Compute & Infrastructure

Asian suppliers now account for approximately 90% of Nvidia's production costs, up sharply from around 65% twelve months ago, according to Bloomberg-compiled data reported by Tom's Hardware. The increase reflects the accelerating complexity of Nvidia's hardware stack — advanced packaging, HBM memory from SK Hynix and Samsung, substrates from Taiwan and South Korea — all concentrated in a geography that sits at the centre of US-China trade tensions and potential Taiwan Strait risk.

The trajectory is particularly concerning because physical AI deployments — robotics, autonomous systems, edge inference — are expected to add further exposure by pulling in additional Asian-sourced sensing, actuator, and connectivity components. A 90% concentration figure is not merely a procurement metric; it is a systemic single-point-of-failure risk for the entire AI hardware ecosystem, given that Nvidia currently supplies the dominant share of training and inference accelerators globally. Tariff escalation or supply disruption would propagate directly into hyperscaler delivery timelines and AI project economics.

Apple has held exploratory discussions with Intel and Samsung about producing its main device processors on US soil, according to Bloomberg. These are confirmed as exploratory — not contracted — but the strategic logic is unmistakable: Apple's near-total dependence on TSMC for its A- and M-series silicon represents the same geopolitical concentration risk that governments and investors are now pricing explicitly. Intel's 18A process node is the operative question here; if it achieves yields competitive with TSMC's N3 family, it becomes a viable candidate. Samsung's US fab presence remains limited in advanced nodes.

This development intersects directly with the CHIPS Act investment thesis. Intel has received federal funding to build out domestic advanced node capacity, and Apple's potential demand signal — even exploratory — is precisely the anchor customer relationship that makes those investments financially viable. For the AI infrastructure stack, the significance is less about Apple's consumer devices and more about whether a viable US-based advanced logic manufacturing ecosystem can exist at all: Apple's process technology demands are among the most stringent in the industry, and if Intel or Samsung can satisfy them, it validates domestic capacity for AI chip production more broadly.

Meta is working with Morgan Stanley and JPMorgan to structure approximately $13 billion in financing for a single data centre campus in El Paso, Texas, according to Bloomberg. This is a confirmed financing process underway, not a speculative announcement. The scale — $13 billion for one facility — reflects the capital intensity of hyperscale AI infrastructure at current GPU and power costs, and the decision to use debt rather than corporate cash signals that even Meta's balance sheet cannot absorb AI capex without capital markets intermediation at this velocity of deployment.

The structuring of data centre assets as financeable infrastructure — similar to how utilities or toll roads access debt markets — is a structural evolution in how AI capacity gets built. It introduces new risk intermediaries (lenders, rating agencies) into the AI infrastructure stack and implies that data centre expansion rates are now partially governed by credit market conditions and debt service economics, not just corporate investment appetite. Starwood Capital's Barry Sternlicht, speaking at the Milken Institute Global Conference, has also signalled active interest in data centre investment, consistent with the broader pattern of private capital flowing into AI infrastructure to bridge the gap between corporate capex capacity and buildout demand Bloomberg.

Microsoft has made a confirmed commitment to double its AI infrastructure within a two-year window, according to Next Platform. This is a stated corporate commitment from Microsoft leadership, not an analyst projection, though execution against it will depend on data centre construction timelines, power procurement, and GPU supply availability — all of which face independent constraints. Microsoft's Azure AI capacity underpins both its own Copilot product suite and its OpenAI partnership obligations, giving the doubling target both internal product and external contractual dimensions.

The commitment functions as a capacity signal to the broader ecosystem: GPU suppliers, cooling vendors, power developers, and colocation operators can treat it as demand signal for procurement and buildout planning. The two-year horizon — targeting roughly mid-2028 — means that land acquisition, permitting, construction, and power interconnection processes need to be at advanced stages now to deliver on schedule. Given that US grid interconnection queues currently run 3-5 years in many regions, the power procurement component is the most plausible binding constraint on this commitment.

Nvidia has participated in a $107 million Series B for DeepInfra, a cloud inference platform aimed at reducing compute bottlenecks in AI serving, according to Bloomberg. Samsung also joined the round. The investment is notable not for its size but for what it signals: Nvidia is actively investing in the software and platform layer of inference infrastructure, not just selling GPUs into it. As the industry's workload mix shifts from training-dominated to inference-dominated — a transition now well underway — controlling the inference optimization layer becomes strategically significant for sustaining GPU demand and pricing power.