The Memory Market Rewired: Bank Trillions, $94 Billion in Contracts, and the End of the Spot-Price Era

The Money Changed First: Wall Street Underwrites the Buildout

The most important change in the AI trade this year is not a chip or a model. It is who pays for the infrastructure. For the first three years of the AI buildout, hyperscalers funded data centers out of their own operating profits. That model put a hard ceiling on the construction, because every capital expenditure cycle in history has ended the same way: the builder’s balance sheet said stop. Investors priced every AI supplier against that ceiling, and memory suppliers, the most cyclical names in the stack, were priced against it hardest.

That ceiling is being removed. Goldman Sachs, Morgan Stanley and Bank of America each concluded that the buildout cannot be stopped and built financing frameworks to fund it. Morgan Stanley launched a $1.5 trillion ten-year infrastructure initiative, and its own research identified a data center financing gap of roughly $1.5 trillion that outside capital has to fill. Goldman Sachs joined a $500 billion AI infrastructure collaboration alongside NVIDIA, Blackstone, BlackRock and other firms. Bank of America committed $250 billion to a critical infrastructure initiative running from January 2026 through July 2027, targeting power, water, transportation and critical minerals as well as data centers, a program detailed in American Banker’s report Inside Bank of America’s $250 billion in data center funding. Global AI debt issuance is running near $570 billion in 2026. Morgan Stanley CEO Ted Pick says 2026 data center capex is tracking near $850 billion against an earlier $575 billion forecast, and the firm has said total AI capex could eventually reach $10 trillion over many years.

Why this matters is simple. The financial system is now structurally inside the trade. Financing frameworks measured in trillions remove the fear that the sheer size of the investments forces the construction to stop. The build no longer depends on any single quarter of hyperscaler earnings, because the capital feeding it comes from dedicated credit structures rather than from operating cash flow alone. When a bank underwrites a multi-year infrastructure program, the demand that flows from that program is committed, and committed demand behaves differently from forecast demand at every point in the supply chain.

The gap being financed is also specific. Software and AI demand move at software speed. Hardware runs into fab construction timelines, energy shortages, and multi-year equipment lead times. Capital is being thrown at exactly that gap, and it lands hardest on the physical inputs that cannot be conjured quickly: fabs, power, and memory. The power constraint is real enough that utilities have become an AI infrastructure trade of their own, a dynamic CleaRank covers in its Utilities Select Sector (XLU) ETF analysis.

“Every prior capex cycle died the same way: the builder ran out of internal cash, the orders stopped, and the suppliers absorbed the correction. The market still prices memory as if that rule applies. Once the banks underwrite the build, the demand signal reaching suppliers stops being a forecast and becomes an obligation, and obligations do not disappear in a soft quarter. That is the single most important repricing input in the memory complex right now.”

Jacob Bakshi, CleaRank Senior Derivatives Strategist

Sandisk Investor Day: $94 Billion Says the Spot Market Is Over

On August 13, Sandisk showed what underwritten demand looks like when it reaches a memory supplier: eight customers, approximately $94 billion in contracted value, and a stock that closed up 13.67%. The event converted the memory structural thesis from an argument into a disclosure.

The core announcement was the New Business Model. Sandisk revealed agreements with eight customers worth approximately $94 billion in total contract value, with roughly $91 billion in remaining performance obligations. The typical contract runs about four years. The agreements are built on committed volumes, enforceable contractual frameworks with minimum financial guarantees, and structured pricing mechanisms rather than spot negotiation. Together the commitments secure roughly half of fiscal 2027 production and about two-thirds of fiscal 2028 bits. The full framework is laid out in the company’s release, Sandisk details growth strategy and long-term financial model at 2026 Investor Day.

The long-term financial model through fiscal 2030 is the part that repriced the stock. Management targets mid-to-high double-digit revenue growth, a non-GAAP gross margin near 80%, an operating margin near 75%, an adjusted free cash flow margin near 50%, and capital expenditure around 5% of revenue. Capital allocation invests in the business first, keeps a healthy balance sheet, and then returns 100% of excess free cash flow to shareholders through buybacks or dividends. On the technology side, the upcoming BiCS10 QLC node delivers roughly 60% higher bit density than BiCS8, which is what lets contracted bit growth arrive without a matching surge in capital intensity.

The reported numbers behind the model are already extreme. Full fiscal year 2026 revenue reached $20.25 billion, up 175% year over year. GAAP net income was $11.43 billion. Diluted EPS came in at $73.76, nearly 24 times the prior year. Data center revenue rose 437% for the year. Guidance for fiscal Q1 2027 of $10.30 to $10.80 billion implies roughly 4.5x year-over-year revenue growth at the midpoint, which means the single next quarter is guided to earn more than half of what the entire prior fiscal year produced.

