Why Chinese AI Labs Rent Nvidia GPUs in Malaysia
Chinese artificial intelligence labs train advanced models by renting American graphics processors hosted in third-party hubs like Johor, Malaysia, completely bypassing physical export restrictions.
Theme
AI infrastructure and compute: what the top podcasts are saying.
Chips, data centers, power, and the cost of running AI at scale. 269 write-ups from 21 shows so far, the newest from November 2026.
269 write-ups21 shows
The short version
Podcasters report that power grid opposition and local permitting freezes restrict the expansion of AI data centers. At the same time, hardware buyers bypass export bans through remote cloud access, while software vendors discover that persistent background computing destroys consumer subscription margins.
Top talking points
Data center construction stalls under local voting and capital constraints
Voters passed data center moratoria in 43 out of 44 local ballot measures. Texas regulators froze new permits, halting up to 15 gigawatts of power capacity and instantly stranding pipelines for $50 billion construction projects.
- Data Center Bans Hand AI Compute Share to Amazon TBPN · Nov 8
- David Sacks on the $50 Billion Data Center Freeze All-In Podcast · Nov 8
- Brad Gerstner on AI Power: Permitting Moratoriums Could Stall 15 GW All-In Podcast · Nov 8
- Why Neocloud Stocks Dropped 50% After the Midterms TBPN · Nov 8
- Vik Uppal on Why Mavik Passes on AI Data Centers Capital Allocators · Nov 8
Persistent background computing ruins fixed consumer subscription fees
Continuous background operations cost $20 to $50 per active user each month. Charging flat monthly fees destroys unit economics, leading Elon Musk to pull compute from Grokbot to sell capacity for $15 an hour.
- Persistent AI Agents Cost $30 a Month to Run 20VC with Harry Stebbings · Nov 15
- Rory O'Driscoll and Jason Lemkin on AI Agent Unit Economics 20VC with Harry Stebbings · Nov 8
- Elon Musk Ripped Compute from xAI to Sell at $15 an Hour TBPN · Nov 8
Spot computing costs compress gross margins across frontier vendors
Frontier laboratories spend heavily on spot capacity, with Anthropic allocating $5 billion to on-demand processors that run far above contracted prices. Pure token reselling carries zero switching costs, triggering immediate price wars.
- Anthropic Did Not Suffer a Demand Problem in Q3 Complex Systems · Nov 8
- Anthropic Reached 50% Compute Margins on 7x Inference Gains Latent Space · Nov 8
- Erik Bernhardsson on Why Selling AI Tokens Fails 20VC with Harry Stebbings · Nov 8
- Frontier AI Labs Need Public Debt to Fund Compute Latent Space · Nov 8
Most interesting insights
OpenAI spent roughly $10 million in compute over an unconstrained one-week experiment to produce a mathematical proof for Navier-Stokes.
From OpenAI Spent $10 Million to Prove Navier-Stokes, Cheeky Pint · Nov 1
Cross-referencing Malaysian and Indonesian trade manifests against physical data center capacity revealed roughly 50,000 advanced GPUs missing from declared sites.
From How 50,000 Nvidia GPUs Vanished in Southeast Asia, Odd Lots · Nov 15
Local public backlash is accelerating: SpaceX representatives Riley and Katie faced hostile pushback in Cochise County over grid capacity, tax breaks, and rising residential power bills.
From David Sacks on the $50 Billion Data Center Freeze, All-In Podcast · Nov 8
From our research reports
The State Of: Q2 2026
AI Agents Drive SaaS to Headless APIs Amidst Worsening Compute Scarcity and Political Resistance
Across 7 shows, there was broad agreement on two critical trends: AI agents are fundamentally reshaping software value, shifting it from traditional user interfaces to headless, API-first architectures. Concurrently, a severe and escalating compute and power crisis, driven by infrastructure limitations, rising costs, and increasing public opposition to new data center construction, threatens AI's growth.
Where they split: A key tension emerged around the necessity versus the environmental and social costs of expanding AI infrastructure. Some argued for the critical need to build more data centers to maintain global AI leadership, while others highlighted the immense energy consumption and local burden of these facilities, fueling public and political opposition.
The State Of: Q2 2026
Elon Musk's SpaceX launches 'Elon Web Services' (EWS) via a $45 billion compute deal with Anthropic, turning existing assets into a major AI infrastructure play
Across 3 shows, there was agreement that Elon Musk's XAI and SpaceX have significantly entered the AI compute market, notably with a multi-billion-dollar deal with Anthropic. This strategic move, referred to as 'Elon Web Services' (EWS) or a 'neo-cloud' offering, positions SpaceX as a critical infrastructure provider by monetizing substantial data center investments and X's data.
