TVE Corporation is the only qualified supplier of safety valves to Japan's pressurized water reactor fleet, a position no domestic or foreign competitor has breached in over two decades. Japan's nuclear policy has structurally reversed, mandating more operating reactors, longer reactor lifespans, and a fixed nuclear share of the grid, each of which compounds demand for TVE's highest-margin recurring inspection and replacement work. The qualified supply chain that must serve this wave shrank through the post-Fukushima decade and cannot be rebuilt quickly, so pricing power accrues to the credential holder. The market prices none of it: net cash is 61% of the market capitalization, the operating business trades at roughly 2x EV/EBIT, and the order book has nearly doubled while management still guides to a full-year profit the first half already exceeded. A formal upward revision is arithmetically unavoidable. As an added bonus, buying the shares also gives you a call option on a nuclear export channel that runs through TVE's controlling shareholder, carried at zero cost.

THE THESIS IN SIX POINTS

• Japan's nuclear policy reversal converts directly into more of TVE's highest-margin recurring O&M work, now and compounding as each unit restarts and ages.

• The qualified supply chain is structurally smaller than the demand it must serve, which hands pricing power to whoever holds the regulatory credentials rather than whoever can ramp fastest.

• TVE holds those credentials and no one else does, protected by three interlocking qualification layers that have kept every domestic and foreign rival out for more than twenty years.

• The valuation prices none of this, with 61% of the market cap in cash, the operating business at roughly 2x EV/EBIT, and a downside floored by a trough the post-Fukushima decade already proved.

• The order book has nearly doubled and the first half already beat the full-year guidance, making an upward revision a matter of arithmetic and timing.

• A US-Japan nuclear export option runs through TVE's controlling shareholder and is carried at zero cost.

Japan's nuclear policy has structurally reversed

Japan's nuclear policy has structurally reversed. The government has mandated a target of 27 to 30 operating reactors by 2040, against 15 today, extended permissible reactor lifespans to 60 years, and committed to nuclear as a fixed share of the national electricity mix.

Each of those decisions directly increases the volume and frequency of TVE's highest-margin work. More operating reactors means more mandatory inspection outages every 13 to 18 months. Longer reactor lifespans mean more replacement cycles on aging valve populations. Higher fleet utilization means shorter windows between outages and more intensive refurbishment at each one.

The policy stack does not create demand for new reactors someday. It creates demand for TVE's recurring O&M revenue now, on the existing fleet, compounding as each additional unit restarts and as each operating unit ages into heavier inspection requirements.

The supply chain is smaller than the demand it must serve

The supply chain that must serve this wave is structurally smaller than it was a decade ago. TVE's own 2022 submission to METI documented the attrition directly: of eleven casting pattern makers in its supply chain, three had already closed and four had no identified successor. Manufacturing headcount had fallen roughly 30%. Skilled valve designers were concentrated above age 55.

These are not cyclical shortages that normalize as orders recover. The subcontractors that closed during the post-Fukushima shutdown are gone. The qualified welders who retired did not pass their certifications to a successor cohort. TVE is the constraint. So is every other qualified supplier in this segment. That makes the supply-demand imbalance structural, and it means pricing power accrues to whoever holds the regulatory qualifications, not to whoever can ramp fastest.

TVE holds the qualifications, and no one else does

TVE holds those qualifications and no one else does. No domestic supplier has achieved Class 1 nuclear safety valve certification since at least 2000. No foreign supplier, including Velan with its 380-reactor global installed base, Flowserve with 200-plus reactors, and Framatome with a Japan office and long-standing utility relationships, has penetrated the domestic market.

The barrier is not a single certification hurdle. It is three interlocking layers: compliance with Japan-specific engineering codes that differ from Western standards in seismic methodology and material specifications, independent vendor registration with each utility under a quality assurance framework that requires years of audited manufacturing track record, and plant-level NRA construction plan approval that is verified reactor by reactor. TVE holds all of it. The moat did not erode during the lost decade. It is intact, and the industry it serves is now growing for the first time in fifteen years.

