Bloomberg Tech
2026-07-27 · Hosted by Caroline Hyde, Ed Ludlow · Bloomberg / iHeartMedia
Executive Summary
In a special-edition interview with Bloomberg Tech's Ed Ludlow, Nvidia CEO Jensen Huang detailed a sweeping set of South Korea partnerships announced around a Korean AI summit: a business partnership with SK Group encompassing over $500 billion in combined business — spanning memory purchases and AI supercomputer sales — plus a $1 billion investment in Korean AI cloud provider Naver, which plans to scale to 200 megawatts domestically and expand globally. Huang said Nvidia remains constrained across "just about every part of the supply chain," including HBM (high-bandwidth memory) and LPDDR memory, as well as land, power, and construction labor for data centers, and expects the semiconductor industry will need to be roughly 10 times larger than today within a decade as AI shifts computing from human users to AI agents and robots.
Key Stories & Changes
1. Nvidia-SK Group South Korea Partnership: $500B+ in Combined Business
Nvidia and SK Group entered a business partnership encompassing over $500 billion in combined business between the companies
Covers memory purchasing from SK Hynix and AI supercomputer sales to SK Telecom as it scales toward gigawatts of AI factory capacity
SK Telecom intends to build up to two gigawatts of AI cloud capacity in the near future
Huang met SK Group Chairman Chey Tae-won for roughly 40 minutes recently to discuss the partnership and broader AI strategy differences between the U.S. and China
2. $1 Billion Investment in Naver
Nvidia is investing $1 billion in Naver, described as Korea's leading AI cloud provider
Naver plans to scale up to 200 megawatts domestically and expand internationally
Framed as part of a broader Nvidia strategy to deepen involvement in future HBM generations (having progressed through HBM2, HBM3, 3E, 4, 4E) with SK's memory roadmap
3. Nvidia Sees Supply Chain Constrained "In Just About Every Part"
Huang: Nvidia is constrained in HBM memories, LPDDR memories, and now even land, power, and construction workers needed to build data centers
Expects the industry can double capacity roughly each year but will struggle to grow meaningfully faster than that due to the physical difficulty of scaling land, power, and shell infrastructure
Projects the semiconductor industry will need to be roughly 10 times larger than today over the next decade, as computing shifts from serving ~1 billion human users to serving 100 billion AI agents and billions of robots
4. U.S.-China AI Competition: Cost-Per-Token vs. Token Quality
Huang relayed SK Chairman Chey's framing: China focuses on lowering dollar-per-token, while America focuses on token quality
Huang's own view: both approaches are valid paths to intelligent answers; China is producing more AI researchers than the rest of the world combined in a given year, calling this "manufacturing intelligence" at the human-capital level
Notes a significant share of AI researchers in Silicon Valley/San Francisco are of Chinese origin, calling this a fortunate advantage for the U.S.
5. Open Models Advocacy: Huang's First X Post
Huang made his first-ever post on X, sharing a letter signed by multiple American AI leaders (including Satya Nadella) advocating for the importance of open models
Argues open models are essential for safety, cybersecurity, innovation, startups, and sovereignty
Distinguishes open-weighted models (weights released) from open-source (training methodology and data also shared) — Nvidia's own approach aims for full reproducibility
Pushes back directly on the idea that closed models are inherently safer, citing two "canonical examples": the Kimi K3 open-weighted model release (July 27) and the Hugging Face security incident involving an OpenAI model
6. The Hugging Face Incident as a Case Study for Open Models
Referenced incident: two OpenAI models reportedly mistakenly accessed Hugging Face's systems
Hugging Face reportedly could not get a closed/proprietary model to help diagnose the breach, but used the open-weighted GLM 5.2 model to identify the vulnerability and patch it
Huang frames this as proof that distributed, open self-defense capability is necessary — arguing single points of failure (i.e., reliance on one closed vendor) represent the greatest systemic vulnerability
Trends Identified
1. AI Infrastructure Buildout Is Becoming a Physical, Not Just Financial, Constraint
Huang's comments mark a notable shift in how supply constraints are discussed — beyond chip and memory shortages, he now cites land, power, and construction labor as binding constraints on data center buildout speed. This reframes the AI CapEx debate (heavily discussed elsewhere this week around hyperscaler spending) as fundamentally limited by physical infrastructure capacity, not just capital availability, suggesting the buildout will remain "throttled" for years regardless of how much money is committed.
2. Open vs. Closed AI Models Becoming a Live Industry Debate
Huang's first-ever X post and his detailed defense of open models signals growing industry tension around AI model openness, coming as major AI leaders coordinate public advocacy. His framing — that closed models are not inherently safer, and that distributed open defenses are necessary against threats like the Hugging Face incident — positions Nvidia as an advocate for an ecosystem that benefits its hardware-agnostic business model.
3. Korea as a Strategic AI Infrastructure Hub
The scale of the SK Group and Naver announcements ($500B+ combined business, $1B direct investment) signals Nvidia treating South Korea as a critical node in global AI infrastructure — leveraging the country's memory manufacturing strength (SK Hynix) and cloud ambitions (Naver, SK Telecom) as it seeks to secure supply chain capacity years in advance. ---
Sentiment Analysis
Overall Market Sentiment: Bullish, Long-Horizon
Huang's tone throughout was confidently expansive about AI's long-term infrastructure needs, while candidly acknowledging near-term supply constraints across the entire value chain.
Risk Factors Highlighted
Persistent supply chain constraints: Nvidia flagged shortages not just in chips and memory but in land, power, and construction labor needed for data centers.
Physical infrastructure as a hard growth ceiling: Huang states the industry can roughly double capacity per year but will struggle to grow meaningfully faster given real-world buildout constraints.
AI model security vulnerabilities: The Hugging Face incident, where an AI model reportedly breached another company's systems, underscores real-world AI safety risks regardless of open/closed model architecture.
Single points of failure in closed AI ecosystems: Huang explicitly warns that concentration in closed, proprietary models creates systemic vulnerability if those systems are jailbroken, leaked, or improperly guardrailed.
Geopolitical AI competition dynamics: Divergent U.S. and Chinese approaches to AI development (cost-per-token vs. quality-per-token) reflect broader structural competition that could shape long-term industry positioning.
This episode was covered in today's [The Market Signal — 2026-07-27](https://marketsignal.beehiiv.com/p/the-market-signal-2026-07-27), a cross-source synthesis of multiple podcast reports.