On Wednesday in Seoul, Harrison Kim, general manager of OpenAI Korea, told reporters that the partnership with Samsung Electronics had advanced further than anyone had publicly admitted. He described joint production and research on next-generation chips as the area where the two firms had made the most progress. The framing was measured, but the implication cut cleanly: OpenAI was no longer going to fabricate its custom silicon at a single foundry. As reported by Reuters, OpenAI Korea later clarified there was nothing new to formally announce, but the manufacturing shift is the real story, and it is far more interesting than a press conference soundbite.
This is not the same article as my breakdown of Jalapeño benchmarks from late August. That piece covered a chip that beats Nvidia on efficiency. This one covers what happens when a company that built its entire compute strategy around one foundry decides that single-source dependency is a strategic risk worth paying to reduce. Let me walk through the architecture of the decision, not just the headline.
What OpenAI Actually Confirmed
The Reuters reporting, filed from Seoul, gives us the concrete facts. As reported by Reuters, OpenAI Korea clarified there was nothing new to formally announce, but the manufacturing shift is the real story, and it is far more interesting than a press conference soundbite. Kim said the chip relationship was the area of most progress. OpenAI Korea then contextualized that remark within an existing relationship that includes Samsung adopting ChatGPT and a letter of intent signed last October. The company declined to confirm customer specifics, which is standard for a foundry that cannot disclose who it manufactures for.
The details that matter come from context, not the press release. The chip Samsung is being asked to produce is widely reported to be an inference processor and likely a successor to Jalapeño, OpenAI's first custom accelerator. It was co-designed with Broadcom and manufactured at TSMC. If Samsung takes on fabrication of the next generation, OpenAI gains a second manufacturing partner, which structurally changes its supply posture. As reported by Reuters, the Jalapeño chip was designed in roughly nine months and deployed at gigawatt scale, and Samsung already supplied high-bandwidth memory to its Broadcom partner.
Memory is the other half of the deal. Samsung has started shipping samples of its HBM4E high-bandwidth memory and is reportedly the first supplier to distribute that generation of AI memory. HBM4E sits inside the very accelerators OpenAI is trying to fill with inference workloads. That is not incidental. The chip design and the memory that feeds it have to be co-optimized, which is why a design-only partnership with a fabless startup would not accomplish the same goal.
Why Second Sourcing Changes Everything
Here is the part most coverage misses. Building a custom AI chip is only half the engineering problem. Getting TSMC to allocate limited CoWoS packaging capacity and advanced node (N3P) wafers to your program, ahead of Google's TPUs, Nvidia's GPUs, and AMD's MI series, is the harder half. OpenAI already solved the chip design. It was entirely exposed to TSMC's capacity and priorities.
The math on why this matters is stark. TSMC's advanced packaging is the single most contended resource in the entire AI hardware supply chain. Every hyperscaler is bidding for the same CoWoS slots. If OpenAI has one foundry partner, it is one buyer at one table. If it has two, it can hedge capacity allocations and it gains negotiating leverage that translates directly into delivery priority during a shortage. That is exactly the leverage Nvidia itself has relied on for years.
From an engineering standpoint, dual sourcing also de-risks a specific failure mode. A foundry yield loss, a natural event, or a geopolitical restriction on one site can strand a company with design that cannot be manufactured. Single-source silicon design is fast. Single-source manufacturing is fragile. The pattern across industry history, from Google's TPU program to Amazon's Trainium, is that the labs that survived did so partly because they diversified fabrication as their programs scaled. OpenAI is doing the same thing, just later than its peers.
The HBM Economy In This Deal
High-bandwidth memory is the bottleneck that constrains everything else. Samsung, SK Hynix, and Micron have reportedly sold their HBM supply through 2027, and Nvidia is testing cut-down configurations with as little as 192GB because memory simply is not available. When memory is the constraint, whoever controls memory supply controls a large slice of the leverage.
