The Compute Crunch: GPU Scarcity, Chinese Rivals, and a Decade-Low in Cash Flow

The AI infrastructure industry faces a defining tension: demand for compute is growing faster than supply, prices are rising in ways that compress margins across the ecosystem, and a Chinese competitor has entered the frontier model market at a price point that challenges the economics of the entire Western AI build-out. Google has capped Meta's access to Gemini AI compute amid what sources describe as a genuine infrastructure crunch — not a commercial dispute, but physical scarcity. When one of the largest AI developers in the world tells one of its largest customers it cannot honor current demand, it reveals where the capacity constraints actually lie.

The financial signal is stark. Hyperscaler free cash flow has hit a decade low while aggregate AI infrastructure spending has crossed $700 billion. Those two facts together describe companies spending at a pace that outstrips their ability to generate cash — not because the business is failing, but because the capital requirements of maintaining competitive position are simply enormous. Elon Musk has described what he called the biggest chip price surge he has ever seen, and Apple has raised Mac and iPad prices by up to $500, citing soaring memory costs from the AI data center build-out. The infrastructure investment is not an abstraction; it is showing up in the price of consumer devices.

SpaceX's data center business reportedly could top $76 billion in contracts by 2029, with Anthropic, Google, and Reflection AI among tenants. The vertical integration of launch capability, satellite internet through Starlink, and terrestrial data center leasing creates a bundle no other company can replicate. The FTC's clearance of Musk's acquisition of Mesh Optical Technologies — which makes high-speed optical transceivers connecting servers inside data centers — adds another layer to that stack. Regulators found no antitrust violation because Mesh is a relatively small player in a market with multiple competitors, but as the pieces of a vertically integrated AI infrastructure empire assemble, the competitive picture may look different.

The Chinese dimension represents the most structurally significant challenge. Zhipu, a Chinese AI company, has released an open-source model that reportedly matches Anthropic's Claude performance at roughly a fifth of the cost. The specific performance claims warrant caution — AI benchmarks can be gamed — but the directional claim is significant. The US has implemented significant restrictions on advanced AI chip exports to China specifically to slow competitive Chinese AI development. If Zhipu is achieving frontier-level performance at low cost on chips available within China, it raises serious questions about whether those export controls are achieving their intended effect. The open-source release compounds the challenge: by releasing the model publicly, Zhipu makes it available globally, including in jurisdictions where US AI companies might otherwise be dominant.

Goldman Sachs is meanwhile directing investment banking clients toward industrial and energy companies, arguing that the next phase of AI economic benefit flows to factories and mines rather than software companies. The logic is that AI deployment is moving beyond the data center phase into industrial automation — predictive maintenance, supply chain optimization, autonomous mining equipment. A Goldman strategist has separately argued that investors should shift from chips to cloud stocks, framing the semiconductor build-out phase as past its peak and the application layer as next.