HBM is the memory storyline in AI compute that the market believes in most. After demand, pricing, yield and capacity have been debated round after round, the argument now is whether memory ultimately reverts to an ordinary inventory cycle, or whether customers de-spec while demand is merely deferred, and three-to-five-year LTAs with locked volumes plus 30% prepayments break the traditional cycle valuation framework. The bull case rests on a static capacity gap. The layer that goes unmentioned works out how excess profit, in turn, induces capital expenditure, process migration and new entrants — that is, the second-order feedback from supply.
The conclusion first: contractualization may genuinely change the bottom of the cycle; but precisely because profits at the top are so high, it may also be manufacturing the next wave of supply. The market already believes in AI memory demand, the HBM4 ramp and long-term agreements. What the share prices trade at this stage is how much longer high profitability can last. A genuine upside surprise is more likely to come from one of two situations: first, product revenue keeps being revised up while free cash flow, after stripping out customer prepayments, joint-venture contributions and investment in new capacity, remains strong, which would say the profit base is structurally rising; second, new capacity turns into qualified supply more slowly than real demand converts, so the supply constraint is more durable than the market expects. Conversely, if shipments and company profits are still growing but adjusted cash conversion deteriorates first, or if new capacity clears qualification faster and enters saleable output, today's high profitability looks closer to a cycle peak. Put back into traditional sell-side modelling language, the valuation-sensitive term should move away from a single point on revenue growth and gross margin, toward "cleanly attributed product revenue × adjusted cash conversion × competitive advantage period." When all three move in the same direction, an earnings upgrade has quality behind it; if only revenue and gross margin are revised up while the cash side starts to lag, it becomes easier to see earnings expectations peak and the multiple compress first. So the more appropriate action is to wait for the next one or two reporting periods to verify the revenue-to-cash-capture gap, with the higher-frequency observable being the pace at which the second-order supply feedback actually shows up.
We have just completed the first version of the Memory data theme for the LunarTulip Research Desk. Working through this set of data and ruling items out one by one, I think the direction most worth verifying right now — and most likely to produce research alpha — has emerged: the changing gap between the HBM revenue and profit narrative and true cash capture.
Among the five feature proposals formed in this version, the other four serve first as constraints on attribution, tenor and supply status; the "revenue-to-cash-capture gap" is the only one promoted to top priority and moved into continuous forward-looking validation as an alpha hypothesis.
The core judgment: the next leg of alpha is not in demand, it is in the cash-capture gap
The market has spent several quarters discussing the logic that AI systems need more memory bandwidth and that HBM4 is shipping at scale. Micron has disclosed that the production ramp of HBM4 12-high is running at roughly twice the pace of HBM3E 12-high, with cumulative shipment revenue above US$1 billion; SK hynix has begun volume shipments of HBM4; Samsung has also disclosed expanding HBM4 sales and has sent HBM4E samples to major customers.
These facts confirm the direction of the industry, but on their own they are unlikely to create further expectation gaps. What has not yet been adequately answered is this:
For each incremental unit of HBM revenue, after deducting customer funding, joint-venture contributions and the capital required for incremental effective capacity, how much repeatable cash is the listed company ultimately left with?
We call this yet-to-be-verified relationship the "revenue-to-cash-capture gap." It can run in two directions.
If product revenue keeps growing, standalone profitability becomes progressively verifiable, cash conversion after stripping out customer funding and capacity-expansion investment remains stable, and the arrival of effective supply is still constrained by qualification and yield, then the market may be underestimating the duration of high profitability.
If HBM shipments and company profits are still rising while adjusted cash capture starts to weaken, with customer prepayments, capital expenditure or joint-venture contributions carrying more and more of the surface prosperity, and qualified supply arrives faster at the same time, then the market may be overestimating the bottom of this cycle.
This is closer to a tradable disagreement than continuing to argue about the rate of demand growth, because what price ultimately discounts is whether growth can be delivered at a sufficiently high return on capital on a sustained basis.
Why can the other directions, used as factors, not support more core alpha for now?
First, shipments and yield have entered consensus territory. HBM4's path from sampling and qualification to volume delivery is still worth tracking, but a commercialization fact that major suppliers have confirmed repeatedly is better used as a research starting point than as a standalone reason to buy.
