Compute & Infrastructure

The National Supercomputing Center in Shenzhen has unveiled LineShine, a 1.54-exaflop system built on 2.4 million Huawei-designed Armv9 LX2 CPU cores — with no GPUs. The architecture draws explicitly from Japan's Fugaku/A64FX playbook, prioritising aggregate core count and memory bandwidth over the GPU-centric approach that defines Western AI infrastructure. This is a confirmed deployment, not an announced plan, and its existence signals that US export controls have catalysed a divergent hardware trajectory rather than a capability gap.

The strategic implication is significant: China is not merely waiting for domestic GPU alternatives from Huawei or Cambricon. It is actively developing a parallel compute paradigm that sidesteps the chokepoint entirely. Whether CPU-only exascale architectures can match GPU-accelerated clusters for transformer training workloads at scale remains technically contested, but for inference and certain HPC tasks the gap is narrowing. The existence of LineShine will likely accelerate US policy debate around whether controls on Arm architecture licensing — currently under FTC scrutiny — should be tightened, which would itself create new supply chain risks for every chipmaker dependent on Arm IP.

Tom's Hardware confirms the Armv9 core design and Shenzhen deployment.

The US Federal Trade Commission has reportedly opened an antitrust investigation into Arm Holdings following the company's launch of its own AGI-targeted CPU product, examining whether Arm is leveraging its position as the dominant instruction set architecture licensor to disadvantage competing chip designers who build on Arm IP. This is a reported investigation, not a confirmed enforcement action, but the probe itself is consequential: Arm's architecture underpins virtually every mobile chip, a growing share of data centre silicon, and — critically — China's emerging indigenous compute stack including the LineShine LX2 processor. Tom's Hardware reports the investigation centres on whether Arm is restricting architecture access to rivals.

The concentration risk here is underappreciated: NVIDIA, Apple, Qualcomm, Amazon's Graviton, and a significant portion of Chinese domestic chips are all Arm-licensed designs. Any regulatory action that restructures Arm's licensing model — or that forces a separation between its IP licensing and chip manufacturing businesses — would create cascading uncertainty across every AI infrastructure roadmap dependent on Arm-based accelerators or CPUs. This probe also intersects with the broader geopolitical question of whether Arm, a SoftBank-owned but UK-headquartered company listed in the US, can continue licensing to Chinese customers under tightening export frameworks.

Alphabet closed a $17 billion bond issuance — its largest ever — and was already marketing additional debt before that deal settled, according to Bloomberg. The company is simultaneously tapping yen-denominated bond markets for the first time, a move that reflects both the scale of AI infrastructure financing requirements and the saturation of US dollar credit appetite for tech-sector paper at current spreads. This is confirmed capital deployment, not speculative: the $17 billion deal is closed.

The structural dynamic this reveals is critical for infrastructure planning: the capital requirements for competitive AI buildout — training clusters, data centre construction, power infrastructure, networking — are now large enough that even the world's most creditworthy technology companies are diversifying funding sources geographically. Anthropic's reported $30 billion raise at a $900 billion valuation, still described as early-stage talks by Bloomberg, reinforces the scale. CME's reported plan to create a futures market for computing power — noted in Bloomberg Tech's May 12 broadcast — would, if confirmed, represent the financialisation of raw compute capacity as a commodity asset class, with significant implications for how data centre operators and cloud providers hedge capacity risk.

Former Treasury Secretary Hank Paulson's public warning — that electricity shortages could become a major binding constraint on US AI leadership — landed in the same week that Bloomberg detailed China's concurrent investments in transmission capacity, renewables, batteries, and power generation at a scale that former US Ambassador Nicholas Burns described as already reshaping global supply chains. These are analyst and official assessments, not confirmed infrastructure completions, but they reflect a growing consensus view among senior policymakers.

The energy asymmetry is structural. US data centre buildout is running into grid interconnection queues that extend three to five years in many markets, combined with local permitting resistance and transmission constraints. China's state-directed energy buildout faces none of these coordination failures. The Oklo-Idaho National Laboratory partnership on AI-assisted advanced reactor design, confirmed by Data Center Dynamics, is a small but directionally significant signal that the US AI industry is beginning to pursue long-cycle nuclear solutions — but Oklo's Pluto reactor remains in design phase, not near-term capacity.

Kioxia's shares were locked in a glut of buy orders on Monday after the NAND flash supplier reported profit growth that trounced expectations, driven entirely by AI data centre demand for storage. The Alger CIO separately characterised AI memory demand as 'insatiable' in Bloomberg commentary, consistent with SK Hynix and Samsung's recent HBM capacity signals. These are confirmed financial results, not projections.

The infrastructure implication is that the AI compute stack's bottleneck is shifting progressively from raw GPU compute — where NVIDIA supply is scaling — toward memory bandwidth and storage density. HBM3E supply remains tight, with TSMC's CoWoS advanced packaging capacity still the primary constraint on how quickly HBM can be integrated into accelerator packages. A new technical paper roundup from Semiconductor Engineering included work on multi-chiplet memory-centric attention serving, an architectural approach specifically designed to address this bottleneck by placing memory closer to compute. These are research-stage results, but the direction of academic and industrial R&D confirms that the memory wall is a recognised and actively-worked problem at the frontier of AI hardware design.