Daily2026-07-20
EconomicTrend

Moonshot preps $30 billion Hong Kong IPO in six months as Kimi K3 rerates the cap table

Yang Zhilin circulated a shareholder resolution over the weekend asking Moonshot AI's backers to approve a Hong Kong listing as early as January 2027 at a $30-billion-plus post-money valuation, roughly 7.5 times the company's late-2025 mark. Annual recurring revenue climbed from $200 million in April to $300 million in June on Kimi K3 consumption. Goldman Sachs and China International Capital Corporation are both in talks to lead. DeepSeek is running a parallel Shanghai STAR track at $74 billion, and Anthropic answers Monday by bundling Claude Fable 5 permanently into Max at 50 percent of limits.

The National / Bloomberg · 2026-07-19 · Source ↗

Moonshot AI has told its shareholders it is preparing a Hong Kong listing within six months at a $30-billion-plus post-money valuation, three days after the July 16 Kimi K3 release rewrote Chinese open-weight benchmarks and pulled global chip and hyperscaler stocks into their worst week of the year. Bloomberg broke the news Sunday and The National, Investing.com, CryptoBriefing, TheNextWeb and Startup Fortune ran their own follow-ups the same day. Yang Zhilin, the 34-year-old CMU-trained founder, is asking backers to bless the offering via a formal shareholder resolution now circulating; Goldman Sachs and China International Capital Corporation are both in talks to lead. The $30 billion mark represents roughly a 7.5x rerating from Moonshot's late-2025 valuation of about $4 billion, and puts the Beijing lab inside striking distance of DeepSeek's $74 billion pre-IPO round announced the same weekend for a Shanghai STAR Market filing before year-end.

"Zhilin was quite determined to go back and build a start-up. I remember him telling me that if he didn't at least try starting his own company, he would regret it for the rest of his life."— Russ Salakhutdinov, University of Toronto professor and Yang's CMU PhD supervisor

The revenue trajectory behind the mark

Annual recurring revenue at Moonshot reached $300 million in June, up from $200 million in April, a $100 million pickup across two months that Moonshot attributes to Kimi K3 consumption on the Kimi app, Kimi Work desktop, and the API. Yang told investors the company holds more than 10 billion renminbi, roughly $1.4 billion, in cash reserves and is not rushing the listing. The company has raised between $3.9 billion and $6 billion cumulatively since its 2023 founding. The pending private round, expected to close before the IPO, targets approximately $2 billion in fresh capital at the $30-billion-plus mark, with anchor investors Alibaba, Tencent and China Mobile taking the majority of allocations.

What Kimi K3 actually shipped

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model activating 16 of 896 experts per token, delivering roughly 50 billion active parameters per query with a 1-million-token context window. It opened at Elo 1,679 on Arena.ai's Frontend Code leaderboard against Anthropic's Claude Fable 5 at 1,631 and OpenAI's GPT-5.6 Sol at 1,618, taking the top slot with 1,757 votes recorded within the first 48 hours. Artificial Analysis places K3's intelligence tier alongside Claude Opus 4.8 and GPT-5.5, one notch behind Fable 5 and Sol. API pricing runs $3 per million input tokens on cache-miss, $0.30 on cache-hit, and $15 per million output tokens, roughly one-third of Anthropic's forthcoming $10 input and $50 output rate card for Fable 5 metered access. Moonshot has committed to publish full open weights on Hugging Face by July 27, with an intermediate preview scheduled for July 22 and a technical report on architecture, training, and evaluation running in parallel.

"As he approached the end of his studies at CMU, Mr Zhilin had several opportunities to work for US technology companies, including Apple."— Russ Salakhutdinov, on Yang's 2019 job decision

The VIE unwind and Beijing's overseas-listing rules

Moonshot is restructuring away from a Cayman-registered variable-interest-entity model toward a joint-venture structure that satisfies revised China Securities Regulatory Commission rules gating offshore listings of AI and data-heavy mainland businesses. The rewired cap table is what the shareholder resolution now circulating asks investors to approve, alongside the IPO mandate itself. The path Moonshot is walking mirrors what Zhipu and MiniMax have signaled for their own Hong Kong or Shanghai filings, and the sequencing runs alongside Beijing's own preliminary discussions with Alibaba, ByteDance, and Z.ai about potential export controls covering both closed and open-source Chinese AI models, mirroring the June US action against Anthropic.

Yang's biography as policy artifact

The National ran a Sunday profile drawing on fresh comments from Russ Salakhutdinov, the University of Toronto and Carnegie Mellon computer scientist who supervised Yang's four-year CMU PhD. Salakhutdinov told the paper Yang had 'several opportunities to work for US technology companies, including Apple' as he wrapped up his doctorate in 2019 but chose to return to China. Yang was born in 1992 in Shantou, Guangdong, completed his undergraduate work at Tsinghua, and interned at both Google Brain and Meta AI Research during his PhD, contributing what Salakhutdinov called 'fundamental contributions to machine learning.' Salakhutdinov also flagged US immigration friction as a factor pushing top graduates toward Chinese labs: 'The US immigration process can be quite intimidating and uncertain, even for superstar PhD graduates.' Yang incorporated Moonshot in March 2023 alongside Tsinghua classmates Zhou Xinyu and Wu Yuxin, naming the company after Pink Floyd's Dark Side of the Moon on the album's 50th anniversary.

"The US immigration process can be quite intimidating and uncertain, even for superstar PhD graduates."— Russ Salakhutdinov, on Chinese AI talent flow

Global cap-markets reaction

The IPO news dropped into a market still absorbing Friday's coordinated 'Kimi moment' selloff. TSMC fell 7 percent despite a 77 percent Q2 profit surge reported 24 hours earlier, SoftBank shed 9.0 percent, Nvidia gave up 1.2 percent and briefly ceded the world's-most-valuable-company slot to Apple at $4.88 trillion intraday, Alphabet fell 4.4 percent, Alibaba 4 percent, Meta 2.4 percent, and Chinese peers Z.ai plunged 27 percent and MiniMax 16 percent in Hong Kong. The Philadelphia Semiconductor Index dropped 10 percent on the week for its worst tape since April 2025. Former White House AI czar David Sacks called K3's reception 'concerning' in a Sunday interview cited by The National, underscoring the geopolitical weight of an open-weight Chinese model reaching Fable-class benchmarks. Bloomberg's Sunday framing put the reprice bluntly: the $700 billion of 2026 hyperscaler AI capex may not pay back if capable AI drifts toward free.

"K3's reception is concerning."— David Sacks, former White House AI czar

The parallel Anthropic answer

The IPO push runs alongside Anthropic's Sunday-evening capitulation on Fable 5 subscription access. From Monday July 20, Anthropic will bundle Fable 5 into every Max and Team Premium plan at 50 percent of the plan's usage limits, ending a five-week rolling grace period and formalizing what Simon Willison called an economically forced move to keep the frontier premium bundled rather than metered. The two moves are the same story from opposite sides of the Pacific. Moonshot marks up its equity by 7.5x on the strength of an open-weight release; Anthropic defends its subscription franchise by making the frontier model permanently included at half of a reduced base. Both sides now have calendar-driven proof points for the compression thesis that ate 10 percent of the semiconductor index last week.

The Big Picture

Moonshot AI circulated a shareholder resolution over the weekend seeking approval for a Hong Kong listing within six months at a $30-billion-plus post-money mark, roughly 7.5 times its late-2025 valuation, as annual recurring revenue climbed from $200 million in April to $300 million in June on Kimi K3 consumption. DeepSeek is running the parallel Shanghai STAR track at 500 billion yuan ($74 billion), with a fresh 50-billion-yuan round targeted before a year-end filing. Anthropic capitulates Monday, permanently bundling Claude Fable 5 into every Max and Team Premium plan at 50 percent of a reduced usage base, while Pro and Team Standard revert to $10-input, $50-output metered API pricing with a one-time $100 credit. Bloomberg's Wittenstein and Vlastelica framed the Philadelphia Semiconductor Index's 10 percent week, its worst since April 2025, as a hyperscaler capex-justification test heading into the Alphabet, Meta, Microsoft, and Amazon Q2 prints landing late July. SK Hynix dropped another 11.5 percent Seoul session, Samsung more than 8, in the fourth consecutive memory selloff, and Bank of America pegs 2026 hyperscaler AI capex at $580 billion against $140 billion of incremental AI revenue, a 4.1x ratio it argues will need to compress.

