The Compliance Deadline: Brussels and Sacramento on the Same Day
The European Commission opened a new phase of AI Act enforcement on 2 August, centred on the transparency obligations of Article 50: systems that interact with people must disclose that they are not human, AI-generated or manipulated content must be marked and machine-detectable, and deepfakes must be clearly labelled [1]. Enforcement powers over general-purpose AI (GPAI) and the full penalty regime kicked in at the same moment — which means the transitional period of guidance documents and voluntary commitments is formally over [2].
The more interesting story, though, is the half that did not take effect. The high-risk obligations originally scheduled for today were pushed back substantially by the Digital Omnibus on AI signed on 8 July: stand-alone Annex III systems now have until 2 December 2027, and AI embedded in regulated products under Annex I has until 2 August 2028 [3]. In effect, Brussels chose to ship the visible transparency layer first and defer the most expensive compliance engineering by roughly eighteen months.
Across the Atlantic, California's SB 942 became operative the same day. Generative AI providers with more than one million monthly California users must now embed C2PA-compatible provenance data in images, video and audio, offer a free public detection tool, and let users attach a visible AI label — with penalties of $5,000 per day per violation [4]. Washington, by contrast, blew past its own 1 August deadline under Executive Order 14409: no notices appeared covering the classified benchmarking process, the voluntary frontier-model disclosure framework, or the federal cyber-workforce plan [4]. The centre of regulatory gravity keeps drifting from Washington toward Brussels and Sacramento.
The Open-Weights Fight: The Debate Kimi K3 Started
Almost every industry argument this week traces back to one release: Kimi K3, published by China's Moonshot AI on 16–17 July with open weights following on 26 July. It is the largest open-weight model to date — 2.8 trillion total parameters with roughly 104 billion activated per forward pass, a one-million-token context window and native vision [8][9]. On GDPval-AA v2, which scores real-world tasks across 44 occupations and nine industries, K3 placed third at 1,687, behind only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8). It reached frontier level on BrowseComp (91.2) and Terminal Bench 2.1 (88.3), and briefly took the top spot in the Arena Frontend coding leaderboard [8][9]. A freely downloadable model closing on the closed-source frontier is, by itself, a policy event.
Jensen Huang's response was unusually direct. On 24 July he used the first X post of his career to circulate an open letter, "Open Weights and American AI Leadership," arguing that open models strengthen cybersecurity, accelerate innovation and competition, broaden access to AI across industries, and underpin national sovereignty — and urging policymakers not to impose premature restrictions on downloadable models [5][6]. The letter draws an explicit parallel to the 1980s, when open-source software came close to being strangled at birth. Signatories doubled from 25 to 50 within a single day, including OpenAI and Google. Conspicuously absent: Amazon and Anthropic [7].
The top of the market is simultaneously fighting on price. Anthropic's Claude Opus 5, released 24 July, still leads Artificial Analysis on both the Intelligence Index (61) and the Agentic Index (55.3), priced at $5 per million input tokens and $25 per million output. OpenAI, meanwhile, cut GPT-5.6 Luna by 80% on 30 July, to $0.20 / $1.20, and trimmed Terra by 20% to $2 / $12 [10]. A crowded frontier, collapsing entry-level prices, and a free 2.8-trillion-parameter model sitting in the middle — that is the market structure as August opens.
Embodied Intelligence: Gemini Robotics 2 Moves the Whole Body
Google DeepMind announced Gemini Robotics 2 on 30 July, built around what it calls whole-body intelligence: rather than driving a single arm through a grasp, the robot coordinates walking, crouching, bending and two-handed manipulation, adjusting in real time when an action fails [11]. The release comes in three variants — the standard Gemini Robotics 2 vision-language-action (VLA) model, Gemini Robotics ER 2, a vision-language model for embodied reasoning and human-robot communication, and Gemini Robotics On-Device 2 for edge deployment. All three are available in early access [11].
The live demonstration ran on Apptronik's Apollo 2 humanoid, showing autonomous whole-body motion alongside real-time task reasoning [12]. Two technical claims matter most: the model can sustain task sequences lasting several minutes and involving hundreds of decisions, and it can adapt to a new robot body within hours [11]. The second is arguably the bigger deal. Embodied AI's chronic bottleneck has never been success rate on any single task — it is that every new hardware platform demands a fresh round of data collection and training. If hours-to-adapt holds up outside the lab, the marginal cost structure of robotics changes materially.
Compute and Capital: $750 Billion and the Circularity Problem
The package of deals Nvidia advanced in late July totals more than $750 billion, and it has reignited concern that AI demand is being inflated by circular financing [13]. The headline item: Nvidia is in talks to provide roughly $250 billion in financing guarantees to help OpenAI lease a 10-gigawatt data centre that SoftBank's SB Energy is building in Piketon, Ohio, slated for 2028. A separate arrangement reportedly under discussion would help underwrite OpenAI's purchase of $350 billion in Nvidia chips [13][14]. Nvidia shares fell roughly 4.5–4.9% on the news, and the market reaction identifies the problem precisely: when the chip supplier is also financing its customer, the demand signal loses credibility [13].