CleaRank flagged the August 13 Investor Day as the decisive catalyst in its Sandisk (SNDK) stock analysis of the memory wall and the HBF standard, where the open question was whether management could convert a technology story into a revenue model. The $94 billion answer exceeded the frame of the question. QLC and High Bandwidth Flash remain the technology pillars underneath the contracts, but the commercial architecture is now the story.

“Before Investor Day, the debate was whether an 84.6% gross margin was a shortage artifact that reverts on the next supply wave. A four-year contract book with minimum financial guarantees converts that question into a different one entirely. The debate is no longer whether margins revert. It is what the contracted floor actually is, and almost nobody on the sell side has a model built for that question yet.”

Jacob Bakshi, CleaRank Senior Derivatives Strategist

Micron: The Quiet Confirmation

Micron confirmed the same structural shift months earlier with far less drama: its HBM output is sold out at fixed prices, and customers are signing agreements that run three to five years. The entire 2026 HBM production is contracted, management says supply stays tight well into 2027, and capacity is effectively fixed until new fabs contribute. Fixed-price, multi-year commitments are a structural break from the quarterly negotiation pattern that historically defined the memory industry.

The pattern read matters more than any single data point. When the number one, number two and number three suppliers in a market all report multi-year committed volumes at the same time, that is not a company story. It is a market structure story. Sandisk’s $94 billion contract book, Micron’s sold-out fixed-price HBM, and the SK hynix commitments covered below all describe the same change from a different seat: buyers are locking supply years forward because they cannot afford to be caught short, and suppliers are accepting term structure because it converts cyclical revenue into contracted revenue.

SK hynix: Rumors, a 15% Drop, and JPMorgan’s Answer

SK hynix lost roughly 15% in a week on rumors that NVIDIA was cutting HBM content and that HBM4 would price at half the expected level. JPMorgan examined the claims and called the pricing report inaccurate. The episode is worth walking through carefully, because it shows how violently sentiment can detach from contracted reality in memory names.

The rumor set was layered. Reports circulated of NVIDIA HBM purchase cuts, a 50% HBM4 price discount, uncertainty over the timing and size of shareholder returns, and a potential subsidiary IPO, all stacked on top of a large infrastructure capex plan that some investors read as margin-dilutive. Each item alone was survivable. Together they compounded into a 15% drawdown in five sessions.

JPMorgan’s response was unusually direct. The bank explicitly stated that the 50% discount report is inaccurate, said the worst is behind the stock, sees no fundamental indicators of weakness over the next six to twelve months, and pointed at Q2 2026 revenue of 79.32 trillion won. The capex plan that spooked the market was already scheduled rather than new. The company’s own actions support that read: SK hynix approved roughly $38 billion in new memory fabs and moved its shareholder return policy announcement forward to before the end of Q3, earlier than the original year-end timeline. Companies shrink disclosure timelines when they are confident, not when they are hiding something.

The balance note belongs here rather than in a footnote. The volatility is a reminder that sentiment can detach from fundamentals in memory names in both directions, up as violently as down, and position sizing should respect that even when the structural case is intact. For ongoing price data and coverage, see the SK hynix (SKHY) stock page on CleaRank.

Intel Wants Back In

The strongest external validation of the memory thesis came from a company that exited the business: Intel is signaling a return. CEO Lip-Bu Tan said in August 2026 that he used to tell people not to invest in memory because it is a commodity business, but that it has now become different. He cited new memory architectures, said CPU and memory have many ways to stack together, and noted Intel is not ready to unfold the projects yet. Reports point to XBM and ZAM technologies and an emerging patent trail, and Intel hired Seok-Hee Lee, the former CEO of SK hynix, as detailed in TrendForce’s report on Intel reportedly hinting at memory re-entry and CPU-memory stacking.

The logic requires no inside information. Nobody re-enters a market they believe is a dying commodity. Re-entry signals are what a market looks like when informed insiders believe the economics have structurally improved, and Intel’s hiring pattern is the costliest form of signaling available. It also confirms the architectural point underneath the whole thesis: memory is becoming critical to the next generation of compute design, stacked against and integrated with the processor, rather than a passive component bought by the bit on a spot market.

“The market is flooded with opinions about whether memory is still a cyclical commodity. The actual participants are answering with actions. Eight customers signed roughly $94 billion with Sandisk. Micron is sold out years forward at fixed prices. SK hynix is committing $38 billion to new fabs. The banks are underwriting the demand side, and Intel is trying to get back into the business it exited. Watch what they do, not what the debate says.”

Jacob Bakshi, CleaRank Senior Derivatives Strategist

The Efficiency Objection, Answered Again

The standing objection is that AI will become efficient enough to need less infrastructure. The observed behavior of every large buyer points the other way. As AI systems get cheaper and more efficient to operate, more organizations and more users can afford to run them. Wider adoption raises total workload, and total workload is what drives the requirement for compute and memory. Efficiency enlarges the market rather than shrinking it, because the cost decline unlocks demand that was previously priced out.

