No. 1 · July 12, 2026 · Sunday
inklede.
The lede of your week.
An estimated 15 minute read.
Disable autoplay and infinite scroll or risk massive fines, EU tells Meta
The European Commission has preliminarily found that Meta's autoplay, infinite scroll, and personalized recommendation features on Facebook and Instagram are addictive, failing to adequately assess risks to users' mental and physical wellbeing, including minors. The EC recommends Meta disable these features by default and adapt its recommender system to be less engagement-driven. Meta disputes the findings, citing Teen Accounts that cap daily screen time at 15 minutes, but the EC says parental controls only work if adults have technical expertise, undermining their effectiveness. Meta faces fines up to 6% of global annual turnover under the Digital Services Act if it doesn't comply, with a final decision expected in coming months. Separately, Meta faces a US trial beginning in August where 29 states may seek up to $1.4 trillion in penalties over child safety claims.
Reporting: Ars Technica
Meta announces its first data center in Canada — estimated $9 billion 1GW Alberta mega facility sees the AI expansion cross the border
Meta will build its first Canadian data center, a $9 billion, 1GW facility near Edmonton in Sturgeon County, Alberta. Construction takes two to three years and will support around 3,000 construction jobs and 300 permanent operations jobs once running. Meta is spending CAD $60 million on local roads and water infrastructure and says it will match the facility's electricity use with 100% clean and renewable energy, covering the costs itself. Alberta was chosen for its infrastructure access, robust grid, talent pool, and lenient regulatory environment. Meta is working with local energy providers including Greenlight Limited Partnership, Altalink, and Capital Power. This is Meta's 33rd data center campus globally, part of a buildout the company expects will cost $125-145 billion in 2026 alone.
Why this matters
It signals Meta's accelerating, capital-intensive AI infrastructure race is now crossing into Canada, with implications for regional grids, jobs, and cloud capacity competition against Amazon, Google, and Microsoft.
Reporting: TechRadar
Lawsuit: Man used Grok to make 7K sex images of stepdaughter, then shot himself
An amended class action lawsuit against X and xAI alleges a stepfather used Grok to generate roughly 7,000 sexually explicit images of his 11-year-old stepdaughter, then took his own life after being arrested. The complaint alleges xAI's safety system only flagged a 'gang rape' prompt to NCMEC, omitted the AI-generated images from its report, withheld the user's IP address, and obstructed the police investigation for weeks. A second victim, Jane Doe 5, was allegedly targeted by a family friend using similar tools. The suit adds Stability AI as a defendant, alleging its open-weight models underpin many 'nudify' apps, citing a study finding Stable Diffusion models account for 42.7% of such images. NCMEC reported over 1.5 million CyberTipline reports tied to generative AI last year.
Why this matters
The case tests whether AI companies can be held liable for failing to cooperate with law enforcement on CSAM, with legal and regulatory exposure that could reshape content moderation obligations across the industry.
Reporting: Ars Technica
Microsoft is selling off four Xbox studios as part of significant gaming cuts
Microsoft is cutting roughly 3,200 Xbox jobs over the next fiscal year, with about 1,600 eliminated immediately, and is spinning off four studios. Double Fine and Compulsion Games return to independence under founders Tim Schafer and Guillaume Provost, while Ninja Theory and Undead Labs are being sold to new owners with funding secured to finish Senua's next game and State of Decay 3. Arkane Studios, developing the delayed and over-budget Blade, is under review in France. Xbox CEO Asha Sharma's internal memo says the division lost 64 cents for every dollar invested in a typical year and operates at margins 3 to 10 times lower than comparable businesses. Mojang and King will now report directly to Sharma, and management layers will be cut to a maximum of five.
Why this matters
The scale of Xbox's restructuring, including selling off acquired studios and admitting steep losses, signals a fundamental strategic retreat from Microsoft's decade-long gaming expansion that investors and industry rivals will be watching closely.
