SK Hynix will ship 900,000 DRAM wafers a month to OpenAI. That more than doubles the entire HBM industry's current capacity.

A Hardware Play, Not a Sales Outpost

OpenAI's Seoul office, which Chief Strategy Officer Jason Kwon inaugurated in September 2025 as the company's 12th global outpost and third in Asia, opened alongside hardware commitments that make the "sales outpost" framing look generous. On October 1, 2025, OpenAI revealed binding letters of intent and memoranda of understanding with Samsung Electronics, SK Hynix, SK Telecom, Samsung C&T, Samsung Heavy Industries, Samsung SDS, and South Korea's Ministry of Science and ICT, a sweep covering memory fabrication, data center construction, floating data center development, and enterprise AI resale in Korea. President Lee Jae Myung met Sam Altman at the Yongsan presidential office in Seoul with Samsung Electronics Executive Chairman Lee Jae-yong and SK Group Chairman Chey Tae-won, a level of government-and-corporate coordination that signals state-level infrastructure alignment, not a commercial side deal.

The numbers expose the real commitment. SK Hynix signed a letter of intent to supply high-bandwidth memory for OpenAI's Stargate project at volumes reaching 900,000 DRAM wafers per month, which SK Group said more than doubles current HBM industry capacity. Samsung Electronics matched that target, committing to scale advanced memory chip production to the same wafer-per-month figure across the two partners. That demand projection is not a forecast for Korean consumer ChatGPT adoption; it is a procurement plan for GPU-attached memory on a scale that only makes sense if OpenAI is designing its Stargate data center memory architecture around Samsung and SK Hynix output. SK Chairman Chey Tae-won said the partnership spans "the full AI stack: memory semiconductors, data centers, energy, and networks," and framed it as the starting point for collaboration on "next-generation AI computing solutions" including new memory-computing architectures. Altman, for his part, said OpenAI intends to "support the sovereign AI needs of Korea" and called the country a future "top-three global AI nation."

SK Telecom signed an MOU to co-develop and operate an AI data center in the southwest region of Korea under the "Stargate Korea" label. Samsung C&T and Samsung Heavy Industries committed to jointly develop floating data centers, an approach Samsung said can ease land scarcity, cut cooling costs, and reduce carbon emissions compared to land-based facilities. Samsung SDS entered a letter of intent to design, build, and operate Stargate data centers while reselling ChatGPT Enterprise to Korean companies. OpenAI also signed a separate MOU with the Ministry of Science and ICT to evaluate AI data center sites outside the Seoul metropolitan area, supporting the Lee government's push for regional balanced growth. The Samsung semiconductor division's press release listed the partnership's scope as covering "advanced semiconductors, data centers, shipbuilding, cloud services, and maritime technologies," a span that reads like a procurement schedule, not a partnership announcement.

This hardware-first structure separates OpenAI's Seoul operation from the company's other Asian offices. The memory supply agreements tie directly to Stargate, the $500 billion joint venture with SoftBank and Oracle that the three firms first announced in January 2025. The Korean government's own AI blueprint, which OpenAI's policy team published in October 2025, explicitly framed Korea's role as an infrastructure and compute hub for the Asia-Pacific. None of this requires a localized Korean-language ChatGPT interface. It requires wafer fab capacity, data center siting, power infrastructure, and cooling engineering, exactly the domains where Samsung and SK Hynix hold dominant global positions. OpenAI's Seoul office is the organizational node that lets the company embed itself in that supply chain rather than buying memory and capacity at arm's length.

When Sovereign AI Meets Export-Control Reality

OpenAI's decision to open a Seoul office landed on the same regulatory fault line that cracked open in June 2026, when the Trump administration forced both OpenAI and Anthropic to restrict their newest models to government-approved customers. The parallel wasn't coincidence. Korea's sovereign-AI buildout and Washington's tightening grip on frontier-model access are converging into a single pressure front, and OpenAI's local presence looks less like a growth play than a survival tactic.

