21: AI runs hot

and

Sep 12, 2025

Hey, it’s Marc.

This week felt like peak AI euphoria: Oracle jumped 36% and added nearly $250B of market value (currently $922B) in a single session. The reason? Revenue projection of $144B in 2030. (Read full story 👇)

Whereas, Tesla’s stock is down 25% this year, despite being last cycle’s “AI stock.”

The takeaway: Narratives move faster than fundamentals. AI companies’ valuations are based on the classic “pay for tomorrow, hope it arrives” framework. The real test will be actual earnings and cash flow, quarter after quarter, along with the scale meeting the demand.

Top AI signals this week:

Microsoft plans to buy AI from Anthropic. Link
OpenAI introduced ‘Projects’ for free users. Link
ByteDance launches Seedream 4.0 AI image model, challenging Google. Link
Claude now creates and edits spreadsheets, documents, slides, and PDFs. Link
Mistral AI adds free memory and enterprise integrations to Le Chat. Link
Boston startup and MIT Media Lab introduced Alterego: the near-telepathic wearable. Link

👉 Get your brand in front of 30,000+ decision-makers — book your ad spot now.

Power Moves

Oracle’s $300B OpenAI deal sparks market frenzy

Oracle’s stock jumped 36% in a day, adding $250B in market cap, after revealing its forecasts of $144B revenue by 2030, and OpenAI signed a $300B, 5-year cloud contract. That’s a 359% jump in backlog. Ellison’s (Oracle’s co-founder and CTO) personal stake surged by $89B. The halo effect lifted energy suppliers (GE Vernova, Vistra), data centre REITs (Digital Realty), and chipmakers (Broadcom). [RELEASE]

So what? OpenAI will burn $115B by 2029 and must raise tens of billions annually just to pay Oracle. “Future contract revenue” sounds huge, but valuations are being pushed higher on forward commitments, and forecasts lay fragile ground.

OpenAI’s gpt-realtime: faster, cheaper, smarter voice

OpenAI just made its Realtime API generally available, along with gpt-realtime, its most advanced speech-to-speech model yet. The model scores 82.8% on Big Bench Audio (vs. 65.6% last Dec), follows complex instructions 50% more accurately, and executes function calls at 66.5% accuracy, critical for customer support, banking, and healthcare workflows where precision matters. New features: SIP phone support, image input, and remote MCP integration. It means agents can now connect to call centres, see what users see, and plug into enterprise systems instantly. [RELEASE]

So what? AI voice isn’t just “IVR (Interactive Voice Response)2.0,” it’s about automating high-cost, high-friction human interactions (think claims processing, loan approvals, clinical triage).

gpt-realtime vs ElevenLabs: OpenAI GPT Real-Time operates as a unified speech-to-speech model that processes and generates audio directly through a single API. Whereas, ElevenLabs functions primarily as a TTS-focused solution that requires external logic or speech recognition layers for full interactivity. Ideal use cases:

gpt-realtime: Interactive voice agents, multimodal experiences
ElevenLabs: Content creation, voice cloning, multilingual content

Web licensing standard targets AI free riders

Reddit, Yahoo, Medium, Quora, and others just backed a new Really Simple Licensing (RSL) standard, an “RSS for the AI age,” that lets publishers encode licensing and compensation terms directly into robots.txt. Unlike today’s block-or-allow setup, RSL adds options like pay-per-crawl (get paid each time an AI scrapes your content) or pay-per-inference (get paid each time an AI response uses your data). [RELEASE]

So what? This could be as transformative as ASCAP was for music, pooling rights into a collective licensing body (RSL Collective) to set floor prices with AI giants. For enterprises, this means two things: (1) cost of training data is about to get priced in, ending the free lunch for AI models, and (2) if you publish digital content, you now have a path to monetize it at scale.

Spotlight: Replit’s Agent 3

Replit just dropped Agent 3, its most autonomous coding agent yet. Unlike earlier versions, Agent 3 doesn’t just write code; it tests, fixes, and even spawns new agents and automations on its own. [ANNOUNCEMENT]

Why it matters: The length of tasks AI agents can complete has been doubling roughly every 7 months. Agent 3 pushes that frontier, going beyond snippets to testing, fixing, and chaining work, a sign that coding is shifting from human-led to agent-driven.

How it works

Self-testing: Agent 3 spins up your app in a real browser, clicks through buttons, forms, and APIs, then patches what’s broken.
Longer runs: From 2 minutes in Agent 1 → 20 minutes in Agent 2 → 200 minutes now. That’s hours of autonomous dev time.
Builds other agents: It can now create bots and automations for Slack, Notion, Dropbox, Outlook, and more, no API key scavenger hunt required.
Flexible modes: Choose between quick prototyping (frontend-only) or full-stack app creation.

Key features

Autonomous app testing & fixing – up to 3x faster, 10x cheaper than “computer-use” models.
Max Autonomy mode – runs unsupervised for extended tasks.
Agent-building ability – creates Slack bots, Telegram schedulers, automated reporting, etc.
Enterprise-ready integrations – GitHub, Notion, Linear, SharePoint, Stripe, Databricks, and more.
Live monitoring – track app progress from desktop or phone.

Top use cases: QA automation, internal tooling, ops automation and prototype-to-production flow

Also: Replit raises $250M at $3B valuation after 50x revenue growth.

Agent3 is an add-on to the tech team: autonomous, iterative, and enterprise-integrated. For companies, this isn’t about marginal productivity gains; it’s about compressing the dev cycle itself.

Top AI Vendors/ Agents This Week

Agent Factory (Azure AI Foundry): Connects AI agents, apps, and data using open standards
ServiceNow Build Agent: Turns plain English prompts into enterprise-ready applications
Amazon Bedrock AgentCore Observability: End-to-end monitoring for trustworthy AI agents
Autonomous Detection Engineer (ADÉ): AI agent that rapidly adapts email threat defenses
Vanta AI Agent: Automates policy, risk, and vendor management workflows

👉 We’re tracking 1000s of AI and blockchain vendors so you can find the right vendors, partners, and bet on the right tech. Sign up for early access.

More from 51:

That’s all for today.

Thanks,

Marc & Team