One new data point on graphics processor unit useful life was provided by Oracle on its most-recent conference call. “Of all the GPUs that came up for renewal in the first quarter of 2026, that capacity was renewed or resold at a 20 percent premium to prior contracts,” said Clay Magouyrk, Oracle CEO. “The majority of those GPUs are four years or older.”
To be sure, that might not apply universally to every supplier of compute as a service using GPUs. Nor does it necessarily reflect the most-recent sales, as pre-sold capacity is included in the figures.
Still, the data suggests high demand for AI compute services, as Oracle's GPU fleet ran at 97.9 percent utilization, meaning nearly every chip was busy serving customers.
That might well suggest that demand absorbed the hardware as fast as Oracle could install it. Other metrics also suggest continued high demand outrunning supply: Oracle closed more than $30 billion of new AI contracts in the quarter.
Oracle’s remaining performance obligations (contracted revenue it hasn't yet delivered) reached $664 billion, up $209 billion year over year.
Computing hardware is supposed to get cheaper as it ages. In this case, that did not happen, suggesting strong demand for public AI “compute as a service” is pushing prices higher. That might not apply to private enterprise capacity, though.
Company / indicator | Latest figure | What it suggests about capacity constraints |
Microsoft Azure | Azure revenue +40% YoY in FY26 Q3; management said demand continued to exceed available capacity | This is probably the cleanest large-scale evidence: Microsoft says customers want more Azure capacity than it can currently provide |
Microsoft — forward capacity | Microsoft expects to remain constrained through at least 2026 | Management explicitly says additional GPU/CPU/storage investment will not eliminate the constraint this year |
Microsoft — capacity expansion | Added another 1 GW of capacity in FY26 Q4; 88 new data centers during FY26 | It is simultaneously adding enormous capacity and still describing supply as insufficient |
Oracle AI Cloud | $30B+ of new AI-cloud contracts in Q1 FY27; RPO reached $664B | Oracle explicitly says AI training/inference demand is growing faster than supply |
Oracle GPU deliveries | 300,000+ GPUs delivered in one quarter, nearly 3× Q4 FY26 | Very rapid physical deployment still isn't catching demand |
CoreWeave | $104.2B backlog in Q2; another >$25B of commitments in early Q3 | AI-specialist cloud is effectively selling future capacity years in advance |
CoreWeave near-term fleet | Management says it remains largely sold out, including older A100/H100/H200 capacity | Particularly important: scarcity has spread beyond the newest Blackwell GPUs |
Google Cloud | Q4 2025 cloud backlog jumped to $240B, +55% sequentially | Huge contracted demand; Google says it operates in a tight supply environment |
NVIDIA | FY27 Q2 Data Center revenue $89B, +117% YoY and +18% QoQ | The chip supplier continues to see extraordinary end-demand as new capacity becomes available |
NVIDIA forward outlook | FY28 revenue expected to grow roughly 70%, explicitly described as a “supply-constrained outlook” | NVIDIA says even its own growth forecast is limited by supply rather than demand |
NVIDIA/AWS | AWS committing to deploy an additional 2M NVIDIA GPUs through FY29 | One hyperscaler is committing to an enormous future inventory of AI compute |
Amazon/AWS | AI revenue reportedly exceeded $25B annualized; Anthropic and OpenAI have made multi-year, multi-GW Trainium commitments | Increasingly large customers are reserving dedicated accelerator capacity years ahead |
Microsoft OpenAI commitment | OpenAI contracted an incremental $250B of Azure services | One customer alone has contracted capacity on a scale that helps explain why Azure remains constrained |
Google/Anthropic | Multi-year, multi-GW TPU commitments | Frontier-model companies are securing enormous future compute supplies rather than buying only spot capacity |
AI GPU prices | H100/H200 pricing remains elevated in constrained markets; CoreWeave says pricing for both new and older GPUs is strengthening | If supply were comfortably ahead of demand, one would expect much stronger price erosion |
Enterprise GPU utilization — counterevidence | Cast AI finds average GPU utilization of only ~5% across tens of thousands of Kubernetes clusters | Shows that owned/provisioned enterprise GPUs can be dramatically underutilized even while cloud capacity is scarce |
The useful life of GPU infrastructure is not the only issue when assessing the value and cost of AI compute as a service businesses. But if Oracle’s experience is broadly the case, the business case is helped.
No comments:
Post a Comment