On June 26, OpenAI said it would limit GPT-5.6 Sol (its newest model) to roughly 20 customers pre-approved by the Trump administration, a restriction the company called a "temporary step" it did not want as "the long-term default." Hours later, Anthropic announced a partial reprieve: the government had cleared its Mythos 5 model for limited redeployment to cyber defenders and infrastructure providers, two weeks after the Commerce Department effectively banned foreign-national use of both Mythos 5 and Fable 5. The sequence, reported by the Associated Press and confirmed across multiple outlets, exposed a new reality. Frontier-model releases now run through a White House filter that is voluntary on paper but, in practice, mandatory for any company that wants to stay in the U.S. market.

Trump's executive order earlier that month formalized the framework: up to 30 days of federal vetting for the most advanced AI systems before public release. Participation is technically voluntary. The framework is not yet fully built. That ambiguity is the point. Companies can refuse, but refusal means watching competitors get cleared while your product sits in review limbo.

For OpenAI, the squeeze is compounded by geography. The administration's export-control regime, aimed squarely at China, restricts not just chips but the models that run on them. Anthropic already drew the Pentagon's designation as a national security risk for raising ethical concerns about AI in warfare; Trump ordered federal agencies to stop using Claude. Anthropic is suing. The legal fight is ongoing, but the chilling effect is immediate: any foreign government or contractor that touches a U.S. frontier model now has to account for the possibility that Washington pulls the plug.

South Korea sits directly inside this contradiction. President Lee Jae-myung campaigned on a 100-trillion-won ($735 billion) sovereign-AI investment, one of the largest state-backed AI infrastructure commitments in the world. The Ministry of Science and ICT is deploying over 50,000 NVIDIA GPUs across a National AI Computing Center and domestic cloud providers including NHN Cloud and Kakao. Samsung Electronics alone committed $230 billion through 2030 for AI infrastructure. NVIDIA's own announcements put the total GPU count above 260,000 across Korean sovereign clouds and AI factories.

That buildout requires frontier models. It also requires the chips those models train and inference on, which means Samsung's HBM4 and SK Hynix's advanced memory packages (both already under U.S. export-control scrutiny for their role in Chinese AI supply chains). Korea's sovereign-AI push is, by design, a bet on American hardware and American models. But the Trump administration's simultaneous restrictions on who can use those models and where the underlying chips can go create a compliance maze that no remote sales team can navigate.

OpenAI's Seoul office, then, reads as a localization move with regulatory teeth. To sell into Korea's sovereign-AI infrastructure (the National AI Computing Center, the government cloud, the Samsung and SK Hynix supply chain), OpenAI needs a legal entity that can satisfy both Korean data-sovereignty expectations and U.S. customer-clearance requirements. A Seoul-based team can handle the paperwork, the government relations, and the on-shore deployment reviews that a San Francisco HQ cannot run across time zones during a 30-day vetting window.

The alternative is ceding the market. Domestic Korean AI companies, among them Upstage, Naver, and Kakao, are building Korean-language models trained on domestic infrastructure precisely because the sovereign-AI mandate demands it. If OpenAI can't clear the regulatory path fast enough, those customers won't wait. They'll buy local.

The timing is brutal. OpenAI is exploring an IPO while its product release schedule depends on the same administration that floated the idea of the U.S. government taking an ownership stake in leading AI companies. Sam Altman spent recent weeks in direct negotiations with Commerce Secretary Howard Lutnick over the Sol release. Those conversations aren't abstract policy talks; they're the precondition for a public offering narrative that depends on predictable international revenue.

Korea is where that narrative gets stress-tested. If OpenAI can make sovereign-AI sales work in Seoul, under U.S. export controls, under Trump-cleared customer rules, against domestic competitors with government backing, it has a template for every allied-nation AI buildout that follows. If it can't, the IPO story loses its international growth pillar before the S-1 ever files.

The Hidden Bottleneck: Memory Bandwidth

Over 50% of attention kernel cycles during LLM inference stall waiting for memory access. The GPU just waits for data. High-bandwidth memory reduces this idle time, and two South Korean companies, Samsung Electronics and SK Hynix, dominate the HBM market so thoroughly that any AI company without a Seoul presence builds blind.

