Beijing Reverses Meta's $2 Billion Manus.

Mazen Al Sabbagh

— 4 min read
Share

TL;DR

  • Beijing Reverses Acquisition: Tencent leads a consortium to repurchase Manus from Meta at the original $2 billion valuation, finalizing Beijing's unprecedented unwind of a completed US acquisition.
  • Apple Sues OpenAI: OpenAI issued its initial response to Apple's trade-secret suit, as new reporting spotlights a "LOL" email at the core of Apple's allegations.
  • Meta's AI Chip Iris: An internal memo confirms Iris, Meta's inaugural in-house AI chip, enters production in September, aiming to double compute capacity to 14 GW by 2027.
  • Memory Supply Crisis: SK Hynix CEO Kwak Noh-jung warns 2027 will bring the worst memory supply shortage in history, with demand projected to outstrip capacity beyond 2030.
  • Frontier Compute Strain: OpenAI lifted GPT-5.6 usage caps following a traffic surge, while Anthropic's extended Fable 5 access window for paid plans closes tonight.

Lead Story: Beijing Reverses Meta's $2 Billion Manus

Tencent leads a consortium to repurchase Manus from Meta at its original $2 billion valuation, a move that would position it as the largest outside shareholder. This initiative, reported by Bloomberg and the FT, finalizes Beijing’s unprecedented reversal of a completed US acquisition. The group includes Manus's original Chinese investors, HSG (formerly Sequoia Capital China) and ZhenFund, alongside company management, reportedly raising about $1 billion for the buyback (Tech Startups).

This buyback culminates an extraordinary regulatory intervention. Meta finalized the Manus acquisition in December 2025, marking its third-largest deal under Zuckerberg. However, China's foreign-investment security review office rejected the transaction outright in April, ordering its cancellation. Officials cited an intent to "hollow out China's technology base." Founder Xiao Hong remains under an exit ban, with Meta commencing its unwind in June.

No completed American acquisition of an AI startup has previously been forcibly reversed in this manner. Beijing now treats agentic AI as a sovereign asset (TechTimes), mirroring Washington's own deal-blocking posture, and occurring days after the US liberalized chip exports to the UAE. Manus anticipates operational independence from Tencent (Tom's Hardware). Key questions remain regarding the scale of Meta's writedown and whether Beijing will target other foreign stakes in Chinese AI firms.

In Other News

OpenAI responds to Apple's lawsuit, as a "LOL" email gains prominence. OpenAI spokesperson Drew Pusateri issued the company's first public statement on Friday's trade-secret litigation: "We have no interest in other companies' trade secrets." New reports from Bloomberg and Fortune detail Apple's claims that engineer Chang Liu, now in OpenAI's hardware division, retained an unreturned MacBook and exploited a software bug for continuous access to Apple file servers, noting to a former colleague: "LOL, I found out I can access the [network storage], so funny." Apple’s complaint alleges theft "at every level," extending to hardware chief Tang Tan (CNBC).

Meta's proprietary AI chip, Iris, is set for September production. An internal memo, reviewed by Reuters, indicates Iris—co-designed with Broadcom and fabricated by TSMC—completed testing within six weeks. The chip is intended to augment, rather than replace, Nvidia and AMD GPUs for ranking, recommendation, and generative AI workloads (Reuters via Yahoo Finance). The same memo projects a doubling of Meta's total compute capacity to 14 GW by 2027, complementing its projected $145 billion in AI infrastructure spending this year (DCD). Meta shares appreciated about 7% on Friday following the infrastructure expansion news.

SK Hynix CEO forecasts the most severe memory shortage in history for 2027. In a Reuters interview coinciding with Friday's Nasdaq debut, widely circulated this weekend, Kwak Noh-jung stated that 2027 "will be the worst year in the industry's history from the supply perspective," with customer demand anticipated to exceed capacity well beyond 2030 despite aggressive expansion plans (Reuters via US News). High-Bandwidth Memory (HBM) for AI accelerators is the primary driver of this squeeze, echoing similar forecasts from Samsung and Micron (Tom's Hardware)—a sobering note following a $26.5 billion listing.

Frontier compute capacity experiences visible rationing this weekend. With GPT-5.6 traffic nearing twice OpenAI's prior peak, the company removed the five-hour usage cap for Plus, Business, and Pro plans, implemented a one-time usage reset, and delivered Sol efficiency improvements equivalent to approximately 10% more usage (BleepingComputer, Windows Report). Concurrently, Anthropic’s extended window for plan-included Claude Fable 5 access—allowing up to 50% of weekly limits on paid plans—expires tonight at 11:59 PM PT. Subsequently, the model will transition to prepaid credits at $10/$50 per million tokens (Digital Applied). Anthropic has not indicated a further extension.

X / Social Pulse

The Musk–Altman feud extended into the weekend, with Musk accusing Altman of "scamming to a whole new level" and invoking the Apple lawsuit (Benzinga). Altman retorted, suggesting Musk is "the one selling public market investors on short-term space datacenters" (DCD). Developer communities are monitoring tonight's Fable 5 cutoff, debating whether to fully utilize remaining included usage or allocate budget for the new credit system. Kwak Noh-jung's "worst year in history" commentary on memory supply is resonating across PC-builder and homelab communities, as users share observations of escalating RAM and SSD prices.

One to Watch

Gemini 3.5 Pro. Google DeepMind is reportedly targeting July 17 for general availability, a delay from its original June projection. This is a ground-up rebuild, departing from the 2.5 Pro architecture, featuring a 2 million-token context window, and integrating Deep Think reasoning (TechTimes). Should the launch hold, Google will enter a market where its primary rivals are visibly managing capacity constraints.

Quick Hits

  • SambaNova Secures Funding: SambaNova closed the initial tranche of a $1 billion Series F at an $11 billion valuation, led by General Atlantic, with JPMorganChase committing as an on-prem inference partner (TechCrunch).
  • Meta AI Detector Lapses: A Reuters test found Meta's Content Seal detector failed to identify 55% of cropped Muse Image outputs, though Meta labels the tool a "preview" (Gizmodo).
  • ICML 2026 Awards: ICML 2026 concluded Saturday in Seoul, with diffusion-model papers dominating the Outstanding Paper awards and DeepMind's A3C receiving the Test of Time award (AI Front Page).
  • China's Companion AI Rules: China's Interim Measures for Anthropomorphic AI Interaction Services, the first regulatory framework for companion AI covering emotional dependence limits and minor protections, take effect Wednesday (Hogan Lovells).
  • Meta's Canadian Data Center: Meta will construct a $13 billion, one-gigawatt data center in Sturgeon County, Alberta, marking its first such facility in Canada (CBC).

A week commencing with Washington easing chip export controls concludes with Beijing reclaiming an AI champion—both capitals now treat frontier AI as geopolitical territory, not solely commercial venture. Simultaneously, the industry's more fundamental constraint manifests globally: insufficient memory, limited inference capacity, and usage metrics dictating access.

Sources

Lock in. M. mazen@thorterminal.com

Read more