Shanghai's WAIC and the Governance Contest
The 2026 World Artificial Intelligence Conference opened in Shanghai today, running through July 20 with more than 140 forums and over 1,100 exhibitors. President Xi Jinping attended in person for the first time since the event began in 2018 — a signal that Beijing now treats AI leadership as a top-tier national priority [1]. The conference's most consequential political move was the proposal to establish a World AI Cooperation Organization (WAICO): 29 countries, including Russia, Brazil, Pakistan, and Indonesia, signed an agreement in Shanghai, pointing toward a China-led architecture for global AI governance [1]. The hardware floor had its own drama: Huawei demonstrated its homegrown Atlas 950 SuperPoD cluster, positioned as an 'Nvidia-class' compute solution built on domestic silicon under US export restrictions, while reports emerged that China is allowing limited imports of Nvidia's H200 — a metered opening in the hardware standoff [1].
The governance chorus is not only Beijing's. Hundreds of experts warned this week that the world is unprepared for AI's economic and social impact, and the UN continues to push for a global governance framework — landing the same week as the WAICO proposal, evidence that the contest over who writes AI's rules is now fully underway [1][7].
The Open-Source Scale Race
Jumping the conference by a night, Moonshot AI released Kimi K3 — a 2.8-trillion-parameter sparse Mixture-of-Experts model billed as the largest open-weight model ever, roughly 75% bigger than DeepSeek V4 Pro's ~1.6 trillion parameters, with native visual understanding, a 1-million-token context window, and an always-on reasoning mode. Full weights are promised for July 27 [2][4]. The scorecard is formidable: on GDPval-AA v2, K3 scored 1,687 — third overall, behind only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8) — and on AA-Briefcase it took second place at 1,527, beating GPT-5.6 Sol Max [2]. API pricing is $3 per million input tokens and $15 per million output tokens, with OpenAI SDK compatibility that squarely targets developers already on Western toolchains [1][2].
The open-source camp had two more entries today: Thinking Machines, founded by former OpenAI CTO Mira Murati, released Inkling, an open-weight model available for local download and fine-tuning — a heavyweight name joining the open-source movement [1] — while DeepSeek pushes toward its V4 stable release at roughly $0.44 per million output tokens, with a July 24 migration deadline for preview users, keeping price pressure on closed models [1].
The Gemini 3.5 Pro Cliffhanger
The day's intended headliner, Gemini 3.5 Pro, instead became its mystery. Multiple reports indicate the rebuilt flagship has now missed its third launch deadline — the reconstructed model failed key reliability standards, including frequent hallucinations, and fell short of GPT-5.6 on some benchmarks. Google is reportedly considering shipping a stopgap, Gemini 3.6 Flash, in the meantime [5][6]. Notably, the widely circulated specs — a 2-million-token context window, Deep Think reasoning, $1.25/$10 pricing — all come from third-party leaks; as of publication, no gemini-3.5-pro listing appears in Google's official API documentation [5]. The scripted drama of a launch colliding with WAIC has, for now, become the awkwardness of a Western flagship no-show.
The rest of Google's ecosystem moved on schedule: NotebookLM was rebranded as Gemini Notebook, gaining a secure cloud environment for code execution, with more than 30 million users across 600,000 organizations [1]; and Android will open to competing app stores starting July 22, ending the Epic Games dispute and opening new distribution channels for AI apps [1]. Among rivals, OpenAI announced it is shutting down its Atlas browser and consolidating features into the ChatGPT 'super app' — a bet that the chat app, not the browser, is the primary interface to the internet [1] — while Microsoft was reported to have instructed its enterprise sales teams to position Copilot as superior to OpenAI's and Anthropic's products, going head-to-head with its own largest investment [1].
Chips and Capital
TSMC reported second-quarter results: $22 billion in profit, up 77% year over year, on $40.2 billion in revenue, up 34%. It raised 2026 capital spending to $60–64 billion and added $100 billion for its Arizona operations, bringing its total US commitment to $265 billion — a single earnings report that captures both the intensity of AI chip demand and the geography of supply-chain hedging [1].
