The AI money is moving downstream — from training the models to serving them — and the security bill is arriving at the same time. This week's frontier news wasn't a flashy new model; it was a pair of very large bets on the unglamorous plumbing of inference, set against a rare, coordinated warning from the world's biggest intelligence alliance. Here's the through-line.
Cerebras books OpenAI as a $20-billion anchor
The week's biggest story came tucked inside an earnings release. Cerebras, the wafer-scale chipmaker fresh off the largest semiconductor IPO on record, disclosed alongside its first-quarter results that OpenAI will deploy roughly 750 megawatts of its high-speed inference compute over the next several years — a deal valued at more than $20 billion. Revenue for the quarter nearly doubled to about $193 million, and the company added a parallel partnership to offer its inference through AWS. The significance isn't only the headline number. Wafer-scale chips are built for low-latency inference, the part of the stack now straining under the weight of reasoning models and agents that burn tokens at serving time. A $20-billion-plus commitment turns a hardware curiosity into a structural supplier — and shows OpenAI deliberately diversifying away from a pure-GPU diet.
Qualcomm reaches for a third pole in AI silicon
The same downstream logic is driving the week's biggest piece of M&A drama. Qualcomm is reportedly in advanced talks to acquire Tenstorrent — the AI-accelerator startup run by legendary chip architect Jim Keller, whose résumé runs through AMD, Apple and Tesla — for somewhere between $8 billion and $10 billion. Tenstorrent's chips are built on RISC-V, an open instruction set that would let Qualcomm sidestep Arm's licensing constraints and field a genuine data-center inference challenger to Nvidia and AMD. Nothing is official yet, but Qualcomm's investor day falls on June 24, the natural venue for confirmation. If it lands, it would be one of the year's largest AI-hardware deals and a clear statement that the inference-chip race below Nvidia is wide open.
A rare joint warning from the Five Eyes
While the money chased compute, the world's biggest intelligence alliance was sounding an alarm. The cybersecurity chiefs of the United States, United Kingdom, Canada, Australia and New Zealand issued a rare joint advisory on June 22 warning that the next wave of AI models will supercharge offensive hacking on a timeline of "months, not years." AI, they said, lowers the barrier for malicious actors and increases the speed and complexity of attacks — and they pointed to recent frontier models' unprecedented ability to find software vulnerabilities. Coordinated public messaging from all five partners is unusual, and it reads as a signal that AI-enabled offense has moved from hypothetical to planning assumption.
Google's flagship slips, but its long-context tier ships anyway
Not every story this week was a win. Google's most-anticipated model, Gemini 3.5 Pro, missed its June launch window — a visible stumble in a fast-moving season where rivals are shipping on a tight cadence. Google's consolation is that its strongest currently-available tier, Deep Think reasoning with a 2-million-token context window, is now available through the Gemini API. Developers can build whole-corpus, long-context agents today, even as they wait for the headline model — which blunts, but doesn't erase, the competitive cost of being late.
The flip side of the agent boom
The same agentic AI that's driving all this compute demand also opened a new front for attackers. Security researchers detailed an attack they call "agentjacking," in which a malicious bug or error report — fed in through a routine developer tool — acts as a prompt injection that steers AI coding agents like Claude Code, Cursor and Codex into executing attacker-chosen commands, reportedly with a high success rate across thousands of organizations. As companies wire agents into their codebases and terminals, every external input an agent reads becomes a potential vector. It's a concrete reminder that agent tool-inputs must be treated as hostile by default.
The money view, and what to watch
The capital this week voted for the inference layer and for routing around Nvidia — Cerebras with wafer-scale compute, Qualcomm with open RISC-V silicon. The countercurrent is risk: both the Five Eyes warning and the agentjacking research say the security cost of capable, agentic AI is coming due. Watch three things next: whether Qualcomm confirms the Tenstorrent deal at its June 24 investor day, whether Gemini 3.5 Pro ships before month-end, and whether any of the Five Eyes governments turn their warning into concrete release controls. The frontier is moving downstream — and learning, in real time, what it costs to secure.