OpenAI cut API prices again. Its $10M consulting engagements tell the real story.

Why Seoul Is OpenAI's Asia-Pacific Ground Zero

OpenAI opened its Seoul office on September 10, 2025, making South Korea its third Asian location after Tokyo and Singapore and its 12th worldwide. The reason was straightforward: Korea is OpenAI's largest market in Asia-Pacific and its second-largest paid subscription market globally, behind only the US.

Weekly ChatGPT users there quadrupled over the past year. Paid subscriptions more than tripled in the same period. Jason Kwon, OpenAI's chief strategy officer, called Korea "a combination of infrastructure, innovative companies and early adopters" at the Seoul press briefing.

But the Seoul office isn't a sales outpost. OpenAI is hiring forward-deployed engineers locally, a signal that the company's Korea strategy runs deeper than subscriber growth. The office sits on top of a partnership stack that includes a February 2025 strategic alliance with Kakao to integrate ChatGPT across its ecosystem, an MOU with Seoul National University signed September 11, and active discussions with Samsung Electronics and SK Group on data center infrastructure and hardware chip collaboration. Kwon met with then-presidential candidate Lee Jae Myung in May to discuss accelerating Korea's AI industry.

An IMF report on AI's impact in Korea found that roughly 50% of Korean jobs are exposed to AI, with the ICT sector leading adoption at 18% in 2022. Larger firms and younger firms adopt faster. The government has committed major funding to AI R&D, and the country's 5G infrastructure and culture of rapid tech adoption create conditions where enterprise AI deployments can move quickly from pilot to production.

OpenAI's Singapore office remains its regional hub, but Seoul is where the subscriber base and enterprise demand concentrate. For frontier AI deployment talent, the message is clear: the work is moving to where the customers are, and right now that's Korea.

What OpenAI Is Actually Hiring For

OpenAI's job listing for Forward Deployed Engineer in Seoul reads less like a typical technical hire and more like a bet that the hardest part of enterprise AI is no longer building models. It's getting them to work inside a real company.

The role asks engineers to "lead complex end-to-end deployments of frontier models in production alongside our most strategic customers." That means owning everything from initial discovery through technical scoping, system design, build, and production rollout. FDEs embed directly with customer engineering and domain teams, write production-grade code in Python or JavaScript, and travel 50% of the time. The Seoul listing requires five or more years of engineering or technical deployment experience, including customer-facing work, and familiarity with deploying systems powered by large language models.

This is not a sales engineer. Sales engineers demo, present, and hand off to implementation teams before moving to the next deal. The FDE role collapses that handoff entirely: the same person who scopes the engagement also builds the system, ships it to production, and stays accountable for whether the customer's workflows actually change. OpenAI measures these engineers on production adoption, measurable workflow impact, and eval-driven feedback that changes product and model roadmaps.

It is also not an ML researcher. The posting asks for full-stack engineering skills (frontend and backend code, not paper submissions). Understanding how model behavior affects product experience matters here, but the output is a deployed system, not a benchmark result.

A separate role OpenAI is hiring in Seoul, the AI Deployment Engineer, sits closer to the traditional solutions architecture function. That position, listed under a "Technical Success" team, focuses on building use-case backlogs with senior customer stakeholders, prototyping applications on the OpenAI API, and codifying best practices into internal and external repositories. It requires six or more years of technical consulting experience and reports to the Head of Technical Success, APAC. The distinction matters: the AI Deployment Engineer is a strategic advisor who scales through knowledge-sharing, while the FDE is an owner who scales through direct delivery.

OpenAI has posted identical FDE listings in San Francisco, New York City, Stockholm, and Singapore, suggesting the role is not a Seoul experiment but a global staffing model. The Seoul posting does not list compensation.

The FDE model borrows from a playbook that Palantir pioneered in defense and enterprise software: embedding engineers inside customer organizations to solve specific problems, then extracting reusable patterns. OpenAI's version applies that approach to frontier AI: instead of shipping a general-purpose product and hoping enterprises figure it out, FDEs start with a customer's specific problem, validate impact, and identify what scales.

For engineers considering the role, the trade-off is clear. You get direct exposure to how the largest companies in Asia-Pacific are trying to adopt generative AI, plus a feedback channel into OpenAI's own research and product teams. You also get 50% travel, ambiguous scoping, and the pressure of owning production outcomes for customers who may not fully understand what the models can and cannot do yet.