Capital-market highlights clustered at the application layer: Neko Health, the preventive-health company co-founded by Spotify's Daniel Ek, closed a $700 million Series C at a ~$7 billion post-money valuation, using AI to analyze non-invasive full-body scans, with a waiting list of over 350,000 people and a US expansion starting in New York [1][8]; India's Emergent, a natural-language app-building ('vibe-coding') platform, raised a $130 million Series C at a $1.5 billion valuation, minting a new unicorn [1]; and Israeli startup Oak stepped out of stealth with $60 million in seed funding to fix the identity-management mess that AI agents are making worse [9]. In aggregate, Asian venture funding hit a multiyear high, showing AI capital formation is no longer a US monopoly [1][10], and consolidation in security is accelerating — 29 AI-security acquisitions in the first half of 2026, nearly triple 2025's full-year total of 10, as enterprises deploy agent fleets faster than their security can keep up [1].
Technology and Research
Today's technical spotlight belongs to Kimi K3's architecture — practically a public textbook on how to make a trillion-scale model both big and fast [3]. The first key is Kimi Delta Attention (KDA): a hybrid linear attention mechanism that delivers up to 6.3x faster decoding in million-token contexts. Standard attention's compute grows quadratically with sequence length; linear attention flattens that curve, and KDA's hybrid design captures the speed dividend while preserving retrieval accuracy. The second is Attention Residuals (AttnRes), an improvement along the model's depth dimension: instead of accumulating representations uniformly across layers, it selectively retrieves them across depth, buying roughly 25% higher training efficiency at under 2% additional cost [3].
The MoE configuration is equally aggressive: 896 experts in total with only 16 active at inference, using Quantile Balancing to derive expert allocation directly from router scores and eliminate hand-tuned hyperparameters. On the training side, quantization-aware training starts at the supervised fine-tuning stage, using MXFP4 weights with MXFP8 activations for broad hardware compatibility — with overall scaling efficiency claimed at 2.5x that of the previous-generation K2 [3]. Combined with $0.30-per-million-token cache-hit pricing and 90%+ cache-hit rates in coding workloads, K3 demonstrates not just a scale record but an entire engineering package that makes ultra-large models economically viable [3]. One more item worth logging: Microsoft's Patch Tuesday this month patched a record 570 vulnerabilities, with AI systems credited for finding and triaging a significant share — the industrialization of AI on the defensive side is already here [1].
References
[1] AI News Today July 17 2026 — Build Fast with AI — https://www.buildfastwithai.com/blogs/ai-news-today-july-17-2026
[2] 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
[3] Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context — MarkTechPost — https://www.marktechpost.com/2026/07/16/moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open-moe-model-with-kimi-delta-attention-and-1m-context/
[4] China's Moonshot AI unveils Kimi K3 model it says rivals OpenAI, Anthropic — CNBC — https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html
[5] Rebuilt Gemini 3.5 Pro Misses Third Deadline: Google Eyes Stopgap Release — Tech Times — https://www.techtimes.com/articles/320736/20260716/rebuilt-gemini-35-pro-misses-third-deadline-google-eyes-stopgap-release.htm
[6] Stopgap Gemini 3.6 Flash May Launch During Gemini 3.5 Pro Delay — Geeky Gadgets — https://www.geeky-gadgets.com/gemini-3-5-pro-delayed-again/
[7] Hundreds of experts warn the world must prepare now for AI's impact — Al Jazeera — https://www.aljazeera.com/economy/2026/7/13/hundreds-of-experts-warn-the-world-must-prepare-now-for-ais-impact
[8] Latest AI Startup Funding News and VC Investment Deals — Crescendo.ai — https://www.crescendo.ai/news/latest-vc-investment-deals-in-ai-startups
[9] Backed by $60M in funding, Oak steps out of stealth — TechCrunch — https://techcrunch.com/2026/07/15/backed-by-60m-in-funding-oak-steps-out-of-stealth-to-fix-the-identity-mess-that-ai-agents-are-making-worse/
[10] The AI Startup Funding Boom Is Not A Global Phenomenon — Crunchbase News — https://news.crunchbase.com/venture/us-ai-startup-funding-boom-data/
[11] Chinese company releases world's largest open-source AI model — People's Daily Online — http://en.people.cn/n3/2026/0717/c90000-20478901.html