraint. Compute is no longer only a procurement function for frontier AI developers.

It now shapes model release timing, serving capacity, product margins, and the number of customers a platform can support. That makes foundry planning part of OpenAI’s product strategy.

Samsung Must Prove It Can Be More Than a Backup

The primary contest is Samsung versus TSMC for a durable place in OpenAI’s processor roadmap, not one company replacing the other immediately.

TSMC manufactures chips designed by many of the largest semiconductor companies. Its advanced manufacturing and packaging capacity sits at the center of the current AI hardware market.

OpenAI’s first processor follows that established route. Broadcom helped transform OpenAI’s architecture into manufacturable silicon, while TSMC fabricated the design.

Jalapeño was developed from initial design through tape-out in nine months, according to OpenAI. Tape-out is the point when a completed chip design is sent for manufacturing preparation.

OpenAI describes Jalapeño as an intelligence processor optimized for large language model inference. The company began testing early chips before planned deployment later in 2026.

The Jalapeño specifications identify Broadcom as the implementation, networking, and connectivity partner. Celestica contributes board, rack, and system expertise.

That arrangement gives TSMC the advantage of an existing, validated design. Samsung must offer more than nominal capacity to win an enduring role.

Manufacturing yield will be one test. Yield measures the share of fabricated dies that meet performance and quality requirements after production.

Low yields can raise costs and limit available chips, even when a factory processes many wafers. Stable yields matter especially for large, complex accelerators.

Power characteristics present another test. Inference data centers must manage electricity delivery, cooling, networking, and memory alongside raw computing performance.

Small differences at the chip level can become significant across thousands of racks. OpenAI will therefore evaluate usable system output, not merely transistor density.

Packaging is equally important. Modern AI accelerators often combine a logic die with multiple high-bandwidth memory stacks through advanced interconnects.

That integration can restrict production even when logic wafers are available. Samsung’s ability to coordinate memory and packaging could strengthen its proposal.

TSMC still benefits from a broad supplier network and extensive experience with leading AI processors. Its manufacturing processes also support Nvidia, AMD, and many custom silicon programs.

Those overlapping customers create both strength and tension. A large customer base validates TSMC’s processes, but it also creates competition for scarce production slots.

OpenAI may see Samsung as leverage against that concentration. A qualified alternative can improve scheduling flexibility and reduce dependence on one manufacturing organization.

Samsung has its own incentives. A visible OpenAI production program would support its effort to attract advanced foundry customers and strengthen confidence in its manufacturing execution.

The opportunity extends beyond one order. Successful production could position Samsung for later OpenAI accelerators, networking silicon, and related infrastructure components.

Yet Samsung cannot win through capacity promises alone. OpenAI needs predictable delivery because data centers are planned around synchronized equipment, power, construction, and networking schedules.

A late processor can leave expensive infrastructure underused. A processor with inconsistent characteristics can complicate cooling, software tuning, and service reliability.

This is why the foundry contest is not simply about headline process nodes. OpenAI will judge the complete route from design rules through deployed model performance.

TSMC retains the strongest position because Jalapeño already uses its manufacturing. Samsung’s opening lies in the generations that follow, where designs and production plans remain less settled.

If Samsung receives a dedicated future design, the relationship could become complementary. TSMC and Samsung might manufacture different processors for separate OpenAI workloads.

If Samsung qualifies the same processor as TSMC, the relationship would look more like direct double-sourcing. That outcome would provide stronger evidence of exceptional volume requirements.

OpenAI has not said which model it is pursuing. Until it does, Samsung should be viewed as a potential second manufacturing pillar, not a confirmed equal supplier.

The Foundry Split Does Not Remove OpenAI’s Bottlenecks

Manufacturing diversification reduces one concentration risk, but it can add engineering complexity while leaving memory, packaging, power, and software constraints intact.

The optimistic interpretation is straightforward. OpenAI expects so much processor demand that one foundry route cannot provide enough capacity or resilience.

A more cautious interpretation is also plausible. OpenAI may be exploring options without committing production to Samsung.

Companies often conduct joint research before selecting a process or signing a volume contract. Not every technical collaboration becomes a shipping product.

Kim’s statement did not identify a factory, process node, processor architecture, tape-out, or production schedule. Those omissions prevent independent verification of Samsung’s expected contribution.

The phrase “next-generation chips” also remains broad. It might describe a Jalapeño successor, a training accelerator, networking silicon, or another specialized processor.

