Grok 4.6 is framed as moving xAI from a distant third to the frontier: it matches GPT 5.6 Soul on one benchmark index, sits one point off Fable 5, and is claimed to be 3-5x cheaper overall, with Grok 4.7 due in 3-4 weeks and Grok 5 by year-end at 10T parameters trained on SpaceX engineering data. The other big AI thread is open-source and infra: DeepSeek V4 Pro is leaking as a frontier cyber model, CoreWeave and Nebius are seeing demand and stock moves as GPU rental prices surge, and Intel is trying to fund AI buildout while the market whipsaws on capex and dilution. The video also ties AI capital to durability in non-AI assets, using the LA Lakers sale as a case for human-authenticated franchises, while OpenAI churn, Anthropic hiring, and an EU watermarking push round out the competitive and policy backdrop. The last stretch widens to eclipses and the Lakers/Disney-OpenAI circle of capital, but the throughline stays the same: AI capability is improving fast, and the bottleneck is shifting to compute, distribution, and durable ownership of scarce assets.
Grok 4.6 is described as a new frontier model from xAI, moving the company from a distant third versus Meta and Gemini to "at the frontier."
On the AAI intelligence index, Grok 4.6 is said to equal GPT 5.6 Soul and land one point behind Fable 5, while Grok 4.6 high scores above Fable 5 on some benchmark-card tasks.
Grok 4.6 is claimed to be 5x cheaper on output than Soul and 8.5x cheaper than Fable, or roughly 3-5x cheaper overall at the general intelligence level.
Grok 4.7 is expected in 3-4 weeks and is expected to beat Grok 4.6.
Grok 4.6 is built around multiple agents, and the same multi-agent prompting is said to be showing up in smaller Meta models and many Chinese open-source models.
xAI’s moat is framed as compute plus data: Elon is said to have amassed the most compute in the game over the last year, with 200,000 coherent GPUs being packaged into a setup about one-third the size of macro hard.
Training is said to be running now on a 10T-parameter model, with plans for 8T-10T parameters and monthly model releases that get incrementally stronger.
Grok 5 is said to arrive by end of year, tied to a 10T-parameter model and trained entirely on SpaceX engineering data, including proprietary rocket-science data and Starship schematics produced with no AI.
GrokBot is pitched as xAI’s version of Open Claw that actually works: fully agentic, hosted on xAI servers, mobile and desktop, with login and chat as the only setup, plus a text-box iMessage-like interface for spawning named, personalized agents.
GrokBot is aimed at knowledge workers automating email admin, calendar scheduling, customer support, and other desktop chores; it spins up cloud VMs, can take over a computer in a "safe" way, and is framed as a direct attack on Apple ahead of Siri AI's launch in about a month.
DeepSeek V4 Pro is said to be an unannounced but leaked open-source frontier model; on Terminal Bench 2.1 it scores 87.9, 0.1 off Fable 5, and on Cyber Gym it is described as slightly better than Fable 5 and a cyber-capability rival.
The open-source risk is framed as rising sharply: Dario’s post is cited as expecting an open-source Mitos-level cyber model in about 6 months, and an open DeepSeek release could give a malicious actor havoc-scale capability.
CoreWeave and Nebius are presented as neo-clouds: they collect GPUs, build their own infrastructure, and rent it for training and inference, with GPU rental pricing moving through the roof.
Nebius is said to be up 25% on the day and CoreWeave up 31% this week, while Nebius revenue is up 454% year over year, quarter-to-quarter growth is compounding by about 50%, and it sits on about $8 billion in cash.
CoreWeave is said to have signed an A100 contract extending to 2029 at an attractive price, carry about $35 billion of debt, have $6.5 billion of capex and just under $6 billion of negative free cash flow, yet still hold a $104 billion revenue backlog plus $25 billion in customer commitments with 50%-60% paid up front for Q3.
Intel asked for $15 billion in common stock to fund AI buildout and growth, the stock fell 5% on the announcement, then demand reportedly came to $100 billion and Intel raised $20 billion in under 24 hours; a major foundry client such as Apple, Anthropic, or OpenAI is expected to be announced within months.
The LA Lakers sale is used as a durability case: the team was sold last year for about $10 billion, then flipped within 12 months for $12.5 billion in what is described as the most expensive U.S. sale and the largest sporting-franchise transaction ever.
OpenAI is said to be dealing with leadership churn, including Brad Lightcap leaving after rising from growth lead to COO, while Sam Altman is said to look forward to working together with him on whatever comes next; at the same time, OpenAI is rumored to be heading toward an IPO.
Anthropic is said to be pulling in more people than OpenAI, yet both are structurally similar in that the research teams keep producing better models.
An EU watermarking rule is described as forcing top AI labs to embed traceable text watermarks tied back to Claude, even after edits, and the rule is questioned for whether it is official and for claiming global reach.
Bob Iger, who acquired Marvel, Lucasfilm, and 20th Century Fox, is used alongside Joshua Kushner and Thrive Capital to argue that frontier-capital investors are buying both AI infrastructure and post-AGI durable assets.