Compute & Infrastructure

Bloomberg reports that Oracle and OpenAI have scrapped plans to expand their flagship AI data centre in Abilene, Texas, after protracted negotiations over financing and OpenAI's evolving capacity requirements. The collapsed talks created an opening for Meta Platforms to consider leasing the planned expansion site from developer Crusoe, with Nvidia facilitating discussions between Meta and the developer. This represents a significant shift in the data centre capacity landscape, as the Oracle-OpenAI partnership had been positioned as a cornerstone of OpenAI's infrastructure strategy. The breakdown suggests either financing constraints at Oracle — which is simultaneously planning thousands of job cuts to manage cash flow pressures from AI data centre spending — or a strategic recalibration by OpenAI regarding its compute needs and supplier relationships.

The involvement of Nvidia as a broker between Meta and Crusoe underscores the chip manufacturer's expanding role beyond hardware provision into infrastructure orchestration, leveraging its position as the dominant GPU supplier to influence data centre capacity allocation. Meta's potential entry as a tenant would represent a consolidation of major AI compute capacity under fewer hyperscale operators, raising questions about whether independent AI developers will face tightening access to frontier-scale infrastructure.

Bloomberg reports that the US Commerce Department has drafted regulations restricting AI chip shipments to anywhere in the world without American approval, extending export control frameworks beyond adversary nations to all markets. This would represent a fundamental shift from targeted controls on China and Russia to a universal permitting regime for advanced AI semiconductors from Nvidia, AMD, and other US manufacturers. The regulatory framework appears designed to give Washington leverage over global AI development by controlling access to the computational substrate, potentially as a tool for enforcing AI safety standards, preventing proliferation to hostile actors, or maintaining American technological leadership.

The implications for semiconductor supply chains are profound. Allied nations that have relied on relatively frictionless access to US AI chips would now face bureaucratic gatekeeping, potentially accelerating efforts in Europe, Japan, and elsewhere to develop domestic alternatives or source from non-US suppliers. For Nvidia and AMD, the regulations could constrain revenue growth in international markets while creating compliance burdens that favour larger, better-resourced customers over startups and research institutions. The draft status means implementation timing remains uncertain, but the signal alone may drive precautionary stockpiling or diversification strategies among international buyers.

Bloomberg reports that HD Hyundai Electric, South Korea's largest power equipment manufacturer, is accelerating US expansion based on expectations that AI infrastructure will drive surging demand for transformers and switchgear. The company is positioning itself to capitalise on what it characterises as an AI 'supercycle' in power consumption, reflecting growing recognition among electrical equipment suppliers that data centre buildout represents a structural shift in electricity demand patterns. This follows mounting evidence that power availability — not just chip supply — is becoming a binding constraint on AI infrastructure expansion, with utilities and grid operators struggling to provision capacity for facilities drawing hundreds of megawatts.

The strategic calculus by a major Asian power equipment manufacturer to expand US manufacturing capacity signals confidence that AI data centre construction will continue at scale despite financial pressures at operators like Oracle. It also highlights the lengthening supply chains and lead times for electrical infrastructure, where transformer manufacturing capacity and grid connection timelines increasingly determine data centre deployment schedules. The power infrastructure bottleneck may prove more intractable than semiconductor supply constraints, as electrical equipment manufacturing lacks the policy focus and investment momentum of chip fabrication.

Bloomberg reports that AI chipmaker Cerebras Systems has retained Morgan Stanley to lead a renewed initial public offering attempt, signalling continued investor appetite for compute infrastructure companies despite recent market volatility and geopolitical tensions. Cerebras, which manufactures wafer-scale AI processors designed to compete with Nvidia's GPU-based approach, previously attempted to go public but faced challenges including scrutiny over revenue concentration among a small number of customers. The renewed IPO push suggests either improved financial metrics, diversified customer base, or investor willingness to overlook concentration risks given the strategic importance of compute capacity.

Cerebras represents a rare credible competitor to Nvidia in AI training workloads, with its wafer-scale engine offering advantages in memory bandwidth and inter-core communication for specific model architectures. However, the company faces significant challenges including limited software ecosystem maturity compared to Nvidia's CUDA platform, dependence on TSMC for fabrication (creating similar supply chain vulnerabilities to other fabless chip designers), and competition from hyperscalers developing custom silicon. A successful IPO would provide capital for expanded production capacity and software development while offering public market investors direct exposure to AI infrastructure beyond Nvidia and the cloud hyperscalers.