Compute & Infrastructure

Samsung's chip division reported a 48-fold year-on-year profit jump for Q1 2026, far exceeding analyst estimates, according to Bloomberg. The driver is unambiguous: AI data center operators are absorbing HBM and DRAM at volumes that have created a genuine supply-demand imbalance, allowing Samsung to extract premium margins. This result follows similar outperformance from Murata Manufacturing, whose passive component business also beat estimates on data center demand, per Bloomberg — suggesting that the demand signal is broad-based across the component stack, not isolated to leading-edge logic.

The strategic implication is that memory — historically a commodity market subject to brutal cyclicality — has entered a structurally tighter regime driven by HBM's role in GPU packages. Samsung, SK Hynix, and Micron collectively control this supply. Any yield issue, geopolitical disruption, or packaging bottleneck at these three firms directly constrains global AI training capacity. The Korean peninsula concentration of this supply chain remains the single most underleveraged risk in infrastructure planning conversations.

All four major hyperscalers reported earnings this week confirming accelerating AI infrastructure spend. AWS posted its fastest quarterly cloud sales growth since 2022 on heavy data center investment, per Bloomberg. Microsoft flagged 'modest' Azure acceleration and committed to continued AI infrastructure spending, per Bloomberg. Alphabet beat estimates on Google Cloud demand, per Bloomberg. These are confirmed revenue outcomes, not projections.

Meta's situation is analytically distinct and more instructive. The company raised its full-year capex guidance to $125–145 billion — roughly $10 billion above prior estimates — and explicitly cited 'higher component pricing' as a factor, per Bloomberg. This is a direct acknowledgment that hardware cost inflation is compressing the return profile of AI infrastructure investment. Separately, Meta signed a multibillion-dollar, multi-year deal with AWS for tens of millions of Graviton5 CPU cores, per Tom's Hardware, signaling that CPU availability — not just GPU — is now a binding constraint for agentic inference workloads that require high core-count general-purpose compute alongside accelerators.

OpenAI has met a key US compute capacity milestone ahead of schedule, per Bloomberg. However, the mechanism by which it achieved this milestone represents a fundamental strategic shift: OpenAI has effectively abandoned direct ownership of Stargate data centers in favor of leasing compute from partners who bear the capital risk, according to Tom's Hardware. The company now describes 'Stargate' as an umbrella term for a broader ecosystem of compute arrangements rather than a specific owned-infrastructure program.

This reframing has significant implications for how the market should interpret Stargate capacity announcements. Confirmed capacity coming online under lease agreements is operationally real — OpenAI can access it. But the capital risk, long-term site control, and stranded asset exposure sit with third-party investors and operators. For infrastructure professionals, this means Stargate's headline numbers are not a proxy for OpenAI's balance sheet exposure, and the distributed ownership structure makes capacity resilience harder to assess.

Cambricon Technologies reported first-quarter revenue more than doubling, with shares surging 14% in Shanghai, driven by Beijing's policy-backed push for domestic semiconductor self-sufficiency, per Bloomberg. Simultaneously, Lisuan Tech became the first Chinese GPU maker to earn Microsoft WHQL certification for its 6nm graphics card, joining NVIDIA, AMD, and Intel as the only four companies to hold this designation, per Tom's Hardware.

These developments are incremental rather than transformative — Cambricon's absolute revenue base remains far below NVIDIA's, and WHQL certification is a driver quality threshold, not a performance benchmark. But directionally, they confirm that China's domestic chip ecosystem is progressing on both the commercial adoption and software ecosystem dimensions simultaneously. The critical remaining gap is process node access: Chinese firms are designing competitive architectures but remain dependent on TSMC-equivalent leading-edge manufacturing that export controls are specifically designed to deny.

Three distinct sovereign compute developments emerged this week. The UK government confirmed it will launch a formal AI Hardware Plan later in 2026, with Technology Secretary Liz Kendall explicitly rejecting the framing that the hardware race has already been won by the US and Taiwan, per Data Center Dynamics. This is an announced policy commitment — no budget allocation or procurement has been confirmed. In Japan, NTT will support Rapidus's liquid-cooled GPU server deployment, per Data Center Dynamics, a concrete operational step linking Japan's domestic chip design and manufacturing effort to real infrastructure deployment. In Canada, Bell Canada is converting a Winnipeg food processing plant into an AI data center for approximately $23 million, per Data Center Dynamics — modest in scale but representative of the opportunistic real estate conversions accelerating across secondary markets.

The UK's position is the most strategically exposed. Announcing a hardware plan without controlling a leading-edge fab, an advanced packaging ecosystem, or a domestic champion chipmaker means the plan's content will determine whether it is a genuine industrial strategy or political positioning. The NTT-Rapidus partnership is more concrete: Rapidus is targeting 2nm production and the NTT deployment gives it a live infrastructure customer to validate its roadmap.