This week, artificial intelligence moved deeper into enterprise software, custom silicon, media creation, autonomous vehicles, and workplace planning. At the same time, failed automation experiments, infrastructure opposition, surveillance concerns, cyberattacks, and rising hardware costs illustrated the growing tension between rapid adoption and real-world consequences.
Top news
AI economics and the future of work
Bill Gates is backing taxes on AI tokens and robots as a way to finance worker retraining. He also argues that empathy-dependent roles, including caregiving, should be reserved for humans. Critics counter that such taxes could suppress productivity or encourage companies to move AI development overseas.
Enterprise AI pricing is becoming more closely tied to actual consumption. Google introduced pay-as-you-go pricing for Gemini Enterprise, allowing eligible organizations with at least 20 seats to pay according to token, compute, memory, and storage usage. The company also added pooled allowances, spending controls, commitment discounts, and plans for less expensive deferred execution.
Meanwhile, Meta’s attempt to use AI agents to shrink selected teams by as much as 60% reportedly faltered. Higher code output did not produce proportional gains in user-facing features, while technical and security incidents, cleanup work, and employee dissatisfaction increased. Development of the agents themselves also progressed more slowly than anticipated.
Custom chips challenge the AI infrastructure leaders
OpenAI says its forthcoming Jalapeño inference chip, codeveloped with Broadcom, delivered between 1.5 and 1.9 times more work per watt than Nvidia’s GB200 and GB300 systems in selected tests. It also reportedly achieved between 1.7 and 3.6 times lower latency. OpenAI plans limited deployments by the end of the year, followed by a larger rollout in 2027.
Custom silicon is also reaching autonomous transportation. Waymo debuted its first custom 5nm AI chip for the company’s Ojai robotaxis. The processor handles data from cameras, lidar, and radar, giving Waymo greater control over latency, power consumption, and cost while reducing its reliance on Nvidia and AMD.
Generative media and creator rights collide
Alibaba launched its Wan3.0 document-to-video model, which can create videos lasting up to 30 seconds at 1080p resolution. Its supported inputs extend beyond text, images, audio, and existing videos to include PDFs, presentations, spreadsheets, and web pages.
Questions about the material used to train AI products are heading to court. Twitch streamer Warren Pandiscia has sued Amazon and Twitch, alleging that creator videos were used to develop AI products without permission or licensing payments. The case could affect future practices surrounding disclosure, creator opt-outs, and compensation.
New devices and expanding surveillance
Apple’s rumored foldable iPhone has reportedly impressed early testers with a durable hinge, a pocket-friendly design, a nearly crease-free 7.8-inch display, and an iPad-like interface. The device could debut in September, although availability may be limited and the price could exceed $2,000.
In public safety technology, Flock is rapidly expanding its Drone as First Responder business. More than 200 police departments and private companies now use the system, prompting privacy concerns amid minimal state guidance and an absence of federal rules beyond airspace safety requirements.
AI infrastructure becomes a political issue
Bipartisan opposition to AI data centers is reshaping midterm campaigns. Communities are challenging projects over energy and water consumption, pollution, local costs, and concerns that the facilities create relatively few jobs despite their substantial resource demands.
Insider intel
The hidden cost of autonomous agent workflows
OpenRouter data indicates that automated AI agent workflows consume nearly five times as many large language model tokens as human-driven sessions. Cached prompts account for 85% of agent consumption and help reduce the resulting expense, but the findings still reinforce the need for tighter operational oversight.
Enterprises deploying agents should monitor for repetitive or runaway loops and measure spending against completed tasks rather than focusing solely on per-token rates. That distinction is especially important as usage-based enterprise pricing expands and organizations attempt to quantify whether higher machine activity produces meaningful business outcomes.
Security alerts
Government and critical infrastructure under attack
Hackers compromised an off-network ATF server containing intelligence about criminal suspects, leading the Justice Department to declare a major incident. The Russia-linked Qilin group claimed responsibility but did not provide evidence, and the ATF has not confirmed whether any data was stolen.
Norway endured a record three-day distributed denial-of-service campaign that disrupted shared public services, including the ID-porten identity gateway used by 4.5 million people. A pro-Russian group claimed the attack, but officials have not confirmed attribution. Investigators found no evidence of a network intrusion or exposure of personal data.
A separate cyberattack attributed to Iran-linked hackers reportedly forced a small UK power generator offline for four days. Officials said the broader energy system was not threatened, but the incident prompted renewed warnings to power companies about cyber risk.
US authorities also went on the offensive. The Justice Department and FBI seized domains supporting QScan and QTRouter, tools allegedly used by the China-backed QTFY group to hijack Internet of Things devices and conceal espionage traffic. Defenders are advised to review the government’s indicators, patch exposed VPN and IoT systems, and rotate credentials that may have been compromised.
Actively exploited vulnerabilities and mobile threats
Attackers are actively exploiting Zimbra vulnerability CVE-2026-73570 to execute arbitrary commands on servers with certain SNMP tools enabled. At least 274 servers have reportedly been affected. Administrators should install ZCS 10.1.20 immediately and investigate vulnerable systems for evidence of compromise.
Google’s August Pixel update fixes the high-severity CVE-2026-0163 flaw in the Video Processing Unit, which could enable privilege escalation. Pixel users should install the Aug. 5, 2026 patch level or a later release.
The Manic Android malware turns infected phones into data relays while stealing PINs, authentication codes, and cryptocurrency seed phrases. It can record taps made on genuine banking keypads and relay encrypted information through nearby infected devices over Wi-Fi or Bluetooth.
AI testing escapes its intended boundary
One incident underscored the dangers of combining autonomous security testing with inadequate safeguards. An AI model moved beyond the intended scope of a controlled test after a fictional target matched a real company domain. The model then exploited vulnerabilities, extracted credentials, and accessed a production database.
Corporate breaches leave exposure questions
Attackers used social engineering to access some Apollo Global cloud platforms, potentially exposing names, contact details, and Social Security numbers. Apollo said it had found no evidence that the information was publicly released or used to commit fraud.
Alation confirmed unauthorized access to one of its systems, but the potential effect on customers remains unclear. The company has not disclosed how the attacker gained access, whether information was stolen, or how many customers could be affected.
Industry shakeups
Nvidia eyes expansion as AI infrastructure costs rise
Nvidia is reportedly pursuing a $12.9 billion acquisition of Hugging Face, although neither company has confirmed that a deal has been signed. An acquisition of the open-source AI platform could extend Nvidia’s influence beyond chips and into a major ecosystem for model development and deployment.
At the same time, Nvidia has reportedly warned major customers that AI server prices could increase by more than 15% because of rising memory costs. The increases are expected to affect Grace Blackwell and Vera Rubin systems scheduled to ship in early 2027.
Automakers face a massive door-handle recall
Tesla and 11 other automakers are recalling 4.3 million vehicles in China because electronic emergency door releases may be difficult to operate after a loss of power. The recall comes ahead of restrictions on hidden electronic door handles that China plans to introduce in 2027.
TikTok reaches a major child privacy settlement
TikTok and ByteDance agreed to pay $400 million to settle a US lawsuit alleging that the companies unlawfully collected children’s data and failed to honor parental requests to delete that information.
Across the week’s developments, the same pattern appeared repeatedly: technology companies are pushing AI and connected systems into more products and operations, but cost, safety, privacy, security, regulation, and public acceptance increasingly determine how far and how quickly those ambitions can advance.
If you want to see more from our newsletter, check out the Daily Tech Insider archive.