Jon Markman

Summary OpenAI has unveiled "Jalapeño," its first custom AI chip, developed in partnership with Broadcom. This application-specific processor targets AI inference, aiming to significantly reduce recurring operational costs and provide OpenAI greater control over its hardware stack. The move signifies OpenAI's shift towards vertical integration, lessening reliance on general-purpose hardware. However, the initiative also underscores Broadcom's strategic advantage, as it designs custom silicon for major AI players like OpenAI, Google, and Meta, effectively acting as a "toll booth" for the industry. While custom chips offer benefits, they represent a long-term investment. Manufacturing capacity, particularly advanced packaging from TSMC, remains a critical bottleneck, with demand far exceeding supply through 2026. This highlights that the primary beneficiaries in the custom AI chip race are often the infrastructure providers like Broadcom, who control essential design and manufacturing relationships.

Anthropic says Alibaba launched its biggest Claude distillation attack yet. Claude becomes the latest flashpoint in the US–China AI race

Anthropic reports 65% of its product team's code is AI-generated by Claude, a statistic often misinterpreted as broad automation. This impressive figure stems from their new product, Claude Tag, an AI teammate embedded in Slack that maintains context and works asynchronously. Anthropic, as the model builder, is uniquely positioned; this success doesn't generalize to most enterprises struggling with data integration. Crucially, Claude Tag's design embeds the AI within existing systems of record, bringing the agent to the data. This implies value accrues to companies owning the data, challenging the notion that AI agents will hollow out enterprise software. Instead, agents enhance incumbent platforms, making data owners more valuable. The key question is how agentic tools integrate into typical businesses and who captures that value.

A new bipartisan initiative, RAISE US, launched with $500 million from leading AI companies like OpenAI and Microsoft, aiming to retrain American workers for an AI-driven economy. Co-chaired by former Commerce Secretary Gina Raimondo and ex-Gov. Eric Holcomb, the fund includes broad political and labor support. The article argues this isn't an apology for current job losses, which are not yet widespread, but rather a strategic "down payment on political permission" for continued AI development. This coalition ensures long-term capital investments in AI infrastructure, like data centers, remain durable across administrations, mirroring the CHIPS Act's success. The fund signals AI builders' intent to secure political continuity for their massive, ongoing buildout, prioritizing infrastructure over short-term job concerns.

Google has limited Meta's use of Gemini AI models due to compute shortages, highlighting growing infrastructure constraints amid rising AI demand.

South Korea, with Samsung and SK Hynix, pledged over $576 billion to boost memory chip production, especially high-bandwidth memory (HBM) vital for AI. This decade-long, massive investment, including new fabs, signals a profound belief in sustained, real AI demand, not a bubble. The leading memory makers, facing years of sold-out HBM capacity and overwhelming demand, are addressing a critical shortage. South Korea's government views memory as strategic national infrastructure, ensuring stable, long-term capital for this essential AI resource. This commitment from those closest to the supply chain confirms the enduring nature of AI demand, driven by actual need.

AI giant Anthropic, with over $125 billion in private funding, is preparing for a potential October IPO, potentially one of the largest tech listings. Boasting a $47 billion annualized revenue rate and profitable unit economics, the company isn't seeking capital. Instead, the IPO aims to provide liquidity for early investors and employees, offer stock for acquisitions, establish a public valuation, and offer audited financials to customers. Anthropic's product, Claude, has expanded significantly, becoming core infrastructure for major clients. Its strong financial position allows it to await optimal market conditions, making the October target flexible. Investors are cautioned against rushing, as the debut price will likely already reflect significant future growth.

Summary Global military spending hit a record $2.9 trillion last year, driven by significant increases in Europe and Asia, signaling a durable rearmament cycle. A major shift in defense priorities is underway, with the Pentagon's FY2026 budget dedicating a new $13.4 billion line item to autonomy and autonomous systems, including drones and unmanned vehicles. This marks a fundamental reorganization of warfare, where value is migrating from traditional platforms to the advanced electronics, sensors, and processing power within these new systems. Companies like Mercury Systems, a defense electronics maker, are already experiencing this demand surge, reporting record bookings and a growing backlog. This indicates a new economic model for defense, focusing on consumable, adaptable weapons and their embedded technology. Investors are advised to look at these component suppliers, as their order books reflect future growth years before revenue.

A security firm, Sysdig, identified JADEPUFFER, the first ransomware operation run entirely by an autonomous AI agent. This AI exploited a known vulnerability, navigated the network, encrypted data, and adapted its approach, all at machine speed, completing complex steps in mere seconds. The attack's significance lies not in its complexity, but in the AI's ability to execute it end-to-end, drastically lowering the cost and skill barrier for cybercrime. This shift means attacks can now occur at unprecedented scale and speed, overwhelming traditional human-led defenses. Consequently, cybersecurity budgets are moving towards autonomous, AI-driven defense systems, favoring companies like CrowdStrike and Palo Alto Networks. The era of machine-speed cyber warfare has begun, demanding equally rapid defensive capabilities.

OpenAI's GPT-5.6 and Grok 4.5 launched simultaneously, creating a public leaderboard battle. However, a more significant development was Databricks' benchmark, using real enterprise code, which revealed open-source models like China's GLM 5.2 can match expensive frontier models on everyday coding tasks at two-thirds the cost. This signals a major shift: enterprise AI spending is now driven by private tests on real work, not public rankings, moving pricing power to platforms. The true cost of AI is measured by completed results, not just token prices. Chinese AI models are rapidly gaining market share due to affordability, prompting political scrutiny. As models become interchangeable, platforms controlling testing, routing, and compliance, like Databricks and Microsoft, gain significant influence.