The $10 Million+ Engagement Model

Forbes reported in July 2025 that OpenAI had launched a custom AI consulting service with a floor of $10 million per engagement, targeting enterprise-grade GPT-4o deployments staffed with forward-deployed engineers embedded directly in client workflows.

That number reframes what "selling AI" actually means at the frontier. OpenAI's standard API pricing tells a different story: as of mid-2026, GPT-5 runs $1.25 per million input tokens and $10.00 per million output tokens, while GPT-4o mini costs $0.15 and $0.60 respectively. At enterprise scale (say, 100,000 requests per day), APIpulse's scenario modeling puts GPT-5 costs at roughly $75,000 per month through the API alone. Even at high volume, you're six figures short of $10 million.

The gap between API spend and consulting spend is the point. OpenAI isn't selling tokens at that tier. It's selling integration: forward-deployed engineers who sit inside a client's organization, restructure workflows around model outputs, handle governance and compliance, and take responsibility for business outcomes, not just uptime. The price tag covers the personnel, the customization, and the accountability wrapper that makes a CFO sign off on generative AI in a regulated environment.

The model is borrowed directly from Palantir. Forbes noted the comparison explicitly: Palantir built its business over two decades on embedding consultants and engineers into client operations, achieving over 40% free-cash-flow margins by tying revenue to mission-critical outcomes rather than software licenses. OpenAI is now replicating that structure. Henri Terho, a partner at the consulting firm Tomoro (which OpenAI later acquired to staff its deployment subsidiary), put it bluntly on LinkedIn: "Models and AI is not the problem, it's the deployment, integration, user access… consulting margins beat API margins."

The margin math supports the shift. Forbes reported that deep-pocket consulting engagements in this category can yield margins of 40%–60%, compared to SaaS-style API margins that compress as models commoditize. OpenAI's own API pricing shows the pressure: GPT-4o mini handles high-volume tasks at $0.15 per million input tokens, a fraction of what GPT-5 charges. The value is migrating up the stack, from the model to the implementation layer.

This acquisition of Tomoro, which brought roughly 150 experienced forward-deployed engineers and deployment specialists into OpenAI's new DeployCo subsidiary, signals the scale of the bet. AIToolsRecap reported that OpenAI launched DeployCo in May 2026 as a $4 billion majority-owned subsidiary.

For the engineers this model demands, the implications are concrete. The role is not research. It is not prompt engineering. It is full-stack deployment work inside client organizations: integration with legacy systems, change management, governance compliance, and measurable ROI delivery, priced at a level that only enterprise clients can sustain.

The $10 million floor is not a price point. It is a hiring specification.

Enterprise AI's Bottleneck Has Shifted From Training to Deployment

The models work. That's no longer the problem.

Enterprise AI adoption accelerated through 2024. McKinsey's survey that year found 67% of organizations using generative AI. But by mid-2025, the trend reversed for large firms. Adoption rates for companies with 250 or more employees began declining, even as smaller firms kept climbing. The reason wasn't a loss of faith in AI. It was the collision between what models can do and what enterprises can actually operationalize.

ISG's 2025 study of 1,200 enterprise AI use cases, representing $2.6 billion in spending across organizations with 1,000 or more employees, found that 31% reached full production, double the prior year. That sounds like progress until you look at the outcomes. Only one in four initiatives met revenue impact expectations. Broad cost savings remained elusive. The use cases delivering the most consistent results were compliance, risk management, and quality control, areas where AI amplifies established processes rather than reengineering them.

The pattern repeats across every major data set. OpenAI's own December 2025 report, drawn from de-identified usage across more than 1 million business customers, found that API reasoning token consumption per organization jumped 320-fold year-over-year. Usage intensity is surging. But the same report noted a widening gap: frontier firms send twice as many messages per seat as the median enterprise, and frontier workers generate six times more messages than the median worker. The technology is being consumed. It's just not being embedded evenly or deeply.

The bottleneck is no longer training better models. It's the unglamorous work of making models function inside real organizations: data integration, workflow redesign, governance, monitoring, cost management, and the human expertise required to stitch all of it together.

ISG's research lays out why in blunt terms. Fewer than 20% of organizations report high data readiness. Over 80% lack mature AI infrastructure (the monitoring, auditability, and control systems needed to run agentic systems safely). Nearly two-thirds of enterprises cite talent as the biggest barrier to production; 57% cite systems. These aren't model problems. They're deployment problems.