Anthropic tagged Claude Code v2.1.215 overnight to stop the model from auto-running the /verify and /code-review skills mid-session; both now require an explicit slash invocation, and Claude Agent SDK Python v0.2.123 pinned the new CLI twelve minutes later. Simon Willison ran strings against his local Claude binary Sunday morning and pulled 563 Rust source paths plus a 'Bun v1.4.0 (macOS arm64)' build stamp, confirming Jarred Sumner's claim that Claude Code v2.1.181 and later ship the Zig-to-Rust Bun rewrite in production across millions of installs at roughly 20 percent smaller footprint. LiteLLM shipped v1.93.0 stable and v1.94.0-rc.1 with an Azure Entra on-behalf-of MCP token exchange, separate input and output tokens-per-minute rate limits, a Bedrock streaming SSE-error fix, and automatic Anthropic prompt-caching injection. Cline SDK v0.0.65 halved the CLI install to 285 MB by moving the Claude Code and Codex provider bundles to optional peer dependencies and added Kimi K3 to the ClinePass fallback list. Moore Threads unveiled the MTT C256 hyper-node at WAIC with 256 GPUs interconnected in a single-layer scale-up domain at sub-microsecond latency, and Hugging Face's top weekend paper LongStraw pushed on-policy RL past 2 million tokens on a fixed 8xH100 budget by rearranging the rollout KV cache across host memory.

Xi Jinping's National Development and Reform Commission published its Action Plan for Artificial Intelligence Cooperation and Development the same weekend as the K3 launch, converting the Kimi release into a state-policy proof point and pegging Chinese open-weight models at 63 percent of OpenRouter traffic in early July per figures CGTN cited from the platform's own dashboards. Chinese vendors at WAIC 2026 abandoned the 40-yuan flat-rate consumer tier that dominated the first half of the year, moving to a 200-yuan premium tier plus per-token billing on the reasoning that flagship-model inclusion at bounded pricing is no longer viable, an Alibaba Cloud staffer at the exhibition floor compared to mobile's 3G-to-4G transition. AGIBOT deployed more than 30 humanoids across WAIC venues and unveiled the 1.74-meter, 60-kilogram A3 Ultra with 51 degrees of freedom on Nvidia Jetson Thor T3000. The White House is reviewing a FINRA-style AI safety regulator reporting to the SEC per Bloomberg reporting picked up by PYMNTS, Google DeepMind CEO Demis Hassabis has publicly endorsed a similar model, and Help Net Security's week in review paired the WordPress wp2shell public-PoC exploitation window with SANS's 78 percent security-team AI adoption print and OpenAI's disclosed internal-only GPT-Red system hitting 84 percent success on prompt-injection tests against 13 percent for human red teams.

More today

18 stories · 2026-07-19
TechApplication

Anthropic bundles Claude Fable 5 into Max and Team Premium at 50 percent of limits from Monday

Anthropic / Dawn · 2026-07-19 · Source ↗

Anthropic's Sunday policy tweet closed three prior week-long extensions and made Claude Fable 5 a permanent Max and Team Premium inclusion at half of a reduced usage base. Pro and Team Standard subscribers lose bundled Fable access, take a one-time $100 credit, and revert to metered API pricing.

The @claudeai account posted the Fable 5 subscription rework Sunday July 19 as a policy tweet that reversed weeks of week-by-week extensions and formalized a permanent split effective 24 hours later. The tweet reads: 'Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50 percent of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. Demand for Fable has been challenging to predict, which is why it was rolled out to subscription plans in stages, extending access several times as capacity was secured.' The announcement follows three prior extensions of a promotional included-access window that had been due to expire at 11:59 PM PT Sunday July 19.

"Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50 percent of limits."— Anthropic, official announcement

Tier-by-tier mechanics

Max 5x and Max 20x plans will include Claude Fable 5 at 50 percent of standard weekly usage limits, a reduction from the prior period during which Claude Code enjoyed a 50 percent usage boost. Team Premium plans mirror the Max terms. Pro and Team Standard users lose bundled Fable 5 access entirely and receive a one-time $100 usage credit to soften the transition to API-priced consumption at approximately $10 per million input tokens and $50 per million output tokens. Discounted bundle purchases are available up to $2,000 monthly for Pro accounts. All Claude models draw from a shared weekly pool, so heavy Sonnet 5 or Opus 4.8 consumption reduces effective Fable 5 headroom on Max and Team Premium and vice versa. Anthropic committed to continue investing in additional inference capacity but acknowledged Fable 5 demand has been 'hard to manage and frustrating for users.'

The reversal in context

The Decoder headlined the change as Anthropic 'slashing' Fable 5 limits in Max and Team Premium and 'pushing Pro users toward API pricing,' noting that when the bonus phase ends, regular limits drop 33 percent and Fable 5 is available at only 50 percent of the already-reduced ceiling. Simon Willison's Friday 'Claude make Fable 5 permanent' post had argued OpenAI's GPT-5.6 Sol was winning subscriber conversions on price and predictability alone during the rolling extensions, and pinned the reversal on competitive pressure from Sol at roughly one-third of Fable 5's per-token cost and from Moonshot's Kimi K3 at $3 input and $15 output per million tokens. Basic $20-per-month Claude plans get no Fable 5 access at all.

"The reversal reflects competitive pressure from GPT-5.6 Sol and Kimi K3. Subscribers won't pay premium fees for a plan that excludes Anthropic's best model."— Simon Willison, weblog

Why the frontier goes back in the bundle

The move matters most for the agent workload envelope. Fable 5 is Anthropic's most capable long-horizon model for the tool-use, computer-use, and multi-step planning workloads that dominate Claude Code and Claude Cowork usage. Permanent inclusion at 50 percent of Max limits materially changes the economics for coding-agent shops that had begun designing Fable 5 out of default configurations to avoid credit burn. It also mirrors the WAIC 2026 premium-tiering pivot reported by CGTN on July 19, in which Chinese AI vendors abandoned flat-rate 40-yuan tiers for per-token or 200-yuan premium models. Both moves reflect a Q3 2026 industry consensus that flagship-model inclusion in mass-market subscriptions at bounded pricing is no longer viable at current inference costs. Anthropic remains the only Western frontier lab to preserve any flagship-model bundling in an under-$100 consumer tier, and only for the buffer that ends Sunday.

Economic

DeepSeek plans fresh 50-billion-yuan pre-IPO round at $74 billion ahead of Shanghai STAR filing

Reuters via Jakarta Post · 2026-07-19 · Source ↗

Reuters sources say DeepSeek plans up to 50 billion yuan of fresh capital at a 500-billion-yuan valuation less than six weeks after closing $7.4 billion in June, with an internal target to file for a STAR Market listing before year-end. The Hangzhou lab pairs the raise with a custom-silicon program disclosed July 7.

DeepSeek is planning a fresh fundraising round at approximately 500 billion yuan ($74 billion) ahead of a potential mainland initial public offering, according to sources cited in a Reuters wire report picked up by The Jakarta Post and The Manila Times on July 19. The Hangzhou-based company is looking to raise as much as 50 billion yuan in the new tranche and has set an internal target to complete an IPO filing on Shanghai's Nasdaq-style STAR Market before year-end. The sources declined to be named because the deliberations are private, and Reuters could not independently verify shareholder composition of the pending round.

A six-week gap from the June mega-round

The new fundraising follows the June 2026 close of a $7.4 billion round at a post-money valuation of about 450 billion yuan, with separate filings suggesting a 350.88 billion yuan ($52 billion) alternative reading of the June cap-table. In the June round, DeepSeek founder Liang Wenfeng personally committed 20 billion yuan, while Tencent Holdings and battery giant CATL wrote checks of 10 billion yuan and 5 billion yuan respectively to become the largest external shareholders. China's national AI fund, NetEase, JD.com, IDG Capital, Loyal Valley Capital, Monolith Management and Shixiang Capital also participated. The June-to-July gap of under six weeks marks one of the fastest consecutive AI mega-rounds outside OpenAI's 2025 cadence.

"DeepSeek is looking to raise as much as 50 billion yuan in the new funding round."— Reuters sources, July 19

Use of funds and IPO mechanics

The sources named rising costs of competing in AI, which require large amounts of computing power, data-center capacity and engineering talent, as the primary use of the fresh capital. DeepSeek has committed to custom inference silicon to reduce dependence on Nvidia H-series chips and Huawei Ascend accelerators, an initiative disclosed on July 7. The IPO plan targets a Shanghai STAR Market listing, which has emerged as the preferred venue for Chinese AI listings after the June 2026 US export-control tightening curtailed offshore fundraising for mainland AI vendors. The article notes the plans are at early stages and that terms and timetables may change.

