OpenAI is preparing to outsource production of at least some of its processors to Samsung Foundry, according to Harrison Kim, General Manager of OpenAI Korea, as reported this week by Tom's Hardware. If the plan proceeds as described, OpenAI would source its AI accelerators from both TSMC and Samsung Foundry — a dual-sourcing arrangement that, as the publication notes, hints at massive volume requirements for the company's custom silicon program.
The disclosure expands OpenAI's relationship with Samsung beyond memory supply and enterprise software into advanced logic manufacturing, putting the ChatGPT maker alongside the small group of companies with chips in production at both of the world's leading foundries. For more context on this story, see our ongoing latest AI developments.
Why OpenAI Builds Custom Chips at All
OpenAI's custom silicon program exists for the same reason other hyperscale AI operators design their own accelerators: frontier training runs and inference at ChatGPT's scale consume enormous amounts of compute, and buying every accelerator from a single merchant vendor concentrates cost, supply risk and negotiating power in one counterparty. A custom chip designed around OpenAI's own workloads can improve performance per dollar and per watt, and Tom's Hardware's reporting on the Jalapeño program indicates those chips are now running in production rather than remaining on a roadmap.
The catch with custom silicon is manufacturing capacity. A designed chip is only as valuable as the foundry slots available to build it, and advanced-node capacity is booked years ahead by the industry's biggest spenders. That is the backdrop against which Kim's comments about Samsung Foundry land.
Which Chips Would Samsung Make?
Exactly which processors Samsung Foundry would fabricate remains unclear. Tom's Hardware reports that OpenAI did not explain whether Samsung's participation concerns a second production source for Jalapeño — the codename for OpenAI's first-generation custom AI chip — a separate processor under development, a chip jointly developed by the two companies, or some other aspect of chip production.
The publication's analysis suggests the first-generation Jalapeño is unlikely to be the chip in question: it is already in mass production at TSMC, and OpenAI's comments referred specifically to "next-generation" processors. Development of Jalapeño's successor is well underway and approaching tapeout, the milestone where a finalized chip design is sent off for manufacturing — meaning its mass production is not far off either, which would make that chip a plausible candidate for a dual-source arrangement.
Why Dual-Source AI Chips
Dual sourcing is standard practice for products with enormous, predictable demand and tight supply constraints. Splitting production between two foundries protects a buyer against capacity bottlenecks, yield problems and geopolitical concentration risk at a single supplier — and it gives the customer leverage on pricing and allocation. For OpenAI, whose data center buildout under the Stargate initiative depends on a steady flow of accelerators, those risks are anything but theoretical.
Samsung and SK hynix already supply memory for OpenAI's Stargate data center initiative, as Tom's Hardware notes, though joint chip development and production have little relation to DRAM supply. Moving into foundry services would make Samsung a far more strategic partner — and mark one of the highest-profile logic wins for Samsung Foundry, which has been competing to close the technology and reputation gap with TSMC.
A Deepening Samsung Relationship
For Samsung, the potential deal would validate years of investment in advanced nodes as it seeks marquee customers for its most advanced process technology. For OpenAI, adding a second foundry reduces dependence on TSMC capacity, which remains heavily contested by Nvidia, Apple, AMD, Qualcomm and virtually every other advanced chip designer.
Kim's comments did not come with a formal contract announcement, and neither OpenAI nor Samsung has published production timelines. The company's custom silicon program has moved quickly, however: the first-generation chip reached mass production this year according to Tom's Hardware's account, and the successor's approach to tapeout suggests any Samsung involvement would need to be finalized soon to influence that chip's manufacturing plan.
What to Watch
Three signals would confirm the shift: a formal announcement from either company, Samsung Foundry appearing in OpenAI's supply chain disclosures, and eventually teardown data showing Samsung-fabricated accelerators in Stargate deployments. Kim's remarks arrived the same week that OpenAI's next-generation chip is approaching tapeout, suggesting any agreement would need to mature quickly to affect that chip's production run.
Until then, the safest reading is the one Tom's Hardware itself offers — the very fact that OpenAI is discussing dual sourcing points to chip volumes large enough that a single foundry may not suffice.
For the broader AI industry, the takeaway is that custom accelerator programs are entering an industrialized phase. The era in which a leading AI lab could satisfy its silicon needs from one foundry's allocation is giving way to supply chains that look more like those of the smartphone giants — parallel sources, locked-in capacity and multi-year commitments.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
