Capital & Industrial Strategy

The memory chip sector has produced two landmark capital markets events in a single week. SK Hynix is targeting a $29.4 billion US depositary receipt listing to capture investor appetite for AI-adjacent hardware, following that announcement with confirmation that its high-bandwidth memory products remain supply-constrained against surging demand from hyperscalers. Simultaneously, Micron reported Q3 gross margins of 84.9% — up from 39% a year prior — and guided Q4 revenue to approximately $50 billion, crushing analyst estimates. The combined announcements triggered what Reuters estimated as a $400 billion rally in AI chip stocks. Bloomberg CNBC

The strategic read here is straightforward: HBM has become the binding constraint in AI accelerator systems, and the two companies that dominate its production are capturing supernormal margins as a result. Kioxia's announcement that it will offer US depositary shares in spring 2027 reflects the same dynamic — memory hardware companies are racing to tap US capital markets while the valuation premium is available. Bloomberg The SK Hynix listing is a confirmed intention; terms and regulatory clearances remain pending.

Qualcomm has confirmed a $3.9 billion all-stock acquisition of Modular, the AI compiler and software infrastructure startup, alongside a projection of more than $15 billion in annual data center chip revenue by fiscal 2029 and the addition of Meta as a customer. WSJ Bloomberg The strategic logic is explicit: Modular's MAX engine and MLIR-based compiler stack allows Qualcomm to offer customers hardware-software co-optimisation across heterogeneous compute — a capability that has been central to Nvidia's moat via CUDA.

Qualcomm's data center revenue is currently negligible, making the $15 billion projection ambitious but not implausible if it can leverage Modular's software to unlock deployment on its Orion server chips. The acquisition is confirmed in terms; regulatory review under current US antitrust posture appears manageable given Qualcomm is not acquiring a dominant incumbent. The deal signals that the next phase of AI chip competition will be fought at the software layer — raw silicon performance is necessary but insufficient without the toolchain to deploy it efficiently across diverse workloads. CNBC

Eight months after announcing a partnership, OpenAI and Broadcom have revealed Jalapeño, a custom ASIC designed specifically for OpenAI's inference workloads. TechCrunch WSJ The chip reflects OpenAI's stated ambition to 'build the full stack' — reducing its structural dependency on Nvidia for inference, where the economics differ fundamentally from training. Inference at OpenAI's scale represents an enormous ongoing cost, and even modest efficiency gains on custom silicon translate to hundreds of millions in annual savings.

For Broadcom, the deal is strategically critical at a moment when its stock has underperformed. Broadcom has positioned itself as the preferred partner for hyperscalers and frontier labs seeking custom AI ASICs — Google's TPUs and now OpenAI's Jalapeño both run through Broadcom's custom silicon design capability. CNBC The competitive implication is clear: as the largest AI labs develop proprietary inference silicon, Nvidia faces structural erosion in its inference market share over a multi-year horizon, even as training remains Nvidia-dominated for now.

Anthropic has formally accused Alibaba and its AI unit of conducting 'the largest known distillation attack' on its Claude models, using thousands of fraudulent accounts to systematically extract model capabilities in what Anthropic characterises as an illicit campaign to replicate Claude's outputs without licensing access. Bloomberg WSJ FT Distillation attacks — training a smaller model on the outputs of a larger proprietary model — represent an increasingly significant vector for IP circumvention that existing legal frameworks are ill-equipped to address.

The timing is politically charged. Anthropic's strained relationship with the Trump White House — reported separately, with CEO Dario Amodei reportedly sidelined from high-stakes meetings in favour of co-founder Tom Brown — means the company cannot easily leverage government support. Wired Meanwhile, Chinese firm 360's concurrent claim to have developed tools matching Anthropic's Mythos model suggests that distillation campaigns may have already yielded competitive capability gains. For capital allocators, this raises the question of how defensible frontier model IP actually is — and whether model weights rather than outputs represent the only truly protectable asset.

Abu Dhabi's MGX has raised close to $50 billion from regional and global investors, positioning itself as one of the largest dedicated AI infrastructure funds ever assembled. Bloomberg The raise is reported by Bloomberg as near-complete from investor commitments, though deployment timing and specific mandates have not been publicly confirmed. Gulf sovereign capital has now become a structural — not opportunistic — force in global AI infrastructure financing, with MGX joining Mubadala and Saudi Arabia's Public Investment Fund as major AI capital allocators.

The strategic implication for the competitive landscape is significant. This capital is not passive portfolio allocation; it is oriented toward securing AI infrastructure — data centres, energy, semiconductor supply chains — that gives Gulf states geopolitical leverage in the AI era. For US and European AI companies, Gulf sovereign funds represent both a critical financing source and a potential conduit for technology transfer that Western governments will scrutinise closely.