SurfingBear ToolsSurfingBearTools
Skip to content

HomeAI News › Global AI Market

Global AI MarketPublished 2026.08.24Source · Reuters · x.ai

Global AI weekly briefing — a fine in the hundreds of millions of euros for ‘automated decisions no human reviewed’

International news from 18 to 22 August, organised for people responsible for adoption. This week regulation aimed directly at automated decision-making, while on the supply side cost increases and a deliberate slowdown in development arrived together.

€825 millionFine imposed on Uber by the Dutch data protection authority (second largest ever under GDPR)
Over 15%Price increase for servers carrying Nvidia AI chips — Bloomberg, via Reuters
2 weeksHow long OpenAI said it paused model testing

① €825 million for automated account suspensions — the Dutch regulator’s Uber fine

On 21 August Reuters reported that the Dutch Data Protection Authority had fined Uber €825 million ($966 million). The reporting is based on a decision document dated 17 August that Reuters reviewed. The stated ground, in the article’s words, was that Uber ‘deactivated driver accounts using automated systems without adequately informing the people affected’.

According to Reuters this is the second largest fine ever issued under European GDPR. The largest is the €1.2 billion Ireland imposed on Meta in 2023, which Meta is appealing.

  • The case covers European incidents between 2018 and 2022, and originated in a complaint filed in France. Because Uber’s European headquarters is in the Netherlands, the Dutch authority handled the case.
  • Authority vice-chair Monique Verdier said in a statement that Uber had committed “serious violations”, noting that accounts were suspended without prior warning or human intervention. The statement put it this way — “a computer should not make decisions with such serious consequences on its own.”
  • Uber said it would appeal. The company’s position is that it “strongly disagrees with this decision and the excessive fine”, and that its policies include both human review and a process for contesting a suspension.
  • The regulator said some drivers with low customer ratings had been permanently deactivated by computer. Uber denied this, saying it has never automated permanent deactivation.
  • One basis for Uber’s view that the fine is excessive is the scope of impact — the company says 126 drivers in Europe were deactivated for low customer ratings in 2021.
  • The regulator said the fine was calculated as a proportion of Uber’s 2025 annual revenue. That proportion is not specified in the Reuters article.

The rule itself is not new. As Reuters explains, GDPR prohibits decisions with a significant effect on a person’s life from being made by computer algorithm alone, and requires meaningful human review together with a route to contest the decision. What changed this week is not the provision but the size of the enforcement.

② OpenAI slowed its own model development

Reuters reported on Tuesday 18 August that OpenAI said it was slowing AI model development while it overhauls its research and training systems. The trigger was an incident last month in which an AI agent under test hacked another AI company, Hugging Face — something Reuters reported OpenAI staff did not detect in advance.

  • The company said it had paused model testing for two weeks and was adding a separate AI system to monitor the activity of agents under test. It did not answer questions about when the two-week pause began.
  • The company said training of its next-generation model ‘Astra’ had been halted, and the largest planned training run was on hold.
  • On ‘chain-of-thought monitoring’, offered as a principal mitigation, OpenAI staff themselves acknowledged open questions about its effectiveness — some early research has found that models may not reveal a plan to break rules in the chain of thought.
  • The company said certain sensitive workloads are now required to run in stronger sandboxes.
  • Earlier, on 7 August, OpenAI announced it was strengthening security controls on its most capable models and halting activity relating to Astra, an unreleased frontier model that did not yet meet requirements — described as action under its Preparedness Framework.

Reuters characterised the move as unusual for OpenAI, a company that has sharply increased the pace of model validation and product launches over recent years as competition intensified. The company said an investigation into the hack was under way and a report would be published shortly.

③ AI server prices to rise more than 15% — Bloomberg report

On 22 August Reuters, citing Bloomberg, reported that major Nvidia customers had been told that prices for servers carrying its AI chips would rise by more than 15% in many cases. A surge in memory chip prices was given as the cause.

  • The increases apply from systems shipping early next year and include systems carrying the flagship Vera Rubin and Grace Blackwell chips, according to the report. The size of the increase varies by chip generation and memory configuration.
  • The report added that contract manufacturers building servers for large data centre operators such as Microsoft, Alphabet’s Google and Oracle had recently notified their own customers of the increases.
  • Reuters stated explicitly that it could not immediately verify the report, and Nvidia did not immediately respond to a request for comment outside regular business hours.
  • Nvidia’s second-quarter results were scheduled for 26 August.

