Compute & Infrastructure
Top Line
Huawei is accelerating its next-generation AI chip launch into 2027, compressing the timeline by several months, signalling that US export controls are failing to prevent China from closing the gap with NVIDIA on domestic AI silicon.
SK hynix and Intel are in reported talks to establish HBM manufacturing in the United States — potentially via a lease of Intel's Ohio fab — marking a critical step toward onshoring the memory supply chain that underpins AI accelerator production.
Jensen Huang publicly assessed that China will develop advanced lithography tooling by 2030, a statement with profound implications for the long-term efficacy of ASML export restrictions as a containment mechanism.
India's government is formally launching its $30 billion chipmaking push at a New Delhi investor summit, with Modi seeking to position India as a credible alternative node in the global semiconductor supply chain.
A new report projects AI data centre e-waste could fill 23 million shipping containers by 2050, introducing a material long-term regulatory and reputational risk for hyperscalers currently racing to expand hardware capacity.
Key Developments
Huawei's Accelerated AI Chip Timeline Challenges the Export Control Thesis
Huawei is pulling forward the launch of its next-generation AI chip into 2027, cutting the original schedule by several months, according to Bloomberg. The chip is positioned as China's most competitive answer to NVIDIA's dominant AI accelerator lineup, targeting both domestic replacement demand and, ambitiously, global markets. This acceleration is notable because it comes despite ongoing restrictions on advanced semiconductor equipment exports — indicating Huawei has developed sufficient process workarounds or alternative tooling pathways to maintain its development cadence.
The strategic context deepens with NVIDIA CEO Jensen Huang's separate public assessment that China will achieve indigenous advanced lithography capability by 2030, telling audiences that such an achievement 'is just a matter of time' and that China is 'already there' in terms of trajectory, per Tom's Hardware. If Huang's timeline is accurate, the window during which ASML's EUV export ban functions as a meaningful chokepoint is narrower than Western policy architects have assumed. These two data points together — Huawei's accelerated roadmap and Huang's candid assessment — represent a significant challenge to the containment-through-equipment-denial strategy.
HBM Onshoring: SK Hynix-Intel Talks Signal a Strategic Restructuring of the Memory Supply Chain
SK hynix is in active discussions with Intel about establishing HBM production on US soil, with options reportedly including leasing Intel's under-utilised Ohio fab or structuring a joint venture with hyperscaler participation, per Tom's Hardware. Neither company has confirmed the talks. The timing is politically charged: South Korea and the US are in ongoing trade negotiations, and SK hynix risks complicating its government's negotiating position by appearing to yield to US industrial policy pressure before a deal is formalised.
The strategic logic is nonetheless compelling. HBM — the stacked memory architecture that is essential to NVIDIA's H100 and successor GPUs — is currently produced almost exclusively in South Korea by SK hynix and Samsung, and in limited volumes by Micron. A US domestic HBM source would substantially reduce the supply chain concentration risk that hyperscalers and the US Department of Defense have flagged as a critical vulnerability. Intel's Ohio footprint, originally planned under the CHIPS Act for logic manufacturing, could be repurposed if Intel's own fab utilisation outlook continues to disappoint, making the lease or JV structure financially attractive for both parties.
India's $30 Billion Chip Bet: Sovereign Ambition Meets Execution Risk
Prime Minister Modi is personally inaugurating a major semiconductor investor conference in New Delhi today, using the event as a platform to attract global chipmakers to India's $30 billion state-backed semiconductor initiative, per Bloomberg. India's programme includes production-linked incentives and attempts to fast-track approvals for fab construction. The Tata Electronics-PSMC joint venture in Gujarat and the CG Power-Renesas assembly and test facility represent the confirmed projects currently underway, though both are at earlier, less advanced process nodes than the frontier capacity being built in Taiwan, South Korea, and the US.
India's play is primarily positioned around assembly, testing, and packaging — the back-end of the supply chain — rather than leading-edge logic fabrication. This is a realistic near-term strategy given the talent, infrastructure, and water availability constraints that make front-end fab construction in India a longer-horizon challenge. The geopolitical rationale is clear: the US, Japan, and the EU all have strategic incentives to see India emerge as a credible third supply chain node outside China and Taiwan. Whether $30 billion in government commitment translates into operational capacity at meaningful scale remains to be demonstrated.
