Compute & Infrastructure
Top Line
Samsung and SK Hynix posted explosive profit surges — 250-fold and 557% year-over-year respectively — while both companies warn memory shortages will intensify into 2027, signalling that AI infrastructure demand is structurally outpacing supply capacity.
DRAM supply to standard module makers could fall below 30% of 2026 levels by 2027 as manufacturers redirect wafer capacity to HBM and server-grade DRAM, creating a bifurcated memory market with severe consequences for non-AI buyers.
A Nvidia employee has been implicated in the Supermicro chip smuggling scandal, and Moonshot AI reportedly used Blackwell chips to train Kimi K3 by circumventing both US export controls and Chinese import restrictions — illustrating that enforcement gaps remain the primary check on restricted compute access.
Blue Owl's Stack Infrastructure is seeking a A$8.5 billion syndicated loan for a single Melbourne data centre project, underscoring how private capital continues to flow into AI infrastructure at scale despite rising interest rate environments.
Intel has formally completed the RAMP-C defense program on its 18A process, closing a chapter on federally-funded domestic foundry development that tested chips from Nvidia and others — a signal of where sovereign semiconductor strategy now stands.
Key Developments
Memory Market in Structural Shortage: HBM Crowding Out Standard DRAM
Samsung reported chip operating income of 89.2 trillion won for Q2 2026, a more than 250-fold year-over-year increase, and explicitly warned that memory shortages will worsen in 2027. SK Hynix simultaneously reported a 557% profit surge on revenue of 79.32 trillion won, though shares fell as market expectations had priced in even stronger forward guidance. Both results confirm that AI-driven demand for High Bandwidth Memory and server DRAM is consuming the bulk of leading-edge DRAM wafer capacity. Sources: Bloomberg, Tom's Hardware.
The downstream impact is quantified by Apacer's CEO, who projects that DRAM chip allocations to module makers — the vendors supplying consumer, enterprise, and DIY PC markets — could drop by more than 70% year-on-year in 2027. That means conventional DRAM buyers will compete for less than 30% of 2026 supply levels. SSD pricing data corroborates this: NAND flash prices have risen approximately 220% over the past year as AI workloads absorb flash storage capacity. Seagate's response — qualifying 50TB HAMR hard drives for 2028 shipment, with most current production already sold out — indicates storage vendors are racing capacity expansions that won't arrive until demand has compounded further. Sources: Tom's Hardware, Tom's Hardware, Tom's Hardware.
Nvidia Hardware Controls Eroding: Smuggling Scandal and Chinese Circumvention
Two parallel developments reveal mounting pressure on US export control enforcement. An Nvidia employee has been detained and had his home and desk searched in connection with the Supermicro chip smuggling scandal, with allegations of forgery and breach of trust. Nvidia has responded by collapsing its approved buyer list to eliminate potential intermediary smuggling chains — a reactive measure that confirms the distribution network itself is a vulnerability. The implication is that Nvidia's own commercial ecosystem had been penetrated by actors routing restricted hardware. Source: Tom's Hardware.
Separately, Moonshot AI reportedly trained its Kimi K3 model on Nvidia Blackwell chips — hardware that falls under US export restrictions to China and, reportedly, was also subject to Chinese import controls at the time of acquisition. If confirmed, this represents a dual-enforcement failure: the chips moved through the US export control perimeter and into China despite both governments' stated restrictions. This is analytically significant because it shifts the policy question from whether controls can be technically enforced to whether they can be enforced at any scale. Source: Tom's Hardware.
Data Centre Capital Flows Accelerate: Private Debt and Cloud Growth Sustain Buildout
Blue Owl Capital's Stack Infrastructure is seeking an A$8.5 billion ($5.9 billion) syndicated loan for a single data centre project in Melbourne — described as potentially one of the largest such financings in Australia's history. This is a confirmed financing process currently underway, not an announced plan. The scale of a single-project loan at this magnitude reflects how private credit markets have become the primary funding mechanism for hyperscale-adjacent infrastructure, filling gaps that public equity and corporate balance sheets alone cannot cover. Source: Bloomberg.
On the demand side, Microsoft's cloud unit grew at its fastest rate since 2022 in the most recent quarter, with management indicating the pace is accelerating rather than plateauing. AWS simultaneously secured a $410 million compute agreement with Recursive Superintelligence, an AI startup that raised $650 million only months ago — illustrating how newly capitalised AI labs are immediately translating funding into cloud compute contracts. Lenovo's expansion of its North Carolina AI server manufacturing line provides further confirmation that supply-side investment is also accelerating domestically, with the Chinese firm building US-native production capacity explicitly to navigate tariff exposure and satisfy procurement requirements. Sources: Bloomberg, Data Center Dynamics, Tom's Hardware.
