Aisummitseoul.com
Asia’s largest applied-AI gathering kicked off Wednesday morning in Seoul with a program that functions less as a technology showcase than as a collective diagnostic: 69 expert sessions, 81 speakers, and a three-day agenda whose dominant subjects — observability, inference architecture, multi-agent failure modes, and context retrieval — map almost precisely to the reasons why enterprise AI agent deployments keep failing before they reach production. The full AI Summit Seoul agenda spans two full conference days and an EXPO running through August 21. AI Summit Seoul & Expo 2026 (AISE 2026), the ninth edition of the private-sector conference organized by DMK Global, KITA, and COEX, opened at the COEX Grand Ballroom and Exhibition Hall B in Gangnam with 118 companies from seven countries across 307 booths — a 66.2% increase in participating firms over last year’s 71 companies and 197 booths, with visitor registration up fivefold over 2025 and more than 25,000 attendees expected over the three-day run.
The scale matters less than what the scale signals. South Korea’s flagship enterprise AI conference has become Asia’s highest-concentration gathering of practitioners who have actually deployed AI in production environments — not researchers studying deployment and not executives approving pilots, but the engineers and operations leaders from companies like DHL, POSCO E&C, Toss Bank, Kakao Pay Securities, Lotte Shopping, HD Hyundai, and KB Kookmin Bank who have spent the past two years discovering, at considerable cost, that building an AI agent and running an AI agent reliably at enterprise scale are fundamentally different problems. IDC research has documented the gap: 88% of agent pilots never reach broad production — a failure rate that traces almost entirely to operational and governance gaps rather than any limitation in the underlying AI models.
The event’s organizers framed this explicitly. AISE 2026 is the first major private-sector AI business event in Korea since the government’s San Francisco AI Declaration of July 24, 2026, in which President Lee Jae Myung pledged before the CEOs of NVIDIA, OpenAI, Anthropic, and Broadcom to make Korea an “irreplaceable” node in the global AI supply chain. “To leap forward as a core country in the global AI supply chain,” KITA’s Jung Hee-cheol said at the opening, “we need not only the development of excellent technology but also a market where technology companies and industrial demand-side companies actually meet, collaborate and create business.” AISE is, in that reading, the commercial translation layer between Korea’s official AI supply chain pledge and tangible business outcomes.
Why 88% of Agent Pilots Never Leave the Lab
The dominant technical theme of Day 1 — agentic AI and its operational limits — was not chosen as a marketing conceit. It reflects an empirically documented failure pattern. Research across enterprise AI deployments in 2026 has established that the hardest part of deploying agentic workflows is not model intelligence but secure, reliable access to production systems: the CRMs, ticketing tools, internal APIs, and data platforms that agents need to interact with, and the audit infrastructure needed to verify they did so correctly. The consensus from enterprise AI infrastructure analysts is that inference cost and operational complexity, not model capability, are now the primary deployment bottlenecks.
The morning opened with Larry Heck, Professor of Electrical and Computer Engineering at Georgia Tech and the co-founder of Cortana, who built Bixby, Cortana, and Deep Dialogue at Microsoft, Samsung, and Google respectively, and who drew the conceptual arc from the narrow voice interfaces of the early 2010s to the current generation of autonomous, goal-directed agents. Heck argued that the shift from AI assistants to AI agents is not primarily a capability jump — it is a reliability engineering challenge. An assistant that gives a wrong answer is annoying; an agent that takes a wrong action in a production system has consequences that may be impossible to reverse.
The session that the conference billed as one of its most anticipated — Kay Zhu’s Day 1 address “The Future of AI Agents: Not Better Models, But Better Workflows” — made the same point from the commercial side. Zhu, CEO of Genspark, the AI research and productivity platform, has argued publicly that the competitive leverage in enterprise agentic AI now lies not in which model a company selects but in how it structures the orchestration, memory, and retrieval architecture around that model. Two teams deploying the same foundation model with different harness engineering, feedback loops, and context retrieval layers will get radically different production outcomes.
