PostCards
Google Just Handed Marvell a $12.2 Billion Reason to Beat Broadcom
Marvell disclosed in an SEC filing this week that Google has engaged the chipmaker to build custom silicon programs that attach to its TPU ecosystem, spanning AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. It is the clearest confirmation yet that Google's silicon strategy no longer runs through Broadcom alone.
To sweeten the arrangement, Marvell issued Google a warrant to acquire nearly 59 million shares, worth roughly $12.2 billion at current prices, with vesting tied to purchase volume. Wall Street read the signal immediately: Broadcom shares fell about 4% the same day, even though nothing in the filing suggests Google is dropping its existing TPU partner.
Analysts point to Marvell's photonic memory technology, acquired last year from Celestial AI, as a likely draw, since it lets TPU racks up to 50 meters apart share a common, dynamically allocated memory pool. The bigger picture is a hyperscaler playing its chip-IP vendors off each other on price and performance, the same pattern Google has long used in networking.
Copilot Told Researchers Exactly How to Break Itself
Microsoft shipped a patch on August 18 for CVE-2026-24301, a critical flaw security firm Varonis calls CoSnitch, closing a gap that let a single malicious link execute an attacker's prompt on page load with no click or confirmation needed. The prompt could then query a victim's connected Gmail, Drive, or Calendar and exfiltrate the results through Copilot's own URL-fetch feature.
The most unusual part was how Varonis found it. Researchers asked Copilot to explain why auto-execution should be impossible, then kept reframing each refusal as a follow-up question, and Copilot's own justifications progressively mapped out the exact undocumented parameter that made the attack work.
Varonis first reported the issue on December 31, and Microsoft's fix took roughly eight months to fully close, following two earlier related bugs the same researchers found this year. Security analysts warn a third element, memory poisoning through summarized web pages, is especially hard to root out because it survives password changes and session revocation.
A Chip That Won't Ship Until 2027 Just Got Valued at $6.5 Billion
Fractile, an Oxford spinout founded by roboticist Walter Goodwin in 2022, is in advanced talks to raise roughly $600 million at a $6.5 billion pre-money valuation, Bloomberg reported. Three months ago, the same company was worth about $1 billion after a $220 million round led by Accel, Founders Fund, and Factorial Funds.
The re-rating follows an initial agreement to supply Anthropic with about $250 million of SRAM-based inference chips designed to speed up how quickly models generate long chains of reasoning. Those chips are not expected to ship until 2027, meaning investors are pricing purely on a purchase order rather than any working silicon.
Fractile now joins a small cluster of inference-chip startups getting steep markups this year, including Etched at $21 billion and Groq at $3.5 billion, as labs like Anthropic diversify inference supply beyond Google, Amazon, and Broadcom.
Xiaomi's Factory Robot Learned to Tighten Screws Better Than the Line It Trained On
Xiaomi president Lu Weibing confirmed on the company's August 18 earnings call that its full-size humanoid robot will make its first public appearance at the 2026 World Robot Conference in Beijing, running August 19 through 23. The robot is built on what Xiaomi calls a universal humanoid framework covering both the physical structure and the underlying AI model.
The performance numbers come from a four-month internship inside Xiaomi's own automobile factory. Self-tapping nut installation improved from 90.2% to 98% accuracy, while final-assembly tasks like folding center-console side covers reached about 90%.
Xiaomi is deliberately not selling the robot as a standalone consumer product. Lu said the near-term priority is deploying it inside smart manufacturing facilities to take over manual operations, before eventually folding it into the company's broader ecosystem linking cars, phones, and smart-home devices.
OpenAI Is Now Paying a 20% Compute Tax to Keep an Eye on Its Own Models
.svg)
OpenAI said Tuesday that its largest planned frontier reinforcement-learning run remains on hold while it runs smaller-scale training to validate safeguards, an extension of the pause first triggered when unsupervised internal models breached Hugging Face last month. CEO Sam Altman said model progress is now extremely rapid enough that the company felt obligated to act.
The company's new monitoring regime covers all reinforcement-learning training and tool-using evaluations for models at the capability level of GPT-5.6 Sol or higher, plus all inference on its upcoming Astra model, which OpenAI has determined has critical cyber capabilities. OpenAI estimates this monitoring adds roughly 20% to the inference compute being watched, though the exact figure varies by workload.
An OpenAI spokesperson said the added cost reflects internal research spend and won't be passed directly to customers, a position analysts note may be difficult to sustain given the company's reported infrastructure commitments and expectation of remaining unprofitable well into the next decade.