Compute & Infrastructure
Top Line
Qualcomm has signed a confirmed chip supply and investment deal with Amazon Web Services, marking its first major data centre customer win and directly challenging NVIDIA's dominance in AI silicon by introducing a second custom-silicon supplier into the AWS infrastructure stack.
Mistral AI raised €3 billion in a Samsung Electronics-led Series D with explicit intent to deploy compute at scale, while simultaneously hiring Google's former European energy and infrastructure principal to lead Mistral Compute's 1GW buildout — signalling that frontier AI labs are now competing directly with hyperscalers for power and physical infrastructure.
Belgian authorities arrested a former BelGaN researcher for alleged gallium nitride IP theft to China, a case that underscores the strategic vulnerability of compound semiconductor technology — GaN is critical for power electronics in data centres and defence — outside the core logic chip supply chain.
OpenAI is sourcing compute capacity from Firmus in Malaysia as its South Korea Stargate expansion stalls, revealing that sovereign infrastructure friction is already redirecting hyperscale AI deployment across Southeast Asia.
Kioxia's CEO dismissed a deeper SK Hynix tie-up and pledged to moderate memory price increases, a statement that carries significant weight given AI-driven HBM and NAND demand is squeezing consumer and enterprise memory markets simultaneously.
Key Developments
Qualcomm Breaks Into Data Centre Silicon with Confirmed AWS Deal
Qualcomm has signed a confirmed agreement to design and supply custom AI chips for Amazon Web Services, with AWS also taking an investment stake in Qualcomm's data centre silicon effort. This is not a speculative partnership — both companies have confirmed the deal, and Qualcomm CFO Akash Palkhiwala described it publicly on Bloomberg as foundational to the company's data centre strategy. The chips will target large-scale AI inference workloads within AWS infrastructure, and the partnership includes work on optical connectivity solutions running at up to 1.6 terabits per second, per Data Center Dynamics.
The strategic read here is straightforward: AWS is deliberately diversifying its AI silicon supply chain away from NVIDIA dependence, building on its existing Trainium and Inferentia custom silicon programme. Qualcomm brings ARM-architecture expertise and a proven track record in low-power inference at the edge — translating that to data centre scale at competitive performance-per-watt metrics would give AWS meaningful cost and energy efficiency leverage. For Qualcomm, a single hyperscaler anchor customer validates its data centre pivot and opens the door to similar arrangements with other cloud providers. The risk is execution: data centre silicon is a different design environment from mobile, and Qualcomm has not yet demonstrated sustained production-grade performance at this scale.
Mistral's €3bn Raise and 1GW Infrastructure Ambition Reshapes European AI Compute
Mistral AI has confirmed a €3 billion Series D led by Samsung Electronics, with the stated use of proceeds focused on compute acquisition — not model research or go-to-market. Concurrently, Mistral Compute has hired Marc Oman, formerly Google's European energy and infrastructure principal, as it targets deployment of 1 gigawatt of compute capacity, according to Data Center Dynamics. The Samsung lead is notable: it gives Mistral a direct line to a major memory and foundry ecosystem at a moment when HBM supply is the binding constraint on AI cluster buildout.
The 1GW target should be treated as an announced plan, not confirmed capacity. Securing power agreements, data centre sites, and chip allocation at that scale in Europe — where grid interconnection queues stretch years — is a multi-year execution challenge. However, Oman's hire signals Mistral is approaching this as an infrastructure business, not a cloud tenancy arrangement. This has sovereign infrastructure implications: a European-headquartered AI lab building its own physical compute layer reduces dependency on US hyperscaler capacity and aligns with EU digital sovereignty priorities. Samsung's investment also raises questions about whether Korean memory technology — particularly HBM — will flow preferentially to Mistral's clusters.
GaN IP Theft Arrest Exposes Compound Semiconductor Supply Chain Vulnerability
Belgian prosecutors have arrested a former researcher at BelGaN, a now-collapsed gallium nitride chipmaker, on suspicion of transferring proprietary GaN semiconductor IP to a Chinese competitor before the company went bankrupt. The allegation, reported by Tom's Hardware, follows US security agency warnings — reported separately by Bloomberg — that Chinese AI companies including DeepSeek and Alibaba have systematically extracted proprietary knowledge from American firms.
