Capital & Industrial Strategy
Top Line
SK Hynix's $26.5 billion US ADR offering — the largest-ever US first-time share sale by a foreign company — confirms that AI infrastructure equities have become a primary vehicle for institutional investors seeking exposure to the compute buildout, even as Taiwan's central bank chief publicly warns of an AI bubble.
Tencent is in advanced talks to become the largest shareholder in Manus after Beijing ordered the reversal of Meta's $2 billion acquisition, a deal that simultaneously illustrates Chinese state intervention as a dealmaking force and the lengths to which Chinese tech giants will go to consolidate domestic AI agent assets.
OpenAI launched GPT-5.6 and confirmed it remains the preferred model powering Microsoft Copilot 365, directly rebutting speculation of a Microsoft breakup — but Anthropic's shift to usage-based pricing for its top consumer model signals that the subsidised-subscription era of frontier AI is ending.
The US government has identified that OpenAI and Google have been supplying AI model access to Singapore-based subsidiaries of Alibaba, Baidu, and Tencent — entities on the US restricted list — creating a significant compliance and regulatory risk for both companies at a moment when the Trump administration is also taking a more active role in reviewing AI models.
Carlyle's sale of a data centre power unit to EQT for a fivefold return, combined with Micron's fresh commitment to US chipmaking investment and Commerce Secretary Lutnick's direct pressure on SK Hynix and Samsung to expand US memory output, signals that AI infrastructure is now as much a political economy story as a market one.
Key Developments
Tencent-Manus Deal Exposes Beijing's Power Over AI M&A and Chinese Capital Flows
Tencent is in advanced negotiations to become the largest shareholder in Manus, the Chinese agentic AI startup, after Beijing ordered the unwinding of Meta's $2 billion acquisition of the company, according to reporting by the Financial Times and confirmed by Bloomberg. This is an announced intention, not a closed deal — terms have not been confirmed. The strategic logic for Tencent is straightforward: Manus represents one of the most credible agentic AI platforms built outside the US hyperscaler ecosystem, and acquiring control of it keeps that capability within the Chinese domestic orbit at a moment when Beijing is actively working to prevent strategic AI assets from passing to foreign ownership.
The episode is significant beyond the single transaction. Beijing's ability to reverse a completed $2 billion acquisition — and then redirect that asset to a domestic champion — demonstrates that the Chinese state operates as a de facto capital allocator in frontier AI, not merely a regulator. For Western acquirers, this should be read as a structural constraint: any acquisition of a Chinese AI company of strategic significance is effectively conditional on state approval, regardless of commercial terms agreed between private parties. Meanwhile, MiniMax is separately raising up to $2 billion through share sales and convertible bonds, and Nexchip's Hong Kong IPO surged 14% on debut, confirming that Chinese AI capital markets remain actively open even as cross-border deal flow narrows.
US AI Model Exports to Blacklisted Chinese Entities Trigger Government Pushback
The Financial Times reports that OpenAI and Google have been supplying AI model access through API and cloud channels to Singapore-based subsidiaries of Alibaba, Baidu, and Tencent — all entities that appear on US government restricted lists, according to FT. The mechanism appears to be geographic arbitrage: Singapore subsidiaries of Chinese groups accessing US AI services, which technically may not trigger the same export controls that would apply to direct China-based access. Simultaneously, the Trump administration is taking a more active role in reviewing AI models prior to commercial deployment, adding a second regulatory vector, per Bloomberg.
The strategic exposure here is significant for both OpenAI and Google. The revenue at risk is real — these are large enterprise API consumers — but the regulatory and reputational risk of being seen to supply strategic AI capabilities to entities on US restricted lists is severe, particularly given the current political environment. The Singapore subsidiary structure is a well-understood workaround in export control regimes, and regulators are clearly now closing that gap. Companies that built revenue projections on broad API access without geographic restriction in their customer base face material compliance restructuring. This also adds friction to the pending IPO narratives around OpenAI and Anthropic, both of which will face investor scrutiny on export control compliance during roadshows.
AI Infrastructure Capital Markets Hit New Intensity — With Bubble Warnings Emerging
SK Hynix's $26.5 billion US ADR offering — confirmed as closed — is the largest first-time US share sale by any foreign company in history, per Bloomberg. The deal was explicitly positioned as an AI infrastructure play, tapping institutional demand for memory exposure in the compute buildout. Italian probe card manufacturer Technoprobe has surged 330% as a direct beneficiary of Nvidia's production volumes, per Bloomberg, representing the pick-and-shovel thesis being applied even to second-order European industrial suppliers. Carlyle's sale of a data centre power unit to EQT for a fivefold return, per FT, confirms that private equity is actively monetising AI infrastructure exposure at peak-cycle multiples. Micron separately announced a new round of US chipmaking investment commitments, with shares rising nearly 5% on the announcement, per CNBC.
Against this backdrop, Taiwan's central bank governor has issued a public warning about AI bubble risk, per Reuters. The tension is structural: capital markets are pricing AI infrastructure as a long-duration growth trade while simultaneously the ROI debate — quantified in a TechCrunch analysis at $3 trillion in claimed value — remains unresolved at the enterprise level. Commerce Secretary Lutnick's direct pressure on SK Hynix and Samsung to expand US memory production, per Bloomberg, illustrates how government industrial strategy is now a direct input into corporate capex decisions in this sector.
