Hey friend. It's Saturday, September 7, 2025 and we're covering: OpenAI's strategic hardware and market plays, advancements in AI agents and robotics, and LLM capabilities and limitations.
Don't keep us a secret: Share the email with friends
💡 Must Know
OpenAI to Make Custom AI Chips with Broadcom OpenAI is partnering with Broadcom to design custom AI accelerator chips, aiming for a 2026 release. This signals a strategic shift to reduce dependency on Nvidia and vertically integrate hardware and software development.
Why it matters: This is a power play to control their own destiny. By designing its own silicon, OpenAI aims to escape Nvidia's hardware monopoly, optimize performance for its specific models, and build a competitive moat that pure software companies can't easily cross, transforming the AI race into a battle of integrated giants.
OpenAI Cash Burn — Projected to Reach $115B by 2029 OpenAI projects its cash burn will hit $115 billion by 2029, driven by massive compute and top-tier talent compensation costs. The spending highlights the aggressive investment required to stay at the forefront of AI research.
Why it matters: The $115 billion figure is a stark declaration that building next-gen AI is a capital game of unprecedented scale. This level of spending is designed to price out all but a handful of competitors, creating a "compute moat" that secures OpenAI's market position and forces the industry to question if the path to AGI is sustainable only through massive financial losses.
💬 Quote of the Day
Optimus will generate 80% of the company's valuation.
⚙️ AI HARDWARE & INFRASTRUCTURE
The physical foundation of AI is being fiercely contested, from raw chip performance and funding to the sheer energy cost of compute.
Groq demonstrated its hardware achieving 400 tokens per second with Claude Code, showcasing the speed of specialized AI processors. [Link]
AI inference platform Baseten raised a $150M Series D, signaling intense market demand for optimizing post-training model deployment. [Link]
The rising electricity consumption of AI data centers is causing significant increases in energy costs, highlighting a major bottleneck for the industry's growth. [Link]
Broadcom's stock is surging as Wall Street increasingly views its AI chips as a leading alternative to Nvidia, intensifying hardware competition. [Link]
🧠 LLM ADVANCEMENTS & LIMITATIONS
While LLMs are advancing with new reasoning techniques, their real-world limitations and potential for misinformation are becoming critically clear.
A pilot study found LLMs struggle to reliably convert therapist notes into structured clinical formats, showing a gap in real-world medical application. [Link]
A new "Matrix of Thought" reasoning method reportedly enhances LLM accuracy and speed on complex questions compared to the standard Chain of Thought. [Link]
A new study reveals that Retrieval-Augmented Generation (RAG) is highly vulnerable to health misinformation when the retrieval system pulls incorrect context. [Link]
NVIDIA has introduced a new framework for building agents with Small Language Models (SLMs), aiming to enable more efficient, specialized applications. [Link]
In a policy simulation, LLMs consistently chose safety-focused policies for homelessness that aligned with expert choices and improved outcomes. [Link]
🤖 AI AGENTS & ROBOTICS
The focus is shifting from single models to coordinated systems, with both physical robots and software agents being deployed at scale.
Chinese startup Astribot has secured a contract to deploy over 1,000 of its humanoid robots in industrial settings. [Link]
LangChainAI released DeepMCPAgent, a new tool designed to help AI agents dynamically discover and use other tools. [Link]
The educational platform QANDA is now using LangGraph to power a multi-agent system for personalized student learning. [Link]
Major Chinese AI lab DeepSeek announced it is developing advanced, self-improving AI agents for a Q4 release, intensifying global competition. [Link]
At Hubspot INBOUND, SuperAGI promoted the concept of Agentic Customer Management (ACM) as a successor to traditional CRM. [Link]
🔬 RESEARCH CORNER
Fresh off Arxiv, this week's papers tackle foundational questions from AGI and hallucinations to the real-world job impact of AI.
PhD Project: An outline for creating AGI by simulating evolution in virtual ecosystems to develop self-modifying neural networks. [Link]
Researchers: A demonstration of using knowledge distillation to train smaller models like GPT-2 for robust, domain-specific SQL generation. [Link]
Microsoft Research: An analysis of Copilot usage data identified 40 high-skill knowledge work roles as highly vulnerable to AI disruption. [Link]
UK Government: A large-scale trial of Microsoft 365 Copilot in the public sector reported no clear productivity gains. [Link]
OpenAI: A study suggests LLM hallucinations often stem from training data that inadvertently rewards models for giving confident but false answers. [Link]
Cheers, Teng Yan. See you tomorrow.