Kimi K3 is the sharpest claim in the video: Moonshot Labs’ new ~2.8T open-weight model is framed as closing the US-China frontier gap, matching or beating GPT 5.6 on coding for early testers, and pushing a new low-cost frontier on the Pareto curve. The other big thread is infrastructure and distribution: OpenAI’s hidden hardware, xAI’s privacy-first coding tool and power buys, and Thinking Machines’ open-weight Inkling all point to context, customization, and cheap inference as the next battleground. The back half widens that theme with voice and screen-aware dictation tools, while DeepSeek’s rumored IPO and another fundraise suggest Chinese AI is also racing toward public-market scale.
Moonshot Labs’ Kimi K3 is the week’s headline: Dario’s open-source vs American frontier gap estimate had been ~8 months, down from ~1 year, and Kimi K3 is claimed to have closed it further.
Kimi K3 is described as ~2.8 trillion parameters, with early testers putting it above GPT 5.6 on coding and calling it ‘almost Fable-worthy’ on some tasks.
The model is said to one-shot visual prompts as well as Fable 5, generate video from a single prompt, and create 3D simulations at Fable 5 demo quality.
Kimi K3 is claimed to have generated a Mario Kart recreation from scratch, with physically accurate wings/physics, but the result still looked more like Minecraft than a finished AA game.
On broader tasks, Kimi K3 is said to fall short in general reasoning, writing, and coding in general, while being very strong on visual prompting and long-horizon generation.
Pricing is the other swing factor: Kimi K3 is said to cost $0.95 per million tokens, with input under $1 and output a couple dollars, implying video generation for roughly $5–$10 and Sonnet-level cost, not Opus-level cost.
The model is framed as a dream tool for long-running agentic workflows because it is cheap, visual, and can reduce sequential prompting by one-shotting artifacts that Anthropic/OpenAI-style tools need multiple prompts to build.
Thinking Machines raised the largest seed round in VC history at $2 billion, led by ex-OpenAI CTO Mira Murati, and shipped Inkling as its first model.
Inkling is described as just under 1T parameters, trained on ~45T tokens across text/image/video/audio in under 9 months, and designed as a completely open-weight omnimodel rather than the strongest general-purpose frontier model.
Inkling is said to be good at reasoning, okay at coding, and far from Fable 5-level general capability, but meant to become much more powerful when fine-tuned for a specific team or function.
Thinking Machines’ Tinker API is pitched as a way to hyper-optimize models for specific use cases with far less data than the ‘terabytes’ traditionally needed for RL, including a hedge-fund-style workflow where data stays local and encrypted.
The clip ties that to CNBC’s Alex Karp complaint that closed frontier tools are too expensive and have murky IP protection, while Thinking Machines is presented as letting users build custom IP without leakage; it also says Inkling uses DeepSeek V3 architecture and MoE for efficiency and customization.
OpenAI’s hardware leak points to a screen-free smart speaker/home computer with camera, environmental sensors, microphones, and a rechargeable battery, aimed at ambient, always-on use rather than manual ChatGPT-style prompting.
The device is targeted for February 2027, priced around $200–$300, with a goal of ~100 million initial units; a lawsuit from Apple over alleged hardware trade-secret theft could delay it, while Apple is also said to be planning AI hardware announcements this year including disk-pendant AirPods with cameras and smart glasses.
xAI’s Grok Build is described as a Claude Code competitor that initially uploaded users’ code to private xAI servers for training; after damage control, xAI open-sourced the product, added zero-retention controls, and turned privacy into a selling point.
Elon Musk’s APR Energy acquisition is framed as a power play: about $1 billion buys a Jacksonville fleet of trailer-mounted gas and diesel turbines with gigawatt-plus capacity, deployable in 15–30 days versus years-long grid interconnection queues.
That power hunt matters because the video says Musk has roughly 1–1.5 million latest Nvidia GPUs but faces a ~5-year wait to connect them to the grid; the claim is that portable power can unlock Colossus and other data centers fast enough to change the race.
Bloom Energy is held up as the modular power incumbent whose stock already went vertical, with Leopold Aschenbrenner said to have taken a big position and Musk framed as potentially wanting to buy it for SpaceX AI.
The voice/UI story shifts toward context collection: Whisper Flow is for speech-to-text by holding a button and dictating, while Farza’s Screen Aware Dictation is described as reading intent, understanding gaps, researching, pulling from personal files, and even building a Google slide deck from another tab.
DeepSeek is described as preparing an IPO, with investors reportedly discussing a new raise, a valuation up almost 1.5x, and founder Liang Wen Feng owning about 94% of the company, implying roughly $26 billion of net worth.
The close ties the week together: the video says OpenAI is releasing three new models in three modes, new open-source models are very cheap, and the market fight is now on price, context, and power rather than raw benchmark bragging rights.