AI Giants Bleed Cash ($13.7B) as NVIDIA, Qwen, and Cerebras Unleash Open-Source Tools

Financial Performance & Business Models

Anthropic and OpenAI Financials: High Cash Burn Amid Rapid Growth

• Anthropic reported a $5.2B cash burn on $9B ARR with ~30M monthly users, while OpenAI had an $8.5B cash burn on $20B ARR serving ~900M weekly users, highlighting the high costs of scaling AI infrastructure.
  • Anthropic underestimated cash burn, -$5.2B on a $9B ARR with ~30M monthly users, while OpenAI had -$8.5B cash burn on $20B ARR serving ~900M weekly users
    • The Information: Anthropic Lowers Profit Margin Projection as Revenue Skyrockets

OpenAI’s Codex API Revenue Surges 20× in 5 Months, Adding ~$1B Annualized Revenue

• OpenAI’s Codex usage grew 20× in 5 months, contributing significantly to its API revenue, now estimated at ~$1B annualized from this segment alone.
  • OpenAI says Codex usage grew 20× in 5 months, helping add ~$1B in annualized API revenue last month

OpenAI CFO Hints at "Outcome-Based Pricing" (Royalties on AI-Generated Work)

• OpenAI’s CFO suggested a shift to outcome-based pricing, potentially taking a cut of profits from discoveries or products created using its AI models, sparking debates over ownership and local AI alternatives.
  • OpenAI CFO hinting at "Outcome-Based Pricing" (aka royalties on your work)? Makes the case for local even stronger.
    • GizmoChina: OpenAI Wants a Cut of Your Profits

New AI Models & Releases

GLM-4.7-Flash REAP (25% Pruned) Released with Strong Agentic Coding Performance

• Cerebras released GLM-4.7-Flash REAP (23B), a 25% pruned version of GLM-4.7-Flash, demonstrating improved agentic coding benchmarks and efficiency. The model is available on Hugging Face.
  • GLM4.7-Flash REAP @ 25% live on HF + agentic coding evals
    • Hugging Face: GLM-4.7-Flash-REAP-23B-A3B

Qwen Open-Sources Full Qwen3-TTS Family (5 Models, 10 Languages)

• Qwen released the Qwen3-TTS suite, including VoiceDesign, CustomVoice, and Base models (0.6B & 1.8B parameters), supporting 10 languages with advanced voice customization.
  • Qwen have open-sourced the full family of Qwen3-TTS: VoiceDesign, CustomVoice, and Base, 5 models (0.6B & 1.8B), Support for 10 languages
    • GitHub: Qwen3-TTS
    • Hugging Face: Qwen3-TTS Collection
    • Qwen Blog: Qwen3-TTS Announcement

AI Products & Tools

NVIDIA Launches PersonaPlex: Open-Source Real-Time Conversational AI Voice

• NVIDIA introduced PersonaPlex, an open-source, real-time conversational AI voice model designed for dynamic interactions. The 7B-parameter version is available on Hugging Face with a Colab demo.
  • Nvidia Introduces PersonaPlex: An Open-Source, Real-Time Conversational AI Voice
    • NVIDIA Research: PersonaPlex
    • GitHub: PersonaPlex
    • Hugging Face: PersonaPlex-7B-v1
    • Preprint Paper