August 28th 2026
Curated AI news and stories.
A federal judge ruled the Pentagon's Anthropic blacklist illegal
Judge Rita Lin of the Northern District of California ruled Thursday night that the Pentagon acted unlawfully when it designated Anthropic a risk in February, finding the move violated the First Amendment as retaliation for protected speech and denied the company due process under the Fifth Amendment. The designation came from President Trump and Defense Secretary Pete Hegseth after CEO Dario Amodei refused unrestricted military use of Anthropic's models, objecting to applications in mass surveillance and autonomous weapons. Lin's 59-page order says the government's actions "were based on a desire to make a public example out of Anthropic for its 'arrogance' in criticizing the government," and that neither the Constitution nor the statute invoked allows penalties imposed "principally on Anthropic's critique." Lin had temporarily blocked the designation and a related executive order in July; this is the final ruling. The government is expected to appeal, and a parallel Anthropic case over a second supply-chain rule is pending before the D.C. Circuit. The editorial takes this one up. SourcesBB
Marvell guided to $18 billion and dropped 8% on the Google fine print
Marvell reported $2.739 billion of July-quarter revenue Thursday, up 37%, with adjusted earnings of 94 cents a share, and raised its full-year outlooks for the second straight quarter: roughly $12 billion for fiscal 2027, up from $11.5 billion, and roughly $18 billion for fiscal 2028, up from $16.5 billion. Data-center revenue grew 46% and is now 79% of the company. The number that moved the stock was the one management would not pin down: the expanded custom-silicon deal with Google, which includes a that could make Google one of Marvell's largest shareholders and which the company says could generate up to $120 billion of revenue through fiscal 2033. CEO Matt Murphy said the deal contributes "much more significantly in fiscal year 2029," and shares fell 8% on the timing. Morgan Stanley's read: expectations were higher precisely because of Google, and the contribution was already largely in the prior guide. A $120 billion headline attached to a fiscal-2029 ramp is a claim about the shape of the custom-chip era, and the market just discounted it at 8%. SourcesBB
Ransomware criminals used Cursor to breach seven companies
Russian-speaking hackers used Cursor, the AI coding assistant SpaceX bought for $60 billion this month, to help break into a Belgian chemical and hygiene-products maker and at least six other firms this year, according to data reviewed by Reuters. Researchers at Gambit Security found the campaign after a new gang calling itself Aur0ra left a server exposed to the internet. The victims map the actual economy rather than the strategic one: Ghent-based Christeyns, German garage-door manufacturer Teckentrup, the Scotland-based Helideck Certification Agency, an Argentine pharmaceutical distributor, an Italian manufacturer, and Bayou Title, which calls itself Louisiana's largest title insurer. Tuesday it was Chinese state groups doubling their tempo with DeepSeek; Thursday it was 116 companies signing a letter about a closing window; this morning it is commodity criminals running the same playbook with an American tool. The mid-market firms on this victim list were never going to buy frontier-lab security advice, which is the population the letter's signatories keep writing about and not to. SourcesBB
Warsh speaks at Jackson Hole this morning, with the AI trade listening
Federal Reserve Chair Kevin Warsh delivers his first Jackson Hole keynote at 10am Eastern, the only scheduled chance before September's meeting to learn how the new chair thinks rather than what the committee decided. He speaks with inflation at 3.4%, the 30-year Treasury near its highest levels since 2007, and his own committee divided in public: two regional presidents spent the week pushing back on rate-cut hopes into a record-setting equity rally. Futures held their breath this morning, the S&P 500 flat and Nasdaq futures down 0.3% with Marvell sinking, after a Thursday session that closed the S&P at a record 7,730.99 on Nvidia's guidance. The 10-year sits at 4.67%. The AI complex has particular exposure: the buildout is financed at the long end of a curve the Treasury has already intervened in this month, and a chair who signals higher-for-longer is repricing every data-center financing signed this summer. SourcesBB
GLM-5.3's open weights are due today, with a countdown running
