Compute & Infrastructure

Nvidia has confirmed it will invest as much as $105 billion to back the first phase of OpenAI's Ports-Pike mega data center campus in Ohio, with the facility set to operate at 4.25 gigawatts of capacity and run exclusively on Nvidia compute and networking hardware. The power figure alone is extraordinary — 4.25GW is roughly equivalent to four large nuclear power plants and exceeds the total AI data center capacity currently operational across most European nations. The SB Energy facility will be leased by OpenAI, with Nvidia functioning as the infrastructure backer and exclusive hardware vendor. Data Center Dynamics and Bloomberg both confirmed the commitment, with Bloomberg noting this is the latest tie-up between two dominant forces in the AI boom.

The strategic architecture of this deal deserves close reading. Nvidia is not simply selling hardware — it is financing and underwriting the infrastructure that will then consume its own hardware exclusively. This creates a vertically integrated demand loop: Nvidia capital builds the facility, Nvidia silicon fills it, and OpenAI's workloads validate the next generation of Nvidia products. The exclusivity clause is particularly significant; it forecloses AMD, Intel Gaudi, and any custom silicon from a 4.25GW customer for the foreseeable future. The $105 billion figure is described as the upper bound of the first phase, meaning total campus investment could ultimately be larger. This is a confirmed announced plan, with construction timelines and grid connection schedules not yet publicly detailed.

Nvidia has disclosed a $21 billion equity stake in SpaceX, following Elon Musk's announcement of an exclusive arrangement to equip SpaceX data centers with Nvidia hardware. Ars Technica reports the investment emerged from a filing, suggesting it may have been structured over time rather than as a single transaction. The pattern here mirrors the OpenAI Ohio deal: Nvidia secures hardware exclusivity in exchange for capital, converting its balance sheet into a mechanism for locking in demand.

The SpaceX angle has infrastructure implications beyond terrestrial data centers. SpaceX's Starlink constellation and its associated ground infrastructure represent a distributed compute and connectivity layer that sits outside traditional hyperscaler control. Nvidia hardware embedded throughout that system — and Nvidia equity entitling it to upside — gives the company strategic exposure to edge and satellite-adjacent compute. This is a confirmed equity disclosure from a regulatory filing, not a speculative announcement, though the operational terms of the hardware exclusivity arrangement have not been fully disclosed publicly.

According to a SemiAnalysis report cited by Data Center Dynamics, Google is rumoured to have partnered with AMD on the design of TPU v10, while shipments of current-generation TPUs have been revised downward. This is a rumour from a research firm, not a confirmed Google disclosure, and should be treated as a credible signal rather than established fact. That said, SemiAnalysis has a strong track record on silicon supply chain intelligence.

If accurate, the implications are significant on two levels. First, it suggests Google's internal TPU programme — one of the most advanced custom AI accelerator efforts in the industry — is encountering design or yield challenges that external expertise is being brought in to address. Second, it positions AMD in a role that goes beyond foundry customer: a co-design partner for hyperscaler custom silicon. This would represent a meaningful expansion of AMD's addressable market and a blow to the narrative that Google's TPU programme is a self-sufficient alternative to merchant silicon. The downward revision to current-generation shipments compounds the concern, suggesting near-term capacity from this programme is below plan.

PC Partner, a major GPU board manufacturer, has warned that graphics card prices will rise further in H2 2026 as memory costs climb and supplies tighten, with entry-level boards facing the most acute shortages. Analyst Jon Peddie, cited by Tom's Hardware, suggests manufacturers are hiking prices beyond what underlying memory cost increases justify — indicating margin extraction on top of a genuine supply constraint. The GDDR6 and GDDR7 memory used in consumer GPUs draws from the same fab capacity serving HBM production for data center accelerators, creating a structural tension between AI infrastructure demand and consumer hardware supply.

The emergence of a Japanese repair shop offering GDDR6 VRAM upgrades — modding RTX 2080 Ti cards to 22GB for $282 — is a vivid symptom of the same underlying constraint. Tom's Hardware frames this as a budget AI solution, but it also reflects a secondary market emerging around memory scarcity. For infrastructure professionals, the relevant signal is that DRAM fabs are allocating capacity toward high-margin HBM for data center AI, and the resulting tightness is propagating through the entire GPU memory supply chain.

A CIC forecast cited by Tom's Hardware projects the data center optical interconnect market growing from $13.7 billion in 2024 to $144.4 billion by 2030, with silicon photonics accounting for 63.7% of revenue driven by co-packaged optics adoption. This is an analyst projection, not confirmed capacity, but the directional logic is sound: as GPU cluster scale increases, copper interconnects become bandwidth-limited and power-inefficient, making optical the architectural necessity rather than a premium option. Co-packaged optics — integrating photonic components directly at the package level — reduces latency and power consumption but requires new manufacturing processes that are not yet at high volume.

Separately, Semiconductor Engineering reports that 800VDC power distribution is emerging as the next architectural standard for AI data centers, with the full conversion path from medium-voltage AC to sub-1V silicon now identified as a system-level bottleneck. This is not a future problem: facilities being designed today are incorporating 800VDC bus architecture because it reduces conversion losses at the scale of multi-megawatt GPU clusters. The implication is that both the internal data movement architecture (optical) and the power delivery architecture (800VDC) of AI data centers are undergoing simultaneous generational transitions, each creating new supply chain dependencies on components and expertise that are currently in limited supply.