Shin Jung-kyu, CEO of Lablup, reveals why he'd create a new programming language if starting a business now and shares insights on AI's rapid evolution, scal...
AI Infrastructure's Future: Why Lablup CEO Would Build a Programming Language Today
Key Insights
- Rapid AI Model Evolution: New foundation models are released every three weeks, with tech giants launching six models simultaneously and reshaping entire business models
- The Scale Challenge: Handling 10x annual growth in GPU clusters requires fundamental infrastructure changes—from managing 100 GPUs to 5,000+ units with exponentially increasing data flows
- Startups vs. Big Tech Paradox: While AI companies compete with internal startup-like divisions from tech giants, two-to-three person teams can now build MVPs in three weeks without needing venture capital
- Building Competitive Moats: Low-difficulty ideas face instant competition, making technically challenging or seemingly unprofitable ventures the only viable paths to sustainable business success
- Multimodal Future: Token generation is exploding (Google increased 100-fold in 14 months), with video and image generation requiring vastly more computational infrastructure than text-based AI
The Rise of Consumer AI Products from Model-Makers
The AI landscape is undergoing a fundamental transformation. Companies that were traditionally positioned solely as model providers—OpenAI and Google—are now launching consumer products at an unprecedented pace. This represents a seismic shift in how the industry operates.
The Acceleration Effect: New business models emerge almost weekly. Startups fail rapidly. More significantly, within major tech companies, hundreds of AI divisions are operating with startup-like independence and agility. This creates an unusual competitive dynamic: external startups must now compete not just with other startups, but with internal AI teams inside billion-dollar corporations that have vastly more resources.
The speed of innovation has become the defining characteristic of this era. Where releasing two major services annually was once considered ambitious, companies now launch dozens or hundreds of experimental services simultaneously, treating them as rapid experiments. This shift in strategy fundamentally changes how businesses approach product development, risk management, and resource allocation.
Why Traditional Startups Face New Challenges: The barrier to entry for AI-powered services has collapsed. What once took months to develop and launch can now be accomplished in weeks or days. This democratization of development tools creates a paradox—while it enables more people to build, it also means that any successful service can be replicated almost instantly by competitors with more resources.
Lablup's Backend.AI: The Invisible Engine Powering AI Services
Most people have never heard of Lablup, yet many have experienced their technology without knowing it. Backend.AI functions as the foundational infrastructure layer that powers AI services across enterprises and cloud providers. To understand Lablup's position, imagine the distinction between seeing a luxury car's sleek exterior and understanding the engineered precision that makes it move.
The White-Label Model: Lablup provides Backend.AI as a white-label solution to cloud companies and large enterprises. This means consumers interact with the platform daily—through various cloud services and AI applications—but the Lablup branding remains invisible. The infrastructure handles GPU optimization, manages multiple users with different requirements, scales computational resources dynamically, and ensures AI services run with optimal performance.
From YouTube for Machine Learning to Enterprise AI Infrastructure: Lablup's journey illustrates the importance of identifying real market needs. Early attempts included creating a YouTube-equivalent for machine learning—allowing users to run code and see results instead of watching videos. While innovative, the market simply didn't want this solution. The team then pivoted to educational platforms, which gained unexpected traction following AlphaGo's victory in March 2016, causing their server to crash from overwhelming traffic.
By 2017, the company committed to developing Backend.AI as a dedicated AI infrastructure operating system, launching the enterprise version in 2019. This focus on a genuine market need—organizations struggling to manage AI infrastructure at scale—proved to be the correct bet.
Why Backend.AI Matters for AI at Scale: The real value of Backend.AI emerges when organizations deploy AI on massive scales. When a company manages 100 GPUs, manual oversight is possible. At 5,000 GPUs, each device constantly transmits telemetry data—temperature readings, utilization metrics, error states. Multiply this across thousands of devices, add thousands of users with varying computational needs, and consider inference scenarios where hundreds of thousands of concurrent users access AI models. The infrastructure becomes impossibly complex without sophisticated orchestration software.
The Scaling Challenge: Growth at an Exponential Curve
CEO Shin Jung-kyu identifies scale as the paramount challenge for AI infrastructure in 2024. The computational demands increase tenfold annually, requiring architectural innovations that conventional approaches cannot accommodate.
From Hundreds to Thousands of GPUs: When managing 100 GPUs, infrastructure teams can potentially monitor each device manually. At 1,000 GPUs, this becomes impractical. At 5,000 GPUs, the sheer volume of data becomes a computational challenge itself. Each GPU generates constant streams of operational data—temperature, clock speeds, memory usage, error states. Multiply this data across thousands of devices, and the infrastructure for monitoring the infrastructure becomes as important as the infrastructure itself.
Geographic Distribution and Heterogeneity: Today's challenge isn't just scaling up; it's scaling across. Organizations need to manage GPU clusters distributed geographically while maintaining unified resource allocation, load balancing, and performance optimization. Different data centers may have different GPU architectures. Some facilities might have older NVIDIA H100s while others have newly released Blackwell GPUs. The software must optimize for each architecture's unique characteristics while presenting a unified interface to users.