Second, company profit cannot be lined up directly into an HBM beneficiary ranking. Micron's CMBU covers both cloud memory and data-center HBM; Samsung's DS division also includes memory, foundry and System LSI; SK hynix's group profit likewise includes other memory products. When the bridge between product, division and company profit is missing, the more precise the ranking, the more misleading it can be.
Third, contract value is not revenue, still less cash profit. RPO, minimum purchase commitments, price floors and caps, customer deposits and revenue recognition are different objects. Adding them together writes contract visibility up as profit that has already been delivered.
Fourth, capacity-expansion announcements are still several gates away from effective supply. A fab investment still has to pass through equipment spend, production start, customer qualification, yield ramp and actual output. The announced amount can change the forward supply scenario, but it cannot add saleable bits on the day of the announcement.
Fifth, a low trough-to-peak valuation carries no direction by itself. Without answering margin, cash conversion and competitive advantage period at the same time, a low P/E can represent either undervaluation or simply the market's refusal to grant a longer duration to peak-cycle earnings.
Once these variables, which cannot yet carry predictive power on their own, are set aside, the "revenue-to-cash-capture gap" becomes the causal path closest to security returns within the data we have.
A control sample: the same company can show both 56.77% and 43.18%
Sandisk does not make HBM, but it provides a very good control sample for the memory cycle: within one and the same set of financial statements, different cash-flow definitions can lead to completely different conclusions about the durability of earnings.
The company's FY2026 operating cash flow was US$11.671 billion and purchases of property, plant and equipment were US$177 million, giving the free cash flow of US$11.494 billion presented by the company. Divided directly by revenue of US$20.248 billion for the same period, the cash flow rate is about 56.77%.
But the company also disclosed US$275 million of net contributions to joint ventures, along with US$2.476 billion of cash impact related to the new business model. After deducting these two items, adjusted free cash flow is US$8.743 billion, or about 43.18% of revenue for the same period.
Both 56.77% and 43.18% can be recomputed, yet they answer different questions. The former describes total free cash flow within the reporting period; the latter is closer to operating delivery after stripping out specific customer funding and joint-venture contributions. Treating the former directly as a steady-state cash conversion rate through the long cycle causes a reverse-engineered valuation to overstate cash capture; treating the latter mechanically as a perpetual margin is equally a case of substituting a single cycle for a long-term judgment.
The implication for HBM is direct: a record high in company or divisional profit does not automatically mean the HBM product line generated a proportionate amount of repeatable cash. The genuinely valuable signal is the direction in which the two definitions move across consecutive reporting periods, and the reason the gap widens or narrows.
Figure 1: Sandisk is a NAND cycle control sample, not an HBM producer. What this cash bridge shows is the research method, not a proxy estimate of HBM product profitability.
How does the Memory data turn this hypothesis into a verifiable signal?
The Research Desk does not directly generate an "HBM alpha score." It follows the real causal order and breaks cash capture into four layers.
Layer one: effective supply. Investment plan, equipment, production start, qualification and saleable output advance in sequence. On 27 August, Kioxia and Sandisk announced investment of more than US$31 billion in Japan through 2032; this is a single joint plan with conditions attached, and until the split between the companies and the actual bit capacity are disclosed, it can only enter the forward supply clock.
Layer two: product commercialization. Samples, customer qualification, volume shipments and recognized product revenue are recorded separately. Micron's more than US$1 billion of HBM4 revenue can advance the commercialization status, but standalone HBM4 profitability remains missing.
Layer three: financial attribution. Product, application, division and company each keep their own boundaries. Samsung's HBM4 sales, DS divisional profit and the contribution of HBM base-die to foundry profitability cannot be counted three times over as three separate HBM profit streams.
Layer four: the cash bridge. Operating cash flow, capital expenditure, customer funding, joint-venture contributions and working capital movements are each kept separate. What the model compares is "what is left after removing non-repeatable sources," rather than picking the best-looking cash figure out of the financial statements.
Figure 2: A plan is not output, a shipment is not attributed revenue, and profit is not repeatable cash either. Only when all four layers of data are crossed at once do we move to a security pricing judgment.
The four layers of data use the same set of companies, periods and points of knowability. An old financial report organized into the database today is not disguised as a new event today; when a company has not disclosed product-level profitability, "unknown" is not encoded as negative either.
