Today's Headlines

  • OpenAI and Broadcom unveil Jalapeño, OpenAI's first custom chip optimized for LLM inference
  • Google integrates computer use into Gemini 3.5 Flash, opening autonomous browser and PC control to developers
  • Microsoft CEO Satya Nadella says picking the "best model" is pointless and that a company's own "learning loop" is what matters

The foundation of AI and the hands of AI advanced on the same day. OpenAI stepped into custom silicon that shapes inference cost, while Google placed an agentic capability — AI that sees a screen and acts on it — inside one of its mainstream models.

At the same time, Microsoft's Satya Nadella waved off the question of which model is strongest, arguing the axis of competition should shift to how an organization is designed to operate. Hardware, agents, and corporate strategy — the question of where to invest in AI moved across several layers at once today.

Today's Top Three

OpenAI and Broadcom Unveil "Jalapeño," a Custom Inference Chip

OpenAI announced Jalapeño, a custom chip optimized for LLM inference and co-developed with Broadcom.

It is OpenAI's first proprietary AI processor. It is designed specifically for inference — running already-trained models rather than training them — with real-time coding among the intended workloads. The partnership itself was announced back in October 2025, and this is the result taking concrete form.

The aim is to reduce dependence on NVIDIA GPUs. With a severe GPU shortage persisting, owning a chip you designed yourself lets you cut both supply constraints and inference cost structurally. OpenAI claims better performance-per-watt than the current state of the art, though the chip is still in a testing phase and broad deployment details have not been disclosed.

The move has OpenAI tracing the vertical-integration path that Google and Amazon have already taken. By controlling everything from chip design through kernels, memory, networking, and deployment, the company positions itself to optimize models for speed, reliability, and cost as a single system. It signals that the contest for control of AI infrastructure now reaches down into the silicon layer.

Gemini 3.5 Flash Gains "Computer Use" and Operates a PC on Its Own

Google announced that it has integrated a computer use capability into Gemini 3.5 Flash.

With it, an AI agent can see a screen, assess the situation, and carry out actions by itself. The scope is not limited to the browser; it spans mobile and desktop environments as well. The key shift is that a capability once offered as a standalone model is now built into the lightweight, fast Flash model itself.

Availability began immediately. Developers can access it through the standard Gemini API, and enterprises through the Gemini Enterprise Agent Platform. It is a step that widens AI's role from something that returns instructions to something that actually moves the controls.

On safety, Google has layered several defenses. The company says it applied adversarial training against prompt injection — attacks that smuggle in malicious instructions from external content. For enterprises, optional safeguards can require explicit user confirmation before sensitive or irreversible actions and automatically stop a task when an indirect prompt injection is detected. Once you hand an AI control of the screen, this kind of guardrail design becomes a precondition for adoption.

Nadella: Picking the "Best Model" Is Pointless — the Key Is Your Own "Learning Loop"

Microsoft CEO Satya Nadella said a company's competitive advantage does not come from choosing a particular AI model.

Nadella first declared that "SaaS is dead," arguing for a shift away from a model where you deliver a service and stop, toward one that keeps learning through use. At the center of that is the idea of a "learning loop": a continuous, two-way cycle in which humans learn from AI and AI learns from human expertise.

Two ingredients underpin the strategy, he says. One is human capital; the other is token capital. Whether a company can combine the two and run its own proprietary learning cycle becomes the differentiator in an era when model performance is converging.

The implication is to move the axis of AI procurement away from "which model performs best" toward organizational design — how a company embeds learning into its own work. The more the gap between models narrows, the more the contest shifts to the operating side.

Other Developments

Models & APIs

Business

  • Agility Robotics, maker of the Digit humanoid robot, announced plans to go public via a SPAC merger with Churchill Capital Corp XI at a roughly $2.5 billion valuation. Orders for its next-generation Digit v5 exceed $300 million, and its investors include Amazon, NVIDIA, and SoftBank Vision Fund 2. Agility Robotics plans to go public via SPAC in a $2.5B deal (TechCrunch)
  • NVIDIA published a Rubin-generation reference design for liquid cooling. A closed-loop system circulating coolant at 45°C can cut data center water use — roughly 2.6 million gallons per megawatt annually — to near zero, and it estimates a 50 MW facility could save over $4 million a year in cooling costs. 45°C cooling design cuts data center water use to near zero (NVIDIA Blog)
  • Kioxia Holdings expanded operating profit roughly 70-fold in a single year. Surging demand for SSDs in AI data centers is driving the growth, with April–June 2026 operating profit around ¥175 billion and full-year growth projected at 74% over the prior year. Why is Kioxia growing so fast? A primer on memory chips (ITmedia AI+)

Products & Features

Policy

Other

Source: Selected by the editorial desk from the AI news inbox (collected June 25, 2026; 28 items, 2 primary and 26 secondary).