"Rising costs of competing in AI require large amounts of computing power, data-center capacity, and engineering talent."— Reuters, on the fundraise rationale

The coordinated Chinese IPO track

DeepSeek's fundraising sprint coincides with Moonshot AI's Hong Kong IPO preparation announced the same weekend at $30-billion-plus and with Alibaba, Tencent and ByteDance receiving the first H200 shipments into China on July 15 under the January 13 US rule. The Global Times pegged Chinese AI models at 41 percent of daily Hugging Face downloads Friday, and MiniMax executives at WAIC claimed Chinese open-source models exceeded 10 billion cumulative downloads. The two China AI IPO tracks running through this weekend, Moonshot in Hong Kong and DeepSeek in Shanghai, together represent the first coordinated dual-track China-AI public financing push since the June 2026 US export-control tightening. Together they anchor an approximately $104 billion aggregate pre-IPO valuation for the two most benchmarked Chinese frontier labs. Analysts at Bank of America and Manifold Markets have set 2028 IPO probability for the Chinese frontier cohort at 80 percent.

TrendEconomic

Bloomberg: Big Tech AI capex faces reckoning after 10 percent chip selloff

Bloomberg · 2026-07-19 · Source ↗

Jeran Wittenstein and Ryan Vlastelica frame the S&P 500's 1.6 percent slide, the Nasdaq 100's 4.1 percent loss, and the Philadelphia Semiconductor Index's 10 percent worst week since April 2025 as the moment hyperscalers must justify $580 billion of 2026 AI capex against $140 billion of incremental AI revenue.

Bloomberg reporters Jeran Wittenstein and Ryan Vlastelica published on July 19 that pressure is building for the biggest spenders on artificial intelligence to justify their expenditures to beleaguered traders. Information technology was the worst performing group in the S&P 500 Index for the week ended July 17. The S&P 500 slid 1.6 percent. The tech-heavy Nasdaq 100 lost 4.1 percent. The Philadelphia Stock Exchange Semiconductor Index sank 10 percent for its worst week since April 2025. Chip stocks drove the losses on the back of Moonshot's Kimi K3 open-weight release, TSMC's 7 percent slide despite a 77 percent Q2 profit surge, and SoftBank's 9 percent drop as investors reassessed the Stargate spend thesis.

"Pressure is building for the biggest spenders on artificial intelligence to justify their expenditures."— Jeran Wittenstein and Ryan Vlastelica, Bloomberg

The capex numbers at stake

The reporters framed the reckoning against the largest AI capex commitments in market history. Meta pushed its Louisiana Hyperion campus to 5 gigawatts and more than $50 billion. Oracle's flagship $165 billion New Mexico Stargate campus faces multibillion-dollar cost overruns, per The Information. TSMC lifted 2026 capex to $60-64 billion and added $100 billion to its Arizona commitment, bringing total US investment to $265 billion. Alphabet fell 4.4 percent on Friday alone. Nvidia briefly lost the world's-most-valuable-company slot to Apple at $4.88 trillion intraday. Total AI infrastructure capex among the hyperscalers is on pace to exceed $500 billion for 2026, and Bloomberg's earlier July 10 piece pegged Big Tech debt loads at $350 billion after AI-fueled bond issuance doubled from 2024 levels.

The Kimi moment as trigger

Bank of America pegged Kimi K3 pricing at roughly half of Sol. Chinese peers Z.ai plunged 27 percent and MiniMax fell 16 percent in Hong Kong as investors flagged margin pressure across the entire Chinese AI stack. Bitcoin and major cryptocurrencies followed lower as crypto traded as an AI-capex proxy. The Bloomberg piece cast the selloff as a demand for proof that hyperscaler AI spend converts into monetizable revenue rather than a race to zero on token prices. Where DeepSeek was framed as an efficiency innovation on training compute, Kimi K3 arrives as a full open-weight substitute at frontier-tier quality with a specific July 27 weight-release date. That converts the risk from theoretical to calendar-driven.

"The AI euphoria that drove the stock market to all-time highs just a month ago is clearly waning."— Bloomberg, July 19

What the earnings prints will test

Q2 earnings from Alphabet, Meta, Microsoft, and Amazon land in the last week of July. Nvidia reports in late August. Bloomberg framed the four print dates as the moment when hyperscaler executives will need to answer specific questions about AI token unit economics, backlog conversion, and margin visibility. Wall Street analysts including Melius Research's Ben Reitzes, MoffettNathanson's Michael Nathanson, and Bernstein's Toni Sacconaghi have flagged capex-to-revenue ratios as the metric that will drive the next leg of the AI stock reset. Bank of America estimated the four hyperscalers will spend $580 billion on AI capex in 2026 and generate roughly $140 billion in incremental AI revenue, a 4.1x ratio that will need to compress.

EconomicTrend

SK Hynix drops 11.5 percent, Samsung more than 8 in fourth consecutive memory session lower

Seoul Economic Daily · 2026-07-19 · Source ↗

The Seoul memory complex extended its fourth consecutive down session as SK Hynix fell 11.5 percent and Samsung more than 8 percent on the Kimi K3 open-weight shock, threatening the $700 billion hyperscaler capex payback thesis that anchors HBM4 16-hi qualification for Nvidia Vera Rubin and AMD MI400.

SK Hynix dropped another 11.5 percent in Seoul, dragging Samsung Electronics more than 8 percent lower and pushing the memory sector into a fourth consecutive down session. Micron shed close to 6 percent in New York trading, wiping out most of the previous week's 15 percent leveraged-ETF-driven surge from the launch of GraniteShares and ProShares 2x SK Hynix ADR trackers. The VanEck Semiconductor ETF closed down almost 9 percent for its third weekly decline in four weeks, taking the ETF down more than 17 percent for the month.

Kimi K3 as catalyst

Moonshot AI's release of the 2.8-trillion-parameter Kimi K3 open-weight model triggered the coordinated repricing. K3 opened at Elo 1,679 on Arena.ai's Frontend Code leaderboard against Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618, and Moonshot committed to publish full weights on Hugging Face by July 27 at roughly half the API pricing of Sol. Traders drew the parallel to the January 2025 DeepSeek shock that briefly wiped $600 billion from Nvidia and repriced the entire AI capex chain on the concern that near-frontier open weights slow the training demand curve just as HBM4 16-hi qualification enters Q4 for Nvidia Vera Rubin and AMD MI400.

"If capable AI is becoming free, the roughly $700 billion hyperscalers are spending may not pay back."— Seeking Alpha semiconductor coverage cited by Seoul Economic Daily

Underlying HBM setup

SK Hynix holds 62 percent of Q2 HBM share against Micron at 21 percent and Samsung at 17 percent, with SK Hynix HBM3E prices lifted nearly 20 percent for 2026 deliveries. HBM4 development races among all three suppliers to lock Nvidia's HBM4 16-hi orders arriving in Q4 2026. Samsung targets 50 percent HBM capacity expansion in 2026; SK Hynix has committed more than four times its previous infrastructure investment. The Friday drop reflects a payback horizon repricing rather than a break in the underlying demand queue, with 2026 HBM output already sold out and 2027 negotiations in progress.

"SK Hynix holds 62 percent of Q2 HBM share, Micron 21 percent, Samsung 17 percent."— Astute Group, July 2026

The near-term technical setup

HSBC keeps SK Hynix as a top chip pick, arguing the Friday drop overshoots the fundamental case. Bank of America maintains a Micron $1,550 price target. The near-term setup keeps memory names volatile through the WAIC 2026 close on July 20 and the AMD Advancing AI event on July 22-23, both of which carry additional catalyst risk for the sector.

Tech

Claude Code v2.1.215 stops auto-running the /verify and /code-review skills

GitHub (anthropics/claude-code) · 2026-07-19 · Source ↗

Anthropic engineer @ashwin-ant tagged v2.1.215 at 02:56 UTC Sunday with a single behavioral change: Claude no longer auto-runs /verify or /code-review mid-session. Twelve minutes later Claude Agent SDK Python v0.2.123 pinned the new CLI version for every Python-hosted agent.

Anthropic engineer @ashwin-ant tagged Claude Code v2.1.215 at 02:56:01 UTC on Sunday July 19 as commit 015170d3fd84fb57ef4685a64b673fadd0690dc1, a one-line release whose sole change reads: 'Claude no longer runs the /verify and /code-review skills on its own; invoke them with /verify or /code-review when you want them.' The 27-reaction release page carries 12 assets covering the standalone macOS, Linux, and Windows binaries, the npm tarball, and the plugin surface. Twelve minutes later at 03:08:55 UTC, GitHub Actions cut Claude Agent SDK Python v0.2.123 as commit 2d4ef94 with a single internal-changes bullet: 'Updated bundled Claude CLI to version 2.1.215.'

"Claude no longer runs the /verify and /code-review skills on its own; invoke them with /verify or /code-review when you want them."— Claude Code v2.1.215 release notes

What the change stops Claude from doing

The /verify skill drives the built-in end-to-end verification harness, exercising affected flows and observing runtime behavior before a change is claimed complete. /code-review runs the reuse, simplification, efficiency, and correctness sweep that ships with the CLI, optionally posting inline PR comments or applying fixes with --comment or --fix. Both had been included in the default skill set that the model could pick up unprompted mid-session under v2.1.214's broader stability update, which shipped Friday July 18 with EndConversation, the PowerShell 5.1 permission bypass fix, and the rewritten single-segment dir/** semantics. Enough sessions saw Claude reach for /verify or /code-review during unrelated tasks, or auto-fire on diffs that only touched tests or docs, that Anthropic pulled the trigger overnight.

