OpenAI Enters the Silicon Game, the $206.5B Agent Economy, Anthropic Nears a Trillion-Dollar Valuation: Thursday Briefing, June 25, 2026

OpenAI and Broadcom revealed “Jalapeno,” an inference-optimized custom chip announced June 24. This is OpenAI’s first foray into custom silicon, and it signals a strategic recognition that owning the inference stack is no longer optional at OpenAI’s scale.

The economics are straightforward: inference is where the money goes. Every ChatGPT query, every API call, every agent loop burns GPU cycles, and at OpenAI’s current usage levels, even small efficiency gains per chip translate into hundreds of millions of dollars in annual savings. By partnering with Broadcom rather than building an in-house semiconductor team, OpenAI gets access to world-class chip design and manufacturing relationships without the multi-year ramp of hiring its own silicon engineers. The move puts OpenAI on the same path Google took with TPUs — except Google started a decade earlier.

Gartner’s latest forecast puts agent software spending at approximately $206.5 billion in 2026, a 139 percent year-over-year increase. More striking than the dollar figure: 57 percent of surveyed organizations report running agents in production environments. Agents have crossed from demo to deployment faster than almost any enterprise software category in recent memory.

The speed of adoption raises legitimate questions about readiness. “Agentjacking”— a newly identified attack class in which adversaries manipulate autonomous agents through prompt injection, tool poisoning, or environmental manipulation — is emerging as a serious security concern. When an agent has access to production systems and can take actions autonomously, the attack surface is fundamentally different from a chatbot that only generates text.

On the research side, VOTP (a preference-learning framework presented at ICML 2026) offers a more principled approach to aligning agent behavior with human intent, but the gap between academic alignment work and production agent security remains wide.

Anthropic’s IPO valuation is reportedly converging around $965 billion, part of a broader AI-listing pipeline that now totals approximately $3.6 trillion in expected market capitalization. These numbers place Anthropic in the company of the world’s most valuable enterprises before it has even gone public.

The valuation reflects both Anthropic’s commercial momentum and a market that is pricing in the assumption that frontier AI companies will capture an outsized share of future economic value. Whether that assumption holds depends on factors that are still deeply uncertain: the durability of model advantages, the economics of inference at scale, and the regulatory environment that is being built in real time around these companies.

GPT-5.6 is being previewed for a late-June release, while GPT-5.5-Cyberposted a record score on the CyberGym security benchmark — suggesting OpenAI is investing heavily in domain-specific fine-tuning for cybersecurity applications.

On the policy front, the EU AI Act reaches full applicability on August 2, creating a hard regulatory deadline that will force compliance decisions across the industry. The convergence of US executive orders and EU legislation means that frontier AI companies now face substantive regulatory requirements on both sides of the Atlantic simultaneously — a first for the industry.