The Inference Explosion: Training models represents one computational workload. Inference—where trained models actually serve predictions to millions of users—represents an entirely different scaling problem. Where 10 people might have been building models in 2022, now 10,000 researchers develop models. But for inference, the scale jumps to hundreds of thousands of concurrent users. This creates unprecedented pressure on infrastructure systems.
Blackwell GPU Stability: NVIDIA's newest Blackwell architecture promises enormous performance improvements, but early deployments have revealed stability challenges. Lablup, as one of the few companies NVIDIA engages closely with in the APAC region, bears responsibility for optimizing these systems, reporting issues, and helping the ecosystem adapt to this new hardware generation.
Why Shin Jung-kyu Would Create a Programming Language
If Shin Jung-kyu were founding a company today, he would create a new programming language. This seemingly unconventional choice reflects a sophisticated understanding of where competitive advantages still exist in the AI era.
The Problem with Current Languages: Programming languages today prioritize human readability. Python became dominant in AI development partly because humans find it approachable. However, as AI systems take over coding tasks, human readability becomes less relevant. Current languages include abstractions like objects and classes—features optimized for human mental models rather than for efficient CPU or GPU execution.
The gap between what humans write and what CPUs actually execute involves significant work from compilers and interpreters. These translation layers introduce inefficiencies. A language designed specifically for AI-accelerated instruction execution—similar to how GPU shader languages (like CUDA) are tailored to GPU architectures—could achieve significantly higher performance.
The Technical Difficulty Barrier: Creating a new programming language requires expertise spanning compiler design, CPU/GPU architecture, machine learning optimization, and system design. It's a technically demanding undertaking that most startups cannot execute. This high barrier to entry means fewer competitors will attempt it, making it an attractive area for building sustainable competitive advantages.
The Complementary Ecosystem: Such a language wouldn't exist in isolation. It would require AI systems capable of programming in this new language and translating code from conventional languages into this optimized format. This creates a self-reinforcing ecosystem where only companies with deep AI expertise and infrastructure knowledge can participate.
The Moat Problem: In most software domains, competitive moats are evaporating. If you create a web-based service, competitors can copy it within a week. If you build a mobile app, the barrier to replication has similarly collapsed. Open-source tools and AI-assisted development mean that ideas alone provide minimal protection. Only technically difficult projects or those that appear initially unprofitable—and thus unattractive to well-funded competitors—maintain defensibility.
The Startup Paradox: How Startups Can Compete with AI Giants
A remarkable shift is occurring in startup dynamics. The traditional venture capital model—wherein founders raise seed funding, develop over months, validate market fit, and eventually seek Series A funding—is becoming obsolete.
The New Three-Week MVP Model: Today's most ambitious startups follow a different pattern. Two or three founders leverage AI development tools to build a minimum viable product in three weeks. They launch immediately to real users. If the concept doesn't resonate, they pivot without significant sunk costs. If the concept gains traction, they're already generating revenue before reaching break-even, eliminating the need for venture capital.
The VC Dilemma: This has created an unusual situation for venture capital firms. Companies that appear promising are already making money and don't need investment. Companies that require capital funding typically aren't attractive investment targets. Venture capitalists report difficulty identifying investment opportunities because the dynamics have fundamentally shifted.
Advantages Are Available to All: Korean startups are discovering they can leverage the same AI advantages as Silicon Valley companies. The tools that enable rapid development in California work identically in Seoul. What matters is rapid iteration, willingness to pivot quickly, and the ability to identify market signals early. Because of this accessibility, startups that adopt AI development practices gain significant advantages over those that don't.
Unclear Winners: The outcome remains uncertain. Will major tech companies dominate the startup ecosystem through their internal AI divisions? Will external startups win by iterating faster and adapting to diverse market segments that large companies ignore? Will tech giants acquire successful startups that emerge, accelerating the consolidation trend? These questions will likely clarify around 2025-2026.
Open Source, AI Collaboration, and the Korean Developer Community
Korea's open-source participation rate is higher than commonly perceived. The perception of lower participation stems from a visibility bias: Korean developers often contribute under English usernames to international projects, making their contributions less visibly attributable to their nationality. In contrast, Chinese developers frequently maintain Chinese development logs, making their origin obvious.
Open Source as AI Collaboration: Contributing to open-source projects develops skills that transfer directly to AI collaboration. Communicating requirements to AI systems requires the same clarity and structure-oriented thinking that open-source development demands. Developers experienced in understanding large codebases, proposing changes, writing clear documentation, and iterating on feedback through code review cycles are naturally skilled at directing AI systems.
Corporate Recognition: Major Korean companies including SK, KT, Naver, and Kakao increasingly value open-source experience in hiring and advancement. These companies are investing in open-source education curricula, recognizing that such experience accelerates adaptation to AI-augmented development workflows.