HBM solves this by stacking memory chips in layers and placing them directly adjacent to the GPU, feeding data at speeds conventional DRAM cannot touch. HBM4, which Samsung and SK Hynix began mass-producing simultaneously in February 2026, runs at 11.7 Gbps, 46% above the JEDEC baseline. A single HBM4 stack costs roughly $700. A single Nvidia Rubin GPU carries 12 of them. That's $8,400 in memory per GPU, and HBM accounts for 55–60% of total GPU bill-of-materials cost.

From Nvidia's Hopper H100 to Blackwell B200, 80% of the generational uplift came from HBM: capacity jumped from 80 GB to 192 GB, bandwidth from 3.35 TB/s to 8 TB/s. Compute rose 128%. Bandwidth and capacity moved in near lockstep. Nvidia engineers know that scaling compute faster than bandwidth wastes silicon.

Counterpoint Research's data shows SK Hynix held a 57% share of the HBM market in Q3 2025, with Samsung at 22% and Micron at 21%. The gap was driven by SK Hynix's lead in HBM3E qualification for Nvidia's Blackwell platform, where it was effectively the sole-source supplier for the 12-high stack configuration through 2024 and most of 2025. That dominance translated directly into financials: SK Hynix posted a record 47.2 trillion won operating profit in 2025, overtaking Samsung for the first time, driven almost entirely by AI memory.

Samsung is fighting back. After repeated qualification failures on HBM3E through 2024 and early 2025, its 12-high HBM3E finally cleared Nvidia testing in September 2025, and HBM4 volume shipments began in February 2026. Samsung co-CEO Jun Young-hyun told employees in a New Year's address that "Samsung is back," citing positive customer feedback on HBM4 performance. The company is targeting 30% HBM market share in 2026. Analysts at SemiAnalysis expect SK Hynix to maintain its lead in HBM4 but acknowledge Samsung will become more competitive than in prior generations.

The supply picture is tight. All three suppliers' 2026 HBM capacity is already sold out, booked by Nvidia, AMD, and Google through the end of Q1. The HBM shortage is forecast to persist through 2028. The constraint isn't wafer fabrication; it's back-end stacking: TSV throughput, thermocompression bonder availability, and yield on 12- and 16-layer stacks. A 12-layer stack with 99% per-layer yield produces roughly 87% total yield; 16-layer stacks drop further. This is why new entrants like China's CXMT cannot replicate HBM production on any near-term horizon, despite aggressive domestic investment.

For OpenAI specifically, the memory bottleneck is existential. The company's custom inference chip, code-named Jalapeño and built with Broadcom, targets the inference economics that GPU hardware cannot justify at OpenAI's scale. Inference workloads are memory-bandwidth-bound, not compute-bound; every token decode step reads both the model weights and the ever-growing KV cache from HBM at terabytes per second. OpenAI's reported push into agentic, long-context reasoning models makes memory bandwidth even more constraining. Without direct access to the HBM supply chain, OpenAI negotiates bandwidth access through intermediaries.

This is why the Seoul office is a hardware play. Samsung and SK Hynix are not potential customers or localization partners; they are the gatekeepers of the memory bandwidth that determines whether OpenAI's models, custom silicon, and inference infrastructure actually ship on schedule.

Inside the Talent War for Korean AI Engineers

OpenAI's Seoul office is hiring across at least 13 listed roles, from Account Director and B2B Marketing Lead to Forward Deployed Engineer and AI Deployment Engineer, Codex, making it one of the company's largest non-U.S. hiring clusters. The job postings reveal the company isn't just staffing a satellite sales team. The technical bar is high: the AI Deployment Engineer role demands six-plus years of technical consulting experience and hands-on generative AI implementation work, while the Codex-specific posting requires eight-plus years in post-sales engineering or solutions architecture. Both roles are hybrid, three days per week in-office in Seoul, with relocation assistance offered, a signal that OpenAI expects to recruit from outside Korea as well as from the domestic market.

OpenAI's compensation for key technical roles sits at the top of the market:

Against Korean salary standards they are aggressive, and that gap matters because OpenAI is entering a market where the talent pool is already shrinking relative to demand.