The observable evidence sits in the capital accounts. Cloud providers are using efficiency gains to expand capability, not to cut infrastructure budgets. Industry data center capex is rising toward $850 billion in 2026, not falling, and it is rising alongside the efficiency improvements rather than in spite of them. The financing structures described in the first section exist precisely because the banks studied the same question and concluded the build continues for years. A bank does not construct a $1.5 trillion ten-year initiative around demand it expects efficiency to erase.

The honest limit is that this argument describes direction, not timing. A supply-driven price correction in any given quarter remains entirely possible, and 80% margin targets will be tested by capacity additions eventually, because that is what capacity additions do. Both the structural demand story and cyclical volatility can be true at once. The difference now is that contracted volume dampens the amplitude: when half to two-thirds of production is pre-sold under minimum guarantees, a downturn compresses results instead of collapsing them.

Financial Snapshot

The table below gathers the reported figures, contract disclosures, and financing commitments described in this article. Read the contract and production-lock lines together, because they are what separate this cycle from every previous memory cycle on record.

What Could Break the Thesis

Contract quality is the first risk. $94 billion in contract value is only as strong as the counterparties behind it and the enforcement mechanisms inside it. Take-or-pay structures in semiconductors have been renegotiated before during downturns, quietly and repeatedly, because a supplier suing its largest customers is rarely a viable commercial strategy. Minimum financial guarantees reduce cyclical exposure. They do not eliminate it, and the market will not know how firm the floor is until it gets tested.

Supply response is the second risk. An 80% gross margin target is a signal flare to every fab planner on earth. SK hynix’s $38 billion program, Samsung’s existing capacity and appetite, a potential Intel re-entry, and China’s memory ambitions all add supply eventually. The question is not whether new capacity arrives but whether contracted demand grows faster than committed supply through the window when it does. Memory history says supply usually wins that race in the end. The contract structure is the industry’s first serious attempt to change the terms of the race.

Financing is the third risk, and it cuts both ways. The same bank money that removes the funding ceiling introduces leverage into the system. If AI monetization disappoints, debt-financed data centers become the transmission mechanism for a much broader correction, one that runs through credit markets rather than just through equity multiples. The financing structures described in this article are new and untested through a downturn. That does not make them fragile, but it does mean nobody can point to a cycle they have survived.

When the Buyers Sign for Four Years, the Argument Is Over

The memory cyclicality debate has raged all year. The participants have now voted. Eight customers signed roughly $94 billion with Sandisk. Micron’s customers locked three-to-five-year agreements at fixed prices. The banks built trillion-scale financing structures so the buyers can keep building. These are not opinions, models, or conference takes. They are contracts, and contracts are the one form of market opinion that carries a penalty clause.

None of this abolishes the cycle. Memory remains a capital-intensive industry where supply eventually answers price, and the risks in the previous section are real. What changed is the amplitude. When half to two-thirds of production is pre-sold under minimum guarantees, the downside case moves from collapse to compression, and a business whose worst case is compression deserves a different multiple from one whose worst case is collapse.

The noise will continue. SK hynix losing 15% in a week on rumors that JPMorgan then called inaccurate is what this market does, and it will do it again. The gap between headline sentiment and contracted reality is exactly where the mispricing lives, and it is the gap disciplined investors should be watching while everyone else trades the rumor cycle.

The larger picture is hard to unsee once it is assembled. The financial system is inside the trade now, with trillion-scale frameworks rather than quarterly enthusiasm. The suppliers are contracted years forward at guaranteed minimums. And the last company to famously call memory a commodity is hiring memory executives and filing memory patents. When the buyers, the banks, and the former skeptics all move in the same direction, the burden of proof shifts to the bears.

Frequently Asked Questions

Disclaimer: This analysis of the memory market, including Sandisk Corporation (SNDK), Micron Technology (MU), and SK hynix (SKHY), is for informational purposes only and does not constitute financial, investment, or legal advice. Memory is a historically cyclical industry, and record margins invite a competitive supply response over time. Contracted volumes and minimum guarantees reduce but do not eliminate downside exposure, and long-term supply agreements in semiconductors have been renegotiated in prior downturns. The bank financing structures described in this article are new and untested through a full credit cycle, and debt-financed infrastructure introduces systemic leverage that could amplify a correction if AI monetization disappoints. Rumors and sentiment can move memory stocks violently in either direction, as the SK hynix episode described in this article demonstrates. Figures for SK hynix are reported in Korean won, and currency movements affect comparisons. CleaRank and its analysts may hold positions in securities mentioned in this article. Past performance is not indicative of future results. Always consult with a licensed financial advisor before making investment decisions.