The AI compute market has experienced a dramatic shift in pricing dynamics, with GPU rental prices rising vertically rather than declining as expected by market participants. According to Silicon Data, the price to rent a Nvidia Hopper H100 chip for an hour has risen to $2.71, up from $1.96 at the end of November, with forward rates compiled by Silicon Data curving upwards into 2027 and 2028. Tech investor Gavin Baker noted that "everyone in 2024 and 2025 thought that GPU prices would decline slowly, if you were bearish, you thought they would decline precipitously. I don't think anyone in '24 or '25 thought that the prices of old GPUs would be going vertical." One rental customer, Baseten, revealed that its cloud service provider will push up rental prices on Nvidia Blackwell B200 GPUs from $2.63 per hour to $5.10 when contracts renew in October. Even older vintage chips are appreciating, with CoreWeave signing contracts to lease Nvidia Ampere A100 chips introduced in 2020 all the way out to 2029 at attractive prices.
Memory has become a critical bottleneck in the AI arms race, with Elon Musk identifying it as the "limiting factor" in infrastructure costs. As reported by multiple sources, memory demand is compounding at 200% per year, significantly outpacing supply growth of just 20% annually. Musk specifically highlighted that "if you've got demand increasing much faster than supply, then Economics 101 would suggest that the price increases. It does not decrease." High-bandwidth memory (HBM) stacked next to GPUs is essential for feeding processors with enormous volumes of data, as modern AI models contain hundreds of billions or even trillions of parameters that must be accessed at breakneck speed. The supply gap is particularly acute because nobody can simply spin up a new memory fab on short notice, creating a situation where Economics 101 would suggest that the price increases.
The scale of AI compute expansion is unprecedented, with hyperscalers planning combined 2026 capex of approximately $860 billion, up 80%, targeting $1.2 trillion in 2027. As reported by Investing.com India, Amazon lifted its spending to $220 billion and warned it "will still not have enough capacity," while Alphabet is at $205 billion with guidance for "significantly higher." SpaceX's compute revenue hit $2.6 billion (+247%) and turned EBITDA-positive, with management targeting closer to 10GW than 5GW by end-2027 - a $200 billion business on their own calculations. Meta runs near 7GW and is doubling toward 14GW, demonstrating that the two companies previously identified as potential market oversupppliers are now among the largest buyers.
The resilience of GPU pricing has catalyzed a major financial market development, with Nvidia CEO Jensen Huang announcing a deal with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital to support AI infrastructure buildout. Blackstone president Jon Gray noted that "people are going to begin to recognize that this is a financeable asset class," comparing it to how banks underwrite home loans while also considering the value of the property. Larry Fink of BlackRock went further, stating "This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s and I look upon this as a next future for financial engineering." The deal's success depends on whether Huang's assessment that his chips are durable proves correct, potentially expanding the scope for financing AI infrastructure significantly.
The scale of infrastructure investment reveals the long-term nature of this build cycle. As reported by Investing.com India, it takes approximately $37.6 billion to stand up one gigawatt, with well over half going to physical plant infrastructure including land, power, transformers, cooling and fiber systems. Building timelines range from 24-36 months for data centers to four-to-seven years for power infrastructure, with transformers requiring over 160 weeks and switchgear extending through 2028. The analysis suggests this represents the beginning of a new cycle rather than a peak, with capital being re-engineered around AI compute through initiatives like Nvidia backstopping GPU residual values and CME/ICE launching compute futures.
