Compute & Infrastructure

Microsoft committed $10 billion over four years to AI and cloud infrastructure in Japan and $5.5 billion to Singapore through 2029, according to Bloomberg and Data Center Dynamics. The Japan investment represents Microsoft's largest commitment to the country and includes data center capacity, AI skilling programs, and research partnerships. Singapore has hosted Microsoft cloud infrastructure since 2010, making this an expansion of existing operations rather than greenfield deployment.

These investments follow a pattern of hyperscalers building sovereign compute capacity partnerships with allied governments concerned about data residency and supply chain security. Japan has explicitly prioritized domestic AI infrastructure through government policy, while Singapore serves as a strategic hub for Southeast Asian cloud services despite land and power constraints that make per-megawatt buildout costs significantly higher than U.S. markets.

Nvidia's share of China's data center accelerator market has contracted as domestic suppliers including Huawei, Cambricon, and others have collectively reached 41% market share, according to Tom's Hardware. The shift accelerated following tightening U.S. export restrictions that continue to evolve, with Tom's Hardware reporting that frequent regulatory changes are creating planning uncertainty for both chip buyers and suppliers globally.

Huawei's Ascend series and Cambricon's MLU accelerators have gained traction among Chinese cloud providers and enterprises that previously relied on Nvidia hardware. While these chips generally lag Nvidia's latest generations in raw performance and software maturity, they provide sufficient capability for many inference workloads and some training applications. The market consolidation around domestic suppliers represents the most significant forced technology bifurcation in the AI era, creating parallel ecosystems with limited interoperability.

Global semiconductor foundry revenue reached $320 billion in 2025, growing 16% year-over-year, driven primarily by AI accelerator demand, according to Counterpoint Research data reported by Tom's Hardware. TSMC's market position strengthened further during the period, reflecting its monopoly on advanced process nodes below 5nm that are critical for leading AI chips. The growth rate, while substantial, represents a deceleration from earlier pandemic-era peaks, suggesting foundry expansion is beginning to catch up with demand after years of severe shortages.

Industry observers note that physical and energy constraints are increasingly limiting further buildout velocity. Data Center Dynamics reported that AI infrastructure is hitting physical limits, with efficiency becoming critical as data center power requirements outpace grid capacity expansion in key markets. This creates a paradox where foundries can manufacture chips faster than data centers can deploy the power infrastructure to run them at scale.

Nvidia invested $2 billion in Marvell Technologies and announced a partnership focused on NVLink Fusion interconnect technology, according to Tom's Hardware. The investment is notable because Marvell's largest customers, including major cloud providers, are actively developing custom ASICs designed to reduce dependence on Nvidia GPUs. Rather than competing directly, Nvidia is positioning NVLink as infrastructure that custom chip designers must adopt to achieve competitive performance in multi-chip systems.

The strategy represents a shift from hardware monopoly to ecosystem control. By making NVLink the de facto standard for high-speed chip-to-chip communication in AI clusters, Nvidia maintains influence even as customers design their own silicon. Marvell's expertise in custom ASIC development and its relationships with hyperscalers make it an ideal partner to embed NVLink across next-generation designs, effectively turning potential displacement into soft lock-in through interconnect dependency.

Security researchers disclosed GDDRHammer and GeForge, new variants of Rowhammer attacks that exploit Nvidia GPU memory to gain complete control of host systems, according to Ars Technica. The attacks manipulate GPU memory bit flips to compromise CPU security boundaries, demonstrating that AI accelerators introduce attack surfaces beyond traditional CPU vulnerabilities. Unlike previous Rowhammer variants targeting system RAM, these attacks leverage the massive memory bandwidth and capacity of modern GPUs used for AI workloads.

The disclosure comes as GPU deployment accelerates across cloud infrastructure and enterprise data centers without corresponding hardening of security architectures. Multi-tenant cloud environments where GPUs are shared among customers are particularly vulnerable, as one tenant could potentially exploit these vulnerabilities to access another tenant's data or compromise the host system. Nvidia has not yet released mitigation guidance, and it remains unclear whether fixes will require firmware updates, driver changes, or hardware revisions in future GPU generations.