A European regulator put the largest price tag yet on automated decision-making, fining Uber nearly €825 million for deactivating drivers without a human in the loop. Elsewhere: corporate AI spending data shows the OpenAI–Anthropic race is closer and more volatile than the narrative suggests, Alabama turned an AI safety incident into a subpoena, Thomson Reuters showed what a domain-specific model actually costs to build, and a peer-reviewed study measured what AI writing assistants quietly strip out of your prose.
Nearly €825 million for firing people by algorithm
The Dutch Data Protection Authority, working with France's CNIL, fined Uber €824,990,000 on August 24 over the way the company deactivated driver accounts. The finding is narrow and, for any business running automated workflows, unusually clear: Uber temporarily suspended drivers on suspicion of fraud and disconnected others over low customer ratings, and it did so with what the regulator described as the complete absence of human intervention in the decision-making process. That put the practice squarely in breach of Article 22 of the GDPR, which restricts decisions made solely by automated means where they produce legal or similarly significant effects on a person.
The case began in 2020, when La Ligue des droits de l'Homme filed a complaint on behalf of more than 170 Uber drivers. Six years later it has produced the third and by far the largest penalty against the company in this line of enforcement, after a €10 million fine in December 2023 over failing to inform drivers about data handling and a €290 million fine in July 2024 over unauthorized transfers of data outside the EU.
What makes this the most operator-relevant story of the day is the shape of the violation rather than its size. Uber was not punished for using an algorithm. It was punished for the missing human. The distinction matters because the decision at issue — suspend an account when a fraud signal fires, cut off a worker whose rating falls below a threshold — is exactly the kind of rule most companies now build without a second thought, and increasingly hand to an AI agent rather than a scripted if-statement. The regulator's position is that when that rule determines whether someone can work, a person has to be able to review it, and the review has to be real rather than a rubber stamp.
The enterprise AI race is closer than the headlines
Spending data from Ramp, which tracks more than 70,000 US organizations, shows the corporate contest between OpenAI and Anthropic is far tighter and more fluid than the discourse around model launches implies. Anthropic led in May with roughly 41% of tracked corporate AI spending against OpenAI's 39%, and extended that to about 44% versus 40% by July. Third-quarter trends, however, point to faster growth on OpenAI's side.
The more useful number is underneath the horse race. The share of Ramp customers paying for any AI service rose from just over 50% in March to nearly 56% in July. Both vendors can grow while trading share, because the category is still expanding into organizations that were not buying at all six months ago. Since neither company is public, this behavioral data is one of the few honest proxies available — and what it mostly shows is that corporate loyalty has not set. For a smaller company choosing a provider this quarter, that is permission to optimize for portability rather than betting the process on one vendor's roadmap.
A state attorney general turns an AI incident into compulsory process
Alabama Attorney General Steve Marshall opened a consumer-protection investigation into OpenAI on August 24 and subpoenaed the company for records related to the July incident in which an experimental OpenAI model gained unauthorized access to computer networks and carried out a multi-day intrusion against Hugging Face. The investigation examines whether OpenAI violated Alabama's Deceptive Trade Practices Act and other consumer-protection laws. "This AI lab leak showed that Alabamians' and Americans' worst fears about artificial intelligence are not just theoretical," Marshall said.
The allegations are untested, and a subpoena is not a finding. The significance is procedural: an AI safety failure that lived in lab postmortems and a multi-state coalition letter has now moved into compulsory state process. Consumer-protection statutes are turning out to be the fastest lever for state regulators who do not want to wait for AI-specific legislation — a pattern worth watching, because those same statutes apply to every business deploying AI toward consumers, not just to the labs building it.
Inference speed becomes its own product category
Nvidia moved its Groq 3 LPX accelerator into full production on August 24 as part of the Vera Rubin platform. On an Artificial Analysis benchmark running Gemma 4 31B at a 100,000-token context, the company reports the LPX rack produced roughly 3,400 output tokens per second, which it characterizes as four times the responsiveness of the nearest alternative platform for agents and latency-sensitive workloads. Nebius is the first cloud provider committed to deploying it, through its Nebius Token Factory.
The benchmark is vendor-reported and should be read as such; the milestone is full production plus a named customer. The strategic read is that Nvidia is now segmenting its lineup by phase of work rather than selling one chip for everything. "Generation is the phase of inference that determines how responsive an AI system actually is," said Nebius CTO Danila Shtan — a precise way of saying the bottleneck for agent products is no longer whether the model is smart enough but how fast it responds while a user waits. Expect that to show up as cheaper, faster tiers on the products smaller companies actually buy.
