Is the baton really changing hands?

Global AI hardware did not downshift in early September.

DDR5 spot quotes re-accelerated through August and server DRAM contract prices kept rising, though the rate of increase narrowed. SK hynix and Micron pushed latest-quarter operating margins to 76% and 80% while their shares sat roughly 43% and 26% below 52-week highs on the 2026-09-01 close. In optics, Innolight grew first-half revenue 182.5% and attributable net profit 241.7%, Lumentum grew revenue 109.3% in its latest quarter, and the midpoint of its next-quarter guidance implies about 24% sequential growth.

Put those numbers on one page and the market is not trading a rotation. It is trading a fork inside a single AI capex cycle: the memory constraint still sits with the upstream manufacturers, optics keeps converting profit, and the Chinese module makers hold the steepest exposure of all — and the one most dependent on inventory timing.

Our conclusion has three layers:

1. Structural priority in memory belongs to the constraint owners: SK hynix, Micron and Samsung. Capacity, process and customer qualification for HBM, server DRAM and enterprise NAND are all concentrated upstream, so tight supply reaches their ASPs, gross margin and cash flow first.
2. Kioxia, SanDisk and the Chinese module makers offer more cycle elasticity, and the quality of that elasticity is not the same. The first two are directly exposed to NAND pricing and data-centre SSD. Demingli and Longsys earn much of their elasticity on the spread between low-cost inventory and realised selling prices, so cash flow and inventory determine how long the profit stays.
3. Optics has not handed anything over. All six core names across China and the U.S. reported strong revenue or guidance; the dispersion sits in valuation, cash conversion and capacity execution. Innolight's post-print drawdown shows the margin for error has thinned. On its own it does not establish that the fundamental leg is over.

That leaves one variable worth watching from September through November: whether profit keeps landing with the companies that carry cleaner balance sheets and tighter customer lock-in.

1. Memory is not one trade. Break it into three.

1.1 Spot is re-accelerating; contract prices are decelerating

On September 1 the TrendForce DDR5 16Gb spot average was $54.083, up 0.31% on the day. The DDR5 32GB RDIMM module average, updated August 24, rose 4.06% week on week, and a week earlier Chinese-market quotes for DDR5 16Gb and 24Gb both rose 14.29% week on week. One qualification belongs here immediately: these are quoted-price series. Transaction volume and demand confirmation through this leg have stayed weak, so spot should be read as quote strength rather than verified demand. Whether volume follows is the only thing in this series worth waiting on over the next few weeks. TrendForce spot and module quotes; TrendForce channel-price update, August 17

Monthly fixed prices describe a different slope. Korea's Ministry of Trade, Industry and Energy reports the DDR5 16Gb fixed price rising from $35.0 in April to $46.5 in August, with a 12.5% monthly increase in July and only 3.3% in August. The monthly increase on NAND 128Gb fixed prices fell from 9.5% in May to 1.3% in August. The same release shows Korean semiconductor exports of $46.65 billion in August, up 209% year over year — aggregate conditions remain extremely strong. Korea August 2026 export data

The divergence matters. Spot reflects marginal tightness and channel restocking at the quote level; contract prices sit closer to the ASP that manufacturers actually settle with large customers. Accelerating spot lifts trading elasticity over a few weeks to a quarter. If contract increases keep narrowing, year-over-year profit can stay very high while sequential acceleration peaks first. Collapsing both lines into a single "memory prices are up" indicator simultaneously overstates how long module profit lasts and understates the pricing power sitting upstream.

The demand anchor is intact, provided the basis is read correctly. NVIDIA reported FY27Q2 revenue of $96.2 billion, including $89.0 billion of data-centre revenue. More consequentially, disclosed future supply and capacity commitments rose from $119.0 billion in the prior quarter to $279.0 billion, with the company noting these relate principally to memory procurement. This is a commitment figure, not paid-for procurement, and not an absolute lock on demand. It spans multi-year arrangements, and the terms leave parts of it adjustable or deferrable. It establishes the scale and direction of purchasing intent; it does not convert into contracted revenue for any single manufacturer over the next four quarters. NVIDIA FY27Q2 CFO Commentary (SEC)

1.2 The structural pool: SK hynix, Micron, Samsung

The core nodes of global memory all reported. A domestic-only lens simply never put them at the centre of the table.

