Sandisk reportedly has approximately $94 billion of NAND contracts at floor prices. The question is whether AI memory remains a cyclical product or is being remade into contracted infrastructure.
That question extends across AI infrastructure, where demand is increasingly expressed through multiyear contracts, prepayments, guarantees, infrastructure debt, and institutional ownership. The investment question is shifting from whether AI demand exists to who retains the cash flow after financing, utilization, pricing, and depreciation.
Contracts Make Memory More Predictable, Not Less Cyclical
Sandisk has reportedly signed eight agreements representing approximately $94 billion in total contract value at floor pricing. The agreements have a weighted-average duration exceeding four years and include $16.5 billion in financial guarantees.
JPMorgan expects them to cover more than 50% of FY27 bits and approximately two-thirds of FY28 bits. Estimated gross margin is approximately 80% even at the floor tier.
If the agreements work as intended, Sandisk gains greater visibility into future volume, pricing, and margins. But contracts do not abolish cyclicality. They redistribute it.
JPMorgan estimated that the NAND total addressable market could rise from approximately $70 billion in CY25 to more than $300 billion in CY26 and approximately $500 billion in CY27, driven by data center inference workloads, including persistent KV cache. Sandisk’s FY28 to FY30 model targets mid-to-high-teens revenue growth, approximately 80% gross margin, approximately 75% operating margin, and approximately 50% adjusted free cash flow margin.
Those are aggressive expectations. A separate memory investor warned that if AI capital spending enters a “digestion phase,” “memory won’t be immune.”
Floor pricing and guarantees can protect economics within the obligations and credit quality of the agreements. They cannot guarantee that the broader industry will avoid excess capacity.
Revenue Visibility Is Not Equity Value
Nebius reported $582 million of revenue, up 454% year over year. Annualized recurring revenue reached $3.0 billion, up 58% quarter over quarter, while four contracts signed during the quarter each exceeded $1 billion in total contract value.
That suggests the neocloud model is moving beyond selling access to scarce GPUs. Nebius is attempting to turn power, customer commitments, and software layers into a platform with more predictable revenue.
But AI campuses require enormous amounts of capital before generating revenue. The financing structure determines how much operating cash flow ultimately reaches shareholders.
CoreWeave illustrates the distinction. Its backlog supports the case that demand is real. A skeptical account, however, pointed to “contractual debt waterfalls” and claimed senior lenders structured facilities around a 1.15x debt service coverage ratio floor. Under that interpretation, as much as 87% of net operating income could be directed toward debt service.
That claim does not disprove the backlog. It raises the more important valuation question: after paying for the infrastructure, how much of the economics belongs to the equity?
Institutional Capital Is Testing Infrastructure-Like Returns
Jensen Huang reportedly called NVIDIA AI factories a “NEW investable asset class.”
Goldman Sachs disclosed a $1.48 billion IREN position representing 33,642,054 shares and 9.40% ownership in a Form 13G filed on August 12, 2026. PSP Investments disclosed 1.8 million IREN shares valued at $82.3 million.
Those filings do not reveal the investors’ theses, hedges, or current holdings. They do suggest that large pools of capital are evaluating whether power-intensive compute can produce infrastructure-like returns.
Scarcity Must Still Be Converted Into Returns
The physical supply chain continues to show genuine scarcity.
Lumentum reported quarterly revenue of $1.01 billion, up 109% year over year, and expects to reach its $1.25 billion quarterly revenue target more than one quarter early. CEO Michael Hurlston said the “demand signal had strengthened” and “we are way behind our shipments on high-powered lasers.”
Coherent CEO Jim Anderson said, “we have seen absolutely no push-out of CPO demand,” and described “exceptional customer demand.” Lumentum, Coherent, and Applied Optoelectronics all described demand as exceeding supply.
Scarcity and execution risk are arriving together. Applied Optoelectronics reported an order book above $200 million, but its target of roughly $471 million in monthly transceiver revenue by mid-2027 requires capacity to rise from approximately 200,000 units per month now to more than 650,000 by the end of 2026 and more than 930,000 by the end of 2027.
Demand may be ahead of supply. Capturing it still requires a major and successful capacity expansion.
Broadcom presents the same equation at a larger scale. Wolfe Research estimated that its planned OpenAI and Anthropic capacity could reach 14GW in 2028 and generate roughly $140 billion to $200 billion of revenue, compared with approximately $245 billion in total 2028 consensus revenue for Broadcom.
Attached to that opportunity is $30 billion of residual-value guarantees if the industry overbuilds.
Contracting Redistributes AI’s Risk
Contracts, guarantees, and debt can accelerate construction and improve visibility. They can also move demand and residual-value risk among buyers, lenders, suppliers, guarantors, and investors elsewhere in the financial system.
One skeptical financing case compared the structure to subprime lending. Its concern was that unprofitable AI borrowers could use securitized debt to buy depreciating GPUs, with those loans eventually held across pension funds, insurers, banks, and other investors.
That is an argument about financial architecture, not proof that near-term AI demand is weak. The evidence from contracts, backlogs, revenue growth, and optical shortages is substantial.
Calling AI compute an asset class does not settle its economics. The unresolved issue is whether durable end-market cash flow arrives before financing costs, falling prices, depreciation, and new capacity absorb the returns.