The valuation prices in none of it

The valuation prices in none of it. Net cash of ¥5,762mm sits against a ¥9,480mm market cap, meaning 61% of what you pay is cash and investment securities before the operating business is valued at all. Strip the balance sheet and the operating business trades at approximately 2x EV/EBIT.

The downside is empirically bounded: through the entire post-Fukushima period, with every reactor in Japan offline, TVE's nuclear segment never generated less than roughly ¥700 to ¥800mm of EBIT from mandatory pre-restart inspection and backfit work alone. That is the floor. It existed without a single operating reactor, without the current policy mandate, and without the restart wave now underway. You are buying the operating business at 2x a number that is materially above the worst-case floor the data has already proven.

The order book says the upswing is already happening

The order book says the upswing is not a forecast. It is already happening. The valve orderbook has expanded for seven consecutive quarters.

FY25 total orders reached ¥13,322mm at a 1.31x book-to-bill, the highest in six years. In 2Q FY26, TVE booked ¥4,207mm of new orders at a 1.28x ratio against the largest quarterly revenue print in the dataset. The backlog grew while the company was converting it at a record pace.

The consolidated half-year EBIT of ¥840mm already exceeds the ¥700mm full-year management forecast published six months ago. The mechanism is completed-contract accounting: orders book on signing, revenue recognizes on delivery and completion, so the order book leads the income statement by one to two years. The income statement is now catching up to an order book that nearly doubled. A formal guidance revision at the 3Q release is arithmetically unavoidable. That revision is the proximate catalyst.

An option that costs nothing to hold

Above the domestic case sits an option that costs nothing to hold. The United States and Japan have signed a framework committing tens of billions of dollars to joint nuclear build, with Mitsubishi Heavy Industries named as a potential equipment and technology supplier alongside Westinghouse on the PWR program. TVE's safety-valve position in the Japanese PWR fleet flows through the MHI supply chain via Seika Corporation, MHI's aligned trading house and TVE's largest shareholder at 21.5%.

If MHI takes PWR equipment scope on a multi-reactor program, TVE's sub-tier supply position is the natural place to source Japan-qualified safety valves. The payoff on a fleet-scale PWR valve package across multiple reactors is measured in multiples of the current market cap. Two observable events would confirm the option is converting: TVE re-pursuing ASME N and NV certifications, or an MHI framework agreement naming TVE. Until one appears, the option is tracked, not modeled. At 2x EV/EBIT on the domestic franchise, you carry it at zero cost.

Conclusion

TVE is a regulatory monopoly on the safety valves a restarting nuclear fleet cannot operate without, in an industry that just reversed fifteen years of decline by government mandate. The qualified supply chain is too small to serve the wave, and TVE holds the credentials that decide who gets paid. The order book has nearly doubled, the first half already beat the full-year guidance, and the revision that closes the gap is a matter of arithmetic and timing. You buy that franchise at roughly 2x the operating business, with 61% of the price sitting in cash and a downside the post-Fukushima decade already floored. The export option is free. The catalyst is already in the prints.

They also trade at a steep discount to every legitimate global comparable despite holding stronger competitive positions. Curtiss-Wright and BWX Technologies trade at 20-30x earnings with explicit nuclear restart premiums. Japan Steel Works, the closest domestic analog with an equivalent monopoly in reactor pressure vessel forgings, has re-rated sharply to 30x+ fwd P/E. The Japanese valve and actuator names have not followed, despite equally impregnable moats built on reactor-type-specific qualifications, utility-managed vendor lists, and 50-plus years of irreproducible operational data.