IBM's research found that computing costs jumped 89% between 2023 and 2025, with 70% of executives citing generative AI as the primary driver. Every executive IBM surveyed had canceled or postponed at least one AI initiative due to cost concerns. Forrester projects 25% of planned enterprise AI spending will be delayed until 2027, not because companies don't want AI, but because they can't prove ROI amid infrastructure costs and deployment delays.

The result is what the industry calls "pilot purgatory." Sixty percent of enterprises evaluate enterprise-grade or agentic AI systems. Twenty percent reach pilot. Only 5% reach full production with material business impact. The gap between experimentation and operational scale is where enterprise AI goes to stall.

This is precisely why OpenAI, Anthropic, and others are hiring forward-deployed engineers, a role pioneered at Palantir and now spreading across frontier AI. These engineers don't train models. They embed with customer teams, translate business logic into AI-executable processes, and build production systems that function in messy enterprise environments. The Wall Street Journal reported in October 2025 that AI startups now treat forward-deployed engineers as a "secret weapon." LinkedIn data shows job postings for the role jumped 800% from January through September 2025. Mid-level positions start around $300,000 in total compensation; senior roles at frontier labs clear $500,000 or more, and companies still can't fill them.

The infrastructure layer matters as much as the human one. Anthropic's Model Context Protocol, introduced in November 2024 and adopted by OpenAI, Google DeepMind, and others, standardizes how AI systems connect to external data sources and tools. Before MCP, every enterprise implementation required custom integrations for each data source, an N×M problem. MCP collapses that to a single implementation. It's plumbing, but plumbing determines what you can build. Cognizant's November 2025 deployment of Anthropic's Claude to 350,000 employees explicitly incorporates MCP to give agents governed access to enterprise systems at scale.

The companies that successfully scale AI aren't the ones with the best models. They're the ones with three things: platforms that orchestrate agents with governance and observability, protocols that standardize tool and data access, and embedded implementation teams that turn capability into outcomes. Everyone else is still piloting.

For engineers watching this shift, the signal is clear. The scarce talent in frontier AI is no longer the person who can train a model. It's the person who can make one work in production, inside a real company, with real data, under real constraints. That's why OpenAI's Seoul office is staffed with forward-deployed engineers, not researchers. The bottleneck moved. The hiring followed.

Who Else Is Hiring FDE-Style Roles Across Asia

Anthropic is OpenAI's most direct rival in the Seoul deployment talent race, and it's moving fast. The company opened its Seoul office on June 17, 2026, its third Asia-Pacific hub after Tokyo and Bengaluru, with six named enterprise deals already signed: Samsung SDS, NAVER, LG CNS, Nexon, Hanwha Solutions, and Channel Corp. It also hired KiYoung Choi, former Snowflake Korea GM, as Representative Director of Anthropic Korea, and signed an MOU with South Korea's Ministry of Science and ICT.

The launch came with a complication that reveals how geopolitical risk now shapes deployment hiring. US export controls suspended Anthropic's Fable 5 and Mythos 5 models on June 12, five days before the Seoul opening. All six enterprise deployments are proceeding without Anthropic's most advanced models. That means the engineers Anthropic is hiring in Seoul are deploying a constrained product lineup into production environments at Samsung, NAVER, and LG, a bet that either the existing Claude family is enough or that the export controls resolve before deployment commitments deepen.

Anthropic's scale ambitions are explicit. The company said in September 2025 that it would triple its international workforce and expand its applied AI team fivefold in 2025. Claude now has more than 300,000 enterprise customers, with nearly 80% of usage coming from outside the US. Anthropic's Economic Index shows that on a per-person basis, adoption in South Korea, Australia, and Singapore has already surpassed US levels.

The hiring overlap with OpenAI is direct. Both companies are recruiting country-level leads and applied AI engineers across the same markets: Korea, Singapore, India, Australia. Both are building teams that sit between research and sales, embedded inside enterprises to productionize frontier models. The difference is in go-to-market philosophy. OpenAI's FDE model ties engineers to million-dollar-plus engagements with a consulting-style delivery structure. Anthropic's applied AI team, growing fivefold, is organized more around vertical specialization (telecom, pharmaceuticals, financial services, government), with Chief Commercial Officer Paul Smith emphasizing domain-specific systems over generic integration support. Anthropic is also investing in 24/7 support and data sovereignty infrastructure, which matters for regulated-sector buyers in Asia who can't send data to US-hosted endpoints.