Alex Triplett of You.com addressed what he called “The Blind Spot of Agentic AI” — the observability gap. Without the ability to audit what an agent did at runtime, in what sequence, with what confidence, and with access to what data, neither governance nor systematic debugging is possible. An agent that appears to work in a demo environment and fails silently in production is worse than an agent that fails visibly, because silent failure produces confident wrong outputs that downstream systems act on.
Michael Albada of NVIDIA carried the technical argument further in his afternoon Track B session, “From AI Agent to AI Engineer: Building Multi-Agent Systems Ready for Real Deployment.” The specific engineering problems that distinguish multi-agent prototypes from production-grade systems include task decomposition strategy, inter-agent communication protocols, failure mode handling (what happens when one agent in a pipeline crashes mid-task?), and tooling choices that determine whether a live multi-agent system can be debugged at all. These are not research problems — they are infrastructure problems, and they are currently the primary reason why enterprise AI transformation projects stall between pilot and scale.
Colin Presly, VP of Customer Engineering at Seagate, added the storage layer to the diagnostic: data architecture, retrieval latency, and pipeline reliability are the constraints that enterprise AI projects most consistently underestimate. “AI-ready” infrastructure is not a marketing term — it means NVMe-speed access to training data and inference-time retrieval stores, structured to minimize the latency spikes that cascade through agent reasoning chains.
Disaggregated Inference: Why AI Inference Is Being Split in Half
One of the more technically specific sessions of the conference comes on Day 2, when Reggie Lu of SambaNova will address an architectural development that is reshaping how enterprise AI inference is delivered at scale.
Traditional LLM inference runs two fundamentally different computational phases on the same hardware. The prefill phase — processing the full input prompt in parallel to generate a key-value (KV) cache — is compute-bound and FLOPS-limited. The decode phase — generating one token at a time using the stored KV cache — is memory-bandwidth-bound and HBM-limited. Running both on the same GPU forces the hardware to do neither optimally, and creates a specific failure mode: a long prefill stalls every concurrent decode request behind it, spiking inter-token latency in ways that are invisible in a demo but catastrophic at scale.
Disaggregated inference solves this by separating the two phases onto specialized, independent hardware pools connected by high-speed networking (InfiniBand or NVLink). Prefill nodes run on high-FLOPS chips (NVIDIA H100 SXM5, B200) while decode nodes run on high-HBM-bandwidth chips (H200 SXM5 with 141GB HBM3e). Each scales independently, eliminating interference and enabling operators to tune time-to-first-token and inter-token latency as separate variables. Independent benchmarks have measured cost cuts of 15–40% and throughput gains of 2–7x in optimized configurations. Jensen Huang validated the architectural direction at GTC 2026; NVIDIA’s deal with Groq was explicitly to acquire LPU-based decode hardware for disaggregated deployments.
The tradeoff is real: disaggregation requires high-speed inter-node networking for KV cache transfer, adds operational complexity, and can hurt performance by 20–30% at small scale when workload sizes are too small to justify the separation. It is a production-scale architecture, not a development-environment solution — which is precisely why it belongs at a conference whose theme is the gap between demos and deployment.
SambaNova’s approach pairs GPUs for prefill with its Reconfigurable Dataflow Unit (RDU) for decode, independently addressing the different hardware profiles of each phase. Lu’s session is framed as “Premium AI Inference: From Ultra-High-Speed Tokens to Next-Generation Infrastructure” — language that signals the audience: inference cost and latency have become the primary bottleneck in production AI as training costs fall and deployment volumes grow.
NASA’s Deep-Space AI and What It Teaches Enterprise Engineers
The Day 1 keynote that drew the sharpest engineering interest came from Steve Chien, Senior Research Scientist and JPL Fellow at NASA’s Jet Propulsion Laboratory. Chien leads JPL’s autonomous AI research, where AI must operate with minimal or zero human oversight due to light-speed communication delays between Earth and spacecraft — delays that make real-time human supervision physically impossible.