The multiples above answer one question only: how much peak profit is already in the price. Micron's 9.5x is market capitalisation divided by four times latest-quarter GAAP net income — a peak-earnings screen. The SK hynix TTM denominator carries large investment gains and is the more distorted of the two. Samsung remains a multi-business group and is not strictly comparable with two pure-play memory manufacturers. None of these can be read as a forward P/E or a valuation conclusion. SK hynix Q2; Micron FQ3; Samsung Electronics Q2

Switch Micron to a forward basis and the picture changes materially. Sell-side consensus captured from Seeking Alpha on 2026-09-01 shows FY2027 consensus EPS near $155, which corresponds to roughly 6.2x forward earnings. Three qualifications belong with it. This is a third-party consensus snapshot, not company guidance. Consensus moves with every print and every revision. And different data vendors treat fiscal-year alignment and GAAP versus non-GAAP adjustments differently, so the same figure will disagree across sources. Its use is to show the distance between current pricing and the market's own earnings expectation, not to supply a valuation that can be copied across.

Of the three, we weight SK hynix and Micron more heavily for directness of supply and share position. SK hynix has begun volume shipments of HBM4, plans to lift 321-layer NAND to roughly half of its Korean domestic capacity by year-end, and holds net cash of KRW 69.4tn. Micron's Cloud Memory and Core Data Center units posted quarterly operating margins of 78% and 83%, and FQ4 guidance calls for $50.0 billion of revenue at roughly 86% gross margin. Samsung's advantages are valuation cushion and an HBM catch-up option; the cost is that handsets, displays and foundry dilute the memory upcycle.

The market has already issued a warning. SK hynix grew operating profit 557% year over year and still sold off hard on results day, because revenue and operating profit came in below elevated expectations and because of the HBM4 shipment cadence. Of its KRW 93.9tn net income, roughly KRW 63.3tn came from gains on investment assets, which is why we prefer operating profit when assessing the operating business. Reuters on SK hynix results and the market reaction

Put differently: single-digit peak-earnings screens and low-teens screen P/Es already embed a substantial normalisation of profit. The more interesting bull question is how high the floor sits after normalisation. The scenario the market does not want is the one where long-term agreements raise demand visibility while capping the incremental ASP elasticity that manufacturers get from rising spot, and 2027 capacity additions then pull today's excess margins down. One supply-side thread remains unresolved: Reuters, citing people familiar with the matter, has reported that Chinese memory maker CXMT has begun small-volume production of HBM3E. The report is not company-confirmed, and we draw no pricing, share or supply-demand inference from it here, recording it instead as an unverified 2027 supply-side risk pending first-party disclosure.

1.3 NAND elasticity: Kioxia and SanDisk

NAND prices are more elastic than DRAM, and enterprise SSD adds a layer of AI data-centre demand on top. Kioxia reported FY26Q1 revenue of JPY 1.767tn, up 76.2% sequentially, with non-GAAP operating profit of JPY 1.326tn at a 75% margin. SanDisk reported FQ4 revenue of $8.97 billion, up 51% sequentially with roughly two-thirds of the increase from price, data-centre revenue up 103% sequentially, and GAAP operating profit of $7.04 billion. Kioxia IR; SanDisk FQ4 (SEC)

As of the 2026-09-01 close, Kioxia and SanDisk sat roughly 55% and 35% below 52-week highs. Screening market capitalisation against four times latest-quarter net income puts Kioxia near 7.9x on a non-GAAP basis and SanDisk near 8.7x on GAAP — two different accounting bases, and not directly comparable with each other. That cheapness rests on extreme prices and margins. Every NAND ASP inflection works in both directions on profit. SanDisk's post-print decline makes the same point: the growth is in the price, and the market now needs continued upside rather than maintenance of a high level.

Kioxia and SanDisk are therefore best read as NAND cycle elasticity, while SK hynix and Micron represent ownership of the AI memory constraint. Both groups can benefit at once; the research basis and the falsification conditions are entirely different.

1.4 Chinese modules: the cheapest optics, the heaviest inventory

Demingli and Longsys have not lost their research value. They simply can no longer stand in for the entire global memory chain.

Both companies wrote the inventory spread straight into the income statement, and wrote the cost of building that inventory into the cash-flow statement. A 7–8x static multiple looks attractive at first glance, but the denominator already contains peak profit formed jointly by cheap inventory, sharply rising ASPs and light impairment. If die prices keep rising, the new purchase cost enters inventory. If shipment volumes or end-market tolerance weaken, margin and cash flow can turn before revenue does. Demingli periodic reports (CNINFO); Longsys periodic reports (CNINFO)

The rising second derivative in spot quotes matters practically for these two: inventory revaluation runs longer, near-term selling prices hold better, and pressure to cut Q3 gross-margin estimates eases. What it has not done is verify operating quality. That requires enterprise-grade product share, controller capability, sustained expansion and cash recovery.

