Google locks in $12.2B Marvell chip warrant as AI silicon race heats up
A pair of deals and a security teardown tell the same story from different angles: the AI infrastructure stack is getting deeper, more expensive, and more contested.
Google and Marvell Technology have expanded their custom silicon partnership in a deal that hands Google the right to buy up to 58.97 million Marvell shares at $206.58 each — a package worth up to $12.2 billion at full exercise. The warrant, reported by Bloomberg and confirmed by Reuters, signals that Google is locking in deep custom chip capacity as the AI infrastructure arms race intensifies. Marvell designs the custom ASICs that hyperscalers use alongside — and increasingly instead of — off-the-shelf Nvidia GPUs, and this deal effectively gives Google a major equity stake in its chip partner's future. The structure is notable: rather than a straightforward purchase, Google gets a warrant, meaning it can exercise the option if Marvell's custom AI chip business continues to grow. For Marvell, the deal validates its position as one of the few companies capable of building the specialized silicon that powers frontier AI training and inference at scale. For Google, it's a bet that custom silicon — not just buying more H100s — is the path to controlling both cost and capability in the AI compute stack.
A criminal AI tool called Kriminal, marketed as "the AI that answers everything" with no filters or guardrails, is mostly just SpaceXAI's Grok running behind a jailbreak prompt, ThreatDown researchers found. The security arm of Malwarebytes reverse-engineered Kriminal's front-end code and discovered that the service charges $12.99 to $99 per month on the clearnet — no dark web required — and routes all chat through Grok with a system prompt that explicitly strips safety controls. What makes Kriminal different from a generic jailbreak is the packaging: four named agent personas (PHANTOM for money laundering, ARCHICTECT for exploit research, ORACLE for intelligence analysis, WRAITH for social engineering) turn a repackaged commercial model into a menu of criminal tradecraft. The service also uses OpenRouter to access Claude and Mistral for specialized tasks, and accepts crypto payments with no identity verification. The finding illustrates a growing pattern in AI-enabled crime: you don't need to train your own model when you can rent a commercial one and wrap it in a prompt that tells it to ignore its own rules.
What to watch: Whether xAI moves to shut down the accounts behind Kriminal, and whether the warrant structure signals Marvell's continued rise as a custom AI chip powerhouse alongside Broadcom.
Should AI companies be required to detect and block jailbreak wrappers that repackage their models for criminal use? Tell us in the comments.
Sources: Bloomberg via Techmeme · Reuters · SiliconANGLE · ThreatDown · Quartz
