Compute & Infrastructure

YMTC overtook Kioxia in NAND shipments during Q2 2026, the first time a Chinese memory vendor has broken into the global top three, according to research cited by Bloomberg and Tom's Hardware. Samsung leads with 25%, SK Hynix follows at 22%, and YMTC now holds approximately 14% share, ahead of Kioxia and Micron. The primary driver is structural: AI server deployments are now absorbing 48% of all NAND flash, creating demand volumes that legacy suppliers cannot fully satisfy and that give domestic Chinese buyers a ready market for YMTC output regardless of Western export controls.

The strategic implication is dual. First, YMTC's rise demonstrates that Chinese semiconductor firms can achieve market leadership in segments where they are not cut off from advanced EUV tooling — 3D NAND at current densities remains manufacturable on older process generations. Second, it further concentrates Western AI infrastructure dependence on Samsung and SK Hynix for high-end NAND, since US-listed hyperscalers cannot legally source from YMTC at scale. If Korean capacity faces any disruption, there is no readily substitutable Western alternative at YMTC's now-proven shipment volumes.

CoreWeave reported $2.58 billion in quarterly revenue — up 112% year-over-year — but the more strategically significant disclosure was its investor warning that shifting away from Nvidia chips would require substantial time and capital, as reported by Bloomberg. This is not a theoretical risk statement; it reflects CoreWeave's physical infrastructure, software stack, and — critically — customer contracts, including an A100 deployment signed through 2029, which Tom's Hardware notes is generating profit nine years post-GPU launch, sustained by power constraints and legacy compatibility that make replacement uneconomic.

The A100-into-2029 contract is a revealing data point: it means that hardware generations Nvidia considers mature are still being contracted at profit-generating rates in an environment where power scarcity makes deploying new, higher-TDP hardware unattractive relative to the capital already sunk. This extends Nvidia's effective revenue tail on each architecture and reduces the urgency for cloud customers to upgrade — but it also means CoreWeave's per-unit economics are increasingly set by Nvidia's pricing power, as evidenced by the RTX PRO 6000 Blackwell doubling in MSRP to $16,000 within twelve months, per Tom's Hardware.

As AI clusters scale toward hundreds of thousands of accelerators, the interconnect fabric — historically a secondary concern — is becoming a primary infrastructure chokepoint. Silicon photonics and optical transceivers operating at 112G and above are now essential for the low-latency, high-bandwidth links that dense GPU clusters require, and the technology is advancing rapidly, as covered by both Tom's Hardware and Semiconductor Engineering. The complication is that China currently dominates optical transceiver manufacturing, having built cost-competitive capacity over the past decade while Western firms focused on higher-margin logic chips.

The US government is reported to be considering restrictions on Chinese optical transceivers in AI data centre deployments, with a bilateral summit creating a near-term decision point. Unlike the GPU supply chain — where TSMC and ASML create clear Western leverage points — photonics manufacturing does not map cleanly onto existing export control frameworks. A ban would force rapid reshoring or qualification of alternative suppliers, a process that typically takes 18-24 months minimum for components embedded in certified data centre builds. Cisco's AI infrastructure revenue projection of $7.5 billion this fiscal year, noted by Bloomberg as a disappointment against $9.3 billion in accumulated orders, may in part reflect supply-side friction in exactly this interconnect layer.

AI data centre developers have begun filing lawsuits against local jurisdictions that have enacted temporary bans or moratoriums on new construction, with claims ranging from officials exceeding statutory authority to due process and equal protection violations, according to Tom's Hardware. At least one county has reversed its position following legal pressure; multiple cases remain in active litigation. The legal escalation marks a strategic shift from developers treating local opposition as a negotiating friction to treating it as a justiciable rights violation.

This matters structurally because siting timelines are already among the longest lead-time items in data centre development — commonly 24-36 months from land acquisition to energisation — and moratoriums compound this by introducing legal uncertainty that prevents capital commitment. The IBM and Together AI deal to deploy an Nvidia HGX B300 cluster, with go-live not expected until Q1 2027 per Data Centre Dynamics, illustrates the baseline timeline even without siting disputes. If litigation becomes a standard developer tool, it may chill municipal opposition in the near term, but it also raises the political cost of AI infrastructure expansion in communities with legitimate concerns about power and water consumption.