His session, “Autonomous AI Beyond Earth: How Satellites Think and Collaborate Independently,” described real deployed systems — including autonomous satellite constellations that coordinate science observations without waiting for ground instructions — and drew out the engineering and governance principles that transfer directly to terrestrial enterprise AI. The constraints that JPL’s AI must satisfy are extreme versions of exactly the constraints that make enterprise AI hard: low tolerance for silent failure, mandatory explainability (a spacecraft that does something unexpected must produce an audit trail a human engineer can follow), and the requirement that autonomous decisions degrade gracefully when the operating environment exceeds the system’s training distribution.
The central argument Chien advanced was structural: the constraints of deep-space AI have been a forcing function for solving the reliability and explainability problems that enterprise AI is only beginning to confront. JPL has been running production AI on spacecraft since 2003, when the Earth Observing-1 autonomous science mission began autonomously identifying and targeting significant events — volcanic eruptions, flooding, wildfires — before scientists were even aware they had happened. The accumulated engineering discipline from two decades of deployed autonomous systems is not a space-specific asset; it is a practical curriculum in how to build AI that cannot be supervised in real time.
Korea as the Context: From Silicon to Supply Chain
AISE 2026 arrives at a specific inflection point for Korean AI. The San Francisco AI Declaration committed Korea to positioning itself as a global AI testbed and production platform. The private sector moved fast to demonstrate commercial readiness: in the weeks surrounding the declaration, NVIDIA CEO Jensen Huang visited Seoul for meetings with NAVER, whose GAK Sejong data center project is scaling toward gigawatt AI infrastructure capacity; SK Hynix announced a multiyear partnership with NVIDIA on next-generation AI memory; and LG Electronics and NVIDIA separately confirmed a 12-year collaboration on robot training data. South Korea’s AI startup ecosystem has grown from roughly 200 scaleups a decade ago to more than 2,100 today, concentrated in the Gangnam and Pangyo AI corridors that surround COEX.
AISE itself has tracked this growth precisely. In 2018, the event’s first year, it drew a handful of domestic exhibitors. Nine years later, it has become a gateway event for international AI companies seeking distribution and partnership in one of Asia’s most AI-intensive enterprise markets — a market that includes, by AISE 2026’s participant count, companies from seven countries across 307 booths.
The Korea-specific context surfaced directly in Day 2 sessions (scheduled to begin Thursday morning in Seoul), which will take up enterprise AI transformation, sovereign AI, and AI infrastructure — themes that map to the structural questions Korea’s AI policy is actively trying to answer. The Day 2 sovereign AI infrastructure track will address the technical and policy conditions for national AI development independent of foreign model dependencies.
Agents That Appear to Work, and Agents That Actually Do
A Day 2 session — “AI Agents: The Gap Between Appearing to Work and Actually Working — Lessons from Robotics and Chip Design” — may be the most precisely titled session in AISE’s nine-year history. The panel brings together Pannag Sanketi (Avolla), Michael Albada (NVIDIA), and Prerit Mishra (DHL) to examine the specific failure modes that distinguish AI demos from production systems. Mishra has articulated the distinction publicly: AI agents that execute tasks are automation with a better interface; AI agents that make consequential decisions require fundamentally different data infrastructure, governance, and operating models. The gap between these two categories is where the production failure rate lives.
The fintech sessions on Day 2 are among the most practically specific of the conference. Jeong I-tae (Axina), Jeong Jae-young (Toss Bank), and Lee Hyeong-ju (Kakao Pay Securities) will examine how enterprises should design the knowledge graph and context retrieval layers that agents actually query at runtime — the data architecture that determines whether an agent’s reasoning is grounded in accurate, current, specific organizational knowledge or in generic model weights that have no idea what a particular company’s actual inventory, policy, or customer record contains. South Korea’s financial regulators are among the most cautious in Asia regarding AI in client-facing applications. The fact that both Toss Bank and Kakao Pay Securities are presenting production deployments signals that the regulatory framework has matured enough to permit meaningful rollouts — a development that other Asian jurisdictions are watching closely.
