Compute & Infrastructure
Top Line
Nvidia and SK Hynix have signed a confirmed multi-year partnership to co-develop next-generation AI memory chips, with Jensen Huang publicly stating that the new Vera CPU line will use SK Hynix memory — a strategic deepening of the Nvidia-HBM supply chain that further marginalises Samsung in the high-value AI memory segment.
Google has signed a confirmed $920 million per month compute deal with SpaceX, securing 110,000 Nvidia GPUs starting October 2026, making SpaceX a significant hyperscale data centre operator ahead of its June 12 IPO and signalling that non-traditional players are competing directly for scarce GPU capacity.
Jensen Huang's Seoul visit produced a cluster of commercial agreements — including a Naver data centre buildout on Nvidia architecture — illustrating South Korea's emergence as a strategically important node in Nvidia's global supply and customer network.
South Korean President Lee Jae Myung vowed to redirect surplus semiconductor tax revenues into broader growth investments, signalling that Seoul intends to leverage its semiconductor windfall for sovereign industrial policy rather than passive collection.
Key Developments
Nvidia Locks In SK Hynix as Strategic Memory Partner, Sidelining Samsung
Nvidia and SK Hynix have formalised a multi-year agreement to co-design future generations of AI memory chips, confirmed by both companies during Jensen Huang's visit to Seoul. The deal is not speculative: Huang explicitly confirmed that the upcoming Vera CPU — Nvidia's first in-house general-purpose processor — will ship with SK Hynix memory, giving the Korean firm a design-in advantage at the platform level rather than merely competing on supply volume. Bloomberg
This is a significant supply chain concentration event. The AI memory market is already an effective duopoly between SK Hynix and Samsung, with Micron a distant third. A Nvidia-SK Hynix co-development pact means future HBM generations could be architected with Nvidia's requirements baked in from the design stage — creating switching costs and qualification barriers that will be difficult for Samsung to overcome even if it closes its current HBM yield gap. The strategic risk is that Nvidia's memory supply chain is now heavily weighted toward a single Korean supplier operating within the same geopolitical exposure footprint as TSMC in Taiwan. Bloomberg
SpaceX Becomes a GPU-Scale Compute Operator with $920M Monthly Google Deal
Google has confirmed a $920 million per month compute agreement with SpaceX, under which SpaceX will provide access to 110,000 Nvidia GPUs beginning October 2026. The annualised value of approximately $11 billion would exceed SpaceX's combined 2025 revenues from Starlink, launch services, and AI — making data centre operations a primary revenue line for the company ahead of its June 12 IPO. This is the second major hyperscale compute deal SpaceX has closed in recent weeks, indicating it has successfully positioned Starship-infrastructure capacity as a credible alternative to the established cloud providers. Tom's Hardware
The deal raises several structural questions. Google securing 110,000 GPUs via SpaceX rather than through its own buildout or direct Nvidia allocation suggests either that internal capacity planning has fallen short of internal demand, or that Google is using third-party capacity to accelerate deployment timelines while its own facilities come online. Either interpretation points to continued acute GPU scarcity at hyperscale. The deal also establishes SpaceX as a new class of infrastructure intermediary — aggregating Nvidia GPU allocations and reselling compute capacity — which creates a new layer of concentration risk in the supply chain between Nvidia and end users.
South Korea Crystallises as a Strategic AI Infrastructure Node
Jensen Huang's Seoul visit produced a confirmed cluster of commercial and diplomatic outcomes: the SK Hynix multi-year memory partnership, a Naver agreement to build data centres on Nvidia architecture, and high-visibility meetings with leadership from SK Group, LG Group, and Naver. President Lee Jae Myung separately pledged to channel semiconductor sector surplus tax revenues into broader sovereign investment in future growth industries. Taken together, these developments represent a deliberate convergence of South Korean sovereign ambition and Nvidia's supply chain strategy. Bloomberg
Naver's commitment to build its data centre infrastructure on Nvidia models is analytically significant beyond its commercial value: it means South Korea's dominant domestic AI platform will deepen dependency on Nvidia architecture at the infrastructure layer, creating a long-duration alignment between Korean digital sovereignty ambitions and US semiconductor supply chains. This is a pattern being replicated across multiple sovereign markets — Japan, the UAE, France — where Nvidia is trading access to its technology stack for preferential positioning in nationally strategic infrastructure buildouts. Bloomberg
Signals & Trends
GPU Scarcity Is Generating a New Class of Infrastructure Intermediary
The SpaceX-Google compute deal is not an isolated transaction — it reflects a structural pattern in which entities with preferential Nvidia GPU allocations (whether through capital, relationships, or manufacturing adjacency) are monetising that access by reselling capacity to hyperscalers who cannot procure directly at the required scale or speed. This intermediary layer — between Nvidia's allocation desk and the end-deploying hyperscaler — introduces pricing opacity, counterparty concentration, and dependency on entities whose core competencies are not data centre operations. If this pattern persists, it suggests Nvidia's allocation pipeline has become a strategic asset class in itself, with downstream implications for how cloud pricing, AI service economics, and infrastructure ownership structures are assessed.
Nvidia Is Executing a Sovereign Infrastructure Playbook That Builds Structural Lock-In
The South Korea visit is the latest instance of a consistent Nvidia commercial strategy: arrive at the national level, sign co-development agreements with local memory or chip firms, secure commitments from dominant local AI platforms to build on Nvidia infrastructure, and position the outcomes as aligned with sovereign digital ambitions. The pattern has been repeated in Japan (with Rapidus adjacency discussions), the UAE (G42 partnership), and France (Mistral adjacency). The strategic consequence is that sovereign AI infrastructure programmes in mid-to-large economies are being architected around Nvidia's technology stack before domestic alternatives are commercially viable — creating a decade-long dependency that is institutionally and politically difficult to unwind even if geopolitical conditions or export control regimes change.
Memory Co-Design Is the Next Competitive Frontier — and It Favours Incumbents
The Nvidia-SK Hynix co-development agreement signals that the competitive battleground in AI memory is shifting from production yield and volume to architectural co-design. When a chip company and a memory company jointly specify the interface, bandwidth, and power envelope of the next HBM generation, the resulting product is optimised for that pairing — and rivals face qualification timelines measured in years, not quarters. This dynamic structurally advantages SK Hynix against Samsung in the near term, but it also concentrates Nvidia's memory supply chain risk onto a single partner. Analysts and procurement teams should treat any geopolitical disruption affecting South Korea's semiconductor industry as having direct, near-term consequences for Nvidia GPU availability and Vera CPU ramp timelines.
Explore Other Categories
Read detailed analysis in other strategic domains