Reporting: The Verge
AI
Netflix AI Team Cuts Wide-Partition Read Latency from Seconds to Milliseconds by Splitting Cassandra Partitions Per ID
Netflix's engineering team published a detailed writeup on how it fixed wide-partition read latency in Apache Cassandra 4.x within its TimeSeries Abstraction platform, which handles petabytes of temporal event data. The team built dynamic repartitioning, an asynchronous pipeline that splits oversized partitions per TimeSeries ID without application changes. Detection happens on the read path via byte counting, which triggers a Kafka event; splitting targets immutable partitions first using a strategy like EventBucketPartitionSplitStrategy. Bloom filters (single-digit microsecond lookups) plus cached metadata route reads to smaller child partitions, with checksums and shadow-mode comparisons guarding correctness. Results: read latency for wide partitions dropped from seconds to low double-digit milliseconds, tail latency fell to about 200ms, and partitions over 500MB stayed available for pagination instead of timing out.
Why this matters
Engineers running large-scale Cassandra deployments get a concrete, production-proven pattern for eliminating tail-latency spikes without costly cluster scaling.
Reporting: MarkTechPost
NVIDIA Releases Nemotron-Labs-3-Puzzle-75B-A9B: A Compressed Hybrid MoE LLM Delivering 2.03x Server Throughput at Matched User Throughput
NVIDIA released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed version of its Nemotron-3-Super hybrid Mamba-Transformer MoE model, shrinking it from 120.7B total/12.8B active parameters to 75.3B total/9.3B active while preserving the original's 88-block layout. Using an iterative neural architecture search process called Puzzle, plus knowledge distillation and RL recovery, the compressed model delivers 1.60x to 2.14x higher total throughput on an 8xB200 node depending on workload, and raises concurrency for 1M-token requests on a single H100 from 1 to 8, since its NVFP4 weights occupy about 44.5GB versus 70GB for the original. The tradeoff is accuracy loss concentrated in agentic and instruction-following benchmarks: Arena-Hard-V2 drops 4.2 points and SWE-Bench drops 2.6 points, while long-context benchmarks like RULER barely move. Checkpoints are available on Hugging Face in BF16, FP8, and NVFP4.
Why this matters
Teams deploying large MoE models get a concrete blueprint for cutting inference cost and boosting concurrency without abandoning the same architecture, with clearly quantified accuracy tradeoffs to weigh.
Reporting: MarkTechPost
OpenAI Releases GPT-5.6 (Sol, Terra, Luna): A Three-Tier Model Family With Programmatic Tool Calling in the Responses API
OpenAI moved GPT-5.6 to general availability with three tiers: Sol (flagship, $5/$30 per million input/output tokens), Terra ($2.50/$15), and Luna ($1/$6). Sol leads the Artificial Analysis Coding Agent Index at 80, versus Claude Fable 5's 77.2, and posts 53.6 on Agents' Last Exam, though OpenAI's own eval table separately lists Sol at 52.7%, a discrepancy the piece flags. The release adds Programmatic Tool Calling, which runs model-written JavaScript in an isolated, network-free V8 runtime, and a multi-agent
Reporting: MarkTechPost
Google Research Introduces SensorFM: A Wearable Health Foundation Model Pretrained on One Trillion Minutes of Sensor Data
Google Research introduced SensorFM, a foundation model for wearable health data pretrained on over one trillion minutes of sensor data from 5 million consented participants across 100-plus countries. The model uses a ViT-1D encoder with a masked-autoencoder objective, ingesting 34 aggregate features from PPG, accelerometer, EDA, skin temperature, and altimeter sensors over 24-hour windows. Four model sizes were tested, with the largest (110.7 million parameters, trained on 5 million subjects) cutting reconstruction loss by 31% versus the smallest and winning 33 of 35 downstream health prediction tasks, covering cardiovascular, metabolic, mental health, sleep, and demographic outcomes.
A novel masking technique called AIM lets the model treat missing sensor data as signal rather than noise, improving imputation accuracy by up to 83.7% over baselines. In blinded physician evaluation across 31 participant profiles, adding SensorFM predictions to AI-generated health summaries improved quality ratings and was statistically indistinguishable from using ground-truth data.
Why this matters
This shows how large-scale wearable sensor foundation models could reshape consumer health screening tools and create new data and compute demands for health-tech and cloud providers.