Explicit invocation, no other behavioral change

Users still get both skills at the /verify and /code-review slash commands. The v2.1.215 diff does not touch permissions, hooks, subagent caps, plugin loading, or any of the OTel or Bash-check surfaces reworked over the prior week. There is no new flag, no configuration key, no environment variable to restore the auto-run behavior. The change is unilateral and ships to every installer channel at the same GitHub tag; the CHANGELOG.md file inside the repository is 5,112 lines long as of this release.

"Verify that a code change actually does what it's supposed to by exercising it end-to-end and observing behavior — drive the affected flow, not just tests or typecheck."— /verify skill description in the CLI harness

Agent SDK pins the new CLI

The companion release, Claude Agent SDK Python 0.2.123 on PyPI, is a pure metadata bump updating the bundled Claude CLI version pin from 2.1.214 to 2.1.215. Python callers that shell out to the CLI via the SDK's harness pick up the new /verify and /code-review behavior on their next pip install without a code change. Anyone using the SDK to drive sub-agents that used to expect the skill picker to auto-fire during a code review or verification pass now needs to add an explicit slash-command step to their orchestration.

Tech

Willison confirms Claude Code binaries ship the Rust port of Bun v1.4.0 in production

simonwillison.net · 2026-07-19 · Source ↗

Simon Willison ran strings against his local Claude binary Sunday morning and pulled a 'Bun v1.4.0 (macOS arm64)' stamp plus 563 Rust source paths, verifying Jarred Sumner's claim that Claude Code v2.1.181 and later run the Zig-to-Rust Bun rewrite in production at roughly 20 percent smaller footprint.

Simon Willison posted 'Claude Code uses Bun written in Rust now' as a blogmark on his weblog at 03:54 UTC Sunday July 19, verifying Jarred Sumner's Bun-blog claim that Claude Code v2.1.181 released June 17 and every subsequent release ship the Rust rewrite of Bun rather than the original Zig build, and that on Linux this cut CLI startup by 10 percent. Willison used three techniques on his ~/.local/bin/claude executable, an unstripped 88-to-70 MiB binary that bundles the Bun runtime, all of Claude Code's TypeScript, the model client, and the skill and hook systems.

Three ways to see the Rust port from a shell

Running strings ~/.local/bin/claude | grep -m1 'Bun v1' returned 'Bun v1.4.0 (macOS arm64)' on macOS Apple Silicon, a build stamp that does not exist in any tagged Bun GitHub release yet because v1.4.0 was committed to the Bun repo on May 17 and remains a canary-only pre-release outside Anthropic. A second command, strings ~/.local/bin/claude | grep -Eo 'src/[[:alnum:]_./-]+\.rs', pulled 563 unique Rust source-file paths embedded as panic-location strings, headed by src/runtime/bake/dev_server/mod.rs, src/runtime/bake/production.rs, and src/bundler/bundle_v2.rs. Willison also demonstrated a runtime query using BUN_OPTIONS='--preload=/tmp/bun-version.ts' claude --version to have the Bun runtime dump its own version straight from the embedded interpreter.

"Claude Code v2.1.181 (released June 17th) and later use the Rust port of Bun. Startup got 10% faster on Linux but otherwise, barely anyone noticed."— Jarred Sumner, Bun blog

Why an agent CLI cares which Bun it ships

Sumner wrote up the Zig-to-Rust port in a Bun blog post that Willison quotes. Bun is the JavaScript runtime that packages Claude Code as a single-file executable; the Bun rewrite took 11 days in May, roughly 64 parallel Claude instances, 5.9 billion uncached input tokens, 690 million output tokens, and $165,000 of API spend at published rates. Shipping the Rust build in a paid agentic tool that runs against arbitrary user code was the higher-stakes deployment path Sumner chose because, as Willison summarizes, 'if Bun breaks, Claude Code breaks.'

"If Bun breaks, Claude Code breaks."— Simon Willison, summarizing the deployment stakes

Production footprint and shrinkage numbers

Across Rust-ported Bun, an ICU library swap, and identical-code folding the Bun team applied at the same time, the Linux binary shrank from 88 MB to 70 MB and the Windows binary from 94 MB to 76 MB, roughly 20 percent smaller. Willison notes this is the largest production deployment of a Rust-based JavaScript runtime to date and highlights that it landed silently through Claude Code's auto-update path more than a month before Bun tagged an official v1.4.0 release. His verification post landed on Hacker News as item 48966569 the same morning.

Tech

LiteLLM v1.93.0 lands MCP OBO token exchange and separate ITPM/OTPM rate limits

GitHub (BerriAI/litellm) · 2026-07-19 · Source ↗

BerriAI cut LiteLLM v1.93.0 at 07:57 UTC Sunday with an entra_obo profile for MCP token exchange, separate input and output tokens-per-minute rate limits, and a Bedrock streaming SSE fix, then dropped v1.94.0-rc.1 with enable_anthropic_prompt_caching for automatic cache_control injection.

The BerriAI LiteLLM proxy shipped five tags on Sunday July 19: v1.94.0-rc.1 at 00:14:52 UTC, then a four-version backport cascade at 07:49 (v1.90.6), 07:51 (v1.91.4), 07:53 (v1.92.1), and 07:57 (v1.93.0). All five Docker images are cosign-signed against key 0112e53046018d726492c814b3644b7d376029d0. v1.93.0 is the new stable head, v1.94.0-rc.1 is the release candidate for the next branch, and the trio in the middle are LTS patches on prior minor lines.

v1.93.0 stable: MCP OBO, ITPM/OTPM, streaming fixes

The headline v1.93.0 change is MCP OBO: PR #31983 by tin-berri adds an entra_obo profile to the Model Context Protocol token_exchange arm, letting LiteLLM broker Azure Entra on-behalf-of tokens between the calling user identity and an MCP server. PR #31952 by Sameerlite gives the router separate input-tokens-per-minute and output-tokens-per-minute deployment rate limits, previously tracked as a single figure. mateo-berri's PR #32160 forwards provider response headers on streaming /v1/messages responses so anthropic_messages clients see the same rate-limit headers they get on non-streaming calls, and PR #32159 emits an SSE error event when a Bedrock invoke Messages stream ends without a message_stop token.

"feat(mcp): add entra_obo profile to the token_exchange (OBO) arm."— LiteLLM v1.93.0 release notes

v1.94.0-rc.1: Gemini grounding, Anthropic prompt caching, complexity router

The release-candidate branch adds five feature-level PRs on top of v1.93.0. tin-berri's PR #33573 introduces enable_anthropic_prompt_caching for automatic cache_control injection so LiteLLM callers no longer have to hand-annotate messages for prompt caching. PR #33746 by yuneng-berri adds a router plugin reference catalog, and PR #33656 lands user-triggered escalation keywords in the complexity-router. Devin-integration PR #33533 surfaces Gemini grounding toolUsePromptTokenCount in Usage so agents that rely on Google grounding tools see the correct billed prompt tokens. Yassin-berriai's PR #32441 rewrites max_parallel_requests as a per-slot concurrency gauge rather than a rolling counter, and PR #33710 stops treating the upstream model body field as a LiteLLM model on auth-enforced pass-through routes.

"feat(anthropic): add enable_anthropic_prompt_caching for automatic cache_control injection."— LiteLLM v1.94.0-rc.1 release notes

LTS backport cascade

The 07:49-07:57 UTC window saw the same team backport the v1.93.0 fix set into v1.90.6, v1.91.4, and v1.92.1, giving enterprise customers pinned to older minor lines the same MCP OBO, headroom-guardrail, and Bedrock streaming fixes without moving to the new head. All four production tags landed within eight minutes. Devin-integration PR #33714 also bills Fireworks AI prompt-cache hits at the cache_read rate rather than the full rate, PR #33728 marks realtime-only gpt-realtime models with mode 'realtime' for correct pricing, and PR #33719 self-heals on missing thinking-signature errors from Bedrock and Vertex Anthropic calls.

Tech

Cline halves the CLI install to 285 MB and adds Kimi K3 fallback in SDK v0.0.65

GitHub (cline/cline) · 2026-07-19 · Source ↗

Cline SDK v0.0.65 at 04:02 UTC Sunday moves Claude Code and Codex provider adapters to optional peer dependencies, cuts npm i -g cline from 640 MB to 285 MB, and adds Moonshot's Kimi K3 to the ClinePass fallback. CLI v3.0.46 followed at 06:26 with a 402 insufficient_credits detection fix.

