ENFR
8news

Tech • IA • Crypto

TodayShortsTop StoriesTopicsAll videosYT channelsCryptoArchivesFavorites

Kimi K3 shakes AI race as Nvidia, Microsoft-Mistral expand

AISunday, August 2, 2026· 6 videos

Briefing

Audio player
0:00 / 0:00

China pushes WICO governance bloc

Beijing is promoting an alternative AI order through the World AI Cooperation Organization (WICO), now backed by 29 countries. The initiative emphasizes open models, lower costs, and influence across developing markets. It signals a refusal to align with U.S.-led governance frameworks. The move could reshape standards-setting and market access in global AI.

Moonshot unveils Kimi K3 model

Moonshot AI introduced Kimi K3, a 2.8 trillion-parameter open-weight model. Benchmarks place it near top systems like Claude Opus 4.8 and GPT-5.6, with an intelligence score of 57. It jumped to first place in several real-world coding tests from a prior 18th ranking. The release narrows the perceived gap between Chinese and U.S. frontier models.

Open pricing undercuts Western leaders

Kimi K3 is priced around $3 per million input tokens and $15 per million output tokens, well below premium rivals. Its open-weight availability lets enterprises self-host and reduce dependency on subscriptions. This puts pressure on OpenAI and Anthropic pricing power. The shift could accelerate enterprise migration toward hybrid or self-managed stacks.

Kimi K3 targets autonomous agents

Kimi K3 is optimized for long-running, agent-style workflows rather than short chats. It is designed to plan, execute, and iterate across complex tasks with minimal supervision. This aligns with a broader industry push toward autonomous systems. The capability raises both productivity potential and concerns about oversight.

ChatGPT tiers split on privacy

ChatGPT Plus (~$20/month), Pro (~€103–€229/month), and Business (~€26/user/month) now define a tiered offering. Business disables training on user data by default and adds admin controls. Pro delivers up to 20× higher usage limits and advanced models. Billing differs between Work (token-based) and standard chat, enabling cost optimization.

Fable 5 beats Soul 5.6

Fable 5 built a playable escape-room game in about 15 minutes, versus 54 minutes for Soul 5.6. The faster model produced coherent mechanics, audio cues, and consistent visuals with minimal fixes. Soul 5.6 eventually delivered a playable result but suffered from latency and iteration friction. The test highlights speed and reliability as key differentiators for coding agents.

Phantom MK1 fuels robot arms race

Foundation Future Industries is developing the armed humanoid Phantom MK1, targeting military use. The company cites roughly $24 million in contracts, largely inherited via Boardwalk Robotics and intermediaries like IHMC. Plans aim for 5,000 to 50,000 units annually, with long-term costs below $20,000 versus about $300,000 today. Experts note prototypes remain fragile, with stability issues likely persisting for years.

Nvidia debuts GPU revenue financing

Nvidia launched a July 1 model combining credit support with revenue sharing on deployed compute. The approach targets capital-constrained neoclouds facing billion-dollar data center builds. Early projects include Sharon AI (up to 40,000 GB300 GPUs) and Firmus in Indonesia (360 MW, 170,000 accelerators). The structure boosts demand while echoing historical risks tied to vendor financing.

Microsoft–Mistral deepen Europe push

Microsoft and Mistral AI expanded a multibillion-dollar partnership on July 21 without equity exchange. Plans include European data centers such as Bruyères-le-Châtel with 14,000 Nvidia GPUs, targeting 1 GW by 2030. Mistral models like Medium 3.5 and OCR4 will be distributed via Azure and Copilot Studio. The deal balances sovereignty concerns with rapid enterprise distribution.

Videos covered

Previous briefings · AI