Where they split: While shows reported on the deal's significance, TBPN highlighted the 'tension' and unexpected nature of the alliance due to Elon Musk 'hurling insults at the Anthropic team' just months prior, raising questions about the underlying motivations for the partnership.
The State Of: Q2 2026
AI's Insatiable Demand Forces Radical Overhaul and Recalitalization of the Global Semiconductor Supply Chain
The 2 shows broadly agreed that the rapid growth of AI is profoundly reshaping the semiconductor industry, creating an insatiable demand for physical hardware and driving a critical need for re-engineering the supply chain. Both recognized significant bottlenecks and rising prices as a direct consequence of this shift.
Where they split: Lenny's Podcast indicates AI is only at the "very, very beginning" of impacting CAD and lacks the foundational understanding for complex engineering design. In contrast, No Priors highlights AI's immediate, practical impact, driving significant efficiency gains in electronic design automation tools for companies like Intel.
People quoted on AI infrastructure and compute
Chase Lochmiller21 quotesEvan Conrad20 quotesJensen Huang19 quotesMackenzie Hawkins19 quotesErik Bernhardsson15 quotesJason Lemkin13 quotesMing-Yu Liu12 quotesPatrick McKenzie12 quotesReiner Pope12 quotesSean Lie12 quotesAndrew Thomas11 quotesDon Wilson11 quotesRene Haas11 quotesWalter Goodwin11 quotesLip Bu Tan10 quotesAlex Zhang9 quotesBrad Gerstner9 quotesDylan Patel9 quotesRoman Chernin9 quotesSundar Pichai9 quotesGavin Baker8 quotesJamin Ball8 quotesAndrew Feldman7 quotesAnjney Midha7 quotes
Latest write-ups
Persistent AI Agents Cost $30 a Month to Run
Ephemeral virtual machines for tools like Replit and Lovable cost $4 to $5 per active user each month by spooling up only when triggered.
Nvidia's Export Probe Escalated After In-House Counsel Called BIS
Nvidia General Counsel Tim Teter contacted the head of the Bureau of Industry and Security (BIS) in late summer to check if the agency played a role in detaining former Megaspeed executive Alice Huang in Singapore.
Sam Parr Says America Has an Unbeatable Setup for the Next AI Boom
Sam Parr argues the United States is entering an expansion matching 1880 to 1970, when Rockefeller's oil and Great Lakes iron ore drove global economic power.
NeoCloud Data Centers Face a 2028 Cliff
Iren saw its stock fall more than 70% after Microsoft demanded a mid-build chip upgrade, showing that new silicon cycles outpace data center construction schedules.
How 50,000 Nvidia GPUs Vanished in Southeast Asia
Megaspeed bought $4.6 billion in Nvidia hardware to become the supplier's biggest buyer in Southeast Asia, four days after incorporating in Singapore via a British Virgin Islands shell.
Joe Fioti: Why Software Decides the AI Chip Race
Luminal raised $25 million to build compilers that run models on alternative chips like Amazon Trainium and Positron Atlas.
Jensen Huang Resigns as Nvidia CEO Over Code of Conduct Breach
Jensen Huang stepped down as Nvidia CEO less than 24 hours after reports revealed an undisclosed relationship with direct subordinate Kimberly Powell, Nvidia's VP of Healthcare.
Nvidia Cannot Escape Federal Scrutiny Over China GPU Access
Jason Lemkin argues that Jensen Huang deserves a Steve Jobs-style pass from federal regulators because American AI infrastructure depends entirely on Nvidia.
DeepSeek Bought 160,000 Huawei AI Chips for Inference
DeepSeek ordered 160,000 of Huawei's flagship 950DT AI chips for a new data center cluster in Inner Mongolia.
Smuggling Two-Ton Nvidia Racks into China
Black market chip smugglers bypass US export controls by placing hardware orders for ghost data centers that do not exist.
Arthur Mensch Says AI Compute Has a Capital Problem, Not an Energy Problem
Energy accounts for only about 10 percent of the total cost of ownership for AI compute infrastructure today.
Shows covering this
TBPN6220VC with Harry Stebbings43Latent Space41All-In Podcast31No Priors20Odd Lots13Cheeky Pint8Dwarkesh Podcast8The Infrastructure Investor Podcast8How I Invest5Lex Fridman Podcast5Practical AI5Complex Systems4My First Million4Capital Allocators3How I AI3Dry Powder2Founders Podcast1How I Built This1Invest Like the Best1Lenny's Podcast1