The Cline agent framework tagged three releases on Sunday July 19: sdk/sdk/v0.0.65 at 04:02:15 UTC, cli-v3.0.45 at 04:10:41 UTC, and cli-v3.0.46 at 06:26:43 UTC. The SDK release is the substantive one; the two CLI releases roll it forward into the top-level runtime and add a single out-of-credits error handler.

SDK v0.0.65 optionalizes provider adapters

The SDK bullet reads: 'Claude Code and Codex provider SDKs are now optional peer dependencies loaded on demand, dramatically cutting install size.' In practical terms this pulls the Anthropic Claude Code adapter and the OpenAI Codex adapter out of the required dependency graph so npm no longer downloads either bundle unless the user actually asks Cline to talk to Claude Code or Codex. The corresponding CLI v3.0.45 release note quantifies the win: 'Smaller install: the Claude Code and Codex providers are now optional and loaded on demand, cutting npm i -g cline from ~640MB to ~285MB.' The 355 MB drop makes CI images and pre-baked base containers roughly 55 percent smaller.

"Smaller install: the Claude Code and Codex providers are now optional and loaded on demand, cutting npm i -g cline from ~640MB to ~285MB."— Cline CLI v3.0.45 release notes

Kimi K3, OAuth retry, team-run tightening

The SDK adds Moonshot's Kimi K3 to the bundled ClinePass model fallback, so a Cline agent that runs out of credit on its primary provider cascades to K3 alongside the existing fallbacks. Runs 'now retry once after refreshing expired OAuth credentials,' silencing a class of spurious failure that hit users whose OAuth tokens expired mid-agent-turn. Two team-run changes tighten multi-agent safety: 'the spawn tool is no longer exposed to teammate agents,' meaning a subagent inside a Cline team run can no longer spawn its own subagents through the tool interface, and 'errored teammate runs now report as failed instead of completed' so a team parent can distinguish a crashed teammate from a clean finish. A Windows shell-command parsing mismatch is also fixed, and the bundled model catalog gets a refresh.

"Team runs: the spawn tool is no longer exposed to teammate agents."— Cline SDK v0.0.65 changelog

CLI patches

cli-v3.0.45 rolls SDK v0.0.65 into the top-level cline binary, exposes hub status output with version numbers, and inherits the SDK's smaller install and Kimi K3 fallback. cli-v3.0.46, tagged at 06:26 UTC roughly two hours later, is a single-bullet patch: 'Fixed out-of-credits detection so the CLI reliably recognizes the Cline API's real insufficient_credits (402) error and shows the add credits card instead of a generic error.' The 402 handler had been swallowing insufficient_credits responses into a generic failure message, giving users no direction on how to recover. The release also ships a new @cline/ui agent-chat component set with Storybook and npm packaging for third-party embedders.

TechTrend

Moore Threads MTT C256 debuts at WAIC with 256 GPUs in a single-layer scale-up domain

Sina Finance · 2026-07-19 · Source ↗

Moore Threads unveiled the MTT C256 hyper-node at WAIC 2026, packing 128 GPUs per cabinet across two racks with sub-microsecond inter-card latency and pitching it as the first Chinese-designed platform to break past the industry-common 64-card single-layer limit.

Moore Threads pulled the wraps off the MTT C256 hyper-node at the 2026 World Artificial Intelligence Conference in Shanghai, positioning it as the first Chinese-designed GPU platform to sustain a single-layer scale-up network across 256 cards. A single 2U node houses eight GPUs; a full-width rack packs 128 GPUs with switched interconnect built into the chassis, and two racks lock together via blind-mate optical connectors to reach the full 256-card domain. The integrated compute-and-switch design collapses what typically ships as separate GPU trays and top-of-rack fabric switches into a single high-density enclosure.

Interconnect and latency

Moore Threads compresses inter-card round-trip latency to sub-microsecond levels within the 256-card domain and cites headroom for cluster expansion from ten thousand to a hundred thousand cards through scale-out layers stacked on top. Bandwidth utilization improves against traditional multi-layer leaf-spine designs because every card sees every peer through a single fabric hop rather than traversing two or three switches. The design breaks past the industry-common 64-card single-layer limit that Alibaba's ALink and Huawei's Lingqu also target with their own architectures.

"First-of-its-kind single-layer scale-up network with 256 cards fully interconnected, and inter-card communication latency compressed to sub-microsecond levels."— Moore Threads MTT C256 technical brief

Three-factory positioning

Moore Threads pitches the C256 across three named workloads. The Model Factory handles pretraining and reinforcement learning for frontier models. The Token Factory serves mixture-of-experts inference with million-token context windows where cross-card KV cache traffic dominates. The Agent Factory targets high-concurrency serving with per-user throughput above 100 tokens per second, betting that agent workloads over the next 12 months will spike concurrent session counts inside enterprise deployments.

"Compatible with existing intelligent computing software and hardware ecosystems, supporting smooth cluster expansion from ten thousand to one hundred thousand card levels."— Moore Threads WAIC 2026 launch briefing

Engineering foundation and ecosystem

The company anchors credibility on cluster-scale RAS work drawn from operational Kua'e training clusters running FP8 mixed precision workloads on the MTT S5000 lineage. Advanced liquid cooling, blind-mate connectors, and cluster management sit as prerequisites for the density claim rather than headline features. The 2U node shape and the compatibility layer keep the C256 addressable through existing software targeting Moore Threads' MUSA runtime, allowing customers already running S4000 or S5000 clusters to graft C256 racks onto in-flight training runs without a full stack rewrite. WAIC 2026 puts the C256 alongside Huawei's Atlas 950 SuperPoD, Alibaba's Panjiu AL128, and Biren's photonic supernode preview as the four Chinese entries in the 2026 super-node race that Digitimes framed as China's pivot from single-chip FLOPS chasing to system-level scale-up.

Tech

Hugging Face top weekend paper: LongStraw pushes on-policy RL past 2 million tokens on 8xH100

Hugging Face Papers · 2026-07-19 · Source ↗

Mind Lab's LongStraw takes the top Hugging Face daily papers slot at 176 upvotes for lifting on-policy RL from 800K to 2 million tokens on a fixed 8xH100 budget, rearranging rollout and gradient computation into disjoint memory partitions with an NVLink KV spill.

Mind Lab's LongStraw (arXiv 2607.14952) posts 176 upvotes across the weekend on Hugging Face and holds the top slot on the daily papers list. The paper attacks a specific bottleneck in on-policy RL for reasoning-heavy language models: the KV cache for million-token rollouts blows through GPU memory long before the gradient update phase. LongStraw rearranges the rollout and gradient computation into two disjoint memory partitions, spills the rollout KV cache to host memory during the update step, and pipelines the two phases across a rolling window so that no single GPU ever holds both the full rollout tensor and the full gradient graph. The technique lifts effective RL context length from roughly 800K tokens under vanilla GRPO to more than 2M tokens on the same 8xH100 node without changing the underlying policy architecture.

"We lift effective RL context length from roughly 800K tokens under vanilla GRPO to more than 2M tokens on the same 8xH100 node without changing the underlying policy architecture."— LongStraw abstract

Results and comparisons

LongStraw reports RL fine-tuning of a Qwen-3.7-32B base to full 2M context on a fixed 8xH100 80 GB node with wall-clock overhead within 12 percent of a 512K baseline. Held-out benchmarks include RULER, LongBench-v3, and the newer InfiniteBench-Reason suite. On InfiniteBench-Reason, the LongStraw-trained model gains 14 to 21 percentage points against SFT-only baselines and 6 to 9 points against short-context RL, with the largest gains on multi-hop chains that require reading and cross-referencing distant portions of the prompt. Against SEED, the sibling self-evolving distillation paper released the same week, LongStraw wins on any task that requires the full context but loses on tasks that fit in 128K, where SEED's simpler pipeline benefits from a stronger teacher signal.

"On InfiniteBench-Reason, the LongStraw-trained model gains 14 to 21 percentage points against SFT-only baselines."— LongStraw results section

Runner-up papers on the same list

The weekend list also elevates Alaya Studio's From Pixels to States (407 upvotes), which recasts interactive world models as event-sourced game engines with explicit state deltas rather than pixel-space diffusion, cutting rollout compute by 5 to 8 times on Genie-style benchmarks. VideoChat3 from the Multimedia Computing Group at Nanjing University lands at 145 upvotes for a fully open 4B video multimodal model that beats proprietary VideoLLM-Online on 7 of 9 video QA benchmarks. Ant Group's SearchOS-V1 (65 upvotes) frames open-domain agent collaboration under a formal robustness axiom set and hits state of the art on the Meta-Search bench.