Google DeepMind is a third competitor, though its Asia footprint skews more toward research than deployment. The company opened a Singapore lab focused on frontier AI development for the region, with Google Cloud pushing Gemini adoption across public agencies and enterprises in Singapore, starting with health and life sciences. ATxSG, Singapore's national AI initiative, includes both Google and NVIDIA as partners. But Google's Asia hiring for forward-deployed or customer-engineering roles at the scale OpenAI and Anthropic are pursuing is less visible in the current data. Its advantage is the existing Cloud install base: CIOs already on GCP can add Gemini without a new vendor relationship, which lowers the deployment friction that FDE-style roles exist to solve.

For engineers choosing between these employers, the distinction matters less than the trend. Three frontier AI companies are now competing for the same scarce talent class across the same Asia-Pacific metros (Seoul, Singapore, Tokyo, Bengaluru), and all three are hiring for roles that didn't exist eighteen months ago. The engineers who can take a model from API call to production system inside a regulated enterprise are commanding premium compensation at every one of them. Seoul is where the competition is most visible, but the talent market this is creating stretches across the entire region.

What the Seoul Signal Means for AI Engineering Career Strategy

OpenAI's Seoul office does something most career advice can't: it puts a real location, a real team, and a real salary band behind the argument that deployment engineering is where the market is moving. For AI engineers weighing a research track against a production track, the signal is hard to ignore.

The pay gap is concrete. FDE Pulse salary data shows that Forward Deployed Engineers at OpenAI earn between $162,000 and $325,000 in base salary, with equity pushing senior total comp past $500,000. In Seoul, the same role pays an average of ₩106,897,467 (roughly $80,000 at current exchange rates) with senior engineers reaching ₩122,961,660, per SalaryExpert. That dollar figure looks low next to San Francisco, but it sits well above the ₩75,315,659 entry-level average for the same role in South Korea, and far above what Korean tech companies typically offer for comparable AI work. The geographic arbitrage works both ways: OpenAI gets US-calibrated talent at a discount; the engineer gets a top-tier Korean salary with relocation support and a path to global-scale systems.

The skill premium is narrower than people think but real. FDE Pulse reports that AI-specific FDE roles, the ones requiring LLM integration, RAG architecture, and model evaluation experience, pay roughly 10% to 20% above non-AI FDE work at the same seniority. That premium exists because the candidate pool is small. Most ML engineers train models. Few can wire one into a customer's production stack, run evals, handle hallucination edge cases, and explain the trade-offs to a non-technical stakeholder in the room. If you can do both, you are rare, and the market prices you accordingly.

The career calculus breaks down into a few clear questions.

Do you want to build models or make them work in production? Research roles at OpenAI, Anthropic, and similar labs remain competitive, but the hiring bar is publication-heavy and the feedback loop is long. Deployment roles ship in weeks. OpenAI's own job posting for the Seoul FDE position lists "own technical delivery across multiple deployments from first prototype to stable production" as a core responsibility, with 50% travel expected. You see the impact of your work directly, and that feedback compounds faster than a paper cycle.

Are you willing to talk to customers? This is the filter most engineers underestimate. The FDE job posting explicitly asks for engineers who "communicate clearly with engineers, product teams, and customer stakeholders" and can "scope work, sequence delivery, and remove blockers early." Palantir built its entire early engineering culture around this model. OpenAI, Salesforce, and Databricks are now scaling it. If customer-facing work drains you, this track will burn you out regardless of the salary.

Where do you want to live and what's your risk tolerance? Seoul is not a default destination for most Western-trained AI engineers. That's the point. The talent pool is thinner, the competition for roles is less saturated, and OpenAI is offering relocation assistance. Engineers who move early into a new office often get outsized ownership over team direction and customer relationships. The trade-off is being far from the core research org in San Francisco, with all the career gravity that implies.

The broader trend is clear from the board data. FDE Pulse tracks 260+ open FDE positions across 50+ companies, with postings growing roughly 800% over the past year. Salesforce has stated a target of nearly 1,000 FDEs for Agentforce deployments. This is not a niche. It is a hiring wave.

For engineers deciding now: learn to deploy models in production, get comfortable in front of a customer, and watch where the offices open, not just where the research papers come from. The Seoul posting is a case study. The next one will be in a different city, and the engineers who moved first will already be running the team.

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