Compute & Infrastructure

The US is reportedly drafting a ban on imports of Chinese data center components, with optical transceiver manufacturers — led by Zhongji Innolight — bearing the immediate market reaction. Shares in Chinese optical stocks slid sharply on the Reuters report, per Bloomberg. This matters structurally: Chinese firms have captured a dominant share of the global optical transceiver market, which is the connective tissue linking GPUs, switches, and storage at scale inside AI data centers. A ban would force hyperscalers and colocation operators to qualify alternative suppliers — a process measured in quarters, not weeks.

Bloomberg separately reports that China's official press has warned Washington against expanding tech curbs, framing any such move as a threat to the fragile trade truce established earlier this year, while assessing that the measures are unlikely to materially damage China's broader export economy Bloomberg. The asymmetry is telling: the supply disruption risk falls harder on US data center operators than on Chinese exporters, at least in the near term. US-headquartered hyperscalers have limited domestic alternatives at the volume and price point currently sourced from Zhongji Innolight and peers.

AMD reported Q2 2026 data center revenue of $6.7 billion, up 107% year-over-year and sequentially from $5.8 billion in Q1, driven by MI300-series AI accelerator shipments and EPYC server CPU momentum The Verge. The magnitude of the year-over-year growth confirms that AI infrastructure spend is broad enough to sustain a second large-scale accelerator vendor — AMD is no longer competing at the margin against NVIDIA but capturing a structurally distinct portion of the buildout, particularly among customers seeking supply diversification.

Despite the headline result, AMD shares fell in after-hours trading on a forward guidance figure that underwhelmed analysts priced for accelerating momentum Bloomberg. The divergence between exceptional historical growth and cautious guidance is analytically significant: it suggests either supply constraints on AMD's packaging capacity — MI300X relies heavily on TSMC CoWoS advanced packaging, which remains the primary bottleneck across the industry — or that large cloud customers are pacing their AMD orders more conservatively than the most bullish projections assumed. Gaming revenue declined, confirming that AMD has effectively reallocated wafer capacity from consumer to data center segments.

SpaceX's inaugural public quarterly earnings revealed that its AI compute division generated $2.6 billion in revenue — more than three times the prior year figure — primarily from leasing GPU capacity to third-party AI companies The Verge. This positions SpaceX alongside CoreWeave, Lambda Labs, and similar neoclouds as a significant GPU-as-a-service operator, but with the distinctive advantage of captive Starlink connectivity infrastructure and the operational scale of a vertically integrated industrial firm. The revenue mix shift — AI now outpacing launch revenue — is a structural signal, not a one-quarter anomaly.

However, SpaceX's aggressive reinvestment into AI infrastructure — buying GPU clusters at a pace that exceeded Wall Street's spending models — caused the stock to fall post-earnings despite the revenue beat Bloomberg. This is the same capital intensity dynamic pressuring hyperscalers: the cost of acquiring and operating current-generation NVIDIA hardware is high enough that even strong revenue growth does not immediately translate to margin expansion. For the broader infrastructure ecosystem, SpaceX's aggressive procurement posture adds another large buyer competing for constrained GPU supply.

Japanese distributor CFD Sales has confirmed that Gigabyte graphics card prices will increase 20-40% effective this month, driven by upstream supply tightening Tom's Hardware. While this is the consumer and commercial GPU tier rather than data center accelerators, the inflation signal is directionally consistent with what is happening at the top of the stack: NVIDIA's allocation of wafer capacity and TSMC's CoWoS packaging lines are being prioritised for H100/H200/B200 data center products, leaving consumer and mid-market GPU supply structurally undersupplied. Price increases in Asia's distribution channel typically presage broader global retail price movements within one to two quarters.

Three distinct neocloud launches or expansions were confirmed this week, each reflecting a different strategic thesis. Foxconn subsidiary Visionbay.ai brought a 5MW NVIDIA-powered AI cluster online in Taiwan, reporting 90% utilisation at launch — a confirmed, operational asset Data Center Dynamics. Volta launched from stealth with plans to lease capacity in Norway from Bitdeer, targeting a named but undisclosed leading AI lab as anchor tenant, with a stated ambition of gigawatts by 2030 — this is an announced plan with a real anchor customer but the gigawatt target remains speculative Data Center Dynamics. Perimeter Compute launched as a distributed edge GPU provider, a model predicated on inference workload decentralisation rather than centralised training clusters Data Center Dynamics.

Separately, Chinese AI lab Moonshot has leased 20,000 NVIDIA GPUs from Alibaba Cloud, representing a significant proportion of its total compute for Kimi model development Data Center Dynamics. This transaction is notable because it demonstrates that Chinese AI labs continue to access NVIDIA hardware through cloud intermediaries at scale, raising questions about the effective enforcement perimeter of US export controls — Alibaba Cloud's GPU inventory presumably predates or circumvents current restrictions.