Compute Cost Curve Becomes Institutional Asset Allocation Signal

$690 billion. That is the combined capital expenditure five companies — Microsoft, Alphabet, Amazon, Meta, and Oracle — plan to spend on AI infrastructure in 2026. Nearly double 2025. More than the entire publicly traded US energy sector spent, combined, in any prior year. And it is only the beginning of a multi-year cycle reshaping how institutional capital evaluates compute.

For most of the last two decades, compute was a cost centre. Cloud providers managed it, startups rented it, and no institutional allocator thought of it as an investment thesis. That has inverted. In May 2026, CME Group announced a partnership with Silicon Data to launch the first regulated compute futures market, pending regulatory review. The product is built on daily GPU benchmark indices tracking on-demand rental rates for Nvidia H100 and B200 chips. Sovereign wealth funds and pension plans have already deployed an estimated $120 billion into AI data centers in 2025–2026, drawn by 15-year leases with investment-grade counterparties. The shift is not hypothetical. It is happening now, and the infrastructure being built today will determine portfolios for the next decade.

The declining cost curve

Compute costs are falling faster than any comparable industrial input in modern history. The Nvidia H100, launched in 2023 at roughly $30,000 per unit on the secondary market, can now be rented for $2.69 per hour through specialist GPU cloud providers — a decline of more than 75% from peak pricing in early 2024. The successor B200 delivers 2.5× the inference throughput at a comparable or lower per-unit cost. Per-token pricing for leading large language models has dropped by roughly a factor of ten over the past eighteen months, from around $15 per million tokens for GPT-4 at launch to below $1.50 for equivalent frontier models in mid-2026. The trajectory is exponential, not linear.

These numbers matter because the cost curve changes the risk profile of compute-enabled assets. A data centre lease that looked marginal at $12 per GPU-hour becomes attractive at $3. The unit economics shift structurally. Institutional capital, which moves in vintage years and multi-year commitment cycles, needs pricing signals that reflect this trajectory rather than spot volatility. CME's compute futures are designed to provide that: a forward curve for compute, analogous to what crude oil futures did for energy markets in the 1980s.

When crude oil futures launched on NYMEX in 1983, the market was opaque, dominated by a small number of vertically integrated producers, and pricing was set through bilateral contracts. Within a decade, the futures market transformed oil from an operational input into a financial asset, creating the pricing transparency that enabled energy as a distinct institutional asset class — now roughly $3 trillion in allocations globally. Compute follows the same arc: from bilateral GPU rental agreements and fragmented spot pricing toward a standardized, exchange-traded benchmark. The difference is speed. CME's announcement came less than three years after the launch of GPT-4. Oil futures emerged more than a century after the first commercial oil well.

Silicon Data, the startup that built the underlying benchmark, now publishes daily indices covering GPU rental rates, inference cost per token, and forward pricing curves. The data is already available on Bloomberg and the London Stock Exchange Group platforms. For the first time, an institutional allocator can evaluate compute infrastructure using the same pricing transparency framework they apply to power purchase agreements or toll roads.

Institutional portfolio construction

The investor base for compute infrastructure has expanded far beyond venture capital. Universal Asset Owners, a sovereign wealth fund research platform, estimates that institutional investors — pension funds, endowments, sovereign wealth funds — committed over $120 billion to AI data centers and digital infrastructure in 2025–2026. Gulf sovereign funds alone accounted for 43% of sovereign wealth fund investment worldwide in 2025, directing capital toward compute corridors and data centre platforms. British Columbia Investment Management Corporation recorded a 7.6% return from its infrastructure and renewable resources portfolio for fiscal 2026, with a record C$4.7 billion in new commitments, partly allocated to digital infrastructure.

The appeal is structural. A data centre with a 10-to-15-year lease from an investment-grade hyperscaler produces contracted, inflation-linked cash flows that resemble a long-duration infrastructure bond, not a technology equity. For institutions managing defined-benefit pension obligations with 30-year horizons, that duration alignment matters more than the AI narrative. As we wrote in July when compute derivatives became an institutional infrastructure asset class, the financialization of compute is creating a new layer between the physical data centre and the allocator's portfolio.

Returns data backs this. Sovereign Infrastructure, a category tracked by the AI Asset X Lighthouse Report, shows expected returns of 12–16% for energy and grid assets, 16–22% for data centres and compute platforms with contracted offtake, and 18–25% for digital state platforms. These return profiles sit between core infrastructure and private equity, a risk-return niche historically underrepresented in institutional portfolios.

GIC, ADIA, Mubadala, and other sovereign funds have been the most aggressive allocators. They saw early that a data center with a 15-year hyperscaler lease does not behave like a technology investment. It behaves like a toll road: contracted revenue, inflation-linked escalation, finite construction risk, and a predictable exit via secondary sale to a core infrastructure fund. The difference is that the toll road's traffic growth follows GDP, while the data center's utilization follows AI inference demand growing at roughly 40% compound annually.

The funding mix is also shifting. The Sovereign Infrastructure model projects 60–70% institutional capital (pensions, sovereign wealth funds, insurers) as the primary source, with 15–20% from multilateral development banks and development finance institutions providing first-loss or guarantee layers, 10–15% sovereign equity, and only 5–10% strategic corporate co-investment. This capital stack mirrors the financing structure of large-scale renewable energy projects — a template institutional allocators already understand and have the internal models to evaluate.

