NVDA is framed as the cleanest beneficiary of AI spend: Vera Rubin is now live, pitched at 10x token efficiency, 10x better models at the same cost, 336B transistors, 22 TB/s memory throughput, and rack-scale deployment, while buyers already burn tens to hundreds of millions a year on inference and training. GOOGL is cast as the direct loser: negative cash flow for the first time, 5% lower on the day despite +24% revenue and +17% Search growth, yet still no frontier model in 4 months and a weak Gemini 3.6 Flash. The middle of the video turns into a competition map: China’s open-source stack, US attempts to limit it, and a wave of routing platforms that could compress model costs before big labs react. TSLA closes as a network bet, with Starlink V5, Starlink V3 on Starship, and Cybertruck integration pitched as a decentralized AI/internet layer.
Companies are already spending tens to hundreds of millions of dollars a year on AI and are running out of money; GOOGL reportedly posted negative cash flow for the first time, signaling the token bill is outrunning revenue.
NVDA’s Vera Rubin was standing up yesterday and is the day’s big hardware news, after Jensen Huang announced it at GTC a few months earlier.
Vera Rubin is claimed to save 10x on tokens, enabling a 10x better model for the same cost, and 10x more tokens per megawatt into a GPU.
Vera Rubin is claimed to have 336B transistors, a 62% jump over Blackwell GB200, on TSMC 3nm, with 22 TB/s memory throughput and a design that glues two maxed-out dies together because single-die silicon ran out of room.
The chip is described as 10x more performant per watt and as having a baseline roughly 4x above GB300, which would make models trained on it plus algorithmic gains a “monster.”
NVDA’s advantage is framed as software plus hardware integration, letting users run a Fable 5 model or local open-source Kimiko 3 at 1/10 the infrastructure cost; the company is said to take 50% to 60% profit margin on every AI token dollar spent.
A full Vera Rubin deployment requires 72 units per rack plus 32 NVDA CPUs, and the CPUs have been shipping since last month; the speculative question is whether NVDA can make $20B this year from CPUs alone.
Production is framed as 1,000 Vera Rubin racks per day, with each rack holding many chips; the transcript also mentions a $630B-per-quarter projection described as a per-share forecast.
GOOGL is identified as NVDA’s direct rival via TPUs, but its new chip is said not to catch up; earnings reportedly included a significant chunk of TPUs spending while the company still lost more cash than it made.
GOOGL’s big week included Gemini 3.6 Flash, a Gemini 3.5 iteration marketed as cheap and cost-efficient but described as worse than ChatGPT 5.6 Luna and GLM 5.2; Chinese open-source Kimi K3 is said to undercut it on cost and flexibility.
GOOGL is said to have no frontier model in the last 4 months despite owning Gmail, Search, Chrome, Android, and G Suite, while DeepMind allegedly had to fight internally for compute instead of getting it as OpenAI and Anthropic do.
The market reaction is tied to forward expectations: GOOGL stock is down 5% even after an amazing quarter, revenue is up 24% y/y, and Search is up 17% on the year, but investors care more about whether it will still be at the frontier 12 months out.
The video argues China was 3 years behind on model development and chip architecture but is catching up fast, with Huawei-linked new chip architecture, a gigantic data center built to run only Chinese chips, and no NVDA chips in sight.
The open-source fight centers on Kimi K3, a line of Chinese models said to be headed to 90%-frontier capability at a fraction of US cost; the US is considering banning open-source models, which is said to hurt young startups, be easy to route around, and be ethically close to banning code/free speech.
Three routing platforms landed over 3 days: Cursor Router, Meta’s internal effort, and Ramp’s platform; Cursor claims “frontier quality at 60% lower cost,” and the risk is that OpenAI, Anthropic, and GOOGL could match routing one feature away.
TSLA is pitched as the next network play: Starlink V5 is said to be 40% smaller and much more efficient per watt, will be integrated into every Cybertruck, and together with Starlink V3 on Starship could provide internet anywhere on Earth while feeding data back to a mainframe and supporting AI/robotics.
The bull case for TSLA is a decentralized network for training, inference, and a new AI economy, plus the world’s largest robot network; a consumer example is better internet on Delta flights for streaming, video, and calls.