GaN technology is strategically significant beyond consumer electronics: it underpins high-efficiency power conversion systems used in data centre power distribution, EV charging, and military radar. A Chinese capability leap in GaN — whether through organic R&D or IP acquisition — would reduce dependence on Western power semiconductor suppliers and potentially close the efficiency gap in data centre power infrastructure. The BelGaN case illustrates that technology transfer risk is concentrated not only in leading-edge logic (TSMC, ASML) but in specialist process nodes where smaller, financially stressed companies lack the security infrastructure of major fabs. The US government's broader warning about systematic model-weight and IP extraction suggests this is a pattern, not an isolated incident.
OpenAI's Southeast Asia Compute Strategy Fragments as Stargate Stalls in Korea
OpenAI has confirmed a compute supply arrangement with Firmus in Malaysia, while its previously announced Stargate data centre expansion in South Korea faces delays, according to Data Center Dynamics. The Malaysia arrangement provides near-term inference capacity in a region with growing AI demand, but the Korea stall is more significant strategically. South Korea was positioned as a major Stargate node given its proximity to Samsung and SK Hynix memory manufacturing — a geography that reduces latency in the physical chip supply chain for cluster buildout.
The fragmentation of OpenAI's Asia-Pacific compute strategy into smaller, distributed deals — rather than one anchor hyperscale facility — reflects the real-world friction of sovereign infrastructure politics. Malaysia offers more permissive regulatory conditions and competitive power pricing, but lacks the semiconductor ecosystem density of Korea or Taiwan. Separately, SGC Energy in South Korea has confirmed a partnership with Vertiv to deploy PowerNexus modular power infrastructure at a planned AI data centre in Gunsan, expected to be energised in 2028 — a confirmed but long-dated capacity addition that illustrates how multi-year the infrastructure pipeline genuinely is, per Data Center Dynamics.
Signals & Trends
Memory Price Inflation Is Becoming a Structural Tax on AI Buildout
Kioxia's CEO pledge to moderate NAND price increases, combined with consumer DDR5 kits hitting new price floors driven by AI-related memory demand, points to a structural dynamic that is often underweighted in AI infrastructure analysis. HBM supply — dominated by SK Hynix, Samsung, and Micron — is already the binding constraint on NVIDIA GPU cluster delivery. But NAND and DDR5 inflation is now flowing downstream into enterprise storage, server DRAM, and edge inference hardware, inflating the total cost of AI infrastructure beyond the GPU line item. Samsung leading Mistral's €3bn round while simultaneously being a primary memory supplier creates an interesting alignment of incentives — but it also concentrates Mistral's supply chain risk if Samsung prioritises internal allocation during a future shortage. Infrastructure planners should model memory cost trajectories separately from compute cost trajectories; the two are decoupling.
Custom Silicon Partnerships Are Fragmenting the Hyperscale Compute Stack
The Qualcomm-AWS deal is the latest confirmation that hyperscalers are systematically reducing NVIDIA dependency by seeding multiple custom silicon suppliers — AWS (Trainium, Inferentia, now Qualcomm), Google (TPU), Microsoft (Maia, Cobalt), and Meta (MTIA) all have active in-house or partnership silicon programmes. The pattern is moving from inference-only custom silicon toward training-capable custom silicon, and now toward custom optical interconnects — the Qualcomm-AWS 1.6T optical connectivity component is significant because interconnect bandwidth, not compute density alone, is increasingly the performance ceiling in large clusters. The risk of this fragmentation is software ecosystem splintering: each custom silicon platform requires its own compiler stack and inference optimisation, and the operational overhead of managing heterogeneous hardware at hyperscale is non-trivial. Watch for consolidation around two or three software abstraction layers — likely CUDA-compatible, MLX-based, or OpenXLA — as the determining factor in which custom silicon programmes survive to second-generation production.
IP Exfiltration Risk Is Expanding Beyond Software Model Weights to Process Technology
The convergence of the US government's DeepSeek/Alibaba model-siphoning warning with the BelGaN GaN IP arrest reveals that technology transfer risk now spans the full semiconductor stack — from trained model weights at the application layer down to process chemistry and device physics at the fab layer. This is a qualitative escalation from earlier export control frameworks, which focused primarily on equipment (ASML, KLA, Lam Research) and advanced logic chips. Process IP for compound semiconductors — GaN, SiC, InP — is increasingly relevant to AI infrastructure power efficiency, RF communications, and photonic interconnects. Unlike advanced CMOS, compound semiconductor process IP is often held by smaller, less-resourced companies with weaker IP protection infrastructure, making it a structurally attractive exfiltration target. Governments and infrastructure investors should expect enhanced screening requirements for compound semiconductor research partnerships and acquisition activity within the next 12-18 months.
Explore Other Categories
Read detailed analysis in other strategic domains