Anthropic's $30T TAM; OpenAI Chip Bests Nvidia? Vampire Bot!

Today's AI Outlook: 🤟🏼

Anthropic Brings Massive Prediction to IPO Table

Anthropic is reportedly preparing to tell potential IPO investors that its total addressable market could exceed $30T, roughly the size of annual U.S. economic output. The company could seek to raise up to $100B at a valuation near $2T, according to reporting cited by AI Secret. The pitch follows SpaceX's own $28.5T TAM estimate earlier this year and shows just how much future growth AI companies expect investors to price in.

Why It Matters

A $2T valuation depends on investors believing AI can capture a huge share of spending across software, coding, customer service, knowledge work and other industries. That makes Anthropic's actual growth trajectory especially important.

The Deets

• Anthropic reportedly sees a $30T-plus addressable market.
• A possible IPO could raise up to $100B at a valuation near $2T.
• SpaceX previously floated a $28.5T TAM.
• AI Secret cited monthly revenue growth slowing from 58% in April to 38% in July.

Key Takeaway

Anthropic is pairing one of AI's biggest potential valuations with one of its biggest market forecasts. Investors will have plenty of zeroes to examine.

🧩 Jargon Buster - Total addressable market (TAM): An estimate of the maximum revenue opportunity available if a company could capture every possible customer in a market.

♟️ Power Plays

OpenAI Turns Up the Heat on Nvidia

OpenAI published the first benchmark results for Jalapeño, a 700-watt custom chip developed with Broadcom to run AI models. OpenAI says it delivers faster inference while using less power than Nvidia's flagship systems, giving the company another way to lower the cost of serving AI products. OpenAI also said Astra and Codex helped move Jalapeño from initial design to manufacturing-ready hardware in just nine months.

Why It Matters

Owning more of the computing stack gives OpenAI tighter control over cost, performance and deployment. Nvidia remains critical for training, but Jalapeño could reduce OpenAI's dependence on outside hardware for the huge inference workloads generated by users.

The Deets

• Jalapeño runs at 700 watts, compared with Nvidia systems cited at up to 1,200 watts.
• OpenAI reported gains of up to 3.6x faster responses and 1.9x more work per watt.
• AI Secret cited gains as high as 104x on certain workloads and a cost near $1.56 per chip-hour.
• OpenAI does not plan to sell Jalapeño.
• Two more generations are already in development.
• Initial deployment is planned for later this year, with production ramping through 2027.

Key Takeaway

Jalapeño gives OpenAI more control over the cost of every answer it serves. Nvidia still dominates the ecosystem, but its biggest customers are steadily building more silicon of their own.

🧩 Jargon Buster - Inference: The process of running a trained AI model to generate an answer, prediction or action.

Perplexity Packs Its Agent Into a $4,699 Desktop

Perplexity and Nvidia launched Portable Computer, a version of Perplexity's Computer agent that runs AI workloads locally on Nvidia hardware. The system installs on Nvidia's $4,699 DGX Spark and keeps more work off the cloud, giving users greater privacy and potentially lower costs.

Why It Matters

As smaller models improve, more AI work can happen directly on owned hardware. Portable Computer lets users keep routine tasks local while calling cloud models only when they need extra horsepower.

The Deets

• Users can run Qwen 3.8 27B or Perplexity's PPLX 27B locally.
• More than 15 cloud models are available when needed.
• Local work uses no Perplexity credits.
• Cloud access requires user approval.
• Support for additional PCs is expected.
• Nvidia is reportedly discussing a multibillion-dollar investment in Perplexity at a valuation above $30B.

Key Takeaway

Portable Computer gives local AI a polished consumer-facing agent, with privacy, cost and control doing much of the selling.

🧩 Jargon Buster - Local AI: AI software that runs directly on a user's device instead of relying entirely on cloud servers.

🛠️ Tools & Products

Meet the Robot After Your Blood

Dutch company Vitestro received FDA De Novo authorization for Aletta, an autonomous robotic blood-draw system that uses near-infrared imaging, ultrasound and computer vision to find a vein and guide the needle. The system targets one of medicine's most common procedures and could let hospitals stretch limited staffing by having one phlebotomist supervise several machines.

