Compute & Infrastructure

Nvidia's quarterly disclosures revealed positions that reframe the company's strategic posture: a $30 billion stake in Intel — built from an initial $5 billion bet that has appreciated six-fold — and a newly disclosed $21 billion stake in SpaceX, alongside existing positions in CoreWeave, Coherent, Nokia, and a complete exit from Arm. Taken together, these are not passive financial holdings. As reported by Bloomberg and analysed in detail by Tom's Hardware, Nvidia is systematically acquiring influence over its clients, partners, and suppliers.

The Intel stake is particularly significant from a supply chain perspective. Intel remains one of very few non-TSMC advanced packaging and foundry options at scale. A financially strengthened Intel — with Nvidia as a major shareholder — could accelerate Intel Foundry's viability as a second-source for advanced packaging, reducing the single-point-of-failure risk that TSMC represents for the entire AI hardware ecosystem. The SpaceX stake ties Nvidia to the leading orbital compute and connectivity platform at exactly the moment orbital infrastructure is being evaluated for AI workloads. Nvidia's exit from Arm is noteworthy: it removes a position that was becoming politically complicated and signals Nvidia is concentrating capital in infrastructure plays rather than architecture licensing.

According to a report from Tom's Hardware, Google is in discussions with AMD to co-design a next-generation TPU incorporating on-package CPU cores, targeting reinforcement learning and agentic workloads that require tighter CPU-accelerator coupling than current training ASICs provide. This is currently a rumour, not a confirmed programme, and should be treated as speculative until either company confirms it.

If accurate, the strategic implications are substantial. First, it would represent AMD's most significant incursion into hyperscaler custom silicon, a market previously dominated by Google's internal teams, Broadcom (as Google's packaging partner), and NVIDIA. Second, the architectural direction — integrating CPU cores on-package with AI accelerators — reflects a genuine workload shift: agentic and RL pipelines have different latency and control-flow profiles than pure transformer training, and current GPU and TPU architectures are not optimised for them. This is consistent with the broader industry signal that inference and agentic compute will require heterogeneous silicon rather than scaled-up homogeneous GPU clusters. Google's reported reorganisation of DeepMind, discussed this week by The Verge, adds context: internal structural pressure may be accelerating Google's willingness to bring in external design partners rather than rely solely on internal TPU teams.

Tom's Hardware flagged this week that optical interconnects have become a major flashpoint in AI data centre buildout, with a cluster of new product announcements, partnerships, and technical publications converging simultaneously. The underlying driver is straightforward: as GPU cluster configurations scale toward 100,000-accelerator SuperPODs and beyond, copper-based interconnects hit fundamental bandwidth, latency, and power density limits. Tom's Hardware Premium described this as 'a big week for optical,' with Coherent — in which Nvidia holds a strategic stake — among the companies central to this transition.

This is a confirmed investment trend, not speculation: hyperscalers and co-location providers are actively qualifying optical transceiver and co-packaged optics solutions for next-generation rack architectures. The supply chain risk is real — the optical components ecosystem is far more fragmented and capacity-constrained than the GPU supply chain, with a small number of specialist manufacturers controlling key elements like indium phosphide laser substrates. A bottleneck in optical interconnect supply could constrain the effective utilisation of GPU capacity that is otherwise coming online on schedule.

A notable case study this week: a former Missouri Farm Bureau president is publicly marketing his land to data centre developers after a $6.3 billion project was blocked by a wave of local opposition — described by Tom's Hardware as part of a 500-jurisdiction moratorium movement with 70% public opposition in affected communities. The story is illustrative of a structural dynamic: as the most permissive jurisdictions fill up, developers are being pushed into markets with active organised opposition, fragmenting the land acquisition and permitting process across hundreds of local regulatory environments.

This is a confirmed operational constraint, not a speculative risk. The buildout math is straightforward — AI infrastructure demands announced by hyperscalers through 2028 require land, water rights, and grid interconnection agreements that take two to five years to secure in contested markets. Energy and cooling demands compound the problem: large GPU clusters require gigawatt-scale power delivery and significant water consumption for cooling, both of which trigger environmental review processes in most US and EU jurisdictions. The emergence of orbital data centre concepts — moving from theoretical to early demonstrations according to Data Center Dynamics — reflects in part an industry awareness that terrestrial siting constraints are not temporary.