The daily what-happened-yesterday-in-AI brief from The Excelsior Group · 2026-07-17

The Read

Yesterday Beijing's Moonshot AI released Kimi K3 — 2.8 trillion parameters, the largest open-weight model ever shipped — and early benchmarks put it within reach of the best US proprietary systems: third on the Artificial Analysis Intelligence Index, per The Information. This morning Xi Jinping stood on the World AI Conference stage in Shanghai — his first appearance since the event began in 2018 — and declared 'open source and open collaboration' China's model for global AI. The West's scheduled answer didn't show: Google's rebuilt Gemini 3.5 Pro missed its third deadline this week, with a stopgap release now reportedly under consideration. Open weights are no longer a hobbyist movement or even just a business model — they are national strategy: give the model away, sell the world the standards, the compute, and the market structure around it. The day-zero read: the price of frontier-adjacent intelligence fell again yesterday, and the advantage keeps shifting to whoever points it at re-founded work fastest.

🌊 Tide

No shift. One strong confirmation of the governance tide, which is now visibly multipolar: Xi proposed a world AI cooperation body from the WAIC stage the same week the CEOs of OpenAI, Google DeepMind, and Anthropic converged in writing on frontier-model oversight. The other three tides hold; Kimi K3's price-performance is fresh evidence for the cost-collapse tide.

Governance goes multipolar: Xi pitches a world AI body as US lab CEOs converge on oversight

This morning Xi Jinping delivered the opening keynote of Shanghai's World AI Conference — his first appearance at the event since it launched in 2018 — telling delegates AI 'should not be a solo performance by a single country, but a symphony of international cooperation,' calling for 'open source and open collaboration' in global AI development, and pitching China as the AI partner of the developing world. On the eve of the conference, roughly 29 countries were reported to have signed the founding agreement of a World AI Cooperation Organization that Beijing wants headquartered in Shanghai (reported; details still emerging). The same week on the US side, Axios documented something without precedent: the CEOs of OpenAI, Google DeepMind, and Anthropic have each published detailed regulatory positions in the past five weeks that converge on outside scrutiny of frontier models before release and certification-style oversight bodies — while splitting on venue, with Anthropic pushing progressively tougher state-level rules and OpenAI pushing a single national framework. Rule-writing is no longer a Western monopoly and no longer resisted by the labs; the fight has moved to who hosts the rulebook.

So what: Two governance stacks are forming to match the two technology stacks. If you operate internationally, regulatory strategy is now bloc-level: compliance with Brussels and Washington no longer implies access to markets that align with Shanghai's rulebook, and vice versa. Give the regulatory-operations owner you assigned this week a map of which bloc's standards each of your markets follows.

CNBC on Xi's WAIC keynote (July 17) · Axios: AI godfathers converge on regulations (July 16) · The Information on the keynote (July 17) · SCMP on Xi's first WAIC appearance

🌊 Waves

Kimi K3: the largest open-weight model ever puts China in the frontier conversation

Moonshot AI released Kimi K3 yesterday: a 2.8-trillion-parameter mixture-of-experts model (16 of 896 experts active) with a 1-million-token context window, built for long-horizon coding and agent workloads — the largest open-source model ever shipped. It launched in two variants (K3 Max and K3 Swarm Max) on Kimi Code and the Kimi app, with API pricing of $0.30 per million cached input tokens, $3 uncached, and $15 output; Moonshot says full weights go public July 27. The Information reports K3 ranks third on the Artificial Analysis Intelligence Index and first on an Arena leaderboard — approaching but not matching the best US proprietary models — and makes the sharper point: US labs have accused Chinese labs of advancing by distilling Claude, but if K3 genuinely rivals top Anthropic models, distillation no longer explains the pace. Landing the day before Xi's open-source keynote, the message writes itself: open weights are Chinese statecraft now. Standard caveat: launch-week vendor benchmarks deserve independent verification.

So what: Queue K3 for evaluation on long-horizon agentic coding when the weights land July 27 — at these prices it resets the floor for that workload. And re-read Wednesday's provenance note with fresh urgency: Chinese open models in your stack are a supply-chain dependency with export-control exposure in both directions, and that stack is getting more capable, not less.

SiliconANGLE on the release (July 16) · The Information: K3 challenges US frontier models (July 17) · MarkTechPost on the architecture (July 16) · VentureBeat on the release

Sovereign physical AI: Japan enlists NVIDIA's stack — and Korea's workers push back

In Tokyo yesterday, Jensen Huang and Japan's economy minister Ryosei Akazawa launched a government-backed Physical AI Initiative, with Fujitsu, Hitachi, Kawasaki Heavy Industries and robotics leaders Fanuc and Yaskawa forming a coalition to build robots and industrial AI on NVIDIA's stack — Huang called it 'the beginning of Japanese AI,' a day after shipping Cosmos 3 Edge, the 4B-parameter world model covered in Wednesday's brief. The same week the other side of the ledger arrived: Hyundai auto workers launched a partial strike naming AI deployment and humanoid robots on the factory floor among their grievances — one of the first major industrial actions to do so explicitly, per Build Fast with AI's July 16 roundup — weeks after South Korea pledged to grow robots to 20% of its market by 2028. And the capital keeps coming: The Information reported yesterday that UK-based robot maker Humanoid reached unicorn status. The physical-AI wave now has states subsidizing it, capital chasing it, and organized labor formally negotiating against it. That last one is new — and it will set the deployment template.