The B2B infrastructure beneath the conference also reflects the commercial maturity of the Korean enterprise AI market. “AI Business Connecting” will facilitate approximately 200 one-on-one consultations between roughly 80 participating AI technology companies and demand-side enterprises and investors. An “AI Startup Forum & Pitching” session on Thursday gives early-stage companies direct access to investors and potential partners. Practitioner-focused “AI Workshop” sessions cover workflow automation and AI-assisted marketing for teams with limited prior exposure to production AI tooling.
Physical AI and Humanoid Robots Take the Hall B Floor
The EXPO running through August 21 in Hall B introduces AISE’s first dedicated humanoid robot demonstration zone, where visitors can observe physical AI technology from AgiBot and Pudu Robotics alongside more than 300 booths across Hall B of AI software, cloud infrastructure, and vertical-market solutions. The Day 2 morning panel on physical AI — featuring LIG, FactoryLab, Seoul National University’s Kim A-young, and Handong University’s Jeong Doo-hee — addresses the convergence of robotics, computer vision, and world models that Korean heavy industry is beginning to adopt at scale. LG Electronics’ 12-year robot training data collaboration with NVIDIA provides the background: physical AI for manufacturing automation is no longer a research horizon but a capital-allocation decision.
A notable feature of this year’s EXPO is a dedicated “China Tech” segment on August 20 Thursday, reflecting the growing presence of Chinese AI companies — including AgiBot, a Shanghai-based humanoid robot developer, and BytePlus, ByteDance’s enterprise AI subsidiary — in the Korean enterprise market. That presence is worth noting in context: Chinese AI companies operating in Korea are subject to China’s National Intelligence Law, Article 7 (2017), which requires all organizations to support, assist, and cooperate with national intelligence work on demand, alongside the Data Security Law (2021) and Cybersecurity Law (2017) and their data localization provisions. These legal obligations apply regardless of where data is physically stored or how a company characterizes its privacy practices. Enterprise technology leaders evaluating deployments from Chinese exhibitors at AISE 2026 should factor these fixed legal conditions into procurement decisions alongside performance and cost comparisons.
Hoifung Poon and the Case for Virtual Patients
Among the Day 2 morning sessions, Hoifung Poon of Microsoft Research’s biomedical AI program will present on “The Era of AI-Powered Scientific Discovery: Virtual Patients and the Future of AI Medicine.” Poon leads a research program on biomedical AI that includes AI systems capable of synthesizing large bodies of scientific literature, extracting structured biological knowledge, and modeling patient populations at scale. His introduction of “virtual patients” — AI-generated patient cohorts that can compress clinical trial timelines — is one of the more concrete examples of agentic AI in which the operational stakes make the reliability gap immediately legible: an AI agent making errors in clinical trial simulation is not an annoyance; it is a regulatory and patient safety problem. The same observability and governance requirements that Triplett and Albada described in enterprise software apply here in a domain where the consequences of silent failure are categorically different.
This is precisely the kind of cross-domain transfer of insight that AISE’s session architecture is designed to produce: the engineering discipline developed for spacecraft autonomy at JPL, applied to the observability requirements for enterprise software agents, applied again to the governance standards for AI in medical research. The production gap is the same problem at every level of stakes; only the cost of failure differs.