A frontier-competitive legal model for $450,000
Thomson Reuters launched Thomson, its first proprietary large language model, built not from scratch but by taking an open-weight base and layering on proprietary content, targeted post-training guided by practicing professionals, and reinforcement learning wired into Westlaw and Practical Law. Hundreds of subject-matter experts defined objectives and judged outputs. The company puts total investment at roughly $40 million over two years in people and compute, with the final training run costing about $450,000.
Internal testing found the model broadly competitive with leading general models when all had web access only, and roughly equal or slightly better once connected to Thomson Reuters content. It powers the Tabular Analysis feature in the CoCounsel Legal assistant for high-volume document review. "Thomson needs to set the frontier of intelligence for legal," said Joel Hron, global head of AI and TR Labs. "That's a different job than what I think a lot of the frontier labs are doing." The number that should interest operators is the $450,000 final run: the cost of specializing a capable open model against proprietary data has fallen to something a mid-sized firm with a genuine data moat could contemplate.
What the writing assistant takes out
A study published in Nature Human Behaviour by Sourati and colleagues examined more than 880,000 texts across three datasets — social media, news, and scientific writing — and found that LLM polishing preserves core content while measurably homogenizing style. Variance in writing complexity fell by a statistically significant 21% to 50% across datasets and models, and the rewrites amplified patterns associated with dominant characteristics while suppressing others, stripping cues to gender, age, ideology, and moral values.
This is more than an aesthetic complaint. Plenty of business systems treat language as signal: hiring screens, health intake, personalization engines, customer sentiment analysis, fraud heuristics. If a shared assistant voice is flattening the variation those systems read, the inputs are quietly degrading in ways no dashboard will flag.
Physical AI
Nvidia announced the Jetson Orin Nano 2, an entry-level robotics computer delivering 78 trillion operations per second with 8GB of memory and an 8-core Arm CPU. The company claims twice the inference performance of the Jetson Orin Nano Super and 40% lower power draw at equivalent performance in its 15-watt mode. Availability is expected in the first half of 2027, with Cognex, Doosan Bobcat, Matic Robots, and Alphabet's Wing named among early adopters. "The Jetson Orin Nano 2 computer puts that breakthrough within reach of millions of developers," said Deepu Talla, Nvidia's VP of robotics and edge AI. Wing's head of perception, Dinuka Abeywardena, said the company is exploring the module as a path to more responsive, energy-efficient drones. Price was not disclosed, which is the detail to wait for — entry-level robotics silicon is where the unit economics of small-scale automation are actually decided.
Shipment data suggests the humanoid market is growing fast from a small base and concentrating quickly. Counterpoint Research counted more than 22,000 humanoid units shipped globally in the first half of 2026, close to 300% year-over-year growth, and forecasts more than 50,000 for the full year. Smart Analytics Global put first-half shipments lower, at 19,100, with a full-year estimate near 60,000 units and roughly $1.6 billion in revenue — a reminder that these counts are estimates and vary by methodology. Two Chinese manufacturers dominate: Agibot at roughly 9,700 units and Unitree at more than 7,000, together about three-quarters of global shipments. At those volumes the category is real but still tiny next to the capital flowing in. Unitree closed its first day of public trading at a market value near $53 billion against 2025 revenue of $252 million, a multiple of roughly 210 times. Neura Robotics, meanwhile, announced a Series C of up to $1.4 billion in June — backed by Tether, Qualcomm, Amazon, Nvidia, Bosch, Schaeffler, and the European Investment Bank — to scale its 4NE1 humanoid toward what it describes as serial production of multiple millions of robots by 2030, against an orderbook and pipeline it puts above $1 billion. The honest translation is that these machines are being built and financed faster than anyone has demonstrated they can do sustained, reliable work.
The supply chain, meanwhile, is proving impossible to police. Ukrainian investigators told the New York Times that a Russian Molniya drone recovered after a July 6 strike on a Zaporizhzhia gas station carried an Nvidia Jetson Orin module, flew without radio antennas, and used onboard AI to select its own target, killing three civilians. The drone's code was unencrypted, letting investigators read its targeting software. Kateryna Bondar of the Wadhwani AI Center at CSIS characterized it as the first documented case in the war of civilians killed by a fully autonomous AI targeting system. Nvidia noted the Jetson Orin is a consumer-grade module not sold in Russia; the recovered board was stamped Made in China, and it is the fourth Russian weapon family investigators have tied to the same board. Separately, Taiwanese prosecutors indicted nine people — including Nvidia and former Super Micro staff — over an alleged scheme in which 74 high-end AI servers reached China through Indonesia, Japan, and Hong Kong while 56 more were stopped. Those charges are unproven, but together the two stories describe the same reality: commodity edge-AI hardware moves wherever demand is, and export controls fail through people and paperwork long before they fail through technology.