Reporting: MarkTechPost
Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks on Meta Model API
Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model for agentic tasks, alongside a public preview of the Meta Model API, marking Meta's first paid, hosted, per-token API for its own frontier model rather than open weights. The model supports a context window Meta lists at up to 1,048,576 tokens, adjustable reasoning effort, and inputs across text, images, video and documents. Developer pricing is $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new accounts; the preview is US-only. On Meta's own benchmarks, Muse Spark 1.1 leads rivals including Anthropic's Opus 4.8 and OpenAI's GPT-5.5 on tool-use tasks like MCP Atlas and JobBench, but ranks third on coding benchmarks such as SWE-Bench Pro and DeepSWE 1.1. The API is OpenAI-compatible, allowing migration via a base-URL change, and Meta highlights context compaction and subagent delegation as key orchestration features.
Why this matters
Meta's shift to a paid, closed API for its flagship model signals a strategic pivot in the AI infrastructure market that engineering and product teams evaluating agentic tool-use platforms need to track.
Reporting: MarkTechPost
China eyes export curbs on its top AI models, and Europe is caught in the middle
China is considering restrictions on foreign access to its most advanced AI models, according to Reuters sources describing talks last month between the Ministry of Commerce and companies including Alibaba, ByteDance and Z.ai. The proposed rules, still under debate, could classify theft or transfer of protected AI technology under national security law and tighten controls on funding for domestic AI startups. An expert panel in May proposed a tiered system: basic open-source tools would need only registration, advanced technologies would require security review, and the most sensitive frontier models might not be released publicly or would be restricted to domestic use.
The move mirrors US restrictions, including the June decision to bar foreign nationals from Anthropic's Fable and Mythos models. For Europe, which depends on foreign providers for more than 80% of digital products and services per a Draghi report, losing access to cheap Chinese models like Alibaba's Qwen or Z.ai's GLM-5.2 would raise costs. The EU's 200 billion euro InvestAI initiative, including 20 billion euros for AI Gigafactories, is behind schedule, with construction not starting until 2027, against a backdrop of the four largest US tech companies expected to spend roughly $700 billion combined on AI in 2026.
Why this matters
Companies and governments relying on cheap Chinese open models for cost-effective AI deployment face a real risk of losing that access just as Europe's own capacity-building lags years behind.
Reporting: The Decoder
Finance
AI puts UK startup funding back in the big leagues with strongest first half since 2022, new data shows
UK startup funding hit its strongest first half since 2022, with HSBC Innovation Banking and Dealroom counting €14.8 billion ($17 billion) raised in H1 2026, up 102% year over year and 39% of all European venture capital. Tracxn separately measured €13.3 billion ($15.3 billion), an 84% rise. AI drove the surge: UK AI startups raised a record €11 billion, over four times last year's figure, led by Isomorphic Labs (€1.8B), Nscale (€1.7B), Wayve (€1B) and Ineffable Intelligence (€960M). Deal count fell (543 to 490 rounds per Tracxn) as late-stage companies captured 68% of funding versus 42% a year earlier. London took 86% of Tracxn's tracked capital. Exits were weaker: only two IPOs versus seven in H2 2025, and acquisitions fell 20% to 167.
Why this matters
Investors and founders need to know capital is concentrating in fewer, larger AI and infrastructure bets rather than spreading broadly across the UK startup market.
Reporting: EU-Startups
The Week’s 10 Biggest Funding Rounds: A Pair Of Billion-Dollar Deals For Cyber And AI Infrastructure Lead
AI and infrastructure dominated the week's largest US venture deals, per Crunchbase News. Keyfactor raised $1 billion in private equity led by Summit Partners for its digital identity and machine identity management software. SambaNova closed a $1 billion Series F at an $11 billion post-money valuation led by General Atlantic, with backers including BlackRock, Intel Capital and T. Rowe Price, for its AI chip and infrastructure business. Quantum startup Oratomic raised a $300 million Series A co-led by Arch Venture Partners, Khosla Ventures and Spark Capital. Other notable rounds: Quaise Energy ($134M, geothermal drilling), Prime Intellect ($130M, distributed AI training), Gauntlet ($125M, DeFi risk software from SBI Group), Norm AI ($120M, AI compliance at a $1.2B valuation), Venus Aerospace ($91M, hypersonic propulsion), EDX Markets ($76M, crypto exchange) and Fore Biotherapeutics ($67.4M, oncology). Non-US highlights included Proxima Fusion's €411M round and Skello's €200M raise.
Why this matters
The scale and concentration of billion-dollar AI infrastructure and cybersecurity rounds signals where institutional capital is placing its biggest bets this cycle.