The cyclical upswing is larger than the restart count implies. Japan went from debating nuclear’s role to a Cabinet-approved target of 20% nuclear share by FY2040, requiring roughly 27-30 operating reactors versus 15 today. The 48th Nuclear Subcommittee in March 2026 put explicit numbers on the replacement gap for the first time: 5.5 GW needed in the 2040s, scaling to 16 GW by the 2050s. Layered on top are three structural upgrade cycles: 80-year license extensions requiring component re-qualification, operating cycle extension from 13 toward 24 months requiring higher-specification parts, and SMR buildout on decommissioned sites. These suppliers atrophied through 13 years of near-zero activity. Demand is now accelerating into a qualification-locked incumbent base with no new entrants in over two decades.

I have been long the Japan nuclear restart theme for some time now. Let me share with you my thoughts on one of my holdings.

Okano Valve (6492.T): A Japanese BWR Monopoly Pricing Six Months as a Full Year

Initial FY9/26 guide ¥966M. H1 actual ¥1,787M. Three consecutive years of guide-then-destroy. Bull-to-bear asymmetry 4.6×.

Six months ago, Okano Valve Manufacturing told the market to expect ¥966M of operating profit for the fiscal year ending September 2026. They just reported half-year operating profit of ¥1,787M, which is 185% of the original full-year guide. Management revised the guide up 102% in April after the first quarter print made the original number indefensible, then reaffirmed the revised ¥1,950M guide on May 14 even though H1 alone already covered 92% of it.

The stock is ¥16,730. It is up roughly 10% since the H1 print and modestly higher on the year. The market is now pricing this name against management’s revised guide, which has been wrong by an average of +115% across the last three fiscal years.

This is the setup.

What Okano Does

Okano Valve Manufacturing makes the main steam safety relief valves (MS-SRVs) and main steam isolation valves (MSIVs) for every Japanese boiling water reactor. The valves are written by model number, pressure setpoint, and lift characteristic into the licensed safety analysis of each reactor.

An alternative supplier cannot be qualified without re-running the reactor’s entire safety case through the Nuclear Regulation Authority. That is a multi-year, operator-funded process no utility will undertake while the incumbent performs. Switching cost is structurally infinite within the existing fleet.

BWR-3 and BWR-4 units carry 4 to 8 MS-SRVs per reactor. BWR-5 units carry 8. ABWR units carry 16 to 18. The MSIVs sit at containment penetrations, eight per unit, and must close within seconds on a loss-of-coolant signal. Both categories are NRA-qualified Class 1 safety equipment.

Maintenance was 48% of FY11/2024 revenue and has been growing faster than manufacturing as the operating fleet expands. Field technicians carry decades of accumulated knowledge of specific reactor units, specific valve serial numbers, and specific historical inspection findings. The embedded service expertise is harder to replicate than the physical valves.

The BWR Fleet Has Tripled in Twelve Months

Japan’s operating BWR fleet went from one unit to three between February 2025 and February 2026. Onagawa-2 restarted first. Shimane-2 followed. Kashiwazaki-Kariwa Unit 6 came online in February 2026. KK-7 is licensed, with restart work actively underway.

The key question has shifted from “will ever BWRs restart” to “how fast can the operating fleet expand.”

Each operating unit creates three concurrent revenue streams for Okano. The first is severe accident (SA) equipment installation, a post-Fukushima regulatory category that did not exist before 2011. The second is periodic inspection work on a 13 to 15 month cycle. The third is parts replacement and ongoing maintenance.

The SA equipment category alone represents a 30%+ content uplift per unit above the legacy maintenance baseline. Three operating units, two more close to restart, six more in the medium-probability pipeline. None of the units beyond KK-7 carry probability weight in our model.

The Fat Tail: Kashiwazaki-Kariwa 1 Through 5

TEPCO operates seven BWRs at the Kashiwazaki-Kariwa site. KK-6 is running. KK-7 is next. Units 1 through 5 are not in any model.