The Memory Architecture Divergence: HBM Dominance Is Not Inevitable at the Edge
Two distinct developments this week point to meaningful structural pressure on HBM's assumed dominance across all AI inference contexts. First, Nanya-backed DRAM designer PieceMakers began trading in Taipei after building its business case around a contrarian thesis: that edge AI inference will not converge on HBM, and that hybrid-bonding-based memory — which fuses a DRAM stack directly onto the processor die — offers a superior cost and power profile for that segment, per Tom's Hardware. Second, Fairphone and other smaller device manufacturers are reporting that memory now accounts for 60% of materials cost in some devices, with AI-driven shortages forcing complete redesigns of motherboard architectures and procurement strategies, per Tom's Hardware.
The Fairphone data point is particularly instructive as an early indicator of supply chain stress radiating outward from hyperscaler AI hardware demand into consumer electronics. The report also notes an uptick in counterfeit memory chips entering the market as shortages intensify — a supply chain integrity problem that adds qualification and testing costs at every tier. Together, these signals suggest that the memory market is bifurcating: a high-bandwidth, high-cost HBM tier serving data centre AI training and inference, and a contested, cost-sensitive LPDDR and hybrid-bonded tier serving the rapidly expanding edge AI device market.
Southeast Asian Compute Buildout: Debt-Financed Capacity Expansion Accelerates
GMI Cloud, an NVIDIA-partnered cloud provider, is seeking a $300 million loan to finance chip procurement for its Thailand facility, according to Bloomberg. Bloomberg notes this is part of a 'growing list of similar deals in Asia' — a pattern that warrants scrutiny. Debt-financed GPU procurement by regional cloud operators reflects both the genuine demand signal in Southeast Asian AI markets and the structural reality that at current NVIDIA accelerator pricing, even well-capitalised operators cannot fund chip inventories from cash flow alone.
This model carries concentration risk: the underlying assets (NVIDIA GPUs) depreciate rapidly as new generations arrive, the debt service depends on sustained hyperscaler or enterprise demand that is not yet proven at scale in these markets, and the NVIDIA supply relationship creates an asymmetric dependency. Thailand's positioning as a data centre hub benefits from relatively stable power costs and government incentives, but the wave of debt-financed compute buildout across the region deserves monitoring as a potential source of financial instability if AI inference demand growth in Asia-Pacific trails the pace implied by current capital commitments.
Signals & Trends
E-Waste Is Emerging as a Hard Regulatory Constraint on AI Hardware Refresh Cycles
A newly published report, covered by The Verge, projects AI data centre e-waste reaching the equivalent of 23 million shipping containers by 2050 — substantially above prior estimates. This is not a distant environmental abstraction: the EU's existing WEEE regulations are already being updated to capture data centre equipment explicitly, and several US states are advancing hardware disposal legislation. For hyperscalers and co-location operators, the strategic implication is that accelerated GPU refresh cycles — currently driven by the rapid cadence of NVIDIA's Hopper-to-Blackwell-to-Rubin transitions — will attract increasing regulatory scrutiny and potentially mandatory take-back costs. This creates a previously underpriced cost in the total cost of ownership calculus for AI infrastructure, and may create competitive advantage for operators who extend hardware lifecycles through inference optimisation software rather than continuous hardware upgrades.
Apple's Potential Server Re-Entry Signals a Broader Trend of Vertically Integrated AI Infrastructure
Reports via The Information, covered by The Verge, indicate Apple is planning a return to server hardware after a 15-year absence, potentially in partnership with NVIDIA. This is analytically significant not primarily as an Apple story but as a pattern: the hyperscalers and platform companies that previously were content to procure infrastructure from Cisco, Dell, and HPE are now pursuing vertical integration into server design and manufacturing — driven by the need to optimise for specific AI workloads and to reduce dependency on third-party supply chains. Apple's move, if confirmed, would follow the logic of Meta's custom AI training hardware, Google's TPU programme, and Microsoft's Maia initiative. The NVIDIA partnership angle is notable because it would represent NVIDIA operating further down the stack into systems integration — a strategic direction that raises questions about NVIDIA's relationship with its existing ODM and OEM partners.
The 2030 Lithography Horizon Is Compressing Strategic Planning Cycles for Western Chip Policy
Jensen Huang's public statement that China will achieve advanced lithography capability by 2030 — framed not as a risk but as a near-certainty — effectively sets a four-year countdown on the utility of equipment export controls as a primary policy tool. Western governments and their semiconductor allies are now implicitly operating on two parallel tracks: accelerating domestic capacity buildout (CHIPS Act, EU Chips Act, India's programme) as a supply chain resilience play that remains valid regardless of Chinese lithography progress, and separately managing a narrowing window during which equipment restrictions meaningfully constrain Chinese frontier chip production. Policy architects who treat 2030 as a hard inflection point will need to have alternative strategic tools — tariff architecture, talent restrictions, software and IP controls — more fully developed before that date.
Explore Other Categories
Read detailed analysis in other strategic domains