Today's News
Moonshot AI via Exponential View
Moonshot AI's Kimi K3 (2.8T parameters) has upended the global AI hierarchy by matching or beating frontier US models while pricing at $6/million tokens—40% below Anthropic's Opus at $10/million. The pricing move appears economically rational: analysis of Google's pricing shows every 10% price reduction drives 12-18% volume growth (elasticity of -1.7), meaning lower token costs will likely expand total AI token revenue rather than compress it, at least while demand grows rapidly and workflows become more token-intensive with reasoning models and verification loops.
The structural implications cut deeper. At 50% gross margins, third-party cloud providers can profitably serve K3 tokens well below Anthropic's price, creating economic incentives for hyperscalers to reallocate GPU capacity away from frontier lab contracts paying $7.8–$11.5/GPU/hr toward open-weight inference earning $23–60/GPU/hr depending on workload type. This margin pressure on frontier labs persists despite growing overall token demand. Meanwhile, K3 now leads Code Arena benchmarks (first Chinese model to do so) just three months after K2.6, while Alibaba's upcoming Qwen3.8 (2.4T parameters) will be open-weight, signaling sustained competitive pressure on frontier labs' product differentiation. Moonshot is capitalizing on this momentum by preparing for a Hong Kong IPO targeting a $50 billion valuation, with final fundraising discussions expected in August.
Uber via The Pragmatic Engineer
Engineering leaders report that AI-assisted coding (Opus 4.5, GPT 5.4) has fundamentally shifted the software development bottleneck from code generation to code review, sparking adoption of specialized tools including CodeRabbit, Greptile, Qodo, and in-house solutions from Uber, Cloudflare, Faire, and HubSpot. At Uber specifically, 84% of developers are agentic coding users, with 65–72% of code AI-generated inside IDE-based tools and Claude Code usage nearly doubling in three months (32% in December to 63% in February). The company built internal infrastructure including Minion (background agent platform), Code Inbox (smart PR routing), uReview (high-signal code review comments), and Autocover (generating 5,000+ unit tests monthly). However, AI-related costs are up 6x since 2024, making token cost optimization a growing priority even at forward-thinking companies.
Bloomberg via Techmeme
Microsoft has begun substituting OpenAI's image generation models with its own MAI models across major products including PowerPoint and Bing. Mustafa Suleyman, Microsoft's model chief, stated the in-house models are "faster, cheaper, higher quality" and "drive better retention," with deployments achieving approximately 85% cost reduction compared to OpenAI's offering. This represents a significant shift in AI stack economics and signals major platform companies are moving toward vertical integration to reduce inference costs and dependency on third-party model providers.
The Information via Techmeme
AMD is collaborating with Cerebras to integrate AMD Helios with Cerebras wafer-scale chips for AI inference workloads. CBRS stock closed up 4.86% on the announcement. This partnership signals competitive infrastructure consolidation in the inference layer and validates Cerebras' hardware approach against traditional GPU-centric alternatives.
Black Forest Labs via AINews
Black Forest Labs released FLUX 3, a unified multimodal foundation model spanning image, video, audio, and action prediction in a single architecture. Early access to video generation is available now with an open-weights dev version coming. The model represents independent reproduction of frontier multimodal capabilities and signals competitive pressure on OpenAI and Google's video/audio stacks.
More significantly, BFL partnered with Mimic Robotics to build FLUX-mimic, a video-action model running on single on-prem GPUs that transfers video generation quality directly into dexterous robot control. The model has been tested and deployed at Audi on production tasks including kitting parts, inserting electronic control units, assembling components, and handling soft materials. BFL's thesis—that learning to generate realistic video requires understanding contact, motion, weight, and cause-and-effect, thereby creating a world model useful for robotics—appears validated: when action prediction was added to the training curriculum, video generation quality initially fell 10% but fully recovered after 3,500 steps, demonstrating that video generation and action prediction share a common backbone without permanent capacity loss.
BigCode / Hugging Face via AINews
The Stack v3, the largest open code dataset ever released, contains 114TB raw data across 224M repositories in 770 languages with approximately 5 trillion tokens of deduplicated and filtered source code. The dataset represents a major upgrade from v2 (2024): 68TB raw → 2TB / ~550B tokens, with significant gains in C++ (15x), TypeScript (7.5x), and Rust (7x). Unlike v2, which shipped file IDs requiring separate content fetching, v3 is self-contained with sources embedded directly. The dataset is explicitly positioned as infrastructure for next-generation open code models and raises the competitive floor for labs building code capabilities without closed ecosystems.