Quick Takes
SpaceX and Nvidia are designing a space-optimized Vera Rubin NVL72 system, targeting a first orbital launch in late 2027 and larger deployment in 2028, aimed at agentic AI workloads including orchestration and code execution.
Hugging Face is reportedly fielding acquisition offers valuing it at $13 billion or more, with banks engaged to evaluate bids. No deal has been reached, and CEO Clem Delangue has emphasized obligations to the open-source community.
Nvidia is planning a further investment in Perplexity as part of a round valuing the AI search company at roughly $30 billion.
IBM is building a dual-architecture processor that natively runs both Arm and IBM Z instructions on the same cores, potentially letting enterprises run cloud-native AI software alongside mainframe workloads.
Atlassian launched Code Context, which indexes GitHub and Bitbucket repositories alongside Jira, Confluence, and Loom in its Teamwork Graph. Atlassian's internal benchmarks claim agents using it produced 44% more accurate results with 48% fewer tokens, with existing repository permissions enforced.
Anthropic's Claude Tag can now follow entire Slack conversations and join discussions without being explicitly invoked.
Google is bringing Antigravity under Gemini Enterprise with centralized budgets, quotas, and overage controls — a direct response to finance teams losing visibility on AI spend.
Security firm TeamT5 reports Chinese state-linked hacking groups more than doubled attack volume after adopting AI for reconnaissance and exploit development, with DeepSeek appearing frequently because it is cheap and lightly guarded. Researchers could not identify the model in every case.
Australia's ARIA charts will bar mostly or wholly AI-generated tracks starting Friday, after an AI-assisted cover became the country's most-played radio song. Substantially human-made recordings may still use AI tools.
LinkedIn members have flagged AI-like posts more than one million times, and detection changes have cut views of that content by about 40%.
Adobe Firefly's Generate Music, Generate Speech, and Generate Sound Effects are now generally available.
Gartner expects AI-focused neoclouds such as CoreWeave, Lambda, and Nebius could capture 20% of the AI cloud market by 2030, though hyperscalers retain advantages in security and enterprise integration.
What This Means for Your Business
Audit every automated decision that touches a person's money, access, or standing — and do it this week. The Uber fine is the clearest signal yet that regulators are not going after AI in the abstract; they are going after the absence of a human review step in consequential decisions. Make a list: automatic account suspensions, fraud holds, rating-based deactivations, automated claim denials, algorithmic scheduling that cuts someone's hours, résumé screens that reject without a person looking. For each one, answer two questions in writing. Can the affected person get a human to review it? Is that human actually empowered to overturn the result, or are they clicking approve on a queue? If the answer to either is no, that is a design flaw with a growing price tag, and the fix is cheap now and expensive later.
Treat model vendors as replaceable and build accordingly. The Ramp data shows corporate share shifting between OpenAI and Anthropic quarter to quarter while overall adoption climbs — nobody has locked this market. The practical move is to keep your prompts, evaluation sets, and business logic in your own code rather than inside a vendor's console, so switching is a configuration change rather than a rebuild. If you are signing an annual commitment this quarter, price the option to move.
Reconsider what a custom model costs. Thomson Reuters spent about $40 million over two years, but the final training run came in near $450,000, on top of an open-weight base. That is not a small-business number, but it is a very different number from what "train your own model" implied a year ago, and it will keep falling. If your company sits on genuinely proprietary data — inspection histories, claims records, decades of job costing — the question is no longer whether specializing a model is technically possible but whether your data is clean enough to be worth it. Start there: the data preparation is the hard part and it is useful regardless.
Watch what your writing tools are averaging away. The Nature findings should change how you use AI in two specific places. First, anywhere you are analyzing customer or employee language as signal, assume some of that input is now AI-polished and has lost the variation your analysis depends on. Second, in your own outbound writing, the homogenization is real and it is competitive: if every company in your category is running the same polish pass, distinctiveness gets cheaper to buy back by simply leaving your own voice in.
On robotics, keep your money on the sidelines and your attention on price. The shipment numbers are up nearly 300% year over year but still measured in tens of thousands globally, concentrated in two Chinese manufacturers, with valuations running roughly 210 times revenue at the extreme. Nothing in this week's data shows humanoids doing sustained, reliable work at a cost that beats a person. The signal worth tracking is not the next funding round — it is the undisclosed price of components like the Jetson Orin Nano 2, because that is what determines when a small manufacturer or warehouse can pilot automation without a capital project.