The passage that follows is an author-defined scenario back-test. It is not a forecast, not a price conclusion, and carries no trading implication. It answers one question: how much would the earnings structure have to change for today's low static multiple to be sustained? Starting from the 2026-09-01 close snapshot, Demingli's market capitalisation is about RMB 94.8bn. Assume market capitalisation rose to ten times that snapshot, roughly RMB 948bn. At a neutral 20x assumption that requires about RMB 47.4bn of sustainable annual profit, roughly four times the annualised H1 2026 attributable net profit of about RMB 12.0bn; even at 25x it requires about RMB 37.9bn. For Longsys, at about RMB 150bn today, the same scenario implies about RMB 75bn of annual profit at 20x and about RMB 60bn at 25x. Profit of that magnitude requires enterprise product share, controller capability, customer stickiness and cash recovery to expand together. A spot-quote impulse supplies only part of the distance.

Research priority on these two therefore depends on four signals appearing together: Q3 gross margin stops declining; operating cash outflow narrows materially; inventory-to-assets and inventory turnover stop deteriorating; and enterprise product revenue grows faster than consumer inventory gains. If any one is missing, the apparent low P/E may simply be an inflated cyclical denominator.

2. The optics story continues, and six companies are separating

Defining the whole optics cycle by one sharp post-print decline at Innolight discards the concurrent information from five other companies.

On basis: the Chinese names use attributable net profit, the U.S. annualised multiples use non-GAAP net income, and non-GAAP operating margin is a third basis that cannot substitute for either. The multiples exist to observe growth, cash and what the market is demanding, not to serve as strictly comparable cross-market valuation. Innolight H1 report (CNINFO PDF); Eoptolink periodic reports (CNINFO); T&S Communications periodic reports (CNINFO); Lumentum FQ4; Coherent FQ4; AAOI Q2

Three conclusions come out of that table.

First, the demand slope is intact. Lumentum guides to a next-quarter revenue midpoint near $1.25 billion, Coherent to about $2.3 billion and AAOI to about $273 million, implying roughly 24%, 12% and 42% sequential growth respectively. 1.6T, lasers, OCS and data-centre optical interconnect are still converting into revenue.

Second, the Chinese and U.S. names are not earning the same dollar. Innolight and Eoptolink sit closer to finished high-speed modules and delivery. Lumentum is more vertical across lasers, components, modules and systems. Coherent spans materials, devices, modules and an industrial business. AAOI is still in the middle of a capacity build and a swing to profitability. Compressing six companies into one P/E ranking erases product mix, accounting basis and capital intensity all at once.

Third, cash conversion has become the valuation dividing line. Innolight and Eoptolink converted only 0.13 and 0.21 of first-half net profit into operating cash flow, against 0.83 at T&S Communications. Coherent generated $79.5 million of full-year operating cash flow against $1.103 billion of capex, with inventory rising from $1.438 billion to $2.581 billion. Rapid growth does require prepayments, inventory and equipment. But once a stock already demands a run-rate multiple of 30x to 100x and above, delayed cash recovery amplifies valuation volatility quickly.

The market reaction has not been uniform either. Innolight fell 7.72% on the first session after its interim report. Coherent delivered strong revenue and guidance and still drifted lower. Lumentum and AAOI rose after reporting. Together they describe a phase in which the reported numbers remain strong while the expectations bar has risen. That phase is difficult to trade, and it is still some distance from a fundamental handover.

Our relative research priority inside optics rests on two dimensions: order book and earnings conversion, and balance-sheet strain. Innolight, Eoptolink and Lumentum sit in the group furthest along on earnings conversion, and now have to validate profit quality through collections and inventory. Coherent's vertical capability and data-centre growth are real, while capex and working-capital absorption warrant a wider discount. T&S Communications has the best cash quality, the slowest growth and the highest valuation base. AAOI's capacity expansion offers the largest upside elasticity together with the most visible execution and financing risk.

3. Humanoid robotics: global public markets still lack a clean profit vehicle

Humanoid robotics is better held as a long-dated technology option than compared with memory and optics on a like-for-like reporting basis. Most leading whole-machine developers are either private or wrapped inside large groups, and listed component suppliers rarely disclose humanoid revenue as a separate line.

The Chinese sample has the same problem. Zhaowei Machinery and Zhongda Leader both grew revenue while earnings fell in the first half, with no separately disclosed humanoid revenue. UBTech's full-size embodied-intelligence humanoid revenue is growing quickly and remains loss-making. Shuanghuan Driveline's disclosed "reducers and other" line mixes in non-robotics business. Between qualified supply, design wins, capacity and sustainable profit sit several more checkpoints. These individual datapoints are not each bound to a first-party disclosure link in this piece, so they are treated as evidence still to be sourced rather than as established facts.