Why It Matters

Blood collection is routine, labor-intensive and hard to staff at scale. A reliable autonomous system could solve a narrow but costly hospital problem without requiring a major overhaul of clinical care.

The Deets

• Aletta achieved a 95% first-stick success rate in trials.
• Median blood-draw time was 1 minute, 49 seconds.
• The haemolysis rate was 0.6%.
• One phlebotomist can supervise three devices.
• FDA authorization creates a regulatory path future competitors may follow.

Key Takeaway

Healthcare automation can create real value by handling one repetitive procedure extremely well.

🧩 Jargon Buster - De Novo authorization: An FDA pathway for novel medical devices without a substantially equivalent product already on the market.

💰 Funding & Startups

A Startup Makes a Big Bet on Physics

Caltech professor Anima Anandkumar and engineer Benedikt Jenik launched Accelerated Understanding, a startup building AI systems that forecast how physical environments change through space and time. The company uses neural operators rather than a traditional transformer and is targeting chip materials, extreme weather and robotics.

Why It Matters

Manufacturing, climate forecasting, materials science and robotics all depend on predicting complex physical systems. That creates room for specialized AI architectures built specifically for those problems.

The Deets

• The architecture uses neural operators.
• The system follows physical events across 3D space and time.
• The company says one test processed 5T data points in a single run.
• Anandkumar and Jenik reportedly passed on senior roles and a 35% stake in Jeff Bezos-backed Prometheus.
• Prometheus has since reportedly raised $12B.
• Accelerated Understanding is focused on enterprise applications.

Key Takeaway

Physics-focused AI is becoming its own competitive category, and Accelerated Understanding is betting specialized architecture can unlock it.

🧩 Jargon Buster - Neural operator: A machine-learning architecture designed to model how physical systems change across space and time.

🧪 Research & Models

MIT Gives AI a Taste for Extreme Scenarios

MIT engineers published research on η-learning, an algorithm that can generate plausible extreme events even when its training data contains no examples of those extremes. It learns from ordinary observations, then creates scenarios for events such as severe storms, floods and other rare risks.

Why It Matters

Organizations often need to prepare for events their historical datasets have never seen. η-learning could give insurers, infrastructure planners and supply-chain teams a new way to stress test rare but costly scenarios.

The Deets

• The research was published in Nature Communications.
• η-learning trains on ordinary historical data.
• It can generate spatial maps of extreme scenarios.
• Potential applications include floods, wildfires, financial crashes and supply-chain disruptions.

Key Takeaway

η-learning could help planners model credible worst-case scenarios even when history offers few examples.

🧩 Jargon Buster - Extreme event: A rare occurrence far outside normal conditions, such as an unusually severe storm, market crash or wildfire.

⚡ Quick Hits

• Anthropic upgraded Claude memory, creating shared memory across chat and Cowork that can save topics during a conversation.
• Apple introduced a new $899 Mac Mini for always-on agentic computing, while AI Secret also highlighted its new 2nm M6 and M5 Ultra chips for more local AI performance.
• Stanford updated its AI labor research, finding young workers in AI-exposed jobs are now 19% below less-exposed peers.
• Stability AI raised fresh funding from EA, Sony Music, Universal, Warner and AMD as entertainment companies invest more heavily in licensed creative AI.
• OpenAI data center chief Chris Malone reportedly left the company, adding another executive departure during a major infrastructure expansion.
• Google Cloud introduced industry-tuned Gemini Enterprise editions for financial services and legal teams, with healthcare and life sciences next.
• Nvidia unveiled Jetson Orin Nano 2, an entry-level robotics computer with doubled inference performance.
• Cisco expanded its Secure AI Factory with Nvidia and Supermicro for enterprise AI clusters.
• IBM unveiled a dual-architecture mainframe processor whose cores can run Arm and IBM Z instructions.

🔧 Tools of the Day

• Claude Voice: Anthropic's voice interface can help users talk through a website idea, develop a design plan and turn it into a prototype. The Rundown AI recommends asking for a "static site" when building portfolios, landing pages or mockups. See the guide
• Keenable: A search API that lets AI agents query and learn from the live web. Keenable
• ChatGPT Stickers: A tool for turning photos into sticker packs for iMessage, because apparently regular group-chat chaos needed an upgrade. See the tool

Today’s Sources: The Internet, The Rundown AI, AI Secret