Compute & Infrastructure

Nvidia's fiscal Q2 results — $96.2 billion in revenue — were already a record, but the forward guidance is the more consequential data point for infrastructure planners. CFO Colette Kress guided for approximately 70% revenue growth in fiscal 2028, against an analyst consensus of 45%. At that trajectory, Nvidia is projecting annualised revenue approaching $650 billion by FY2028, a figure that implies hyperscaler and sovereign capex commitments are not merely holding but accelerating. The market reaction was muted — a sign investors had partially priced in strength — but for capacity planners, the signal is unambiguous: GPU demand is not plateauing on any near-term horizon. Bloomberg

Simultaneously, Nvidia is working to extract more compute from constrained power envelopes. At Hot Chips 2026, the company presented its DSX MaxLPS site power management system, framing power as a hard ceiling rather than a soft constraint. The presentation modelled a fixed 100MW data centre facility and demonstrated how Vera Rubin NVL72 rack configurations maximise flops-per-watt within that budget. This is not incremental optimisation — it reflects a strategic acknowledgment that power availability, not chip supply, is increasingly the binding constraint on deployable AI compute. Tom's Hardware

The US government's investigation into Singapore-based Apex Logistics for alleged Nvidia chip smuggling represents a deliberate shift in enforcement strategy. Prior enforcement actions targeted distributors and end-buyers; moving against a transportation company signals that regulators are now mapping and pressuring the full logistics chain rather than just the endpoints. Apex's role — if confirmed — would illustrate how export controls on advanced AI semiconductors create arbitrage opportunities that third-party logistics firms can exploit across entrepot jurisdictions like Singapore. Bloomberg

This development coincides with Nvidia's disclosure that it sold a small number of H200 chips into China during Q2 — within the bounds of Trump administration licensing — marking the first H200 sales to Chinese customers. The coexistence of licensed H200 sales and an active smuggling probe underlines the bifurcated reality of the China chip market: a narrow licensed channel alongside persistent grey-market flows. The licensed volume fell short of the permitted ceiling, suggesting Chinese buyer hesitancy or logistics constraints rather than US-side supply rationing. Bloomberg

Kioxia has confirmed it is building a new chipmaking facility in northern Japan to address AI-driven storage demand. This is a confirmed capital commitment, not an announced plan, and adds to Japan's growing position as a semiconductor manufacturing hub alongside TSMC's Kumamoto fabs and Rapidus's logic ambitions. For the AI infrastructure supply chain, NAND flash capacity matters not just for consumer storage but for the high-density SSDs used in inference clusters and training checkpointing — a segment where demand has outpaced supply. Bloomberg

On the HBM side, Nvidia's NVHBM announcement at Hot Chips 2026 is strategically significant beyond the spec sheet. By developing a custom base die and PHY — rather than using commodity HBM4e from SK Hynix, Samsung, or Micron — Nvidia is inserting itself into the memory stack in a way that could alter purchasing leverage with DRAM suppliers. NVHBM will be exclusive to NVLink Fusion partners, meaning third-party chip designers integrating into Nvidia's ecosystem gain performance advantages that independent alternatives cannot match. This deepens vertical integration at the memory-compute interface. Tom's Hardware

MTN Group and UAE-based Tarek Al Ashram have agreed to develop AI data centres across Africa, a confirmed partnership between the continent's largest mobile operator and Gulf capital. The strategic logic is layered: MTN contributes connectivity infrastructure and market access across 19 African countries; Al Ashram brings capital and regional deal-making experience. For Africa, where latency to European or US cloud regions renders real-time AI applications economically impractical, in-region compute capacity is a genuine infrastructure gap. The announcement does not yet specify site locations, capacity targets, or timelines — these remain to be confirmed. Bloomberg

Separately, Papua New Guinea has launched an AI data centre and sovereign cloud facility — a small-scale deployment by global standards but notable as part of a broader pattern of Pacific Island and frontier-market sovereign cloud initiatives. Governments in these regions are motivated by data sovereignty concerns, latency, and the geopolitical signalling value of domestic compute capacity. The PNG deployment joins similar announcements from Kenya, Saudi Arabia, and Indonesia over the past 18 months, reflecting a genuine globalisation of AI infrastructure investment beyond the US-Europe-East Asia core.

The US Environmental Protection Agency is advancing a rule change that would remove the requirement for states to seek public comment on air pollution permits, with AI data centres explicitly in scope. Currently, major new diesel generator installations and cooling systems at hyperscale facilities trigger public notice requirements that have enabled local communities to challenge or delay projects. Removing that requirement would materially accelerate permitting timelines — a genuine bottleneck — while eliminating a procedural mechanism that has functioned as an early warning system for environmental load on local grids and air quality. Tom's Hardware

This is a proposed regulatory change, not yet finalised. But its direction is consistent with the broader federal posture toward AI infrastructure: treat compute buildout as critical national infrastructure deserving expedited treatment. The practical effect, if implemented, would be to shift opposition from the permitting stage — where it has been most effective — to post-construction legal and political challenges, which are structurally harder to mount.