Limitations and open questions

Mind Lab flags three limitations. First, the host-memory KV spill costs 3-4 GB/s of NVLink and PCIe bandwidth in the steady state, which caps the technique for cost-optimized inference-heavy clusters that lack CPU-GPU bandwidth headroom. Second, the paper only validates on Qwen-3.7 and DeepSeek-V4 base models; results on Llama-4 and GPT-OSS families remain untested. Third, the analysis treats speculative decoding as out of scope, and integrating LongStraw with tree search or diffusion-style decoding remains an open problem. The paper releases model checkpoints under Apache 2.0 and a fork of TRL with the memory reshaping patches under a permissive license.

TechTrend

Gizmochina walks through Kimi K3's 896-expert MoE, KDA attention, and the free consumer tier

Gizmochina · 2026-07-19 · Source ↗

Gizmochina catalogs Kimi K3's 896-expert mixture activating 16 per token for ~50B active parameters, the Kimi Delta Attention hybrid linear-attention scheme, native video and image inputs, MXFP4 quantization-aware training, and free consumer access via the Kimi apps and web.

Gizmochina's July 19 explainer put the widest technical lens yet on Kimi K3 for a consumer audience, mapping every architectural choice Moonshot has disclosed since the July 16 debut of what the company calls the largest open-weight AI model ever shipped. K3 packs 2.8 trillion total parameters into a mixture-of-experts topology with 896 experts, activating just 16 experts per request. That yields roughly 50 billion active parameters per token, keeping serving costs within striking distance of models one-tenth the size while giving the model the raw specialization to compete with Fable 5 and GPT-5.6 Sol.

Kimi Delta Attention and long-context handling

The context window is 1 million tokens, matched to Claude Fable 5 and Google's Gemini 3.5 Pro, and Kimi K3 runs a new attention scheme called Kimi Delta Attention, a hybrid linear-attention design that trims compute across long conversations without dropping accuracy on retrieval or reasoning at depth. The Hugging Face MXFP4 quantization blog from Moonshot puts the raw weights at 594 GB in BF16 and roughly 1.4 TB when unpacked for inference, with community GGUF Q4 quantizations likely to land in the 300 to 400 GB range for hobbyist rigs. Practical single-node deployment needs at minimum 4 to 8 H100 80 GB GPUs; a full production cluster runs 8 nodes of 8 H100s each for 5.12 TB of aggregate memory.

"Kimi K3 is our most capable model to date, with 2.8 trillion parameters."— Moonshot AI announcement

Free access and pricing rails

Consumers can hit K3 without a developer account through the Kimi iOS and Android apps, the kimi.com website, and the Kimi Work desktop app, all on the free tier with usage caps. Developers pay $3 per million input tokens on cache-miss, $0.30 per million on cache-hit, and $15 per million output tokens on the Kimi Open Platform API. Those rails put K3 at roughly 30 percent of Claude Fable 5's forthcoming $10 input and $50 output usage-credit rates, and 50 percent of GPT-5.6 Sol's roughly $6 in and $30 out band. Bank of America analyst Alex Liu wrote in a client note cited by CNBC that K3 shows large-scale pre-training combined with architectural work can still deliver step-change gains for Chinese flagships despite compute constraints.

"K3's pelican only has one reasoning effort right now, 'max' — and it shows. The pelican cost 25 cents."— Simon Willison, July 16 SVG benchmark

Benchmark landing zone

Artificial Analysis places K3's intelligence level in a tie with Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, behind only Fable 5 and GPT-5.6 Sol. On LMArena's Frontend Code Arena, K3 opened at Elo 1,679 across 1,757 votes, taking the top slot and topping Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. On the general text leaderboard, K3 posted 1,486 Elo across 3,026 votes, with 1,530 on coding-specific prompts and 1,506 on hard prompts. GDPval-AA v2 came in at 1,687, GPQA-Diamond at 93.5 percent, MMMU-Pro at 81.6 percent, SWE Marathon at 42.0, and DeepSearchQA F1 at 95.0.

ApplicationTrend

Chinese AI vendors abandon the 40-yuan flat tier for 200-yuan premium plans at WAIC 2026

CGTN · 2026-07-19 · Source ↗

MiniMax, iFlytek, Moonshot and Alibaba Cloud told CGTN at WAIC that the 40-yuan monthly subscription tier is unsustainable. An Alibaba Cloud staffer compared the shift to mobile's 3G-to-4G transition, and MiniMax product leads said flat-rate subscriptions no longer cover inference costs.

At the World Artificial Intelligence Conference in Shanghai on July 19, Chinese AI vendors publicly pivoted from the price-war posture that defined the first half of 2026 to a market-segmentation model that reserves flagship-model access for premium tiers while pushing budget users onto pay-per-token pricing. CGTN reported the shift the same day, documenting the collapse of the 40-yuan (approximately $5.50) monthly all-you-can-eat plan across MiniMax, Zhipu, iFlytek and multiple coding-agent vendors.

Moonshot leads the premium reset

Moonshot AI's Kimi K3, launched two days before the WAIC opening, prices at $3 per million input tokens and $15 per million output tokens with a $0.30 cache-hit rate, roughly half of Claude Fable 5's rate card but three times MiniMax's prior consumer tier. K3 topped the LMArena Frontend Coding leaderboard on launch at Elo 1,679 against Fable 5's 1,631 and GPT-5.6 Sol's 1,618, and Moonshot has explicitly positioned the model as a premium offering rather than a discount competitor.

"Models keep improving. GPT-3.5 cost $20 a month back in the day."— Bai Chuanxu, MiniMax product architect, WAIC exhibition floor

MiniMax, iFlytek and Alibaba Cloud sketch the new architecture

MiniMax product architect Bai Chuanxu told CGTN reporters that the 40-yuan flat-rate tier is an unsustainable transitional artifact. iFlytek product manager Ru Huiya highlighted a new agent desktop product that offers a daily free quota for light users but converts intensive users to per-token billing. An Alibaba Cloud staffer at the WAIC booth compared the pricing evolution to mobile telecom's 3G-to-4G transition, assessing that the 40-yuan tier is unlikely to persist while a 200-yuan (approximately $27.50) tier has better commercial prospects. The Solo App Geek Camp offered free AI courses co-organized by Douyin, Bilibili, Alibaba and Jimeng, but positioned these as consumer on-ramps to premium paid tiers.

"The 40-yuan tier is unlikely sustainable; the 200-yuan tier has better prospects."— Alibaba Cloud staffer, WAIC 2026 exhibition floor

Market implications

The pricing reset follows a first-half 2026 in which many early 40-yuan coding plans were quietly discontinued or made difficult to purchase, and coincides with the Hong Kong lock-up expiry that dragged Zhipu and MiniMax shares down 12 percent and 3.3 percent respectively on July 17. CGTN framed the shift as a Chinese AI industry maturation moment, with premium models carving a high-performance tier, budget users moving to pay-per-token, and fixed-price subscriptions shrinking. Anthropic's July 20 permanent bundling of Claude Fable 5 into Max and Team Premium at 50 percent of usage limits, announced Sunday, mirrors the Chinese premium-tiering move at the international frontier.

ApplicationTech

AGIBOT unveils A3 Ultra humanoid on Jetson Thor and deploys 30-plus robots across WAIC venues

Robotics Tomorrow · 2026-07-19 · Source ↗

AGIBOT unveiled a 1.74-meter, 60-kilogram A3 Ultra humanoid with 51 degrees of freedom on Nvidia Jetson Thor T3000 and deployed more than 30 robots across WAIC 2026 venues. Co-founder Peng Zhihui reframed the industry metric from one-shot demos to reliable real-world integration.

AGIBOT (Innovation) Technology unveiled four new embodied-AI products at the World Artificial Intelligence Conference in Shanghai on July 19 and deployed more than 30 robots across the Shanghai World Expo Exhibition and Convention Center and West Bund venues to handle visitor guidance, information services and live demonstrations of table tennis, boxing and beverage service.

A3 Ultra headlines the launch

The flagship AGIBOT A3 Ultra is a 1.74-meter, 60-kilogram full-size humanoid with 51 active degrees of freedom, 5-kilogram payload per arm and up to 8 hours combined operating time. The perception stack combines 3D LiDAR, RGB-D cameras, fisheye cameras and binocular vision, with positioning via GPS, RTK and UWB. Compute runs on the NVIDIA Jetson Thor T3000 announced two days earlier at Nvidia's WAIC keynote at 865 FP4 teraflops. Charging supports direct charging, hot-swap battery replacement and autonomous docking. Alongside A3 Ultra, AGIBOT introduced the Lingxi X2 Edu educational-tier humanoid, the G2 Max heavy-payload industrial robot, and the OmniHand 3 Ultra-M direct-drive dexterous manipulation system.

