Compute & Infrastructure

Sharetronic, a Chinese AI data centre company and Nvidia cloud partner, procured 300 servers designed for H100 GPUs worth $92 million despite US export bans, according to publicly available documents reported by Tom's Hardware. The revelation follows a Super Micro smuggling arrest and caused Sharetronic shares to plummet. This represents substantial volume bypassing controls — 300 servers likely containing multiple GPUs each translates to significant compute capacity reaching restricted entities.

The incident demonstrates that tightening paper restrictions without robust enforcement mechanisms creates arbitrage opportunities rather than effective containment. Companies operating in jurisdictions with weak compliance frameworks can acquire strategic hardware through procurement channels that nominally comply with local law while violating US export policy. The fact that a public Nvidia cloud partner was involved suggests either lax partner vetting or sophisticated circumvention using legitimate business relationships as cover.

Memory pricing continues escalating beyond typical cyclical patterns, with Corsair silently increasing its Strix Halo AI Workstation 300 flagship model by $400 to $3,399 within two months, explicitly attributed to AI-driven RAM and storage price spirals, reports Tom's Hardware. High-capacity DDR5 kits are reaching extreme price points, with 96GB configurations normally exceeding $1,100 before promotional discounts. This pricing environment reflects supply-demand imbalances created by data centre operators competing with consumer and workstation markets for the same DRAM capacity.

The sustained price increases indicate memory fab capacity additions have not kept pace with AI training and inference demands. While HBM3 and HBM3E production for GPU integration receives investment priority, standard DRAM used in servers and workstations faces allocation pressure. This creates a bifurcated market where hyperscalers with long-term supply agreements secure preferential access while smaller AI infrastructure buyers face spot market volatility and inflated costs that make capital planning difficult.

Taiwan's National Security Bureau reports China is intensifying illicit efforts to recruit semiconductor process engineers and acquire chip-related intellectual property from Taiwan as international export restrictions tighten, according to Tom's Hardware. The explicit acknowledgment from Taiwan's national security apparatus indicates activity has reached a scale and sophistication warranting public disclosure, likely signalling both deterrence messaging and request for allied counterintelligence support.

The talent acquisition campaign represents a logical response to equipment restrictions — if China cannot legally purchase ASML EUV systems or Applied Materials deposition tools, recruiting engineers who understand process integration and troubleshooting provides alternative pathways to capability development. Taiwan's position as home to TSMC's most advanced fabs creates unique vulnerability, as engineers working there possess knowledge of bleeding-edge nodes that mainland Chinese fabs have not yet achieved. This makes Taiwan semiconductor employees disproportionately valuable intelligence targets compared to engineers at older-generation facilities.

DRAM modules remain vulnerable to Rowhammer and Rowpress attacks despite new refresh commands introduced as countermeasures, with permanent fixes still years away, according to Semiconductor Engineering. The article characterises the situation as "whac-a-mole" security where each mitigation creates new attack surfaces or performance penalties. Rowhammer exploits allow attackers to flip bits in adjacent memory rows through repeated access patterns, potentially compromising data integrity and enabling privilege escalation in shared infrastructure environments.

For data centre operators running multi-tenant AI workloads, this represents ongoing exposure that refresh-rate mitigations only partially address. Higher refresh rates impose performance penalties and increase power consumption — problematic when memory subsystems already face thermal and energy constraints. The multi-year timeline to architectural fixes means current data centre buildouts will deploy hardware with known vulnerabilities that require software-layer defences and workload isolation strategies. Cloud providers hosting third-party AI training runs face particular risk if attackers can exploit memory vulnerabilities to exfiltrate model weights or training data from adjacent customer workloads.