Reporting: Crunchbase News
Apple announces chip deal with Broadcom worth more than $30 billion
Apple announced a chip deal with Broadcom worth more than $30 billion, under which Broadcom will design and produce custom silicon and wireless connectivity technology for Apple products. The agreement follows a June 6 SEC filing revealing a multiyear deal running through 2031. Apple says more than 15 billion chips will be built in the US as part of its American Manufacturing Program, and Broadcom will invest $1.5 billion to modernize its Fort Collins, Colorado facility. Tim Cook called it a deepening of the companies' partnership and commitment to US manufacturing. Broadcom stock has risen more than 35% on AI-related chip demand, while Apple's stock is up 47% over the past year, aided by strong iPhone sales and an updated Siri unveiled at WWDC.
Why this matters
A $30 billion multiyear chip agreement signals concrete capital commitments in US semiconductor manufacturing and reshapes the supply relationship between two major tech companies.
Reporting: Yahoo Finance
Kraken Technology Group makes waves as Europe’s newest maritime defence unicorn with €152.9 million Series B
Kraken Technology Group, a London-based maritime defence company, has raised a €152.9 million ($175 million) Series B round at a valuation of nearly €874.1 million ($1 billion), making it Europe's newest defence unicorn. The round was led by Digital Transformation Capital Partners and backed by the British Business Bank, NATO Innovation Fund, Rheinmetall, Inocea Group, HICO, Thesiger Capital Group, BOKA Capital, Supernova Invest and Hakluyt Capital, with earlier investors NSSIF, SmartCap, Notion Capital and Speedinvest converting to equity. Founded in 2021 by CEO Mal Crease, Kraken makes uncrewed surface vessels, including the K3 SCOUT, K5 Kraken and K4 MANTA platforms, for maritime awareness, strike and anti-submarine missions. The company has secured contracts with the UK Ministry of Defence, NATO European partners and USSOCOM, and has manufacturing partnerships with Rheinmetall, Anduril Industries and Davie Shipbuilding, with plans to expand into the Middle East and Indo-Pacific.
Why this matters
A nearly €1 billion valuation for an uncrewed maritime defence startup signals accelerating institutional and NATO-aligned capital flowing into European defence tech, a sector investors are watching closely amid rising geopolitical tension.
Reporting: EU-Startups
North American Startup Funding Shattered Records In First Half Of 2026, Driven By AI
North American venture investment hit record highs in the first half of 2026, per Crunchbase data, totaling $392 billion, driven by megarounds for AI leaders. Q2 alone brought $137.2 billion, the second-highest quarterly total ever, with Anthropic's $65 billion raise (including a $50 billion May round led by Altimeter, Dragoneer, Greenoaks, and Sequoia, plus $5 billion from Amazon and $10 billion from Google) accounting for about half. Anduril raised $5 billion in a Series H led by Thrive Capital and Andreessen Horowitz. Early-stage funding hit a three-year high at $31 billion, led by Prometheus's $12 billion round. About 80% of Q2 investment went to AI startups. SpaceX's $75 billion IPO was the largest ever, and its $60 billion acquisition of Cursor was the largest startup M&A deal on record. Eli Lilly also acquired Kelonia Therapeutics for up to $7 billion.
Why this matters
These record-shattering funding and exit figures signal how thoroughly AI is reshaping capital allocation across venture, public markets, and M&A, with direct implications for valuations and competitive positioning.
Reporting: Crunchbase News
Elon Musk SEC settlement raises 'red flags,' judge says
A federal judge approved a settlement between the SEC and Elon Musk over his 2022 Twitter purchase, despite voicing serious reservations. U.S. District Judge Sparkle Sooknanan levied a $1.5 million penalty against Musk, roughly 1% of the $150 million the SEC alleged he saved by failing to disclose his growing stake in Twitter before completing the $44 billion acquisition. The settlement lets Musk deny wrongdoing. The SEC filed the original suit in 2025, shortly before Trump's inauguration, and later filed an amended complaint adding a revocable trust as a co-defendant. Sooknanan wrote she had
Why this matters
The ruling highlights how lightly the SEC penalized a major disclosure violation, raising questions for compliance officers and institutional investors about enforcement consistency under the current administration.
Reporting: Yahoo Finance