So what: Physical AI is becoming national-industrial policy, which means subsidies, procurement preferences, and labor politics will shape deployment economics as much as capability does. If you run manufacturing or logistics operations, treat workforce agreements as a first-class workstream in any automation pilot, not an afterthought.

CNBC on NVIDIA's Japan expansion (July 16) · NVIDIA on the Japan ecosystem (July 16) · Build Fast with AI on the Hyundai strike (July 16) · The Information: Humanoid becomes a unicorn

Microsoft preps 'Project Perception' — the AI-finds-it, AI-fixes-it security product

The Information reported this morning that Microsoft is preparing an AI security product, internally codenamed Project Perception, that could debut as soon as this month: it will use a combination of models from Anthropic, OpenAI, and Microsoft itself to find software vulnerabilities the way Anthropic's Mythos does — and automatically fix them. It is the product-shaped consequence of everything this wave has been building toward: last week's Gold Eagle disclosure clearinghouse, the security-unit overhaul The Information reported Wednesday, and a July Patch Tuesday that addressed a record 570 flaws with AI systems helping identify and prioritize a large share of them, per this week's roundups. Note the vendor-neutrality: Microsoft's flagship security product will run its rivals' models where they are best. Model routing has reached the security stack.

So what: The 'AI discovers, AI repairs' category hits procurement this quarter. Pilot one vulnerability-repair product against a known backlog and measure time-to-patch — that number is about to become a board question, and vendors who can't move it will hide behind dashboards.

The Information: Microsoft preps Mythos-like bug finder (July 17) · The Information on the security overhaul (July 15)

🌊 Ripples

Gemini 3.5 Pro misses its third deadline; Google eyes a stopgap

The rumored July 17 launch didn't survive to July 17: TechTimes reported Wednesday that the rebuilt Gemini 3.5 Pro has missed its third deadline — after the original I/O commitment and a June 30 GA target — and that Google is weighing a stopgap release, with coverage citing prediction markets that now favor early August. As of this morning there is still no model card, no pricing page, and no API listing. Wednesday's brief said either outcome would be information; this is the outcome, and it lands the same day China's largest-ever open model starts its benchmark lap.

So what: Plan H2 evaluations around models that have shipped. If million-token context is your blocker, test what exists — K3's 1M window when the weights land, or current GPT-5.6 — rather than holding roadmap space for an unannounced launch. A rebuilt model missing three dates is data about organizational state, not just scheduling.

TechTimes on the third miss (July 16)

Valar Atomics in talks to raise ~$1B — datacenter nuclear reprices again

The Information reported yesterday evening that Valar Atomics, the three-year-old small-reactor startup, is in talks to raise about $1 billion; The Information's headline puts the valuation at $6 billion while its report text says around $5 billion — reports differ, but either figure is roughly triple the $2 billion Bloomberg reported at its $450M round in late March. The proof point behind the premium: Valar ran an NVIDIA Blackwell chip off its nuclear microreactor on July 1, a reported US first. Same-day company: The Information's enterprise letter, titled 'Nobody Is Immune From the GPU Crunch,' detailed startups decamping from AWS to neoclouds for capacity. Electrons and compute remain the binding constraints, and the market keeps pricing it in.

So what: Behind-the-meter power keeps attracting billions — Monday's $5.34B Williams deal, now this. When vendors promise 2027 inference capacity, ask where the electrons come from; a capacity contract without a power story is a press release.

Crypto Briefing on the July 1 microreactor demo · Bloomberg on the March $2B valuation

A 27-billion-parameter model now runs on an iPhone

PrismML's Bonsai 27B, released July 14, compresses a 27B model built from Qwen3.6-27B to 3.9GB using 1-bit and ternary quantization — running on an iPhone 17 Pro at about 11 tokens per second while keeping over 90% of full-precision performance, free under Apache 2.0. Pair it with Liquid AI's Antidoom method (open-sourced July 7), which cut a small Qwen model's reasoning doom-loop failures from 22.9% to 1%, and the edge story sharpens: small local models are crossing from demo to deployable. The cost-collapse tide doesn't stop at cheaper APIs — its logical end state is routine intelligence at zero marginal cost on hardware you already own.

So what: Before renewing per-token contracts for high-volume, low-complexity workloads — summarization, classification, drafting — price an on-device or small-model route. The right architecture is increasingly a router: frontier cloud for hard reasoning, local for the volume.

MarkTechPost on Bonsai 27B (July 14) · 9to5Mac on the iPhone claim (July 14)

China's memory champion CXMT files for an $8.6B Shanghai IPO

Memory chipmaker CXMT filed Wednesday for an $8.6 billion Shanghai listing, per The Information — set to be the biggest tech IPO in China's domestic market. Memory has been the quiet constraint under the AI buildout (SK Hynix's HBM-driven listing already proved the thesis), and a heavily capitalized domestic supplier is a missing leg of a self-contained Chinese AI stack: models (K3, GLM, DeepSeek), chips (Huawei is scheduled to unveil its Atlas 950 computing cluster at WAIC, which runs July 17-20), and now memory. The two-stack world keeps getting more literal.

So what: Watch HBM supply and pricing into 2027 — memory sets the ceiling on inference capacity growth for everyone. And add it to the export-control watchlist: memory is the next lever either bloc can pull.

The Information on the CXMT filing · SCMP on WAIC and the Huawei unveil

Read every edition: https://excelsiorgroup.ai/insights/signal/