Stockcast Market Briefing

Musk Announces Terafab: History's Largest Chip Manufacturing Project Elon Musk announced integrating resources from SpaceX, xAI, and Tesla to launch a project called Terafab, aiming to deploy 1 TW of AI computing capacity in space. #Why Earth's Production Capacity Isn't Enough Global AI computing annual output is currently around 20 GW, while Terafab's demand is 1 TW. Musk said: All the wafer fabs on Earth combined only represent 2% of what we need. He stated he has personally told Samsung, TSMC, and Micron: We will buy all their chips. But manufacturers' expansion speed is far insufficient. Either we build Terafab, or we have no chips. #One Austin Building Handles All Processes The first step is building an advanced technical wafer fab in Austin, Texas, integrating complete capabilities in logic chips, memory, packaging, testing, and in-house photomask manufacturing. As far as I know, integrating logic, memory, packaging, testing, and photomasks all in one building doesn't exist anywhere in the world. He believes this rapid iteration mechanism can accelerate chip improvement by an order of magnitude compared to existing methods, and allows them to attempt some very crazy new physics concepts. #Two Types of Chips: Robots and Space Terafab plans to produce two chip categories. The first is optimized for edge inference, primarily for Optimus humanoid robots. Musk estimates humanoid robot annual production will eventually reach 10 billion units, far exceeding automobiles' 100 million. The second is high-power chips designed for space environments, accounting for harsh conditions like high-energy ions and electron accumulation, designed to operate at higher temperatures to reduce radiator weight. He estimates ground chip annual production capacity at around 100-200 GW, while space could reach 1 TW. #Space Deployment More Cost-Effective Than Ground Musk noted space solar efficiency is at least five times that of ground, with no atmospheric attenuation or day-night cycles, and requires no heavy frameworks, making costs actually lower. I think the cost of sending AI chips to space could be lower than ground deployment in just two to three years. Ground power infrastructure becomes increasingly difficult to build, with costs only rising. Completing Terafab requires 10 million tons annual orbital capacity, calculated at 100 kW per ton. Musk said this requires no new physics, with Starship V3 payload increasing from 100 to 200 tons. xAI previously built the first GW-scale computing cluster at record speed,

Memory structural super-cycle has emerged in various companies' financial reports and stock prices. Micron's last quarter earnings far exceeded market expectations, with revenue up 196% year-over-year to ~$738M, gross margin surging to 74.9%, and profit up 686% year-over-year. Next quarter's outlook is even more dramatic with revenue projected up 271% year-over-year and gross margin reaching 81%. EPS increased 10x year-over-year to $19.15. Due to the severe supply shortage currently facing the memory industry, Micron has decided to significantly increase capital expenditure, with this quarter's ~$155M capex projected to increase to ~$217M next quarter. The company will launch an unprecedented capacity expansion plan, raising full-year capital expenditure for fiscal 2026 from ~$620M to ~$775M, with fiscal 2027 capex expected to increase further significantly. More detailed earnings and guidance: https://www.patreon.com/posts/153388027?pr=true Vocus version: https://vocus.cc/article/69bb43b7fd8978000194dbb1

Besides the 2026 GTC conference, Huang also participated in an analyst Q&A where he answered many common market questions. These included Is Nvidia taking too much profit from the industry, Who will pay for AI capital expenditures, Clear timeline roadmap for copper cables and optics, Incremental opportunities from Groq, Future business models for software companies, and more https://www.patreon.com/posts/153327731 Since Huang had more information, he mentioned some things outsiders cannot see, including the remarkable growth rates of OpenAI and Anthropic, as well as the current rankings of token consumption in the model market. He said first place is OpenAI, second is the sum of various open-source models, third is Anthropic, and fourth is xAI. More detailed content can be found in the column's summary.

Old Huang painted a new computing economy centered on tokens at GTC 2026, with the core proposition: how AI factories convert electricity into revenue. He also devoted considerable time to discussing the OpenClaw lobster agent trend. He pointed out that computing demand has exploded 1 million times over the past two years, driven by three inflection points: generative AI changed computing methods, reasoning AI (o1/o3) enabled AI to decompose problems, and Claude Code allowed AI to truly execute tasks. The inflection point of reasoning has arrived. He revised his outlook from last year's $500 billion to at least $1 trillion in infrastructure demand by 2027. The most essential part of the speech was #Token Economics. Your data center used to store files, now it's a factory producing tokens. He illustrated tiered pricing: free tier attracts customers, mid-tier around $3 to $6 per million tokens, premium tier $45, top-tier service up to $150. The smarter the AI the higher the price, but the lower the throughput, so architecture optimization determines everything. Grace Blackwell achieved 35 to 50x power efficiency improvement compared to Hopper, with revenue per data center increasing 5x. The newly released Vera Rubin provides another 10x improvement at the most valuable inference tier, with revenue multiplying 5x again. After integrating Groq LPU, ultra-high-speed token generation improved another 35x. In a 1 GW factory, token generation rate increased from 2 million to 700 million within two years. He summed up the cost logic in one sentence: A 1 GW factory amortized over 15 years is $40 billion, even with nothing on top. You better make sure you put the best systems on it. #Product Roadmap: Rubin Ultra is in tape out, paired with Kyber racks connecting 144 GPUs. Next generation Feynman includes new GPU, LP40 LPU, Rosa CPU, supporting both copper wire and chiplet optics. Committed to releasing new architecture annually. Another major announcement is #OpenClaw. Jensen Huang defined it as the operating system for agent computers: Just as Windows let us create personal computers, OpenClaw lets us create personal agents. He predicted every SaaS company will transform into AaaS (Agent as a Service), with token budgets becoming Silicon Valley's new hiring tool. More detailed content can be found in the column article at https://www.patreon.com/posts/153212614