Supply, meanwhile, keeps expanding. Nvidia's Rubin platform reaches partners in the second half of 2026, with AWS, Google Cloud, Microsoft and OCI among the first clouds to deploy Rubin instances [15]. The supply chain's geography is shifting too: as Taiwanese server makers expand in Mexico, Mexico has become the second-largest supplier of servers to the United States, at $46.9 billion in year-to-date sales.
Cost pressure is closing in from the other side. Epoch AI estimates the major labs are currently spending roughly two to three times their revenue, and the specific tier of compute labs need — isolated enough to protect weights and customer data, large and flexible enough for good utilisation — is where prices are rising fastest [16]. Revenue curves are steep; the compute cost curve may be steeper. That is the most concrete variable hanging over AI business models in the coming year.
Security: Automated Attacks and the "Lots Found, Little Exploited" Problem
Palo Alto Networks' Unit 42 published a genuinely instructive case study. A Chinese-speaking operator tracked under the aliases knaithe and KnYuan wired DeepSeek into the open-source Hermes Agent framework; after a single initial instruction over Telegram, the agent went on to enumerate internet-facing systems, select public exploits and launch attempts against more than 460 targets, with investigators finding no further human input in that session [17]. The division of labour was clean: Hermes Agent supplied orchestration — terminal access, Telegram-based command and control, a skills system — while DeepSeek acted as the reasoning engine for code generation, vulnerability assessment, target selection and decision-making [17].
The outcome, however, is less alarming than the headline. Across roughly 460 attempted targets, Unit 42 confirmed only three successful compromises, all involving memory data exfiltration from Citrix NetScaler (CVE-2026-3055) [17][18]. The manner of discovery is more telling still: Hermes Agent accidentally started a Python HTTP server, publicly exposing the operator's model configurations, API keys, exploit scripts, target lists and shell history [18]. Automation lowered the barrier to entry without raising the hit rate — and magnified the cost of a single operational mistake.
Defenders face a mirror-image signal-to-noise problem. VulnCheck examined 1,061 vulnerabilities attributed to AI-assisted discovery in the first half of 2026 and found confirmed exploitation for just 14 of them, about 1.3%. Anthropic's Project Glasswing generated more than 23,000 findings, which yielded 126 published entries and exactly one confirmed attack [19]. When exploitation does happen, though, it happens faster: median time to exploitation fell from 120 days to 80 year-over-year, and 23% of flaws were attacked on or before disclosure day [19]. As researcher Patrick Garrity puts it, "the sheer volume of findings tells defenders very little about actual risk." A second story the same day is the natural footnote: Apple has begun imposing submission caps and cooling-off periods on Feedback Assistant, overwhelmed by a flood of AI-generated vulnerability reports [4].
Technology and Research: Dissecting Kimi K3, and Serving Diffusion LLMs
The Kimi K3 technical report (arXiv:2607.24653) is the week's most rewarding read, because it explains the engineering choices hiding behind the headline "2.8 trillion parameters" [20]. Three changes carry most of the weight. The first is Kimi Delta Attention (KDA), an efficient mixing mechanism designed for long sequences, interleaved with Gated MLA layers at a 3:1 ratio. Intuitively, KDA handles the bulk of long-range information flow cheaply, while the periodically inserted Gated MLA layers preserve full global interaction — a handful of expensive-but-exact layers compensating for many cheap-but-approximate ones. The second is Attention Residuals (AttnRes), which operate at block level and let each layer selectively draw on the representations of all preceding layers. The classic failure mode of very deep models is that information gets diluted as it is passed layer to layer; AttnRes essentially builds an express lane along the depth axis [20][21].
The third is Stable LatentMoE. K3 expands the routed expert pool to 896 while activating only 16 per token — extreme sparsity, and a notoriously unstable training regime in which routing tends to collapse onto a handful of dominant experts. Moonshot's answer combines normalisation, SiTU-GLU and Quantile Balancing to keep optimisation stable [20]. Together with refined training and data recipes, these changes deliver a claimed ~2.5× improvement in overall scaling efficiency over K2, with MXFP4/MXFP8 quantisation used throughout [20][21]. The significance is strategic as much as technical: under US compute restrictions, the Chinese team's chosen path is to convert the same FLOPs into more intelligence rather than simply stacking more accelerators.
A different problem is emerging on the serving side. As diffusion language models become practical, the existing inference stack no longer fits — their bidirectional attention rules out the exact KV caching that autoregressive models rely on, which makes mature vLLM/SGLang-style infrastructure hard to reuse directly. Sangam (arXiv:2607.04206), published 5 July, proposes a deficit token-budget scheduler that admits in-flight decodes first and lets prefill overflow onto decode workers, using the same scheduling mechanism to protect decode latency [22]. In measurements, colocated serving cut mean latency by 9–20% versus hybrid execution on decode-heavy workloads (LLaDA-8B / ShareGPT), while hybrid execution led by 8–20% on prefill-heavy workloads (Dream-7B / arXiv) [22]. The conclusion is not that one strategy wins, but that the serving strategy has to switch dynamically with the workload — precisely the stretch of road that gets underestimated when a new architecture tries to reach production.
