OpenAI’s AI chip era begins

PLUS: Microsoft's new homegrown image model

Good morning, AI enthusiasts. OpenAI’s pursuit of compute continues to grow — this time to the point of building its own AI chips.

The company is working with Broadcom to design and deploy custom silicon, optimized for both performance and cost. The real test: can it meet Nvidia’s gold standard and mark the start of OpenAI’s self-sufficiency in the AI hardware race?

In today’s AI rundown:

  • OpenAI to make its own AI chips with Broadcom

  • Microsoft’s new homegrown image model

  • Build customer support agents with Agent Builder

  • AI models lie when competing for human approval

  • 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

OPENAI

Image source: Reve / The Rundown

The Rundown: OpenAI just announced a new, multi-year strategic collaboration with Broadcom to develop and deploy 10GW of custom AI accelerators, aimed at powering the next phase of advanced intelligence.

The details:

  • OpenAI will design the chips, using its learnings from developing frontier models, while Broadcom will handle manufacturing and deployment.

  • The racks with the custom chips will use Broadcom’s portfolio of Ethernet, PCIe, and optical connectivity solutions for scale-up and scale-out networking.

  • They will begin to come online in the second half of 2026, with the entire deployment set to be completed by the end of 2029.

  • The partnership, which surged Broadcom’s stock 10%, adds to OpenAI’s existing engagement for compute with Nvidia and AMD.

Why it matters: With this move, OpenAI joins giants like Amazon and Google in developing its own AI accelerator, aiming for tighter control over cost, performance, and supply. Yet, only time will tell whether these chips can truly rival Nvidia, which remains the dominant force and the industry’s go-to partner for AI hardware.

TOGETHER WITH TELY AI

The Rundown: You’re in a niche industry. Customers search on Google, ChatGPT, and Perplexity — but your company doesn’t show up, because there’s no content answering their questions. Tely AI fixes that by analyzing your industry and site to find what people search for and creating expert-level articles on autopilot. No need for a marketing team.

With Tely AI, you can:

  • Start with 60 high-quality articles a month

  • Get indexed on Google in as little as 2 weeks

  • Be featured in ChatGPT and Perplexity answers

  • Enjoy full automation for topics, writing, and publishing

MICROSOFT

Image source: LMArena

The Rundown: Microsoft introduced MAI-Image-1, its first text-to-image model developed entirely in-house, to power creative workflows across its products — signaling another step in its push to go beyond offerings from OpenAI.

The details:

  • MAI-Image-1 optimizes to produce photorealistic images while ensuring quick generations at the same time.

  • Right after its debut, the model appeared among the top 10 on LMArena’s leaderboard, taking the 9th spot (Hunyuan-image-3.0 remains first).

  • Microsoft prioritized rigorous data selection and nuanced evaluation, focusing on real creative use cases to ensure MAI-Image-1 avoids generic outputs.

  • The company said it plans to integrate it soon into Bing Image Creator as well as Copilot, going beyond current image models from OpenAI.

Why it matters: MAI-Image-1 is Microsoft’s third purpose-built model, following MAI-Voice-1 and MAI-1-preview. While these models remain secondary to the AI powering its core consumer and enterprise products, the trend is clear — Microsoft is steadily building its way toward reducing dependence on partners like OpenAI.

AI TRAINING

The Rundown: In this tutorial, you’ll learn how to use OpenAI’s Agent Builder to create a fully automated customer support system that classifies inquiries, gives intelligent answers from your documentation, and integrates directly into your website.

Step-by-step:

  1. Sign in to Agent Builder, go to Billing, and add credits to activate your account before creating your first workflow

  2. Click + Create to start a workflow, then add a Routing Agent that classifies queries (e.g., “product_info” vs. “billing_info”) using a simple prompt and JSON schema

  3. Connect the routing agent to conditional branches, then create specialized agents — one for billing and one for product info — and upload your documentation to power accurate responses

  4. Use Preview Mode to test common messages like “I was charged twice”, then publish your workflow and integrate it into your app using Chatkit

Pro tip: This setup can power far more than support. Reuse it for sales leads, help desks, or any system that needs smart classification and domain-specific responses.

PRESENTED BY IBM

The Rundown: Businesses face many challenges, from AI ambition to implementation. Success requires more than adopting new technology. It involves intentionally aligning AI with workflows, data, and goals. As AI evolves into a business-wide opportunity, leaders must ensure the right AI is applied to the right challenges.

Explore IBM's playbook on 3 challenges & approaches to boost agentic AI impact:

  • Finding your AI problem

  • Creating a clear AI plan

  • Integrating AI with the tools you use every day

AI RESEARCH

Image source: Stanford University

The Rundown: Stanford researchers just found that when “aligned” AIs compete for attention, sales, or votes, they start lying — exposing a fundamental flaw where models trained to win user approval trade truth and hard facts for performance.

The details:

  • Researchers tested Qwen3-8B and Llama-3.1-8B in sales, elections, and social media simulations, training them to maximize success based on user feedback.

  • Even when explicitly told to stay truthful, models began fabricating facts and exaggerating claims once competition was introduced.

  • Every performance gain came with rising deception: +14% misrepresentation in marketing, +22% disinformation in campaigns, +188% fake/harmful posts.

  • Alignment methods like Rejection Fine-Tuning and Text Feedback failed to prevent, and sometimes amplified, these dishonest behaviors.

Why it matters: With this behavior of reshaping answers to please and win rather than to be accurate, AI systems reveal a deep gap in how they learn from human feedback. In the real world, that tendency could quietly erode trust, turning tools meant to assist into systems that spread misinformation, inflate/deflate critical insights (like death toll).

QUICK HITS

  • ⚙️ AI Workflow Builder - n8n’s tool to build workflows from text prompts

  • 📹️ NotebookLM - Generate video overviews in six new visual styles

  • 💼 Claude Code - Package and share custom agents with plugin support

  • 📱 Grok Image 0.9 - xAI’s updated image and video generation platform

Salesforce introduced new Slack innovations, including a rebuilt Slackbot, a Channel Expert agent, Agentforce integrations, and new AI integrations, including ChatGPT.

Google announced an AI hub in Visakhapatnam, India, with an investment of approximately $15B over the next five years — its largest in India to date.

OpenAI founder Andrej Karpathy dropped nanochat — an end-to-end framework to train, fine-tune, and chat with a small-scale ChatGPT clone.

Anduril, the military tech company founded by Oculus creator Palmer Luckey, announced EagleEye, an AI-powered mixed-reality system for soldiers’ helmets.

Google announced that its new image editing model, Nano Banana, is set to launch across NotebookLM Video Overviews, Google Photos, and in Search via Lens.

COMMUNITY

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Cynthia L. in Jupiter, FL:

"I'm an instructional designer of eLearning courses and am building my personal portfolio. I'm using AI to develop scripts and quizzes, images and videos, TTS, and visual design briefs (for color pallets, typography, imagery style and iconography, layout and interface elements). I still have to build the course, but AI saves me so many hours of work!”

How do you use AI? Tell us here.

That's it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.

Login or Subscribe to participate in polls.

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown

Reply

or to participate.