More than a hundred companies — spanning frontier AI labs, cybersecurity vendors, and financial institutions — signed an open letter published August 27 urging governments and industry to jointly prepare for a coming wave of AI-driven cyberattacks. Signatories include OpenAI, Anthropic, Google, and Microsoft on the AI side, alongside CrowdStrike, Okta, and Fortinet among the security firms.
The letter argues that AI-enabled attacks will grow far more widespread and sophisticated in the coming months as frontier models keep improving, and it calls on governments at the local, national, and international level to coordinate new forms of cyber defense rather than treat the problem as any single company's to solve.
The timing is notable: it lands the same week reports surfaced of coding models independently discovering critical software vulnerabilities, underscoring that the industry itself now sees offense and defense accelerating together, not in sequence.
Data shared with Fortune shows Google DeepMind's share of elite research and engineering hires in Europe falling sharply this year, as OpenAI, Anthropic, and a new wave of well-funded startups pull ahead in the race for talent. Current and former staff point to aggressive poaching, frustration over Google's standing in the AI race, and the appeal of pre-IPO equity at rivals.
The departures piled up fast this month: chief scientist Jeff Dean, senior fellow Sanjay Ghemawat, and researchers Oriol Vinyals and Quoc Le all left within weeks of each other to launch a new venture called Discovery Loop, following Demis Hassabis's own move from CEO to chairman on August 8.
The exits land as Google tries to close a performance gap with OpenAI and Anthropic under new DeepMind operational chief Koray Kavukcuoglu, who now has to rebuild momentum with a thinner senior bench than the one that shipped Gemini 3.
A SemiAnalysis deep dive published this week details Jalapeño, OpenAI's first custom inference chip, taped out with Broadcom on TSMC's N3P process in only sixteen months. The B0 stepping delivers 13.4 PFLOPs of MXFP4 compute at 700 watts, well under the 900–1,150 watts Nvidia's Rubin chips draw, and pairs it with HBM4 memory running at 15.4 terabytes per second.
On real workloads, OpenAI's own benchmarks put Jalapeño at 1.5 to 1.9 times more work per watt than Nvidia across GPT-OSS, DeepSeek R1, and Kimi K2.5, with throughput north of 1,400 tokens per second per user on GPT-OSS according to The Verge's separate reporting.
The chip doesn't outrun Rubin on raw performance, but it doesn't need to: for a company burning through power budgets at every data center it builds, efficiency per watt is now as strategically important as peak speed, and it's the clearest sign yet that OpenAI intends to own more of its own hardware stack rather than rent all of it from Nvidia.
Greenko Group's Hyderabad-based AM Intelligence has placed a binding order for roughly 9,000 Nvidia Vera Rubin NVL72 rack-scale systems, with delivery set for the first quarter of 2027. The order positions the company as one of Asia's first operators of a frontier-scale Vera Rubin cluster.
AM Intelligence plans about $8 billion in capital spending to bring 200 megawatts online in the near term, scaling toward a full gigawatt of compute-as-a-service capacity spread across India, the United States, Finland, and Malaysia. Founder Mahesh Kolli said a US customer has already reserved initial capacity.
The Hyderabad facility is engineered to deliver around 450 exaFLOPS of NVFP4 inference compute, backed by Greenko's low-cost renewable power — a combination that could make South India a genuinely competitive location for AI infrastructure rather than just a market for it.
A Pew Research survey of 3,488 US adults finds that 34% now use AI chatbots for at least one health-related task, led by looking up quick health information, trying to understand what's causing a symptom, and finding low-cost information they might not otherwise access.
Roughly 47% of users call the answers extremely or very helpful, but only 29% say they're very comfortable sharing personal health data with these tools — a gap between usefulness and trust that hasn't closed even as adoption climbs. Asian American adults (56%) and adults under 30 (44%) show the highest uptake.
A companion Pew report released the same day adds a sharper edge: Americans are more likely to say chatbots hurt than help people who turn to them for loneliness, depression, or stress, suggesting the comfort people feel with AI for facts doesn't extend to AI for feelings.