Meta's Custom Silicon Strategy and Vertical Integration Ambitions Accelerate
Meta will begin production of its new AI chip in September, according to an internal memo reported by Reuters and confirmed by TechCrunch. The design philosophy is explicitly modular, reflecting Meta's recognition that model architectures are evolving fast enough that rigid custom silicon risks obsolescence before deployment at scale. The chip is part of Meta's stated goal to double its computing capacity. Simultaneously, Meta has launched Muse Spark 1.1, its entry into the AI coding market, targeting large agentic workloads and enterprise code migration — a direct challenge to Anthropic and OpenAI's commercially successful coding products, per CNBC.
Meta's strategic posture is now clearly one of vertical integration across the AI stack: custom silicon to reduce Nvidia dependency, open-weight models to commoditise competitor advantages, proprietary applications to capture enterprise revenue. The coding market entry is notable because it represents Meta moving from infrastructure provider to direct enterprise software competitor — a segment where Anthropic's Claude and OpenAI's products currently command premium pricing. The modular chip approach also directly addresses the core risk in custom silicon programs: that by the time fabrication completes, the target workload has changed. This is precisely the problem that has made Nvidia's programmable GPU architecture so durable.
Pending AI IPOs Dwarf Historical VC Exit Benchmarks — Pricing and Monetisation Models Now in Flux
Analysis from TechCrunch estimates that Anthropic, OpenAI, and SpaceX are collectively expected to generate more exit value than the entire universe of US VC-backed exits since 2000. These are analyst estimates based on current private valuations, not confirmed IPO terms. The scale of these pending liquidity events has direct implications for VC fund return calculations, LP re-commitment decisions, and the capital recycling that will fund the next generation of AI startups. Anthropic's appointment of former Fed Chair Ben Bernanke to its independent trust, per CNBC, is a deliberate governance signal ahead of its public market debut — institutional credibility-building for an organisation that will face intense scrutiny on safety governance from public market investors.
Anthropic's simultaneous move to usage-based pricing for Claude's top consumer model, per Wired, is commercially significant. The flat-subscription model was a customer acquisition strategy; the shift to consumption pricing is a margin optimisation strategy. This transition will be closely watched by enterprise buyers who have built cost models on predictable subscription fees, and by investors trying to model sustainable unit economics for frontier AI companies. Palo Alto Networks CEO Nikesh Arora's public call for a 90% reduction in token costs, per CNBC, reflects the tension between current pricing levels and the economics required for broad enterprise adoption at scale.
Signals & Trends
China's Five-Year Plan Silence on Jobs Is an Implicit Industrial Policy Signal on AI Labour Displacement
China's omission of any numeric urban job creation target from its Five-Year Plan — the first time this has happened in at least three decades — is not an administrative oversight. Read alongside the rapid domestic AI deployment across manufacturing, logistics, and services, it is a quiet acknowledgement that the state no longer believes it can set credible employment targets in an era of accelerating automation. For capital allocators, this matters in two directions: it removes a political constraint that previously moderated the pace of AI adoption in Chinese enterprises, and it signals that Chinese industrial policy will prioritise AI-driven productivity over employment preservation in the next planning cycle. The downstream implication is that Chinese enterprises face less regulatory friction in deploying automation at scale than their Western counterparts, where labour market politics constrain the speed of AI adoption. TCS CEO K. Krithivasan's separate projection that AI could reach 20% of TCS revenue within four to six quarters — with explicit acknowledgement of job role restructuring — suggests that the same dynamic is playing out across Asian IT services, where management is now publicly quantifying AI's displacement effect in revenue terms.
The Nvidia Competitive Paradox: Custom Silicon Proliferation Is Both a Threat and a Validation
The TechCrunch analysis on Nvidia's competitive position frames a structural irony: by proving that compute is the scarce resource in AI, Nvidia has motivated every major hyperscaler and AI lab to develop custom silicon to reduce dependency on it. Meta's September chip production start, combined with the broader custom silicon programs at Google (TPUs), Amazon (Trainium), and Microsoft (Maia), represents a coordinated — if not coordinated — effort to commoditise the one layer of the stack Nvidia controls. The Computacenter share price surge, driven by hardware resale margins on data centre infrastructure, and Technoprobe's 330% surge as a probe card supplier to Nvidia's fabrication process, illustrate that second and third-order beneficiaries of the compute buildout are capturing significant value without bearing frontier model risk. The signal for portfolio construction is that the most durable AI infrastructure returns may not be in the hyperscaler or model layer, but in the industrial supply chain — power, cooling, test equipment, and networking — where competition is less intense and switching costs are higher.
The AI Pricing Ceiling Is Becoming Visible — And It Will Define the Enterprise Adoption Curve
Three separate signals this week converge on the same structural constraint: token economics as the binding limit on enterprise AI adoption at scale. Palo Alto's CEO publicly stated that token costs need to fall 90% for broad enterprise deployment to be viable. Anthropic is moving its best model to usage-based pricing, which will make cost modelling more complex for enterprise buyers. And the $3 trillion ROI question — whether enterprise AI deployment is generating returns commensurate with capital invested — remains contested. The dynamic is familiar from prior enterprise software cycles: early adopters absorb high unit costs as a competitive investment, but broad deployment requires cost curves to fall to a level where the ROI is legible to a CFO on a standard business case. The current pricing environment is still in the early-adopter phase for most enterprises outside financial services and software. JPMorgan's AI portfolio allocation agents — which reportedly beat the 60/40 benchmark in backtests — represent the leading edge of an industry that has both the analytical sophistication and margin structure to absorb current token costs. Most other enterprise sectors do not yet meet that threshold.
Explore Other Categories
Read detailed analysis in other strategic domains