Training and inference place different demands on hardware. Training creates models by processing large datasets, while inference uses trained models to produce outputs.

Jalapeño targets inference. OpenAI has not publicly committed a Samsung-manufactured training processor.

That distinction affects the competitive analysis. A future inference design would extend the company’s current program, while a training accelerator would challenge Nvidia in a different workload.

Software compatibility adds another uncertainty. OpenAI must make applications run predictably across its custom processors and the GPUs already operating in its infrastructure.

Models, kernels, compilers, schedulers, and monitoring systems all influence real performance. Hardware advantages can disappear when software cannot use them efficiently.

Supporting processors from separate foundries could create additional variation. Even similar designs may have different frequency, voltage, thermal, or packaging characteristics.

OpenAI can manage those differences through abstraction and workload scheduling. However, doing so requires mature internal infrastructure and extensive validation.

The company must also preserve model development flexibility. A highly specialized chip can become less useful when new architectures demand different memory access or numerical formats.

Nvidia’s general-purpose GPU platform remains attractive partly because it supports changing workloads. Developers can use a broad software ecosystem instead of targeting one customer-specific accelerator.

OpenAI’s processor strategy therefore complements Nvidia before it replaces any substantial share of Nvidia hardware. Custom inference chips can absorb stable, high-volume workloads.

GPUs can continue handling research, training, unusual models, and tasks that require greater programmability. The balance will depend on OpenAI’s software and deployment results.

Supply concentration also extends beyond the processor foundry. Both TSMC and Samsung depend on specialized equipment, materials, intellectual property, and global logistics.

Memory availability could remain limiting even with two logic suppliers. OpenAI’s Samsung and SK hynix agreements acknowledge the scale of that separate requirement.

Advanced packaging might become the binding constraint instead. A completed logic die cannot serve a model until memory, substrates, networking, and cooling are integrated.

Power is another hard boundary. OpenAI can order processors faster than utilities and builders can deliver grid connections, substations, generators, and cooling systems.

Stargate addresses these interconnected requirements, but coordination becomes harder as the program expands. Each component must arrive near the correct site and schedule.

Commercial discipline presents a final risk. Custom silicon can lower operating costs, but only after development, manufacturing, deployment, and software expenses are included.

An accelerator that performs well in controlled tests may behave differently under mixed production traffic. Utilization and reliability will determine the actual economic result.

OpenAI has published performance claims for Jalapeño, but the company remains the source of those claims. Independent, production-scale evidence is still limited.

The same standard should apply to future Samsung chips. A foundry announcement would establish intent, not prove competitive performance or dependable volume production.

The OpenAI Samsung chips plan becomes significant only when named products reach qualified manufacturing. Until then, the verification gap is part of the story.

Nvidia Faces Pressure From Ownership, Not Immediate Replacement

OpenAI’s deeper change is its attempt to own more of the inference stack, while Nvidia remains essential to its broader computing capacity.

OpenAI historically obtained much of its AI computing through Nvidia accelerators deployed by Microsoft and other infrastructure partners. That arrangement enabled rapid growth without owning every hardware layer.

It also left important economics outside OpenAI’s control. Accelerator availability, networking architecture, software dependencies, and supplier margins all affect the cost of serving models.

Custom silicon gives OpenAI another lever. The company can design features around its models, serving systems, kernels, and expected production workloads.

Broadcom provides experience in custom application-specific integrated circuits, known as ASICs. An ASIC performs a defined class of work instead of serving as a general-purpose processor.

Google offers the clearest industry precedent. Its tensor processing units support internal AI services and cloud customers alongside third-party accelerators.

Amazon and Microsoft have also developed custom AI processors for their cloud platforms. These programs show that custom chips can coexist with large Nvidia purchases.

OpenAI follows a related path but begins from a different position. It operates major AI products without owning a hyperscale cloud platform comparable to those companies.

That makes its partnerships more important. Broadcom, TSMC, Samsung, memory suppliers, rack builders, cloud operators, and data center developers must function as one system.

Samsung adds another potential production route, but it also expands the coordination burden. OpenAI must align designs and deployments across corporate boundaries.

Nvidia still possesses several defenses. Its software tools, developer familiarity, product cadence, networking portfolio, and installed base make replacement difficult.

A custom chip does not need to replace Nvidia everywhere to matter. It only needs to handle a large, repeatable portion of OpenAI’s inference demand effectively.