"The industry should no longer judge embodied AI by whether a robot can complete an impressive demonstration once, but by whether it can be reliably integrated into real operating environments."— Peng Zhihui, AGIBOT co-founder and CTO

Peng Zhihui reframes the metric

The 30-plus robot deployment across WAIC venues represents the largest single-venue humanoid rollout ever attempted at a Chinese AI conference, with the fleet visible to WAIC's more than 200,000 expected attendees over the July 17-20 conference. On July 19, AGIBOT also hosted the WAIC 2026 Embodied AI Forum at the Golden Hall of the Shanghai World Expo Center.

"Over 30 AGIBOT robots are operating across WAIC venues providing visitor guidance, information services, and demonstrations."— AGIBOT WAIC 2026 briefing

Commercial and geopolitical context

AGIBOT was singled out in the WAIC exhibition catalog as the 'Gem of the Exhibition' and is positioning A3 Ultra for both domestic Chinese and international sales. Techtimes reported the same weekend that AGIBOT's global export push runs alongside China's spy-law obligations for domestic robotics vendors, complicating US and European enterprise procurement. Competing Chinese humanoid vendors including Unitree Robotics, UBTech and LimX are pursuing near-term IPOs on the Hong Kong exchange, with CNBC on July 13 noting the sector's aggressive listing pipeline. Tesla's Optimus Gen 3 production ramp at Fremont was still targeted for late July to August 2026 per Elon Musk's own guidance, and Figure AI's Figure 03 was extending its BMW Spartanburg deployment. AGIBOT's WAIC scale-out places the Shanghai vendor ahead of Western peers on venue-scale deployment demonstrations.

Trend

Yang Zhilin turned down Apple to build Moonshot, CMU advisor confirms

The National · 2026-07-19 · Source ↗

Carnegie Mellon and Toronto professor Russ Salakhutdinov told The National that Moonshot founder Yang Zhilin had multiple US tech job offers including Apple as he finished his 2019 PhD but chose to return to China. He also flagged US immigration friction as a factor pushing top PhDs home.

The National published a profile Sunday, July 19, of Yang Zhilin, the 34-year-old CEO of Moonshot AI, drawing on fresh comments from Russ Salakhutdinov, the University of Toronto and Carnegie Mellon computer scientist who supervised Yang's four-year PhD program at CMU. Salakhutdinov told the paper Yang had 'several opportunities to work for US technology companies, including Apple,' as he wrapped up his doctorate in 2019, but chose to return to China. 'Zhilin was quite determined to go back and build a start-up,' the professor said. 'I remember him telling me that if he didn't at least try starting his own company, he would regret it for the rest of his life.'

"Yang made fundamental contributions to machine learning during his time at CMU."— Russ Salakhutdinov, on Yang's PhD work

The CMU years and industry internships

Yang was born in 1992 in Shantou, Guangdong, and completed his undergraduate degree at Tsinghua University before moving to Pittsburgh to work under Salakhutdinov and William Cohen. He finished the doctorate in four years and made what Salakhutdinov called 'fundamental contributions to machine learning,' and interned at both Google Brain and Meta AI Research during his studies. Salakhutdinov also flagged the US immigration process as a factor pushing top graduates toward home.

"The US immigration process can be quite intimidating and uncertain, even for superstar PhD graduates."— Russ Salakhutdinov, on the talent-flow dynamic

Moonshot's founding and Pink Floyd connection

Yang incorporated Moonshot in March 2023 alongside Tsinghua classmates Zhou Xinyu and Wu Yuxin, naming the company after Pink Floyd's Dark Side of the Moon, his favorite album, and launching on the record's 50th anniversary. Moonshot was valued at approximately $20 billion in a May 2026 round that raised around $2 billion, and is now finalizing a pre-IPO round that could push the mark above $30 billion, according to Bloomberg's parallel July 19 IPO scoop.

Kimi K3 and the frontier gap

The profile ties Yang's biography directly to Kimi K3, the 2.8-trillion-parameter open-weight mixture-of-experts model Moonshot released on July 16, which activates 16 of 896 experts per token and runs a 1-million-token context window. Per Artificial Analysis, K3 sits at intelligence parity with Claude Opus 4.8 and GPT-5.5 and behind only Claude Fable 5 and GPT-5.6 Sol on broad measures. The Kimi K3 API costs $3 per million input tokens and $15 per million output, roughly one-third the price of Fable 5's $10/$50 usage-credit rate that takes effect July 20. The Salakhutdinov quotes echo similar public reflections from Meta chief AI scientist Yann LeCun and DeepMind's Demis Hassabis about the flight of PhD talent to Chinese labs after 2023.

Policy

White House reviews FINRA-style AI safety regulator reporting to the SEC

PYMNTS · 2026-07-19 · Source ↗

White House Chief of Staff Susie Wiles is reviewing a Bloomberg-reported proposal to establish an independent AI safety regulator modeled on FINRA and reporting to the SEC. Google DeepMind CEO Demis Hassabis has publicly endorsed the model; the plan targets cybersecurity risk disclosure obligations.

PYMNTS on July 19 confirmed a Bloomberg News report from July 17 that the White House is weighing a plan to establish an independent regulator focused on artificial intelligence safety. The proposed body would report to the Securities and Exchange Commission and would model its structure on the Financial Industry Regulatory Authority, the self-regulatory organization that supervises broker-dealers. White House Chief of Staff Susie Wiles is reviewing the plan. If the proposal advances, it would represent the most substantive federal step toward AI-specific oversight since President Trump signed Executive Order 14409 on June 2, and it would arrive as the December 2025 executive order challenging state AI laws faces implementation questions.

"An independent regulator focused on artificial intelligence safety, reporting to the Securities and Exchange Commission, modeled after the Financial Industry Regulatory Authority."— Bloomberg News, July 17, cited by PYMNTS

Silicon Valley pressure and government restrictions

The proposal follows a series of federal actions that Silicon Valley companies have called inconsistent. Anthropic faced export controls that affected access to its Fable 5 and Mythos 5 models under national-security review earlier in July. OpenAI made modifications to its GPT-5.6 Sol model at government request. Founders and investors argued that ad hoc restrictions from Commerce and other agencies leave frontier labs unable to plan capacity, pricing, or research roadmaps. A FINRA-style regulator would provide one point of contact for model testing, incident reporting, and dispute resolution, replacing the current mix of Commerce, FTC, and executive-branch directives.

"A U.S.-led standards body to independently test advanced AI models prior to release."— Demis Hassabis, Google DeepMind CEO

Wall Street and cybersecurity angle

The plan is designed to satisfy Wall Street firms concerned about AI-related cybersecurity risks, including the emergence of autonomous AI attackers documented by Check Point and Hugging Face during the past month. Placing the regulator under SEC jurisdiction ties model risk to disclosure obligations for public companies that deploy frontier systems in customer-facing workflows. FINRA supervises roughly 3,400 broker-dealer firms and about 620,000 registered representatives with a $1.4 billion annual budget funded by member fees. A parallel structure for AI would draw registration fees from Anthropic, OpenAI, Google DeepMind, xAI, Meta AI, and smaller frontier labs.

Industry backing and open questions

Google DeepMind CEO Demis Hassabis recently called for a US-led standards body to test advanced AI models before release, proposing a similar FINRA-based organizational model. His remarks provided industry cover for the Wiles review. Anthropic co-founder Dario Amodei has endorsed California SB 53 and New York's RAISE Act as bridges to a federal framework. OpenAI has pushed for federal preemption via the Great American AI Act. The two labs' divergent lobbying positions would collapse if the FINRA-style body absorbed most inspection and disclosure functions from state regulators. The PYMNTS piece did not identify the White House staffer who drafted the proposal, whether Treasury and Commerce concur, or whether Congress would need to authorize the new body via statute. FINRA operates as a private nonprofit rather than a federal agency; replicating that structure for AI would require frontier labs to fund the regulator through membership fees.

PolicyTrend

China's NDRC publishes AI Cooperation Action Plan alongside the Kimi K3 launch

CGTN · 2026-07-19 · Source ↗

Beijing's NDRC released the Action Plan for Artificial Intelligence Cooperation and Development on July 19, converting the K3 open-weight release into a state-policy proof point and citing CGTN figures putting Chinese models at 63 percent of OpenRouter traffic in early July.

CGTN reported on July 19 that China's National Development and Reform Commission released the Action Plan for Artificial Intelligence Cooperation and Development, together with a case-study collection titled China Intelligence, Global Benefits (2026). The action plan positions Beijing as an advocate of what CGTN characterized as a fairer and more inclusive global AI governance system that benefits all humanity rather than select entities. The document arrived one day after the World AI Cooperation Organization opened in Shanghai with 29 signatory states, and it followed Xi Jinping's July 17 WAIC keynote in which he pledged 5,000 AI training placements for developing countries.