On the machine side, this research hypothesis can be broken into continuous observations: whether the supply plan has genuinely crossed the next gate, whether product capture status has been upgraded, whether attributed revenue has expanded, and whether adjusted cash conversion has improved. Once enough real time cross-sections have accumulated, we can then test the relationship between these changes and earnings revisions and excess returns.
Why does this alpha hypothesis deserve priority as a factor suited to fundamental quant tracking?
It meets three conditions.
First, market attention is high while data friction remains large. Information on HBM demand is extremely crowded, yet product profitability and cash attribution are scattered across financial reports, earnings calls, contracts and joint-venture disclosures. Ordinary financial databases are good at standardizing company financial statements; they are not good at explaining how much company cash a single hot product actually captures.
Second, it produces repeated observations. Every reporting round updates revenue, cash flow, contract performance, capital expenditure and product status. Compared with a one-off industry story, this path can accumulate into a genuine point-in-time series.
Third, it allows both positive and negative outcomes. The data does not presume the HBM bulls are right. If cash capture proves more durable than the market imagines, it supports a higher cycle bottom; if cash weakens first, it can expose the risk while the demand narrative is still strong.
So, unlike a narrative version such as "long-term growth in AI memory," a return hypothesis whose success and failure conditions can be written down in advance is better suited to being written into a factor framework.
Over the next one or two reporting periods, we are watching only five things
- Whether Micron, SK hynix and Samsung provide a more verifiable bridge for HBM product revenue, cost or profit;
- Whether, as product revenue grows, cash conversion after stripping out customer funding and incremental investment moves in step;
- Whether RPO and long-term agreements are delivered within the same contract cohort, or whether we see deferrals, amendments or deterioration in cash quality;
- Whether profit improvement at Kioxia and Sandisk is accompanied by healthy movements in receivables, joint-venture cash and capital investment;
- Whether new fabs and expansion plans move from announcement into equipment, production start, qualification and effective output.
If revenue rises while adjusted cash capture weakens for consecutive periods, we will lower the structural re-rating assumption; if the slope of effective supply runs ahead of real deployment and contract delivery, we will shorten the competitive advantage period. Conversely, if product profitability becomes progressively verifiable, contracts are delivered steadily, cash conversion holds up under adjustment, and incremental supply remains constrained by technology and capital intensity, then the "duration of high profitability" gains new evidence.
Why does the Research Desk need both a Research Note and an API?
Human investors need to read a conclusion: the highest-priority HBM alpha hypothesis right now is whether revenue growth can be converted into repeatable cash.
Machines, meanwhile, need to know when each piece of evidence became knowable, which product and company it belongs to, which financial definition it uses, which status gate it has crossed, and what conditions would falsify it.
The Research Note and the API are both research output from the Research Desk: the former as a report an investment manager can read, the latter as factor data readable by machines and systems. First the raw alternative data required across different kinds of research is compressed into a judgment; then the same judgment is broken back down into recomputable candidate research features — claim, metric, event, entity, state and timestamp — and, with direction, weights and the lag in market reaction plugged in, put through systematic backtesting to validate its forward-looking power.
A closing note
The memory cycle is facing the core pricing contradiction of its current position in the cycle: the divergence between prosperity at the level of the financial statements and true free cash flow.
Our current judgment is this: at the moment when HBM consensus is strongest and earnings are most easily extrapolated in a straight line, the revenue-to-cash-capture gap is the most likely place for expectation errors to surface first. It does not inherently point to the bull case, nor does it presume a reversal; its value is that it tells us earlier than the profit headline whether this round of high earnings is genuinely turning into a higher long-term base, or is being borrowed against customer funding and a fresh round of capital expenditure.
This will be the first alpha direction in the Research Desk Memory theme to enter forward-looking validation. From here, the data is responsible for recording changes in status, the Research Note is responsible for explaining how the judgment is updated, and in the end a real time series decides whether this hypothesis survives.
Information as of 10 September 2026. This article presents a preview of the research methodology behind the LunarTulip Research Desk Memory theme. The data features are still in forward-looking validation, and external release of the API still requires completion of independent content review and data rights gating. This article does not constitute any securities recommendation or investment advice.