The repository for GLM-5.3, Z.ai's flagship coding model, now shows a public countdown to release, an expected date of August 28th, and 749 subscribers waiting for the notification. The page's tagline says what the wait was for: "Frontier Coding with Emergent Cyber Capabilities." Z.ai released the model through its coding service on August 14th and promised within two weeks, tying the delay to safety evaluation of its unusually strong performance on vulnerability-finding tasks, and it shipped the smaller GLM-5.3-Flash under MIT yesterday as a warm-up. If the countdown holds, the strongest open-weights coding model lands with a cyber-capability caveat its own maker chose to print on the label, days after OpenAI's breach report described what an unconstrained coding model did to a production network. SourcesAB
India ordered 9,000 Vera Rubin racks, one of Asia's first clusters
AM Intelligence, the Hyderabad AI-infrastructure company affiliated with renewable-power producer Greenko Group, placed a binding order for about 9,000 Nvidia Vera Rubin rack-scale systems, one of the largest disclosed orders for Nvidia's newest platform and among the first Vera Rubin clusters anywhere in Asia. The order anchors a plan to spend more than $8 billion on AI compute toward a gigawatt of capacity, with first systems delivering by the first quarter of 2027 into a 30-megawatt initial phase at Hyderabad and customers planned across India, the US, Finland and Malaysia. The company's edge is the parent's cheap renewable power, which is the scarce input everywhere else on this page. A week after Nvidia's guide assumed zero China data-center revenue, the order shows where the substitute demand is being found: sovereign-adjacent buyers with power, in markets American export policy still permits. SourcesBB
Apple shipped the first 2-nanometer chip while nobody was watching
Apple introduced the M6 on Tuesday, the first consumer chip built on a process, alongside the M5 Ultra, its most powerful silicon ever, with both debuting in a new Mac mini and Mac Studio on September 22nd. The M6 carries a 12-core CPU Apple calls the fastest single-threaded processor in the world, a 12-core with a neural accelerator in every core, and a dual 16-core Neural Engine, with up to 32GB of at 170GB/s. The M5 Ultra is Apple's first quad-die package, two dual-die M5 Max chips fused together, and Apple claims 4.5 times the AI compute of the M3 Ultra. The strategic read: while the industry fights over HBM allocation and data-center power, Apple keeps pushing serious inference capacity onto the desk, on the most advanced process node in production, in machines that run 60-billion-parameter models locally. The AI buildout has two ends, and the quiet one just moved first to 2nm. SourcesAB
Skild's robot model learns a ten-minute job from one video
Skild AI released S1 this week, a robot that learns a new task from a single human demonstration video, with no fine-tuning and no task-specific training. On multi-step tasks absent from its training data, potting a plant, cooking pancakes, making pour-over coffee, assembling a kit, S1 hit 66% success against 9% for an equivalent policy prompted with language, and Skild estimates one video prompt does the work of roughly 380 task-specific demonstrations. That is arriving in robotics: the task is defined by showing, not telling, the way text models started learning from examples in a prompt in 2020. Deployment starts with industrial partners including ABB Robotics, Universal Robots and MiR. The claim to test is generalization breadth, four demonstration tasks is a demo reel, and the 66% number is Skild's own. But if task acquisition really has dropped from hundreds of demonstrations to one, the economics of deploying a robot on a new line just changed by that ratio. SourcesAC
Google will cap Android apps' memory, because AI ate the RAM
Google Play will impose memory-use thresholds on Android apps starting February 2027, requiring apps to meet limits on dynamic memory and bitmap usage and to shrink and optimize code, a response to a global RAM shortage the company traces to AI data centers absorbing memory-fab output. Memory makers have shifted capacity to for AI accelerators; the consumer-grade that phones and laptops use is in short supply; analysts expect phone bill-of-materials costs to rise 15% or more, with entry-level devices losing RAM outright. So the platform is rationing: the operating system that runs on three billion devices is telling its developers to use less memory because the memory went somewhere else. Amazon repriced its device shelf 60% on Monday and blamed memory. Nvidia gave up a point of to it Wednesday. Now the squeeze reaches the app store's terms of service, which is the version ordinary consumers will actually feel. SourcesBB
Meta removed an Iranian network using AI to pose as Americans