The judgment for September to November is therefore simple: until announcement-grade orders, separately disclosed revenue and gross margin appear, do not let media paraphrase answer a question that belongs to the financial statements. The group can generate high-elasticity event trades. It has not yet formed a profit pool in global public markets that stands alongside HBM or 1.6T optics.

4. The September–November odds table

The entire table below is research judgment and an observation framework, not a statement of fact. The "current research judgment" and "what the market has priced" columns are the author's view and will move with evidence. The "next confirmation" and "primary falsifier" columns are verifiable observation items.

The calendar

• Weekly: DDR5 spot transaction volume, the RDIMM/UDIMM spread, and the directional gap between spot and contract. Prices rising while volume keeps shrinking counts as quote strength, not demand confirmation.
• September: Micron's next print and guidance, testing whether roughly 86% gross margin holds, and whether earnings revisions across global memory move from post-peak decline to stabilisation.
• October: SK hynix, Samsung and the A-share Q3 reports. For the Chinese module names, watch whether gross margin, operating cash flow and inventory share improve in the same direction.
• October–November: The next quarter of optics delivery across China and the U.S. Whether Lumentum, Coherent and AAOI convert elevated guidance, and whether cash conversion improves at Innolight and Eoptolink.
• November: Execution against NVIDIA's FY27Q3 supply and capacity commitments, inventory and customer payment terms, which determine whether the AI hardware demand anchor still holds.

5. The final judgment

"Memory takes the baton and optics hands it over" is too tidy. The global reporting season did not produce that picture.

Industry conditions in memory remain the strongest of the group, and structural profit reaches SK hynix, Micron and Samsung first. Kioxia and SanDisk amplify the NAND cycle. Demingli and Longsys amplify inventory timing. The rising second derivative in spot quotes improves near-term odds for the latter two groups without removing the constraints from decelerating contract prices, inventory financing and cash recovery.

In optics, the margin for valuation error is narrowing while earnings conversion still leads most of AI hardware. Innolight's single-day decline shows how high expectations sit; the concurrent prints from Lumentum, Coherent and AAOI show global demand has not stopped. The more accurate response is to bet less on the sector label and pay more attention to company cash flow, capacity delivery and customer lock-in.

From September through November, our research priority sits with constraint ownership and verifiable profit. NAND and the Chinese module names stay in the observation set as high-elasticity, low-visibility exposures. Optics remains in the earnings-conversion group, with cash conversion determining how much weight the evidence carries. A genuine handover requires relative performance, earnings revisions, breadth and reported results to arrive together. For now, the profit pool has forked and the baton is still in the air.

Method and limitations

This piece prioritises government, regulatory and company disclosure. Prices and market capitalisations are taken at the 2026-09-01 close as a point-in-time snapshot and were not individually reconciled to fully diluted share counts. Accounting standards, fiscal calendars and currencies differ across these markets. The "market cap ÷ annualised latest-quarter net income" and "annualised H1 net profit" measures used here are peak-earnings screens. They are not forward P/E and they are not price conclusions. For the U.S. names, the denominator of the annualised multiple is non-GAAP net income, which is a different basis from the non-GAAP operating margin shown alongside it; the two cannot be used interchangeably. Lumentum's GAAP net income is affected by a one-time, non-cash loss on debt extinguishment, so operating comparisons use the company's disclosed non-GAAP net income and operating margin. SK hynix's net income includes large gains on investment assets, so only its operating margin is discussed and the screen P/E is not used for formal valuation. Any cross-market multiple should be normalised for profit, net cash, tax rate and share count before it is used in formal valuation.

The third-party consensus figures — Micron FY2027 consensus EPS near $155 and roughly 6.2x forward earnings — come from Seeking Alpha, captured on 2026-09-01. That is a vendor-aggregated sell-side consensus snapshot, not company guidance and not our own forecast. Differences in fiscal-year alignment and GAAP versus non-GAAP treatment across vendors will change the number.

On CXMT and HBM3E: the claim originates from Reuters citing people familiar with the matter, is not company-confirmed, and is not used here to infer pricing, share or supply-demand effects. It is recorded as an unverified supply-side risk. The Chinese sample companies referenced in the humanoid robotics section are likewise treated as evidence still to be sourced, since each datapoint is not individually bound to a first-party disclosure link.

Market price sources include Google Finance: MU, Google Finance: SK hynix, Google Finance: Samsung Electronics, Google Finance: Innolight, Google Finance: Eoptolink, Google Finance: T&S Communications, MarketWatch: Kioxia and exchange closing data for September 1. Financial data for the A-share companies follows the periodic reports filed on CNINFO.

This note reflects Lunartulip Research's views compiled from public information. It is not investment advice. Markets carry risk and decisions must be made independently.