The LP response so far is telling. Allocations to digital infrastructure funds have grown from a niche within infrastructure to a standalone mandate at several large pension plans. The reason is structural: defined-benefit plans need long-duration, inflation-linked cash flows with low correlation to public equities, and AI data centers deliver precisely that profile when structured correctly. The open question is whether the asset class can absorb the volume of capital seeking entry without compressing returns to infrastructure-bond levels, where the premium over core infrastructure becomes too thin to justify the construction risk.

The $725 billion capex cycle

The current capex cycle has no precedent. A.L. Capital Advisory, a CFA-affiliated research firm, estimates the Big-5 hyperscalers will spend roughly $725 billion on AI infrastructure in 2026. That is about 5% of projected US GDP. Compare that to the fiber-optic build-out of the late 1990s, which peaked at 2% of GDP. The electrification of the US economy in the 1920s peaked at 1.5%. The current cycle is 2.5 times the fiber overbuild and 3 times the electrification peak.

The structural difference from prior technology waves is the shift from cash-funded to debt-funded expansion. Morgan Stanley and J.P. Morgan project the technology sector needs roughly $1.5 trillion in new debt over the next three years to sustain the build-out. The hyperscalers now spend 45–57% of revenue on capex, ratios previously seen only in capital-intensive utilities. Amazon's $200 billion capex guidance for 2026 alone exceeds the combined annual capex of the entire publicly traded US energy sector. That number bears repeating: $200 billion, one company, one year, one category of spending.

Aggregate investment-grade bond issuance from big tech AI companies surged past $200 billion in 2025. The sector became the dominant new source of supply in global credit markets. Meta priced a $30 billion investment-grade bond in October 2025, the largest single corporate bond financing that year and among the largest on record. This debt-funded model turns hyperscaler balance sheets from cash-rich technology platforms into leveraged infrastructure operators, with direct implications for credit analysts and fixed-income allocators.

Asset allocation framework for compute

The framework for sizing compute within an institutional portfolio is still taking shape, but several features distinguish it from traditional real asset categories.

First, correlation profile. Compute infrastructure cash flows — contracted leases with hyperscalers, power purchase agreements with utilities, GPU rental pools — show near-zero correlation to broad equity markets over the typical measurement window. The demand driver is build-out of AI inference capacity, which follows model release cycles and enterprise adoption timelines, not consumer spending or interest rate cycles. For a pension fund managing liability-driven investment strategies, this decorrelation is the primary portfolio construction rationale. It is the same logic that drove pension capital into core infrastructure and private credit over the past two decades, applied to an asset class built around chips and kilowatts instead of toll booths and transmission lines.

Second, inflation sensitivity. Data center leases typically include annual rent escalators tied to CPI or power cost indices. The Sovereign Infrastructure benchmark data shows 12–16% expected returns for energy and grid assets and 16–22% for data center platforms — both with contractual revenue visibility. In a regime of persistent inflation or regime change between monetary eras, these structures act as a natural hedge, unlike fixed-rate bonds whose real yields compress when inflation surprises to the upside.

Third, vintage and scaling. Unlike venture capital, where returns follow a power-law distribution, compute infrastructure generates returns that are more normally distributed and more predictable. The key variable is not selecting the right GPU or data center developer but securing power availability and permitting timelines. A fund that locks in a 10-year lease at the right power cost has a fundamentally different risk profile from a fund betting on a specific AI model winning. This is the distinction between infrastructure and technology equity that institutional allocators are beginning to price. The two asset classes require different diligence frameworks, holding periods, and return expectations, and mixing them up is the fastest way to misallocate capital.

McKinsey's Global Private Markets Report 2026 says private equity is now a mature industry where alpha is less likely to emerge from market dynamics alone. It increasingly depends on operational value. Compute infrastructure sits at the boundary between these two worlds: the return comes from operational execution — permitting, power procurement, construction timelines and grid interconnection — not from multiple expansion or market timing. For allocators accustomed to the structural growth of infrastructure as an asset class, which has grown from roughly $500 billion in assets under management in 2010 to over $3 trillion today, compute represents the next logical category expansion. The infrastructure playbook already exists. Compute merely writes a new chapter.

Limits: energy, regulation, overbuild

The constraints on this cycle are physical. Data center electricity consumption could reach 945 terawatt-hours by 2030, roughly equivalent to Japan's total current electricity use, according to the International Energy Agency. Energy costs already represent 30–40% of data center operating expenses, and a single Nvidia GB200 NVL72 rack can draw more than 120 kilowatts of continuous power — more than the average American home uses in a week. In regions with limited grid capacity — Northern Virginia, Singapore, Amsterdam — regulators have already imposed moratoria on new data center construction. These physical limits cap supply growth, which from an allocator's perspective supports the pricing power of existing assets.

The overbuild risk differs from the 2001 telecom bust. In that cycle, fiber was laid but never lit — stranded assets with no revenue. Today's data center capacity is leased before it is built, with hyperscalers signing 10-to-15-year commitments locking in utilization before breaking ground. The risk is not empty data centers but compressed returns: if compute pricing falls faster than the amortization schedule of the underlying infrastructure, the yield will disappoint, even at high utilization.

The compute futures market solves this. A data center operator with a long position can hedge against declining GPU rental rates, locking in a minimum return the same way a wheat farmer locks in a harvest price. If the market develops the liquidity its proponents expect, it becomes a risk-transfer mechanism for the largest infrastructure build-out in economic history.