Geopolitical Memory Supply Lines: US Senators Target Apple Over Chinese Chip Sourcing
A bipartisan group of US senators has formally warned Apple against purchasing memory chips from CXMT Corp. and Yangtze Memory Technologies, both of which are on US entity lists. The letter frames the risk as strategic dependency on a US adversary for components central to Apple's product lines. This is a confirmed legislative pressure action, not yet a regulatory mandate — Apple has not confirmed any procurement arrangements with either company, and neither CXMT nor YMTC has confirmed supply discussions. Source: Bloomberg.
The intervention is notable in the context of the broader memory shortage. With Samsung and SK Hynix redirecting capacity to AI applications and DRAM supply to standard buyers tightening sharply, Apple faces a constrained supplier landscape for the volumes of DRAM and NAND it requires at competitive cost. Chinese suppliers offer an alternative supply line at potentially lower cost — which is precisely why the senators perceive a real commercial incentive exists for Apple to explore the option, and why they acted pre-emptively rather than reactively.
Intel RAMP-C Completion and the State of US Domestic Foundry Ambition
Intel Foundry has formally completed the RAMP-C program, a US Department of Defense initiative awarded in 2021 to establish a domestic leading-edge chip manufacturing ecosystem on Intel's 18A process node. The program paid Nvidia and other companies to tape out test chips on 18A, with the objective of proving that a secure, US-based advanced process was viable for defense applications. Completion means the production pilot phase is closed — what remains to be determined is whether Intel 18A advances into volume production and whether defense customers commit to it as a procurement pathway. Source: Tom's Hardware.
The closure comes at a complicated moment for Intel Foundry. RAMP-C demonstrated that the 18A ecosystem can be stood up, but Intel's broader foundry competitive position versus TSMC remains a live question. The program's strategic value was always as much about proving US government willingness to fund domestic advanced manufacturing as it was about the specific chips produced. With CHIPS Act funding still being disbursed and Intel navigating its own internal restructuring, RAMP-C's legacy will depend heavily on whether 18A secures commercial and defense volume orders in the next 18 months.
Signals & Trends
Memory Tiering Is Splitting the Global Tech Economy Into AI and Non-AI Supply Chains
The data points from this briefing collectively describe a structural bifurcation in the memory industry that is moving faster than most supply chain analyses anticipated. HBM and server DRAM now command premium wafer allocation at Samsung and SK Hynix, consumer and module-maker DRAM is being squeezed toward a 70%-plus supply reduction by 2027, SSD pricing has risen 220% in a year, and Seagate's high-capacity drives are sold out through 2028. This is not a cyclical shortage — it reflects a permanent reallocation of manufacturing priorities. The strategic implication is that enterprises, governments, and OEMs outside the hyperscaler procurement tier will face structurally elevated component costs for the remainder of this decade, and supply chain strategies built around just-in-time commodity memory procurement are no longer viable.
Export Controls on AI Compute Are Functioning as Revenue Constraints, Not Access Denials
The Moonshot AI Blackwell circumvention and the Nvidia employee implication in the Supermicro smuggling case, taken together, suggest that determined actors with capital can access restricted compute hardware within a timeline relevant to model training cycles. This matters because US export control policy has been premised on the assumption that denying access to leading-edge hardware would meaningfully delay Chinese AI development. The emerging evidence indicates the delay is being compressed by smuggling networks sophisticated enough to compromise hardware vendors from the inside. The policy-relevant question is no longer whether controls can stop access entirely but whether the friction they impose is sufficient to maintain a meaningful compute gap — and the current evidence is not encouraging.
Private Credit Is Becoming the Structural Backbone of AI Infrastructure Finance
The Stack Infrastructure A$8.5 billion single-project loan in Australia is one data point in a broader pattern: private credit funds, not public equity markets or hyperscaler balance sheets alone, are increasingly originating the debt that funds large-scale data centre development outside core US markets. This matters for infrastructure risk analysis because private credit carries different refinancing, covenant, and liquidity characteristics than investment-grade corporate debt. As interest rates remain elevated and project scale increases, the concentration of AI infrastructure debt in private credit vehicles creates a potential systemic exposure that is largely opaque to public market observers — and that has not yet been stress-tested by a significant AI demand slowdown.
Explore Other Categories
Read detailed analysis in other strategic domains