Moving stable workloads onto Jalapeño could reserve Nvidia systems for training and flexible tasks. It could also give OpenAI better negotiating leverage with hardware providers.

The strategic pressure comes from ownership of workload knowledge. OpenAI sees its own model behavior, traffic patterns, latency targets, and serving bottlenecks.

It can encode that knowledge into hardware and systems. Nvidia must serve many customers, so its processors must remain useful across broader workload categories.

That creates a tradeoff between specialization and flexibility. OpenAI can optimize tightly, while Nvidia can spread development costs across a much larger market.

Samsung and TSMC compete one level below that contest. They manufacture the processors that allow OpenAI’s specialization strategy to reach physical data centers.

A dual-foundry roadmap could make OpenAI less vulnerable to production shortages. It could also accelerate iteration if each foundry supports different product generations.

However, splitting designs can dilute engineering resources. Maintaining one competitive custom accelerator family is already a demanding task.

Maintaining separate variants, toolchains, and qualification programs raises the bar. OpenAI must show that expected scale justifies those additional costs.

Developers and enterprise buyers should care because infrastructure choices eventually shape product behavior. Serving capacity affects response times, availability, model access, and release schedules.

Processor economics can also influence which tasks OpenAI prioritizes. Efficient inference makes sustained agent workflows and high-volume applications easier to support.

Teams evaluating AI services should still focus on measured product outcomes. Foundry partnerships do not guarantee faster models, lower customer costs, or improved reliability.

They do indicate that OpenAI expects infrastructure ownership to become a competitive advantage. That shift will influence rivals, cloud providers, and semiconductor suppliers.

Three Signals Will Show Whether Samsung Has a Real Production Role

The next decisive evidence will come from a named chip, a qualified manufacturing milestone, and visible production deployment.

The first signal is product identification. OpenAI or Samsung must connect the collaboration to a specific processor family and define its intended workload.

A named Jalapeño successor would confirm continuity in the inference roadmap. A training processor would reveal a broader attempt to control OpenAI’s computing stack.

A separate networking or specialized chip would narrow the competitive impact. It would show deeper collaboration without establishing Samsung as an alternative accelerator foundry.

Product identification would strengthen the double-sourcing thesis only if the same processor family also remains connected to TSMC. Otherwise, OpenAI may simply be allocating different designs.

The second signal is a manufacturing milestone. Investors and customers should watch for a process selection, tape-out, sample delivery, qualification, or volume production schedule.

A tape-out would show that engineers have committed a design to Samsung’s manufacturing rules. Working samples would move the project beyond design-stage collaboration.

Qualification would matter even more. It would indicate that the processor, memory, packaging, and system components meet defined production requirements.

Volume production would provide the clearest validation. At that point, Samsung would become an operating part of OpenAI’s compute supply rather than a research partner.

The third signal is deployment evidence. OpenAI should identify when Samsung-manufactured processors enter data centers and what workloads they handle.

Useful evidence would include system availability, sustained throughput, latency, power use, failure rates, and the share of eligible traffic placed on the hardware.

Company benchmarks would offer an initial view, but independent measurements would improve confidence. Production behavior matters more than isolated laboratory results.

Deployment at multiple sites would further support the scale argument. It would show that OpenAI has integrated the chips into repeatable rack, networking, cooling, and software configurations.

Failure to produce these signals would weaken the story. Continued references to broad collaboration without products or milestones would suggest an exploratory relationship.

Clear milestones would strengthen it. They would show that OpenAI is building a multi-foundry processor roadmap around long-term infrastructure demand.

The OpenAI Samsung chips disclosure already tells the market something important. OpenAI does not want its custom silicon ambitions limited by one manufacturing relationship.

What remains unknown is whether Samsung becomes a genuine second source or receives a distinct processor assignment. That distinction will determine who faces the greatest pressure.

A second source would challenge TSMC more directly and confirm unusually large expected volumes. A separate design would show a wider OpenAI hardware portfolio with more specialized supply chains.

Either outcome would deepen OpenAI’s commitment to owning its inference infrastructure. Neither outcome guarantees that the resulting processors will outperform broadly available alternatives.

Developers, enterprise buyers, and AI product teams should watch deployment results rather than partnership language. The relevant question is whether custom silicon improves services under real demand.

Track named products, manufacturing milestones, and production workloads over the coming months. Those signals will reveal whether Samsung is a negotiating option or OpenAI’s second foundry pillar.