Meta disrupted an Iran-based influence operation that used AI to generate content targeting US audiences, removing four Facebook accounts and 31 Instagram accounts that posed as everyday Americans, Axios reported from the company's latest threat report. The accounts attracted thousands of followers with AI-made memes carrying anti-Republican messaging and content on the Israel-Palestine conflict and immigration. The scale is small; the direction is the story. Wednesday OpenAI banned a Russian operation running a fake Israeli think tank, and the pattern across the takedowns is convergence on the same workflow: generative tools for the content, small account networks for the seeding, authenticity as the product being counterfeited. The midterm cycle is fourteen months out, and the platforms are already publishing this genre of report on a monthly cadence. SourcesBB
OpenAI hired Meta's India and Southeast Asia chief
Sandhya Devanathan, Meta's vice president for India and Southeast Asia, is leaving after more than a decade to join OpenAI, where she will be based in Singapore, report to Asia-Pacific managing director Kiran Mani, and oversee consumer growth, enterprise adoption, partnerships and regulatory engagement across Southeast Asia and Australia. The hire lands days after Prabhjeet Singh joined OpenAI as its India head. OpenAI is assembling a senior operating bench for the markets where ChatGPT's next half-billion users live, and it is staffing it from Meta, the company that spent two decades building exactly that distribution. Meta, meanwhile, faces growing scrutiny in India over online safety and content moderation, and just lost the executive who managed that relationship. SourcesB
Barret Zoph landed at Google, his fourth lab in two years
Barret Zoph, who co-founded Thinking Machines Lab with Mira Murati and served as its CTO, is now a vice president of research at Google, TechCrunch reported. The itinerary: OpenAI through 2024, Thinking Machines co-founder, a January departure back to OpenAI that was later revealed to be a firing from the startup, five months at OpenAI running enterprise-sales AI, out in June, and now a return to Google, where he was a research scientist from 2016 to 2022. One of the field's best-regarded researchers has now cycled through a $50 billion startup and two frontier labs since October 2024. The market for elite AI researchers clears at nine figures and the tenure keeps shortening; Google just bought experience with everything Thinking Machines and OpenAI tried this year, at the price of a VP title. SourcesB
Intel's answer to the AI-rack CPU slipped to 2027, and grew
At Hot Chips, Intel detailed Xeon 7, Diamond Rapids, and conceded its timing: the processor that was supposed to ship this year with 192 cores now arrives in 2027 with up to 256 performance cores. The package stitches 22 chiplets, sixteen compute dies on Intel's 18A-P node bonded to base and fabric dies with Foveros, carrying 1.28GB of last-level cache, PCIe 6.0, doubled memory bandwidth and support in its matrix extensions for on-CPU . The engineering is real and the calendar is the problem: Nvidia's Vera, its first ground-up CPU, began production shipments last quarter inside racks Nvidia sells whole, and every Vera socket is a socket Intel's server business never sees. Diamond Rapids will be the most capable x86 chip ever shipped into a market that is being repackaged, rack by rack, around a competitor's CPU. SourcesBB
A third of American adults now ask chatbots about their health
34% of US adults have used an AI chatbot for at least one health-related task, Pew Research found in a survey of 3,488 adults fielded in June: 28% for quick health information, 25% to understand what is causing symptoms, 22% for low-cost information in place of a visit. Adoption runs highest among Asian Americans, at 56%, and adults under 30, at 44%. The satisfaction data carries the warning: 47% call the answers extremely or very helpful, while only 29% are very comfortable sharing personal health data with the tools, and a companion Pew report finds Americans more likely to say chatbots hurt than help people using them for loneliness, depression or stress. A third of the adult population is now getting medical guidance from products whose makers put "not a substitute for professional advice" in the footer, and the usage arrived before the evidence, the liability framework, or the business model for being wrong. SourcesAB
Google's search agent now tracks fares and books the hotel