One more paper belongs on the same reading list. "Scaling Laws Meet Model Architecture" folds architectural information into the Chinchilla framework, proposing a conditional scaling law and an architecture search procedure, validated by training more than 200 models spanning 80M to 3B parameters, with the explicit goal of optimising accuracy and inference efficiency together [23]. From K3's architecture, through Sangam's scheduling, to this work on scaling laws, the through-line of late 2026 is clear: the frontier is no longer about making models bigger, but about converting each unit of compute into more usable intelligence.
References
[1] Commission starts enforcing AI Act rules and new transparency requirements on 2 August — European Commission — https://digital-strategy.ec.europa.eu/en/news/commission-starts-enforcing-ai-act-rules-and-new-transparency-requirements-2-august
[2] EU AI Act: What Actually Applies on 2 August 2026 — Technology.org — https://www.technology.org/2026/07/17/eu-ai-act-what-actually-applies-on-2-august-2026/
[3] EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes — Gibson Dunn — https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
[4] AI News Today, August 2 — Top AI Stories & Live Updates — AI Weekly — https://aiweekly.co/ai-news-today
[5] Jensen Huang just used his first ever X post to warn the AI industry — Fortune — https://fortune.com/2026/07/24/jensen-huang-open-source-letter-nvidia-kimi/
[6] "Open Weights and American AI Leadership" — A Letter by Jensen Huang — StorageNewsletter — https://www.storagenewsletter.com/2026/07/28/open-weights-and-american-ai-leadership-a-letter-by-jensen-huang-ceo-and-cofounder-nvidia/
[7] Huang's Open Weights Letter Doubled To 50 Without Amazon And Anthropic — Forbes — https://www.forbes.com/sites/sandycarter/2026/07/25/huangs-open-weights-letter-doubled-to-50-without-amazon-and-anthropic/
[8] China's Moonshot AI releases Kimi K3, the largest open-source model ever — VentureBeat — https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems
[9] Moonshot releases 2.8-trillion-parameter Kimi K3 — Tom's Hardware — https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3
[10] AI Updates Today (August 2026) – Latest AI Model Releases — LLM-Stats — https://llm-stats.com/llm-updates
[11] Gemini Robotics 2 brings whole body intelligence to robots — Google DeepMind — https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
[12] Google DeepMind unveils Gemini Robotics 2 as Apptronik humanoid demonstrates whole-body AI — Robotics & Automation News — https://roboticsandautomationnews.com/2026/07/31/google-deepmind-unveils-gemini-robotics-2-as-apptronik-humanoid-demonstrates-whole-body-ai/103802/
[13] Nvidia's $750 Billion in Deals Reignite Circular AI Fears — Bloomberg — https://www.bloomberg.com/news/articles/2026-07-27/nvidia-s-750-billion-deals-revive-fear-of-ai-circular-financing
[14] Nvidia plans $250bn push to bolster OpenAI's infrastructure ambitions — Al Jazeera — https://www.aljazeera.com/economy/2026/7/27/nvidia-plans-250bn-push-to-bolster-openais-infrastructure-ambitions
[15] NVIDIA Kicks Off the Next Generation of AI With Rubin — NVIDIA Newsroom — https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
[16] Why compute might get 10x more expensive in coming years — Dwarkesh Patel — https://www.dwarkesh.com/p/why-compute-might-get-10x-more-expensive
[17] Chinese-Speaking Threat Actor Harnesses AI Models for Autonomous Cyberattacks — Unit 42, Palo Alto Networks — https://unit42.paloaltonetworks.com/autonomous-ai-cyber-attack-campaign/
[18] Hacker uses DeepSeek AI to autonomously attack vulnerable servers — BleepingComputer — https://www.bleepingcomputer.com/news/security/hacker-uses-deepseek-ai-to-autonomously-attack-vulnerable-servers/
[19] AI finds plenty of security flaws, but almost none of them get exploited — The Decoder — https://the-decoder.com/ai-finds-plenty-of-security-flaws-but-almost-none-of-them-get-exploited/
[20] Kimi K3: Open Frontier Intelligence — Technical Report of Kimi K3 — arXiv:2607.24653 — https://arxiv.org/pdf/2607.24653
[21] Kimi K3 Tech Blog: Open Frontier Intelligence — Moonshot AI — https://www.kimi.com/blog/kimi-k3
[22] Sangam: Efficiently Serving Diffusion LLMs with the AR Stack — arXiv:2607.04206 — https://arxiv.org/abs/2607.04206
[23] Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs — arXiv:2510.18245 — https://arxiv.org/abs/2510.18245