Meta Needs To Rethink Manus

China's National Development and Reform Commission (NDRC) has ordered Meta and the AI startup Manus to unwind their $2 billion acquisition, ruling that the deal violated foreign investment rules and technology export requirements. The decision carries implications well beyond a single transaction. Manus was founded by Chinese engineers, had a Chinese parent company, and maintained offices in Beijing and Wuhan—but was incorporated in Singapore, a structure known in the industry as "VIE restructuring" or, informally, "Singapore washing." The premise was that registering a holding company offshore placed the entity outside Beijing's regulatory jurisdiction. The NDRC's ruling explicitly rejects that premise: where a product is built matters more than where the holding company is registered. The discussion frames this as a deliberate policy signal on multiple levels. First, it closes what analysts describe as an "offshore arbitrage window"—the mechanism by which Chinese founders accessed Western venture capital (Manus had backing from Benchmark) while building products free from Chinese domestic content restrictions. Second, it arrives just weeks before a planned Trump-Xi meeting, suggesting deliberate timing. Third, it mirrors Washington's existing restrictions on Western investment in Chinese AI companies; as one analyst quoted in the discussion puts it, "there is no more middle lane." The practical consequences are significant. Meta has described the two teams as "deeply integrated," with Manus personnel working alongside Meta colleagues in Singapore—making an unwind operationally complex. More broadly, the scrutiny is expected to deter Chinese AI researchers from pursuing similar offshore structures and will make it harder for Chinese startups to attract foreign capital. A Singapore-based consultancy executive quoted in the coverage warns that the decision will make it "increasingly hard for Chinese AI founders who started in China to sit on both sides."

Analyst Ming-Chi Kuo reports that OpenAI is developing smartphone chips in partnership with Qualcomm and MediaTek, with Luxshare Precision Industry serving as the exclusive system co-design and manufacturing partner. Mass production is targeted for 2028, with specifications expected to be finalized by late 2026 or early 2027. The strategic logic, as the discussion covers it, is vertical integration for AI-native devices. Rather than raw processing power, the chips are expected to prioritize on-device AI inference performance—analogous to Google's Tensor chips in Pixel devices. By partnering with established chip designers rather than developing CPUs and GPUs from scratch, OpenAI avoids the most time-intensive and capital-intensive phases of chip development. OpenAI is separately developing data center AI chips with Broadcom. The broader vision described is an AI-agent-first smartphone ecosystem: instead of discrete apps, users would rely on AI agents for tasks, requiring tight integration between on-device and cloud AI. The hardware would serve as a platform to drive subscription growth and developer ecosystem development—a model explicitly compared to Apple's approach. The announcement moved markets: Qualcomm shares jumped 14% in pre-market trading, Luxshare gained 10%, and Apple fell nearly 2%.

Google Cloud CEO Thomas Kurian disclosed that Google controls approximately 25% of global AI computing capacity, comprising roughly 3.8 million TPUs (Tensor Processing Units—Google's proprietary AI accelerators) and 1.3 million GPUs, according to estimates from research firm Epic AI. Microsoft ranks second with approximately 3.2 million Nvidia GPUs. Kurian's argument is that vertical integration—owning the chips, the models (Gemini), and the data center infrastructure—produces structurally better margins than hyperscalers that resell third-party technology. The claim: for every dollar of revenue, Google does not transfer 80% to a model or chip provider. Google Cloud reported 48% revenue growth in Q4 2025 and is projected to exceed $70 billion in 2026, up from $43 billion in 2024. The discussion notes a pointed counterargument from Nvidia CEO Jensen Huang, who criticized Google for not submitting its TPUs to independent benchmarking and suggested that Anthropic accounts for the bulk of TPU demand—implying Google's infrastructure advantage may be narrower in practice than the headline figures suggest. This disagreement remains unresolved in the coverage.

A structural change in how enterprise software is priced is accelerating. By end of 2025, 79 of the 500 software companies tracked by analyst Kyle Poyar—including HubSpot, Adobe, and Salesforce—had introduced usage-based AI fees, more than double the figure from 2024. ServiceNow, Workday, and Atlassian are among those actively transitioning. The driver is straightforward: flat-fee AI subscriptions became economically untenable as customers increased usage, raising costs for vendors. A secondary concern is that AI agents may reduce the number of human seats required, eroding the per-seat licensing model that has underpinned SaaS economics for two decades. Usage is now measured in data processed (tokens or equivalent units) rather than user accounts. The buyer-side consequences are already visible. An AI consultant quoted in the coverage states that most clients find consumption-based pricing leads to costs that "go through the roof really quickly." Uber's CTO disclosed that the company exhausted its full-year AI budget within a few months of 2026.

A one-week internal experiment at Anthropic—Project Deal—allowed AI agents to conduct a classifieds marketplace for 69 employees. Agents using Claude Opus 4.5 (the frontier model) outperformed agents using Claude Haiku 4.5 (the smallest model) in measurable ways: Opus agents closed roughly two more deals on average, and when the same item sold through both model tiers, the Opus agent captured $3.64 more per transaction. Critically, participants represented by weaker agents rated deal fairness almost identically to those with stronger agents—they did not know they were losing out. Anthropic acknowledges this as an "uncomfortable implication" for AI agent commerce at scale and flags additional risks including prompt injection and jailbreaking in higher-stakes commercial environments. --- Key takeaways: - Beijing's NDRC ruling against the Meta-Manus deal effectively ends offshore jurisdiction arbitrage for Chinese AI startups; where a product is developed now determines regulatory exposure, regardless of holding company location. - OpenAI's smartphone chip initiative with Qualcomm and MediaTek signals a move toward full-stack hardware-software integration, targeting an AI-agent-first device paradigm with mass production expected in 2028. - Google's claim to control 25% of global AI compute rests on proprietary TPU infrastructure and vertical integration, but Nvidia's public challenge to TPU benchmarking leaves the performance case contested. - Enterprise SaaS is undergoing a pricing model transition from per-seat subscriptions to consumption-based fees, with documented cases of buyer cost overruns and accelerating vendor adoption across major platforms. - Anthropic's internal agent experiment demonstrates that AI model capability asymmetry in commercial negotiations produces measurable economic disadvantage for users of weaker models—a dynamic participants cannot detect without external disclosure.