AI-related job postings in South Korea more than tripled in two years, according to a joint analysis by The Seoul Economic Daily and recruitment platform Wanted Lab, even as overall hiring demand nationwide fell to historic lows. AI was the single sector reporting hiring difficulty in a down market. The Korea Chamber of Commerce and Industry identified talent outflow, capital-region enrollment caps, and a mismatch between university output and industry needs as the primary structural causes. A report cited by The Korea Economic Daily placed South Korea fifth globally in overall AI competitiveness but just 13th in talent competitiveness. The Ministry of Science and ICT secured 1.9 trillion KRW in supplementary budget for the AI sector in May 2025, explicitly framed as funding to close the talent gap.

The competition for that talent is not limited to Western entrants. SK Hynix brought on 3,201 new employees in 2025, roughly 3.4 times the 942 it hired the previous year, driven by the ramp-up of its M15X fabrication plant in Cheongju. Domestic AI startups and scale-ups like Toss, Reflection, and FuriosaAI are posting forward-deployed and applied AI roles on Korean LinkedIn at a pace that mirrors OpenAI's own list. Google, Databricks, Mistral, Cohere, and Notion all have active Seoul listings for similar forward-deployed AI engineering roles, the same function OpenAI is hiring for, often in the same candidate pool.

The roles OpenAI is listing (Forward Deployed Engineer, Solutions Engineer, AI Success Engineer, Developer Experience Engineer) map onto a specific deployment model. These are not ML researchers training models in San Francisco. They are the people who sit inside Korean enterprises and hyperscaler partnerships (the AWS-specific Partner AI Deployment Engineer posting confirms that channel) and make OpenAI's models work on local infrastructure, with local memory and compute constraints, on local regulatory terms. That function requires engineers who understand both model behavior and Korean corporate procurement cycles, a rare profile that commands a premium precisely because the supply curve is near-vertical.

For anyone tracking AI talent dynamics in the Asia-Pacific, the implication is straightforward: OpenAI's Seoul hiring push is a leading indicator of how fiercely frontier AI companies will compete for deployment engineers in markets where the talent bottleneck is not model capability but the ability to ship it into production. If you have that profile, the market just got wider.

How the IPO Delay Reframes the Seoul Office

OpenAI's reported decision to delay its IPO until 2027, confirmed by a New York Times report in late June, reframes the Seoul office from a market-entry move to a pre-IPO revenue build. Kalshi traders now price a 59% chance of an official IPO announcement by March 1, 2027, and 73% by June 2027. The company confidentially filed its S-1 on June 8 but signaled it had no fixed timeline. Sam Altman has told people at OpenAI that any public valuation below $1 trillion is a nonstarter, according to the Times report. That number, which would place OpenAI among the world's largest tech firms at debut, requires a revenue story that looks nothing like a U.S.-only API company.

The delay is partly a reaction to SpaceX's post-IPO trajectory. SpaceX's stock climbed past $225 after its debut, then fell back to $151. OpenAI's advisors read that as a signal that retail appetite for mega-cap tech IPOs is weaker than expected, and that rushing out in 2026 risks pricing below Altman's threshold. Sam Altman has told people at OpenAI that any public valuation below $1 trillion is a nonstarter. The company's March post-money valuation was $852 billion, OpenAI's own statements indicated. Anthropic, which filed confidential IPO paperwork on June 1, was last valued at $965 billion, a rival breathing down the timeline.

Staying private buys time, and time is exactly what OpenAI needs to build international revenue that public-market analysts will scrutinize. OpenAI is projected to generate $12.7 billion in revenue in 2025, more than triple the prior year's $3.7 billion. International markets are growing faster than the core U.S. business; paid subscriptions outside the U.S. grew 110% over the last year, according to one industry estimate. Licensing to large tech companies outside Microsoft contributed an estimated $950 million in 2025. Those are the numbers that need to compound before an S-1 goes live.

Korea sits at the intersection of both requirements: a sovereign-AI market willing to spend on infrastructure, and a hardware supply chain OpenAI cannot bypass. The Seoul office gives OpenAI a local entity to book enterprise revenue, negotiate data center partnerships around Samsung and SK Hynix HBM, and embed in Korea's government procurement pipeline, all line items that show up in an IPO roadshow as "international diversification" rather than "U.S. concentration risk."