The Faster Adaptation Factor: People with strong open-source backgrounds consistently demonstrate faster adaptation speeds when using AI coding assistants. This likely reflects that fundamental communication skills—whether explaining ideas to humans or machines—are being cultivated through open-source participation.
Common Mistakes in AI-Assisted Development
Even experienced developers make characteristic mistakes when working with AI code generation tools. Understanding these patterns can significantly improve development outcomes.
The Trust Problem: AI systems like Claude and Gemini generate code that appears correct but often contains subtle errors. AI might claim to have implemented a feature when it actually didn't, or it might misunderstand requirements and build something unexpected but plausible-sounding. The solution involves adopting test-driven development practices—writing test cases first, then having AI implement code to pass those tests.
Mechanical Verification Systems: The most effective approach involves cross-checking AI output using multiple systems. For instance, code generated by Claude can be reviewed by Gemini operating as a dedicated code reviewer. The reviewed code then undergoes unit testing to verify the suggested improvements actually work. This creates a mechanical verification loop where human judgment guides the process while automation handles repetitive testing.
The Testing Reluctance: Developers transitioning to AI-augmented development sometimes resist the additional discipline of comprehensive test coverage. However, in AI-assisted development, tests serve as the critical feedback mechanism by which developers communicate errors to AI systems. Without clear, failing tests, AI cannot understand what went wrong and improve its code.
Starting Point Matters: For developers new to AI-assisted coding, the easiest starting point is having AI write test cases for code the developer already wrote. This provides immediate value—catching mistakes—without requiring developers to trust AI to write production code from scratch. As developers gain experience understanding how to guide AI systems, they can gradually increase the scope of AI responsibilities.
The Multimodal Explosion: Infrastructure Needs for 2025-2026
Token generation statistics from major AI companies reveal accelerating growth that suggests infrastructure is already stretched. Google increased token generation 100-fold over 14 months. Anthropic increased generation 40-fold. These statistics predominantly reflect text-based AI, yet video and image generation are now driving the majority of recent growth.
The Text Baseline: Text-based token generation represents the smallest computational load. A human user might generate 200,000 tokens maximum through typing. However, generating a single AI image requires approximately 25,000 tokens. Creating a video requires generating 30 frames per second—representing orders of magnitude more computational demand than text.
The Historical Precedent: This situation resembles the transition from newspaper reading to radio listening to television watching. When LTE networks first launched, skeptics questioned their necessity—"What would people use all that bandwidth for?" The answer came with Netflix and YouTube, which converted network capacity into entirely new consumption patterns.
The Coming Video Era: If 2024 demonstrated multimodal AI's potential through image and video generation, 2025 will be the year video generation becomes mainstream. Current infrastructure cannot handle this demand. From major tech companies' perspectives, recent massive AI investments represent insufficient rather than excessive spending. This appears as overinvestment only to external observers; internally, it represents necessary preparation for demand that doesn't yet exist but will soon dominate.
Advice for Younger Generations Facing an AI-Transformed World
The emerging reality is that every field will eventually be affected by AI. Expert professions may be impacted even more severely than general occupations. Yet this doesn't mean that studying, research, or intellectual pursuits lose relevance.
The Science Transformation: Scientific research itself will become AI-driven. Where humans previously conducted science directly, AI systems will increasingly lead research efforts. However, the fundamental nature of research—asking important questions, evaluating evidence, building understanding—will persist. The tools will change; the underlying human need to understand the world will not.
Adaptation Through History: The world has transformed multiple times for people of Shin Jung-kyu's generation. Technology has consistently moved toward assisting humans and expanding their capabilities. AI represents the first technology attempting to replace intellectual capabilities directly, making this transition feel uniquely shocking. However, humans and technology have co-evolved throughout history. Once this transition period passes, new fields of study will emerge, new research directions will open, new business models will develop, and new forms of human collaboration will form around the transformed technological landscape.
The Permanent Human Role: Five years from now, researchers will work with AI as naturally as today's developers use Google Search or GitHub. The shock of transition will fade. What matters is developing the ability to ask the right questions, to understand where AI can help and where human judgment remains essential, and to collaborate effectively with both human and artificial intelligence.
Conclusion
Shin Jung-kyu's insights reveal an AI industry in rapid transition. Lablup's position as an invisible infrastructure provider illuminates how the most valuable companies often operate unnoticed by end users. The shift from monthly model releases to weekly product launches demonstrates AI's accelerating impact. The startup paradox shows that advantages are accessible to anyone willing to move quickly and iterate relentlessly.
The most important lesson for anyone navigating this transformation: sustainable advantages come from technical difficulty, deep expertise, and the willingness to tackle challenges that appear initially unprofitable. The age of simple idea-based businesses is ending. The age of technically demanding, infrastructure-heavy, intelligently architected solutions is beginning. For those willing to invest in depth rather than breadth, opportunity has never been greater.
원문출처: “지금 창업한다면 새로운 프로그래밍 언어 만들 겁니다” (래블업 신정규 대표)
powered by osmu.app