AI Server Rental Costs Declining, Challenging the Compute Scarcity Narrative
Miles Bennett
AI server rental prices keep sliding while orchestration frameworks and open-source models close the performance gap on far less compute — the scarcity narrative faces a two-front challenge, and Goldman's trading desk says hardware gets hit first.
Why are server rental prices the key signal to watch?
Goldman Sachs 1-Delta desk head Rich Privorotsky treats server rental prices as the core gauge for whether the AI hardware investment thesis still holds.
This means → if rising supply keeps pushing rental costs down, it directly undercuts the claim that compute is scarce — and hardware is the first sector to feel the pressure.
Rental costs are already on a clear downtrend. But Privorotsky notes the current hardware trading logic will persist until Token spending patterns — how much companies pay for AI inference — shift fundamentally.
What is the market watching next?
Privorotsky flags that market attention is shifting to hyperscaler pricing moves — Microsoft Azure, AWS, Google Cloud.
This means → any strategic signal from these cloud giants — order cuts, price compression — would force a reassessment of the entire AI investment cycle's underlying logic.
Semiconductor ETFs logged unusually high inflows last week, confirming the market is still betting on compute expansion. But that crowded positioning amplifies drawdown risk if the narrative turns.
How does the Fugu framework outperform giant models with less compute?
Fugu, built by Japanese AI lab Sakana, is not a single large model. It is a dynamic orchestrator — in plain terms = a dispatcher that splits a task into parts, routes each to the best-suited frontier model for parallel processing, then merges the outputs into a result better than any single model alone.
On the SWE-Bench Pro coding benchmark, Fugu scored 73.7, beating Claude Opus 4.8's 69.2 and GPT-5.5's 58.6.
This reflects a fundamental shift: frontier-level performance may not require more training compute. Privorotsky calls it "model orchestration and fusion." If this approach is validated more broadly, it challenges the core logic of the compute arms race.
How close are open-source models to the best closed-source ones?
Zhipu's GLM-5.2 scored 74.4 on the FrontierSWE long-horizon coding benchmark — within roughly one percentage point of Anthropic's top closed-source model Opus 4.8 at 75.1, and above GPT-5.5's 72.6.
In plain terms = open-source models have nearly matched the strongest closed-source coding performance, at a price roughly 72% to 82% lower.
GLM-5.2 ships under the MIT license — meaning anyone can use, modify, distill, or reproduce the model for free — which will compress Token costs further.
Can falling rental prices and rising capex coexist?
Privorotsky notes the current incentive structure still points toward more capex, not less — the hyperscalers are still building data centers at full speed.
Yet falling rental prices, rising orchestration frameworks, and open-source catch-up all say the same thing: the value of each unit of compute is being diluted.
This means → whether these two forces can coexist long-term is the central unresolved question for the current AI compute expansion cycle. If they cannot, hardware will be the first sector to deliver the answer.
Content is for reference only, not financial advice.