Every quarter between now and a 2027 filing is a quarter OpenAI needs to convert its $852 billion private valuation into auditable, recurring revenue streams that justify a $1 trillion-plus public price. Korea is one of the few markets where AI infrastructure spending, memory supply chain access, and government willingness to pay converge at scale. The Seoul office isn't a flag in the ground. It's a revenue line that needs to exist when the S-1 drops.

What This Means for AI-Hardware and Defense-Tech Operators

OpenAI's Seoul landing is a signal, not a press release. If you work at the intersection of AI inference, edge hardware, or allied-nation defense infrastructure, the move reframes three assumptions that have governed the hiring and investment climate for the past 18 months.

**The chip-software boundary in defense is dissolving faster than expected. ** Korea Aerospace Industries signed a memorandum of understanding with Samsung Electronics in November 2025 to co-develop defense-grade AI semiconductors, with Samsung's foundry division providing integrated capabilities from design through mass production. The goal is domestic AI chips for manned-unmanned teaming systems, AI pilot technology for the KF-21, and communication satellite applications. This is not a prototype phase; Samsung is committing its process ecosystem and production capacity. For hardware engineers, the implication is straightforward: defense-semiconductor roles in Korea are shifting from component procurement to full-stack co-design with a company that has no intention of staying on the merchant-market sidelines.

**Edge AI for defense is a Korea play, not just a U.S. one. ** Samsung's existing edge-AI semiconductor line targets real-time processing in industrial sensors, smart cameras, and autonomous platforms without cloud dependency. KAI's on-device AI chip project, backed by the Ministry of Trade, Industry and Energy, is explicitly designed for inference at the edge in contested environments where communications links are unreliable or denied. The Bank of Korea reported that semiconductor exports drove over 80 percent of export growth from January through September 2025, with HBM and AI server chips leading. If you are building AI inference hardware for defense applications and you do not have a Korea-facing supply chain or engineering presence, you are designing around a bottleneck you do not control.

**Allied-nation AI infrastructure is becoming a Five Eyes procurement priority. ** The U.S., UK, Canada, Australia, and New Zealand agreed in May 2025 to operationalize Project Arcadia, a shared AI and command-and-control data infrastructure designed to build a Common Operating Picture across the alliance. The U.S. Navy's Project Overmatch reached a milestone through a Five Eyes agreement in February 2025. South Korea is not a Five Eyes member, but its defense AI posture (Defense Innovation 4.0, the Defense AI Center with 110 civilian and military personnel, and the Army Tiger Demonstration Brigade) is designed for combined operations with U.S. forces. The U.S. Space Forces-Korea, established as the first forward operating combined joint space command center in 2022, gives Korea a role in allied space-domain awareness that no other Asian partner matches. For engineers working on multi-national AI defense systems, this creates a bifurcated hiring market: roles that require U.S. security clearances and roles that require Korea-literate hardware-software integration skills. The overlap between those two pools is thin and getting thinner.

**The talent timeline is compressing. ** South Korea's government committed roughly $34.1 billion in a fund for AI, semiconductors, batteries, and robotics in 2025, with $79.1 million directed into 132 projects in the first year (including micro-suicide drone systems and amphibious reconnaissance robots). Defense robotics companies like Hanwha, LIG Nex1 (which acquired U.S.-based Ghost Robotics in 2023), and Hyundai Rotem are delivering operational systems, not slide decks. Hyundai Motor Group secured 50,000 Nvidia GPUs during the APEC summit in Gyeongju, a procurement that signals the conglomerate's defense AI compute capacity is scaling in parallel with its civilian autonomous driving programs.

For operators watching where AI-hardware convergence meets allied defense infrastructure, the Korea node is no longer optional. OpenAI understood that. Samsung and KAI understand that. The remaining question is whether your organization's workforce plan accounts for a market where the AI memory supply chain, the edge inference chip design pipeline, and the allied interoperability standards are all being built in the same Seoul corridor, by the same companies, hiring from the same engineering pool, under the same sovereign-AI mandate.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at OpenAI, and the people building the field.