Practical Information
AISE 2026 runs through August 21. The conference (COEX Grand Ballroom, Level 3) is scheduled Wednesday through Thursday, with Day 2 sessions set to open at 9:00 a.m. Korean Standard Time (8:00 p.m. ET Wednesday). The EXPO in Hall B (Level 1) runs through Friday, with public hours of 10:00 a.m. to 6:00 p.m. on Thursday and 10:00 a.m. to 4:00 p.m. on Friday. COEX is directly connected to Samseong Station (Seoul Metro Line 2, exits 5 and 6). Full program and exhibitor directory: AI Summit Seoul full program and aisummit.co.kr.
Frequently Asked QuestionsWhat is the production gap in enterprise agentic AI, and why does it matter?
The production gap refers to the documented chasm between AI agent pilots — which typically work in controlled demo environments — and agents that perform reliably in live enterprise systems. IDC research found that 88% of enterprise AI agent proofs-of-concept never reach production deployment. The failure mode is almost never the underlying AI model; it is the surrounding infrastructure: observability tooling that lets engineers audit what an agent did and why, governance frameworks that define what an agent is authorized to do, context retrieval layers that give agents access to accurate organizational data at inference time, and failure-handling mechanisms that prevent one agent crash from propagating through a multi-agent pipeline. AISE 2026’s session architecture — dominated by agentic AI, observability, inference infrastructure, and data architecture — is a direct map of where these failures are concentrated, confirmed by Gartner agentic AI failure research and independent enterprise deployment research.
What is disaggregated inference, and why is it significant for enterprise AI?
Disaggregated inference is a serving architecture that separates the two phases of large language model inference — prefill (compute-intensive: processing the full input prompt to create a key-value cache) and decode (memory-bandwidth-intensive: generating tokens one at a time) — onto different, specialized hardware pools. Traditional inference runs both phases on the same GPU, creating interference: a long prefill stalls all concurrent decode requests, spiking latency in ways that are invisible in small-scale testing but become critical at production volumes. Disaggregation eliminates the interference, enables each phase to scale independently, and has produced throughput gains of up to 7x and infrastructure cost reductions of 15–40% in benchmarked configurations. The architecture is now supported by NVIDIA Dynamo, vLLM, SGLang, and SambaNova’s RDU platform, and is increasingly the default choice for enterprises serving AI agents at scale.
Why is Korea hosting Asia’s largest enterprise AI conference, and what does the San Francisco AI Declaration mean for technology buyers?
Korea’s position is structural. The country supplies most of the world’s HBM (high-bandwidth memory) — the chip architecture that enables large-scale AI inference — through Samsung and SK Hynix, giving it an indispensable role in the global AI hardware supply chain even before any software or services ambition is considered. The Korean government’s San Francisco Declaration of July 24, 2026, committed the Korean government and its largest technology companies (Samsung, SK, Hyundai, NAVER) to extending that hardware position into AI data centers, physical AI, and enterprise AI services — with $950 billion in announced joint investment frameworks with NVIDIA, OpenAI, Anthropic, and Broadcom. For international technology buyers, Korea’s emergence as an AI testbed means an accelerating availability of production-grade enterprise AI deployments to study, partner with, or adopt — a density of real-world case studies that AISE 2026’s 73 corporate practitioner speakers represent directly.
What should enterprise technology leaders know before evaluating AI products from Chinese companies at the EXPO?
Chinese AI companies operating internationally — including those exhibiting at AISE 2026 — are subject to China’s National Intelligence Law, Article 7 (2017), which requires all organizations and citizens to support, assist, and cooperate with national intelligence work on demand. The Data Security Law (2021) and Cybersecurity Law (2017) additionally impose data localization requirements and government access provisions. These legal obligations are fixed conditions of operating under Chinese jurisdiction and apply regardless of where a company’s data centers are physically located, how its privacy policy is written, or whether it has a Western incorporation address. Enterprise technology leaders should treat these obligations as a baseline condition — not a risk to be weighed against price, but a structural legal fact to factor into procurement decisions about data that may be subject to compelled government access. Independent security audits of specific products, where available, are the appropriate next step after establishing this baseline.