South Korea commits $585B to semiconductor complex as Samsung, SK Hynix build four new fabs

ChatPRD's Live Testing Reveals Open-Weight Models and AI-Native Development Reshaping Product Velocity

Two detailed case studies from Lenny's Newsletter demonstrate how open-weight models and AI-as-core-engineer are compressing traditional development timelines. GLM-5.2, Beijing-based Z.ai's open-weight model, benchmarks near Claude Opus 4.8 on SWE-Bench Pro with a 1M-token context window and reasoning mode, delivering production-grade capability in code generation, autonomous bug triage, and long-running agentic tasks at $3.36 per 6M tokens through Open Router—structurally cheaper than frontier models for extended-context workloads. Separately, Gusto's CTO Eddie Kim led a five-person team that shipped a tier-one product (Gusto Co-founder, an AI agent for HR and payroll automation) in 10 weeks using only Cloudflare Workers and Vercel AI SDK, with no Jira, no Figma, no specs, and no standups—replacing process overhead with a persistent Zoom room and Claude Code as the primary engineer. Both stories signal that open-weight models are shifting from hobbyist curiosity to viable alternatives that eliminate vendor lock-in, and that process-light teams treating AI as a core contributor from day one can compress product development timelines while maintaining quality.

South Korea Orchestrates $880B Chip and Data Center Buildout, Samsung and SK Hynix Commit $590B to Four New Fabs

South Korea announced at least 1.35 trillion won ($880 billion) in coordinated private-sector investment into chips and data centers, with Samsung and SK Hynix planning to build two chipmaking plants apiece in the southwest for 800 trillion won total. The country aims to double its memory production capacity within five years and secure world-class manufacturing capabilities to maintain its lead in memory chips crucial for AI. The investment represents approximately 5% of South Korea's 2024 GDP and dwarfs prior commitments: in 2021, Samsung and Hynix pledged $450 billion over a decade. Samsung intends to accelerate fab construction in the Seoul metropolitan area and develop Gwangju as a new memory manufacturing hub, while also investing in humanoid robot deployment at its Gumi fabs. The scale underscores how seriously major memory suppliers are betting on sustained AI infrastructure demand and suggests memory supply constraints may ease faster than consensus anticipated, though analysts caution the massive capacity expansion also raises future oversupply risk.

Mobile Gaming Reaches $113.3B in 2025, Now 56% of $201.6B Global Games Market as Fortnite Returns to iOS

The global games market crossed $200 billion for the first time in 2025, reaching $201.6 billion (+9.1% YoY), with mobile revenues hitting $113.3 billion (+10.7% YoY) despite declining download volumes, signaling growth through deepening monetization rather than expanding reach. Mobile now represents 56% of the total market, with 345,000+ games built on Unreal Engine's UEFN. Fortnite's return to mobile distribution unlocks access to the fastest-growing segment at a moment when Epic has built creator tools and monetization infrastructure, positioning the platform as an operating system for gaming rather than a single title. PC delivered its strongest annual growth rate on record at $43.6 billion (+12.0% YoY), while console grew modestly at $44.7 billion (+2.8% YoY). Newzoo forecasts the market will reach $234.4 billion by 2028 at a 5.1% CAGR, with Grand Theft Auto VI as the period's decisive demand catalyst, though memory chip cost pressures and rising user acquisition costs ($0.56 per install, up 30% YoY) present near-term headwinds.

DeepSeek's DSpark Speculative Decoding Achieves 30.9% Higher Accepted Length vs Eagle3, Becoming vLLM Standard

DeepSeek's DSpark speculative decoding technique achieves 30.9% higher accepted length compared to Eagle3 and is now deployed in DeepSeek-V4-Flash and V4-Pro preview engines, with the vLLM community already integrating it as a new single-GPU state-of-the-art decode path. The inference optimization demonstrates measurable gains on Qwen3-4B and reflects the shift in production inference bottlenecks from raw FLOPs to memory-bound KV-cache optimization.

Korea to Invest $585B to Build Semiconductor Complex in Southwestern Region

Korea Times reported on South Korea's government-backed megaproject announcement involving Samsung Electronics and SK hynix committing 800 trillion won ($519B) to build a new semiconductor fab complex. Main takeaways:

• Samsung and SK hynix each building two memory chip fabs in Gwangju (southwestern region) as part of regional diversification away from capital region concentration
• SK Group, GS Group, and Naver investing 550 trillion won to establish AI data centers targeting 18.4 GW capacity by 2035 (starting at 8.4 GW)
• Government pledging 30+ trillion won over 15 years in next-generation chip R&D and support systems
• South Korea positioning physical AI as national strategic industry with goal of top-3 global AI robot power by 2030
• Total SK Group investment commitment reaches 1.1 trillion won across semiconductors, data centers, and physical AI infrastructure

LongCat-2.0: 1.6T Parameter MoE Model Trained on AI ASICs with Long-Context and Agentic Capabilities

Longcat team in Longcat (independent AI research organization) published a report titled "LongCat-2.0: 1.6T Parameter MoE Model Trained on AI ASICs with Long-Context and Agentic Capabilities". Main takeaways:

• Model trained on 50K+ AI ASICs across 35+ trillion tokens with zero rollbacks, demonstrating frontier-scale training viability on alternative hardware platforms
• Introduces LongCat Sparse Attention (LSA) with 3x efficiency improvements over DeepSeek Sparse Attention indexer; achieves 30.9% higher accepted token length vs Eagle3 in speculative decoding
• Achieves 59.5% on SWE-bench Pro (vs 58.6% GPT-5.5, 57.3% Claude 4.6), 70.8% on Terminal-Bench 2.1, and 48.7 on proprietary Text2SQL benchmark (exceeding Gemini 3.1 Pro at 47.9)
• Deployed in production with Claude Code, OpenClaw, and Hermes integrations; demonstrates ASIC infrastructure maturity for long-context agentic workloads at scale

Build + Host vs Claude Managed Agents — Cost Model

Anthropic in Anthropic (via shared spreadsheet) published a report titled "Build + Host vs Claude Managed Agents — Cost Model". Main takeaways:

• At pilot scale, Managed Agents dominates because compute is negligible and engineering cost is the driver; self-hosted paths require six-figure infrastructure builds before validating agent value
• At 5M agent-hours/year, Managed Agents remains cheaper than all self-hosted options in base case, primarily due to active-CPU billing that pauses during idle waits while agents wait on inference and tool I/O
• Cost crossover to self-hosted only occurs under specific conditions: existing platform team (sunk cost), agents running actively most of wall-clock time, and reserved EC2 with efficient packing
• Billing model is the structural driver—Managed Agents, Cloudflare, and Vercel charge active CPU only; E2B, Fargate, and EC2 charge provisioned resources for full wall-clock time
• Public beta Managed Agents rate is $0.08/session-hour; AWS totals commonly underestimate real bills by 33-50% when including NAT, data transfer, logging, and load balancing