#大佬观点

AWS CEO Matt Garman 与 a16z 的 Raghu Raghuram 对谈：全球最大云提供商如何为智能体时代重构： 2026 年 2200 亿美元资本支出，订购 200 万块 NVIDIA GPU，AWS 自研 AI 芯片明年之前售罄，还推出无需信用卡、30 秒即可开通的 AWS 账户，让智能体立即动起来。 0:55 营收 1700 亿美元，增长 37% 2:40 30-40% 的营收最初来自初创公司 4:55 初创公司想要面向智能体的云 9:10 30 秒开通 AWS 账户 12:25 智能体不需要坚不可摧的数据库 16:30 为什么 AWS 不会把所有 GPU 卖给前沿实验室 22:00 为什么 Matt 不担心泡沫 24:45 为什么 AWS 自建电厂 27:45 为下一次供应短缺做规划 30:25 为一个县每位居民节省 5000 美元 33:10 AWS 如何打造自研芯片 40:05 是什么阻碍了企业级智能体 45:35 为什么 Bedrock 从不把你的数据发给模型 49:30 AWS 的新 AI 安全工具 51:50 为什么 AWS 的 10 人团队如今只需 3-4 人 YouTube: https://www.youtube.com/watch?v=rn_afJaPldg @mattsgarman @awscloud @RaghuRaghuram

Lip-Bu Tan: Intel will continue working with Terafab. $INTC

Meta unveiled Muse Charm at Connect. Tamagotchi for 2026. Our Snapdragon Summit note sees personal AI devices as a new growth market. We expect Snapdragon inside Muse Charm, though Meta has not disclosed the chip. (1/4)🧵

#1 company contributing code to @vllm_project right now? @AMD: 502 commits to core vLLM in 90 days. 13% of all org contributions, almost 4x Nvidia. Also grateful to Red Hat, IBM, Embedded LLM & Inferact for building it with us. Open source wins when hardware has a choice.

Someone used Claude Opus 5.5 to make a 3D animation of Nvidia Blackwell GPUs (from server down to an atom). Took 1 hour and full animation “runs from single HTML file, directly in browser. No vid editor. No pre-rendered 3D sequence. Just browser-based experience.” Very cool.

We cut our RAM in half for the Tesla AI5 chip (now 72GB of LP5) and 1/3 for AI6 (now 144GB of LP6). This was the only way to get enough volume for Optimus production and greatly reduces cost. As it turns out, we think this will have a negligible effect on Optimus performance, as memory bandwidth is a bigger limiting factor than total memory storage (bandwidth was held constant).

Fentanyl Grade Compute: Why a GB300 Rack Will Out-Price Blow by Weight in 2030 A GB300 NVL72 weighs roughly 1,580 kg fully populated and recent purchase orders put it at $5M per rack. That is $3,165/kg. Strip out the 1.5 tons of busbar, manifold and coolant and the GPU packages alone are well into gold territory, but we're pricing the rack, because that's what you actually take delivery of. Where that sits on the illicit commodity curve today ($/kg): $2,400 Cannabis flower $3,165 GB300 NVL72 $3,500 Fentanyl $28,000 Cocaine $65,000 Heroin $138,000 Gold So today NVIDIA ships a product that is denser in value than weed, roughly fentanyl-grade, and still an order of magnitude short of cocaine. For now... GB200 NVL72 was $4M, GB300 is ~$5M, and Vera Rubin NVL72 is already quoted at up to $8.8M at essentially the same rack mass. Rack weight is constrained by the datacenter floor loading; rack price is constrained by nothing. That's ~1.75x $/kg per generation with HBM, CoWoS and power all supply-limited through the decade. Cocaine: Colombia's new president has pledged a hard-line security crackdown with $1B in US aid behind it Every prior crackdown has consolidated the industry into fewer, better-capitalized operators with better logistics and higher yields per hectare. Consolidation is deflationary. We model US wholesale drifting from ~$28k/kg toward ~$12k/kg by 2030 as the supply chain professionalizes. **Crossover: 2030.** Our extrapolated NVL72-class rack hits ~$17k/kg while wholesale cocaine falls to ~$12k/kg. At that point the rational cartel pivots to smuggling racks, except a rack has a fixed 500+ kW power draw, a 3,300 lb forklift requirement and an export-control regime that is actually enforced. Cocaine has none of these problems, which is why it will remain the superior product for anyone without a substation. Key risks to the thesis: US retail coke is still ~$60–200/g, i.e. $60k–200k/kg so the crossover only holds at wholesale. Retail compute (H100 hour on a neocloud) is also marked up, so we consider this an apples-to-apples wholesale comparison. Full model available to Coke Research subscribers.

Grok 4.7 is a strong combination of intelligence, speed & low cost

TrendForce explains why 4-Hi HBM isn’t enough. (Personally, I think memory makers wouldn’t produce 4-Hi HBM even if it were enough.) https://open.substack.com/pub/trendforceinsights/p/4hi-hbm-edge-ai-data-center?r=4k4uta&utm_medium=ios

SemiAnalysis says Google's TPU beats NVIDIA by 50% per dollar. Its own dashboard, each chip at its best, says NVIDIA by 9.7x. The trick: run Blackwell with its four biggest advantages off and call it apples to apples. Free isn't cheap enough anon https://x.com/i/article/2098431457442287617