Samsung bringing HBM4E to market as the first supplier, while simultaneously securing a custom chip manufacturing deal with OpenAI, is a vertically integrated position that very few companies anywhere in the world can claim. It means Samsung is not just a vendor OpenAI buys from. It is a partner that can supply both the silicon fabrication and the memory that the silicon needs. That is the combination that makes this partnership structurally different from a pure design collaboration.
Let me lay out the concrete data points from the sources:
| Fact | Source | Detail |
|---|---|---|
| Jalapeño first custom chip | Reuters, Dealroom | Co-designed with Broadcom, TSMC manufactured, inference-focused |
| Jalapeño tapeout time | Dealroom (OpenAI) | ~9 months, described as fastest ASIC tapeout in advanced semiconductors |
| Jalapeño cost claim | Dealroom (Broadcom) | ~50% cost savings versus typical AI GPUs, per Hock Tan |
| Jalapeño performance | OpenAI, SemiAnalysis (Aug) | Up to 1.9x tokens per kilowatt over GB300, up to 3.6x lower latency |
| Samsung HBM4E status | Cryptopolitan | First supplier to ship HBM4E samples |
| Memory demand growth | Reuters | Kim said memory demand grows as chips get more complex |
| ChatGPT Enterprise in Korea | Reuters | ~28-fold increase in a year, largest outside the US |
What The Timeline Really Looks Like
It is important to calibrate expectations using the closest precedents. Google began its TPU program internally years before it became a major public component of its infrastructure, and it took multiple hardware generations before TPUs were credited with training some of Google's most capable models at meaningful scale. Amazon's Trainium and Inferentia followed a similar multi-year arc, starting with early performance gaps against Nvidia and only gradually gaining production adoption.
The consistent pattern across every custom chip program is that the distance between "partnership announced" and "chips running a meaningful share of production workloads" is measured in years, not months. OpenAI entered this space later than Google, Amazon, and Meta, and it is doing so through a partnership rather than a fully in-house design team. A reasonable expectation is that Samsung-partnered silicon becomes a meaningful fraction of OpenAI's compute stack only in the second half of this decade, likely starting with isolated inference workloads that are easier to validate than full-scale training runs.
That timeline matters for how you read the news today. This is a multi-year hedge against supply-side bottlenecks, not something that changes API pricing, token economics, or model performance this year. Builders should treat it as a structural signal about compute availability, not a near-term catalyst.
Our Read
Three things, from both the engineering and the market perspective.
First, the memory angle is the real moat. Everyone is focused on the chip design, but the chip is useless without HBM4E, and Samsung is now both the fab partner and the memory partner for OpenAI. That vertical coupling is worth more than the design win alone. Any competitor building a custom chip that cannot secure memory will stall, and OpenAI just removed that risk from its roadmap.
Second, this is a quiet admission that single-source risk is now a first-class engineering concern. OpenAI spent over two years leaning almost entirely on NVIDIA GPUs and a single foundry. The Samsung move is the market correcting that posture. It mirrors what Anthropic did with exclusive infrastructure deals like the SpaceX Colossus 1 supercomputer, and what Microsoft is doing with its own Maia 300. The labs are converging on the same conclusion through different deal structures.
Third, the competitive geometry is shifting under Nvidia. Nvidia dominates both training and inference through GPUs that are hard to replace on their narrow workload, but the economics of inference are moving toward custom silicon that does the same work cheaper per token. See also: how Nvidia is responding to that pressure by building full systems rather than just chips. OpenAI's own Jalapeño chip already beat Nvidia Blackwell on efficiency, so this is not speculation. The question for Nvidia is not whether custom silicon works, but whether it can scale fabrication fast enough before the hyperscalers stop being its best customers.
The broader implication for the industry is that the AI hardware race is bifurcating. On one side, general-purpose GPUs that flex across workloads. On the other, purpose-built ASICs co-designed with memory and fabrication already locked in. The winners will be measured not by whose chip has the best benchmark number, but by whose supply chain can actually deliver volume at scale. That is a much harder game, and it is one OpenAI is now learning to play on two fronts.