Compute & Infrastructure

Nvidia and SK Hynix have formalised a multi-year co-development and supply agreement covering next-generation memory technologies for Nvidia's upcoming GPU platforms, according to Tom's Hardware. The agreement is explicitly designed to address extended development cycles — a recognition that HBM generations (HBM3E, HBM4 and beyond) now require multi-year co-engineering between chip designer and memory manufacturer, not just procurement lead times.

Strategically, this deepens an already high-concentration supply arrangement. SK Hynix currently supplies the majority of HBM used in Nvidia's H100 and H200 series. Locking in co-development means Nvidia's next-platform roadmap is now architecturally coupled to SK Hynix's process capabilities. Samsung and Micron, both qualifying HBM supply for Nvidia, face a steeper path to displacing SK Hynix if co-development milestones are embedded in platform specifications. For infrastructure buyers, this is a second-order constraint: if SK Hynix faces yield issues, geopolitical disruption, or capacity shortfalls, Nvidia's ability to ramp GPU supply is directly impaired — and no near-term alternative is qualified at equivalent scale.

Two significant deployments confirmed in South Korea this week: Naver Cloud is deploying 55MW of Nvidia hardware at a domestic facility, with stated ambitions for hundreds of megawatts globally in coming years, while LG Uplus is building a dedicated AI factory around Nvidia infrastructure, per Data Center Dynamics and Data Center Dynamics. Both are confirmed announced plans; physical deployment timelines were not specified in available reporting.

The pattern is analytically significant: South Korea is pursuing sovereign AI compute capacity, but the hardware stack is entirely Nvidia-dependent. This creates a paradox for Seoul's industrial policy — domestic compute sovereignty is being built on foreign silicon. Given that SK Hynix supplies Nvidia's HBM and Korea hosts these deployments, the country has unusual upstream leverage, but no domestic GPU alternative. The concentration of Nvidia deployments in a single geography also raises questions about how Nvidia allocates constrained hardware across competing national customers — a prioritisation dynamic that will intensify as European and Middle Eastern sovereign buyers also queue.

A new analysis reported by Tom's Hardware finds that approximately two-thirds of 809 planned U.S. AI data centres are located in areas that have experienced drought conditions over the past year. The 809-project figure represents planned and announced projects — a mix of confirmed construction and speculative pipeline. The geographic concentration in drought zones reflects the same land, power, and tax incentive factors that drove earlier hyperscale site selection in the Southwest and Mountain West, now compounding under accelerating AI buildout.

The operational risk is twofold: evaporative cooling systems in air-cooled facilities consume millions of gallons of water daily, and in drought conditions, local water authorities have legal authority to curtail industrial users. Several jurisdictions — including parts of Arizona and Nevada — have already imposed or are reviewing restrictions on new data centre water permits. The shift toward direct liquid cooling (DLC) and immersion cooling in newer GPU-dense deployments partially mitigates consumptive water use, but the majority of the 809 projects in the pipeline will use conventional cooling architectures. This is a planning-horizon risk that will materialise as permits are contested, not a distant scenario.

Industry experts interviewed by Tom's Hardware identify AI agent deployments as the primary driver of a surge in data centre CPU demand. The mechanism is architectural: agentic AI systems require substantial CPU-side orchestration — task scheduling, tool call management, memory retrieval, and API routing — that runs alongside GPU inference. As agent frameworks scale from pilot to production, hyperscalers are discovering that their CPU-to-GPU ratios, optimised for training and batch inference, are undersized for agentic workloads.

This is a meaningful supply chain signal. AMD's EPYC and Intel's Xeon server lines stand to benefit, and both companies have been capacity-constrained on advanced server CPUs. TSMC's N4/N3 node capacity — already contested between GPU, CPU, and mobile demand — faces additional pressure. For infrastructure planners, the implication is that GPU procurement plans do not capture total compute cost for agentic deployments: CPU, DRAM, and network fabric must be re-evaluated against agent-specific workload profiles rather than legacy inference benchmarks.

Cipher Digital raised $810 million through a high-yield bond sale to fund construction of a data centre with Amazon as the anchor customer, per Bloomberg. The transaction is part of a broader pattern of sub-investment-grade debt entering AI infrastructure financing, reflecting both the urgency of capacity deployment and the fact that many data centre developers lack the balance sheet strength to self-fund at the required velocity. The Amazon offtake agreement provides revenue visibility that makes the bonds marketable, but the junk rating reflects construction, operational, and counterparty risk.

The financing structure matters for infrastructure resilience analysis. If AI revenue or hyperscaler capex commitments soften — as happened with some cloud deals in 2022-2023 — leveraged data centre developers face refinancing stress on assets that are illiquid and operationally complex. Simultaneously, on the connectivity side, Fujikura is raising prices on fiber-optic data centre cables citing sustained demand, per Bloomberg, which adds input cost pressure to developers already carrying high debt loads. The combination of leveraged construction financing and rising materials costs is a margin compression dynamic that bears watching.