Kimi K3 as policy proof point

The CGTN piece framed Moonshot's July 17 Kimi K3 release as the practical embodiment of the NDRC policy. Kimi K3 contains 2.8 trillion parameters, which Moonshot AI called the world's largest open-source AI model to date. The model topped Arena rankings for AI coding tools within hours of release and became the first Chinese AI model to reach the No. 1 position on the platform. Moonshot AI's own framing tied capacity to state policy. The K3 open weights are scheduled to land on Hugging Face by July 27, converting the state slogan of open cooperation into a redistributable artifact.

"The vitality of technology lies in openness and circulation."— Liu Tieyan, president, Zhongguancun Academy

The market share numbers

CGTN cited three data points that anchor the policy pitch. Chinese open-source models have held top positions on developer leaderboards for nine consecutive months. Around 80 percent of companies adopting open-source AI solutions used Chinese AI models, according to international venture capital institutions cited in the piece. Chinese models captured 63 percent of traffic on OpenRouter, a global AI platform, in early July 2026, surpassing US competitors for the first time. Those figures give the NDRC action plan a market foundation it can point to when pitching the 29 WAICO signatories on Chinese model adoption.

"More parameters mean a higher ceiling for capabilities and can deliver more intelligent performance."— Moonshot AI, cited by CGTN

The US-China contrast

The CGTN report placed the Chinese policy against a US backdrop where the White House was reviewing an FINRA-style AI regulator, Anthropic and OpenAI faced federal export restrictions, and the December 2025 Trump executive order was targeting state AI laws. Where the US posture emphasizes controlled release and preemption of state variation, the NDRC plan emphasizes open weights and multilateral cooperation via WAICO. The action plan and Kimi K3's timing convert what could be a diffuse policy statement into a concrete offer: developing-country partners can download frontier-tier weights on July 27 and run them under the WAICO training-placement framework. Bank of America pegged Kimi K3 pricing at roughly half the cost of GPT-5.6 Sol, which reinforces the affordability pitch aimed at Global South buyers.

Safety

WordPress wp2shell exploit circulates as SANS pegs security-team AI adoption at 78 percent

Help Net Security · 2026-07-19 · Source ↗

Help Net Security's week in review paired the WordPress wp2shell public-PoC exploitation window (CVE-2026-63030 + CVE-2026-60137) with SANS's 78 percent AI adoption figure across security teams and OpenAI's internal-only GPT-Red hitting 84 percent success on prompt injection against 13 percent for human red teams.

Help Net Security's July 19 week-in-review lead item covered WordPress 7.0.2, which addressed one critical and one high-severity issue disclosed as wp2shell. CVE-2026-63030 is a REST API batch-route confusion vulnerability introduced in WordPress 6.9. CVE-2026-60137 is a SQL injection in the author__not_in parameter of WP_Query. Chained together, the two flaws take an anonymous request to code execution against a default install. Affected versions include WordPress 6.9.0 through 6.9.4 and 7.0.0 through 7.0.1. WordPress 6.9.5 and 7.0.2 ship the patch and WordPress maintainers enabled forced updates. Searchlight Cyber researcher Adam Kues discovered the chain. More than 500 million WordPress sites were in range before the patch. Public proof-of-concept exploits landed on GitHub within 24 hours of the patch release, and Help Net documented in-the-wild exploitation beginning July 18.

"The attack has no preconditions and can be exploited by an anonymous user in a stock install."— Adam Kues, Searchlight Cyber

AI attack surface items

Help Net grouped four AI-specific safety items in the week's coverage. First, OpenAI's GPT-Red automated red-teaming system outperformed human testers on prompt injection, hitting 84 percent success on prompt-injection tasks against 13 percent for human red teams. OpenAI is keeping GPT-Red internal and will not release it. Second, attackers used AI to run exploitation workflows autonomously, generating thousands of commands across dozens of sessions with minimal human direction, echoing Check Point's 2026 AI Security Report. Third, AI vision-language models were shown vulnerable to overthinking denial-of-service attacks, where crafted prompts force models to expend disproportionate reasoning tokens. Fourth, Hugging Face's own disclosure of an autonomous-AI-driven intrusion into part of its production infrastructure landed the same week, though the intrusion was contained without impact on public models, datasets, or Spaces.

"GPT-Red succeeded in 84 percent of prompt injection attacks, compared with 13 percent for human red teams."— OpenAI internal disclosure, cited by Help Net Security

Defensive AI adoption

The week-in-review also flagged four defensive posture items. Context bombs frustrated AI-driven attacks by forcing the attacker's model into unproductive reasoning. Cynative shipped an open-source agent that refuses write operations by default. SingGuard-NSFA rolled out a guardrails framework specifically for agentic AI safety. The SANS AI Survey found 78 percent of security practitioners integrate AI into daily work in 2026, up from 50 percent in 2025. The 28-percentage-point year-on-year jump reflects the mainstreaming of AI-assisted triage, LLM-based anomaly detection, and agent-driven forensics that Hugging Face used against its own attacker.

"Active AI use in cybersecurity reached 78 percent of practitioners in 2026, up from 50 percent one year prior."— SANS AI Survey 2026

Why this roundup matters

The Help Net compilation frames July 12-19 as a pivot week in three ways. On the offensive side, the wp2shell chain gives anonymous attackers pre-authentication code execution against a WordPress footprint that runs more than 40 percent of the public web, and public PoC availability compresses the exploitation window to hours. On the AI-attack front, OpenAI's GPT-Red disclosure showed that automated red-teaming has crossed the 80 percent success bar on prompt injection, which suggests attackers with similar capabilities are already probing production models. On the defensive front, the SANS 78 percent adoption figure signals that AI in the SOC is no longer a pilot topic. Enterprise incident-response teams now run LLM-based triage, and Hugging Face's disclosure showed the same tooling can cut hours out of forensic timelines.

TrendEconomic

Japan Times syndicates Bloomberg: Kimi upends conventional wisdom on US AI lead over China

The Japan Times · 2026-07-19 · Source ↗

The Japan Times syndicated Bloomberg's analysis Sunday framing Kimi K3 as a second DeepSeek moment that erased the year-long assumption of a US frontier lead. Chinese labs match or exceed Fable 5 and Sol on several benchmarks despite operating without top-tier Nvidia GPUs.

The Japan Times ran the Bloomberg-syndicated 'Moonshot's Kimi upends conventional wisdom on US lead over China' analysis on Sunday, July 19, keeping the story alive in Asian markets over the weekend after Friday's global equities selloff. The piece frames Kimi K3 as the second 'DeepSeek moment' of the past 18 months, a Chinese open-weight release that matches or exceeds American frontier proprietary models on several benchmarks and forces investors to reprice the assumption that Nvidia GPUs are the only path to state-of-the-art AI.

The K3 benchmark case

K3 opened at Elo 1,679 on LMArena's Frontend Code Arena, taking the top slot ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618, and Artificial Analysis has K3 tied with Claude Opus 4.8 and GPT-5.5 on overall intelligence. Bank of America analysts led by Alex Liu wrote that K3 shows large-scale pretraining combined with architectural work, in this case Moonshot's Kimi Delta Attention hybrid and 896-expert mixture, can still deliver step-change gains for Chinese flagships even without unrestricted access to H100 and B200 chips.

"If capable AI is becoming free, the ~$700 billion hyperscalers are spending may not pay back."— Bloomberg syndicated analysis

Friday's coordinated selloff

The Philadelphia Semiconductor Index dropped up to 5.7 percent Friday, extending a decline that put the 30-stock benchmark down more than 20 percent from its late-June record, marking a fresh bear market for semiconductors. TSMC fell 7 percent despite reporting a 77 percent Q2 profit surge just 24 hours earlier, SoftBank shed 9.0 percent, Nvidia gave up 1.2 percent and briefly ceded the world's-most-valuable-company crown to Apple at $4.88 trillion intraday, Alphabet fell 4.4 percent, Alibaba 4 percent, Meta 2.4 percent, and Chinese peers Z.ai plunged 27 percent and MiniMax 16 percent in Hong Kong.

"Chinese open-source models have crossed 10 billion cumulative downloads."— MiniMax executive, WAIC 2026 keynote

The DeepSeek parallel and compute constraint framing

The Japan Times piece explicitly draws the parallel to DeepSeek's January 2025 R1 release, when a Chinese open-weight reasoning model performed comparably to US rivals at a fraction of the cost and briefly wiped $1 trillion off US tech market capitalization on a single trading day. The Kimi K3 architecture works around Nvidia H100 shortages by running MXFP4 quantization-aware training from the supervised fine-tuning stage, which trims memory footprint roughly 4x versus FP16 and lets Chinese labs deploy the 2.8-trillion-parameter model on Huawei Ascend, AMD MI400, and Nvidia Blackwell targets in parallel. Alibaba's T-Head open-sourced its SAIL AI software stack for its Zhenwu accelerators at WAIC the same week, and Biren previewed a photonic 1,024-GPU supernode designed to challenge Huawei's Atlas 950 SuperPoD.