When the world’s largest asset manager predicts a futures market for computing power and an exchange launches one within a week, the signal seems clear: compute has become the AI era’s scarce commodity. Yet the same industry straining against that shortage is engineering its way out of it, targeting cost reductions of up to two orders of magnitude.
Introduction
When the world’s largest asset manager predicts a futures market for computing power and an exchange launches one within a week, the signal seems clear: compute has become the AI era’s scarce commodity. Yet the same industry straining against that shortage is engineering its way out of it, targeting cost reductions of up to two orders of magnitude.
New asset classes typically follow a pattern: a critical resource becomes scarce, demand rises, and financial markets build instruments to manage that scarcity, as with oil futures, then carbon credits. In 2026, the same dynamic is taking shape around AI computing power, as demand for AI infrastructure continues to outpace supply.
1. The Shortage Is Real, Not Just One CEO’s View
At the Milken Institute Global Conference in May, BlackRock CEO Larry Fink argued that demand for compute had grown so intense that “a new asset class will be buying futures of compute,” citing shortages in capacity, chips and memory across the US. He dismissed AI bubble concerns, pointing instead to supply shortages outpacing demand. BlackRock (NYSE:) is backing that view with capital, including a ~$40 billion acquisition of Aligned Data Centers and a $10.7 billion deal for power producer AES. Brookfield CEO Bruce Flatt echoed the scale of the shift, describing a decade-long “rewiring of the global economy.”
Source: Bloomberg
Ground-level research corroborates this. A June 2026 Silicon Valley trip by tech-equity investors Brook Dane and Sung Cho found a compute-constrained environment extending beyond GPUs to ASICs and memory chips, in what Cho called an unprecedented up-cycle. The strain is driven by demand concentration: the top 5% of enterprise users consume roughly three times the tokens of the median company, with inference now overtaking training as the dominant volume driver.
Source: Labor Department
This structural mismatch is what makes the shortage investable: if supply can’t meet demand within a year, capex must stretch over three to five years, supporting Dane’s view that valuations across the supply chain remain “very attractive.”
2. A Tradable Market Is Already Being Built
Within a week of Fink’s remarks, on 12 May 2026, CME Group announced what’s billed as the first futures market for computing power. CEO Terry Duffy summed up the logic: “Compute is the new oil of the 21st century.” The contracts let companies and traders hedge or speculate on GPU rental prices, much as airlines hedge fuel costs, offering enterprises a way to fix forward costs, and investors a new route to AI infrastructure exposure.
The speed of the launch is itself telling: an exchange only lists a contract when the underlying is scarce, volatile and significant enough to generate hedging demand and liquidity. That a major derivatives venue reached this judgment within days suggests compute-as-commodity has moved from metaphor to consensus.
But a futures market embeds a directional bet: that today’s scarcity persists long enough to sustain multi-year hedging demand. That assumption is exactly what the industry’s own engineering roadmap is working to undermine.
3. Will the Scarcity That Justifies the Market Persist?
Engineering analysis of the AI inference stack suggests costs per token could fall sharply, by up to two orders of magnitude. Levers already in deployment include model quantisation and pruning (85 to 95% cost reduction), FP16-to-INT4 precision shifts (2 to 4x throughput gains), advanced 3D packaging (80 to 90% reduction), custom inference ASICs (70 to 80%), and co-packaged optics (50 to 65% lower energy per bit). These aren’t distant gains: hyperscalers are committing over $700 billion in 2026 capex, much of it toward industrialising these efficiencies.
This creates real tension: futures markets bet compute stays scarce; the engineering roadmap bets unit economics collapse. The two can coexist short term, but point in opposite directions long term. There’s a further wrinkle: efficiency gains could let countries with older chip technology, notably China, run optimised models cheaply, turning a disadvantage into a competitive edge and complicating the “US capacity shortage” narrative.
Conclusion
The futures curve won’t just price compute. It will reveal the market’s honest view on how long scarcity lasts. A steep multi-year premium backs a structural shortage; a flat curve bets on efficiency winning out. Either way, the contract becomes a live referendum on whether compute remains AI’s defining constraint or becomes a commoditised input.