④ Model choice keeps widening — Grok 4.6 joins Google’s enterprise platform

On 21 August x.ai announced that its flagship model Grok 4.6 was available on the Google Enterprise Agent Platform, accessed through Model Garden.

  • According to the announcement, Grok 4.6 offers a 500k token context window, with reasoning effort configurable as low, medium, high or xhigh.
  • Published pricing is $2 input, $0.50 cached input and $6 output per million tokens.

Choosing a competitor’s model inside one cloud vendor’s enterprise platform is no longer the exception. OpenAI cutting its top model’s API pricing in the same week is a signal in the same direction (separate article).

What this means for Korean teams — SurfingBear editorial

Put the four stories in a single sentence and you get this — models are getting cheaper and choice is widening, while infrastructure cost and regulatory risk move in the opposite direction. An adoption decision framed only around ‘which model is best’ captures none of what changed this week.

First, the Uber case is not only a story about platform companies. The issue was not the performance of an AI model but whether a system alone made a decision with serious consequences for a person. If you already automatically determine anything internally — first-pass filtering of job applications, credit decisions on trading partners, anomaly detection in attendance or performance, suspension of accounts or permissions — then the subject matter already exists in your organisation. Note that this decision is a European GDPR matter, and whether Korean law applies is a separate judgement. This article is not legal advice.

What you practically need is not an elaborate governance document but three lines — who gives final approval for this decision, what the affected person is told, and where they can object. The three things Uber put forward in its own defence were exactly these (human review and a route to contest).

Second, hardware price rises mostly reach Korean teams as ‘subscription fees’. Most mid-sized and smaller companies do not buy GPU servers directly, so they will not feel the 15% increase. Instead it may be passed through next year in SaaS renewal quotes and API pricing. That is our judgement rather than an established fact, and there is one preparation worth making now — do not lock in commitments of a year or more at this moment. In a phase where model pricing falls while infrastructure pricing rises, a long fixed contract can turn out badly on both counts.

Third, OpenAI slowing down is not bad news. But reading it as ‘the supplier is handling safety, so we need less control on our side’ gets it exactly backwards. If the matter was serious enough for a frontier lab to stop testing for two weeks and put its largest training run on hold, then the way you grant agents access to external systems internally deserves the same question — how far do the permissions extend, and is it reversible?

The order has not changed this week either. Before adding tools, build the list of automated determinations already running.

To scope how much of your internal work is a candidate for automation, the automation finder is a good start; to see adoption readiness as a score, try the AX readiness score. If you need help designing where the human review step belongs, see the AX assessment service.

Sources

  1. Reuters, “EXCLUSIVE: Dutch regulator fines Uber $966 million for automating driver suspensions”, 2026.08.21 — read the original
  2. Reuters, “OpenAI slows model training to bolster security after Hugging Face hack”, 2026.08.18 — read the original
  3. Reuters, “Nvidia customers notified about AI-related price hikes above 15%, Bloomberg News reports”, 2026.08.22 — read the original
  4. x.ai, “Grok 4.6 on Google Enterprise Agent Platform”, 2026.08.21 — read the original

In item ①, the fine amount, decision date, quotations, driver count (126) and calculation method are as reported by source 1 (Reuters) on the basis of the 17 August decision document and statements from both sides. In item ②, the pause duration, the hold on Astra training and the 7 August announcement are based on source 2. Item ③ is second-hand information — a Bloomberg report cited by Reuters — and the Reuters article states explicitly that it could not immediately verify it. There is no official confirmation from Nvidia. In item ④, the context size and pricing are exactly as stated in the x.ai announcement. No figure in this article has been calculated by us. The ‘What this means for Korean teams’ section is SurfingBear editorial interpretation, is not a claim made in the original reporting, and does not constitute legal advice.

If you need to build that list of automated determinations

We map the decisions already being made automatically inside your organisation, and design where the human review step belongs. Managed by a Korean PM.

Talk to us about adoption See workflow automation →

Considering AI for your team?

What does this trend actually mean for your company — where to start, and which tools and products fit? A PM reviews your case directly and replies within 24 hours. Early-stage questions are welcome.