#AI #OpenSource #Nvidia #Regulation #USChina 🧠 🌐
⚡️ US and China schedule first AI risk talks for September - and 200 startups are fighting hard against Washington's open-source ban
The US and China have scheduled their first official AI risk negotiations for September 2026. The American delegation will be led by Treasury Secretary Scott Bessent, with the meeting expected ahead of Xi Jinping's US visit on September 24.
Three items on the table:
• Regulation of frontier models
• Restrictions on Chinese open-weight models
• Sanctions for IP theft
1️⃣ What lit the fuse: Kimi K3
Moonshot AI released Kimi K3 - 2.8 trillion parameters, 1M-token context window.
White House Chief Science and Technology Officer Michael Kratsios publicly accused Moonshot of distilling an Anthropic model while circumventing US chip export controls.
The accusation handed the Trump administration fresh grounds to revive the push to ban Chinese open-weight models outright.
2️⃣ 200 startups to Washington: this ban kills us, not China
Little Tech Association - Proton, Y Combinator, and roughly 200 companies - sent an open letter to Trump and Commerce Secretary Howard Lutnick pushing back on the proposed restrictions.
Chinese open-weight models are already globally distributed. A ban won't stop them - it will gut the American startups building on top of them.
The stated trigger: Kratsios accused Moonshot of distilling an Anthropic model and evading Nvidia chip export restrictions in the process.
3️⃣ Jensen Huang: use them, don't ban them
Nvidia CEO Jensen Huang took the opposite position from the White House in a direct interview with Axios:
🟢 Called Chinese open-source models "excellent" - said American companies should "absolutely" use them
🟢 Dismissed the distillation accusations: "distillation is a fundamental principle of intelligence"
🟢 Semiconductor stocks shed >20% from June peaks following the Kimi K3 release
4️⃣ China: using a mixer = presumed laundering intent
China's Supreme People's Procuratorate has proposed treating the use of crypto mixers and privacy coins as sufficient legal grounds to presume criminal intent to launder money.
Blockchain analytics reports would be formally accepted as court evidence. In 2025, Chinese networks processed $16.1B in laundered crypto - 20% of global volume.
5️⃣ Memory supercycle: no diplomatic solution to a wafer shortage
Zephyr analysts verified the figures: SK Group's chairman said clients are requesting 4-5x more memory than the industry can produce - and capacity will only double by 2030.
The physics is the constraint: one bit of next-gen HBM requires 4x as many wafers as DDR5.
Expected classic DRAM shortage: around 25% - even without accounting for Chinese supply restrictions.
✨ Washington wants to ban, invoking national security. Nvidia and 200 startups are pointing in the opposite direction. AI has become geopolitical infrastructure - and no rulebook for it exists yet.
The Bessent-Beijing September talks are a first attempt to draft that rulebook. Think of it as negotiating the playing field for the next decade of tech competition.
The real question isn't whether Kimi K3 is dangerous. It's whether America can afford - economically, not just technically - to cut itself off from Chinese open-weight models. The industry is already voting: open AI beats closed monopoly.
But the physical bottleneck doesn't care about diplomatic goodwill. Wafers, fabs, HBM scarcity. No meeting in September fills a 25% memory gap.
➡️Crouton.digital | About us⬅️
⚡️ US and China schedule first AI risk talks for September - and 200 startups are fighting hard against Washington's open-source ban
Washington and Beijing are finally sitting down to negotiate AI risk. But the sharper battle is playing out inside the US itself - and its outcome will determine what models the next generation of products gets built on.
The US and China have scheduled their first official AI risk negotiations for September 2026. The American delegation will be led by Treasury Secretary Scott Bessent, with the meeting expected ahead of Xi Jinping's US visit on September 24.
Three items on the table:
• Regulation of frontier models
• Restrictions on Chinese open-weight models
• Sanctions for IP theft
1️⃣ What lit the fuse: Kimi K3
Moonshot AI released Kimi K3 - 2.8 trillion parameters, 1M-token context window.
White House Chief Science and Technology Officer Michael Kratsios publicly accused Moonshot of distilling an Anthropic model while circumventing US chip export controls.
The accusation handed the Trump administration fresh grounds to revive the push to ban Chinese open-weight models outright.
2️⃣ 200 startups to Washington: this ban kills us, not China
Little Tech Association - Proton, Y Combinator, and roughly 200 companies - sent an open letter to Trump and Commerce Secretary Howard Lutnick pushing back on the proposed restrictions.
Chinese open-weight models are already globally distributed. A ban won't stop them - it will gut the American startups building on top of them.
The stated trigger: Kratsios accused Moonshot of distilling an Anthropic model and evading Nvidia chip export restrictions in the process.
💭 Suhail Doshi, founder of Particle:
"There will be hundreds of companies that instantly die. This is great for Anthropic - we'll all have to spend money on it."
3️⃣ Jensen Huang: use them, don't ban them
Nvidia CEO Jensen Huang took the opposite position from the White House in a direct interview with Axios:
🟢 Called Chinese open-source models "excellent" - said American companies should "absolutely" use them
🟢 Dismissed the distillation accusations: "distillation is a fundamental principle of intelligence"
🟢 Semiconductor stocks shed >20% from June peaks following the Kimi K3 release
4️⃣ China: using a mixer = presumed laundering intent
China's Supreme People's Procuratorate has proposed treating the use of crypto mixers and privacy coins as sufficient legal grounds to presume criminal intent to launder money.
Blockchain analytics reports would be formally accepted as court evidence. In 2025, Chinese networks processed $16.1B in laundered crypto - 20% of global volume.
5️⃣ Memory supercycle: no diplomatic solution to a wafer shortage
Zephyr analysts verified the figures: SK Group's chairman said clients are requesting 4-5x more memory than the industry can produce - and capacity will only double by 2030.
The physics is the constraint: one bit of next-gen HBM requires 4x as many wafers as DDR5.
Expected classic DRAM shortage: around 25% - even without accounting for Chinese supply restrictions.
✨ Washington wants to ban, invoking national security. Nvidia and 200 startups are pointing in the opposite direction. AI has become geopolitical infrastructure - and no rulebook for it exists yet.
The Bessent-Beijing September talks are a first attempt to draft that rulebook. Think of it as negotiating the playing field for the next decade of tech competition.
The real question isn't whether Kimi K3 is dangerous. It's whether America can afford - economically, not just technically - to cut itself off from Chinese open-weight models. The industry is already voting: open AI beats closed monopoly.
But the physical bottleneck doesn't care about diplomatic goodwill. Wafers, fabs, HBM scarcity. No meeting in September fills a 25% memory gap.
➡️Crouton.digital | About us⬅️