Last week we covered Cerebras's valuation pressures and whether AI infrastructure valuations could hold under scrutiny. This week, the stress is no longer theoretical. Amazon signed a $17.5 billion delayed draw term loan with Citigroup, JPMorgan, Wells Fargo, HSBC, and BofA Securities just two days after reportedly raising $14 billion through a Canadian bond sale. That brings Amazon's new financing to roughly $31.5 billion in 48 hours, earmarked for general corporate purposes—which in this environment means AI infrastructure. The delta: debt is climbing faster than revenue visibility, and the infrastructure layer is doubling down on a singular thesis that demand for compute will justify any level of capital deployment.
Amazon is hardly alone in the borrowing spree. Google parent Alphabet announced plans to raise $80 billion through a stock sale to fund investments while retaining a healthy balance sheet. Meta disclosed its largest bond sale ever at $30 billion. The scale is striking even by Silicon Valley standards, and the rhetoric is uniform: companies frame the spending as necessary to remain competitive in the AI arms race. What's missing from every announcement is the return horizon. Investors and analysts are no longer asking whether the spending is necessary—they're asking whether the returns will ever justify it.
If the early AI trade was compute, the new question is whether model efficiency improves faster than demand grows—leaving billions in GPU investments facing rapid obsolescence.
The bull case rests on exponential demand. Industry observers estimate that AI labs OpenAI and Anthropic are targeting roughly 3 to 3.5 gigawatts of capacity by the end of 2025, with OpenAI discussing plans to reach 30 gigawatts by 2030. Assuming those two labs represent 30 to 40 percent of new builds, the total market could add more than 150 gigawatts of capacity over the next four and a half years. At roughly $50 billion per gigawatt—70 percent for chips and compute, 20 percent for datacenter shell and power infrastructure, the remainder for cooling and land—that translates to $7.5 trillion in total spend, or approximately $1.7 trillion annually. That figure represents roughly 5 percent of current global GDP, concentrated in a single infrastructure category.
Adjacent markets are signaling the same capital intensity. Tesla, Ford, and GM are all pushing into energy storage to support datacenter expansion, with annual stationary battery installations expected to exceed 110 gigawatt-hours by 2030, double current levels. GM rolled out a new sodium-ion battery chemistry aimed at the heart of the market. Data center energy demand is expected to nearly triple by the end of the decade, driven not just by AI but by broader electrification across transportation, manufacturing, and HVAC. Base Power raised a $1 billion Series C in October; Lunar Energy raised $232 million for residential batteries. The infrastructure thesis is metastasizing beyond compute into the entire energy stack.
The counter-thesis is equally compelling: efficiency may outpace demand growth. Microsoft recently added a new metric to model release cards—average token usage. The company's latest model hit 71.6 on SWE-Bench Verified using roughly a third of the tokens Claude Haiku 4.5 consumed to achieve similar performance. Benchmarks are now measured on two dimensions: overall performance and cost to achieve that intelligence. This is a structural shift. Uber reportedly capped employee AI spending after blowing through its budget in four months. Salesforce is spending $300 million on Anthropic tokens and has frozen engineering hires. Microsoft cancelled Claude Code licenses across its Experiences and Devices division after engineering usage outran budgets.
The efficiency gains are not incremental. Skill distillation—where frontier models author procedural workflows that smaller, local models execute—fundamentally changes the cost structure. A frontier model like Opus 4.7 or GPT-5.1 writes and refines skill files; a smaller model running locally, such as Qwen 35B or Gemma 26B, executes them. This is different from classical knowledge distillation or instruction tuning. The smaller model retrieves procedures rather than learning compressed weights, and it runs on consumer hardware. If this pattern generalizes, the edge computing thesis gains traction, and centralized GPU clusters face utilization risk. The frontier model becomes a teacher; the lightweight model becomes the workhorse.
Memory and storage constraints are already emerging as the next bottleneck, and they carry depreciation risk. SK Hynix is up over 300 percent in the past year; Samsung up over 200 percent; Kioxia up roughly 1,000 percent; SanDisk up over 1,200 percent. The surge reflects a global shortage in NAND flash memory supply, which was foreseeable but insufficiently provisioned. GPU supply has begun to normalize relative to two years ago, but memory and storage now feel like GPUs did then. The question is whether this constraint persists or whether oversupply follows as manufacturers ramp capacity. If the latter, the memory boom becomes the next cycle of overcapitalization, and datacenter operators holding depreciated hardware face write-downs.
The second-hand GPU market is the canary. Neoclouds—alternative compute providers offering bare metal, managed Kubernetes, or full cloud services—are betting that chips retain value after a four- to five-year useful life. That assumption underpins their unit economics. If model efficiency improvements render current-generation GPUs obsolete faster than the depreciation schedule assumes, neoclouds face margin compression or stranded assets. The market is beginning to price in this risk. Depreciation curves are steepening, and buyers are demanding discounts for older hardware. Whether this is prescient caution or premature skepticism will determine whether the current buildout is visionary or a historic overcapitalization.
The infrastructure layer is now a leveraged bet on a single variable: whether token consumption grows faster than cost-per-token declines. If demand scales exponentially—if AGI requires orders of magnitude more compute, if every enterprise workflow becomes agentic, if inference becomes the dominant workload—the $7.5 trillion datacenter buildout will look conservative in hindsight. If efficiency improves faster—if distillation and edge computing collapse inference costs, if models plateau in capability, if regulation or energy constraints cap growth—billions in GPU and datacenter investments face rapid obsolescence. The debt is already on the balance sheet. The return horizon remains unknown.
What's clear is that efficiency is now a competitive advantage, not a footnote. Application layers are shifting from dollars-per-token to dollars-per-outcome, pricing closed tickets, shipped pull requests, and resolved support cases rather than API calls. Model companies must compete on both performance and cost; infrastructure providers must compete on both capacity and utilization. The era of subsidies and tokenmaxxing is over. The question is whether the infrastructure already financed can adapt quickly enough to remain economically viable as the workload evolves. The next twelve months will reveal whether the current capital addiction is building the foundation for the next platform or financing the next wave of write-downs.