Google moved AI Mode from answering travel questions to transacting on them: flight-price tracking is live in more than 180 countries, hotel booking is rolling out in the US with Booking.com, Expedia, Marriott and Priceline completing the checkout through Google Pay, and loyalty-program pricing, what a flight costs in miles rather than dollars, is available globally. The user describes the trip; the agent watches fares, assembles options and hands off a finished booking. Each piece is incremental and the sum is not: the default search surface for a billion travelers now runs the full loop from intent to purchase, with the partners reduced to fulfillment. The online travel agencies signed up anyway, which tells you what they think happens to partners who refuse. SourcesAB
Hugging Face will sell you a $399 robot duck
Hugging Face and Pollen Robotics opened pre-orders for Microduck, a 25-centimeter open-source bipedal robot at $399, aimed at developers who want to experiment with on real hardware. It packs 15 motors, a wide-angle camera, LiDAR, microphones, two IMUs and Wi-Fi, ships with seven pre-trained behaviors from walking to roller skating, uses its articulated beak as a gripper, and publishes its SDK, MuJoCo simulation environment and RL training stack under , so every shipped behavior can be inspected and retrained. Deliveries target the US and UK before the end of 2026. The price is the announcement: hobbyist robotics with a full modern learning stack now costs less than a games console, from the company Nvidia is reportedly paying $12.9 billion to own. The developer funnel for physical AI just got a mascot. SourcesBB
Seoul gave back the Nvidia pop
The Kospi fell 1.16% Friday to 6,831.99, surrendering Thursday's Nvidia-driven gains and more, as foreign money took profits in the chip complex and the won weakened to 1,380 to the dollar. Samsung Electronics and SK Hynix both fell steeply, two days after the Bank of Korea's second straight rate hike put the local cost of capital on the other side of the trade. The index that touched 6,996 at Thursday's open ends the week below 6,900: the memory suppliers got their demand confirmation from Nvidia's own margin, and the market spent one session celebrating before returning to the question of what a central bank hiking into a boom does to a rally built on it. SourcesB
SemiAnalysis checked OpenAI's chip math and mostly agreed
SemiAnalysis published its teardown of Jalapeño, the inference chip OpenAI announced with Broadcom on Tuesday, and the headline is the endorsement: a hyperscaler-designed can now match or beat Nvidia's Blackwell-class GPUs on inference efficiency. The details fill in OpenAI's own numbers: a reticle-sized die in roughly sixteen months on TSMC's N3P, delivering 13.4 petaflops at 700 watts, with the design optimized against the prefill and communication bottlenecks that dominate serving costs. OpenAI's models reportedly assisted in the chip's design loop. Independent validation is the thing first-party cannot buy, and Nvidia's inference business now has a competitor designed by its largest customer, validated by the analysts the market actually reads. The test that remains is production traffic, which no benchmark table settles. SourcesBB
Moonshot's $50 billion close has not been confirmed
The final close of Moonshot AI's pre-IPO round, which the company told investors was scheduled for yesterday at a $50 billion , had produced no confirmation on English-language wires by this morning. The company has denied filing its Hong Kong listing application in August and holds to its own September 30th target, with a listing by early 2027. Nothing about a mega-round closing quietly is unusual in China's private markets, and nothing about this one is routine: the round prices a company negotiating a first-of-its-kind revenue share with three American while under a threatened Treasury blacklist. Until a close is confirmed, $50 billion remains what Moonshot told its investors, not what its investors have paid. SourcesBB
Nature's verdict on the Robot Games: fast feet, useless hands
Nature published its assessment of the World Humanoid Robot Games, and the pattern it draws from Beijing is the one that matters for the industry: broke records in athletics, Tiangong Ultra ran 8.64 seconds in the 100-meter final and two robots went sub-9 in the semis, while struggling at practical events like hammering nails. Locomotion, the problem that defined the field for decades, is yielding to reinforcement learning at pace; manipulation, the thing factories would actually pay for, is not. The medal table tracked the industry's pecking order closely enough to trade on, and the games' real result is a map of where the capability frontier sits: a robot that outruns Usain Bolt cannot yet reliably drive a nail. Skild's one-video learning result this week is aimed at exactly that gap. SourcesB