The binding constraint on KK-1 through 5 has been TEPCO’s balance sheet. TEPCO is currently running a process to bring in a private equity capital partner. KKR, Bain Capital, JIP, and JIC are the named candidates. Selection is expected late 2026 or early 2027.

Once a PE firm takes a stake, the capital constraint loosens materially. Community precedent at the Kashiwazaki-Kariwa site is already established by KK-6. Five additional TEPCO BWRs, each carrying sole-source Okano MS-SRV content, sit on the table at zero probability today.

The Sandbag Pattern

Three consecutive fiscal years of conservative initial guides followed by material beats:

The pattern is two layers deep. Management issues a conservative initial guide. They are forced to revise upward after the first or second quarter when the actual print makes the original number indefensible. Then they hold the revised guide even as H1 alone tracks to overrun the prior year’s full-year actual.

The sandbag isn’t a one-off. It runs every year. The mid-year revision is itself the tell. Management only revises when H1 alone is tracking to overrun the prior full-year number.

Applying the historical 2.4× beat ratio to the original ¥966M FY9/26 guide produces ¥2.3B of full-year operating profit. Current H1 annualized is already ¥3.6B.

The Per-Day Regime Has Broken

Why per-day, not absolute? Okano changed its fiscal year-end from November to September in 2025, making FY9/25 a 10-month transition period. Absolute FY25 revenue of ¥6.93B against FY24’s ¥8.62B looks like a 20% contraction. Per-day, FY25 was ¥23.05 versus FY24 at ¥22.32, up 3%. The same normalization matters quarter to quarter: Q1 FY26 had 92 calendar days, Q2 FY26 had 90 days plus the New Year holiday compression. The per-day series strips both effects out and produces a single comparable line across the FY change and the intra-year day-count drift.

Quarterly sales per calendar day strip out seasonal noise and reveal the underlying operating cadence. The metric is clean and the history is long.

Ten years of post-Fukushima data sit in a ¥17 to 23 per day band. Thirty consecutive quarters. The prior all-time single-quarter peak was 2Q FY9/2025 at ¥30.3/day.

The two most recent quarters broke the range consecutively and decisively:

• 1Q FY9/26 (Oct to Dec 2025): ¥24.4/day. At the ceiling of the prior range. This print pre-dates the KK-7 inspection work that flowed in 2Q.

• 2Q FY9/26 (Jan to Mar 2026): ¥40.0/day. +32% above the prior all-time peak. +75% above the top of the historical range.

Two consecutive quarters above the band mechanically excludes the single-lump-timing interpretation. Something structural lifted the floor.

Management explicitly cites KK-7 pull-forward as a contributor to the H1 print. That is credible. It is not sufficient on its own to explain the 1Q number, which pre-dated KK-7 inspection and already sat at the ceiling of the historical range.

A defensible read: 2Q contained a KK-7 acceleration that will moderate in subsequent quarters, but the structural floor has lifted from ¥17 to 22/day to ¥28 to 32/day. Peak inspection quarters print ¥38 to 40+/day on top of that. Full year FY9/26 sales land at ¥11 to 11.5B and operating profit at ¥3.1 to 3.5B.

Management’s implied 2H run-rate of ¥22.8/day requires the SA equipment work at KK sites, the Shimane-2 inspection cycles, and Tokai-2 content to disappear simultaneously between April and September. None of those things are scheduled to disappear.

Valuation: Three Scenarios

Three scenarios on FY9/26 EPS:

• Bear: ¥530 EPS x 17x P/E = ¥9,010 target. Implies FY26 operating profit of approximately ¥1.4B, a 28% miss against the already-revised ¥1.95B guide. Downside 46%.

• Base: ¥1,090 EPS x 25x P/E = ¥27,250 target. Implies FY26 operating profit of ¥2.9B, +49% versus the revised guide. Upside 63%.

• Bull: ¥1,743 EPS x 30x P/E = ¥52,290 target. Implies FY26 operating profit of ¥4.7B, +141% versus the revised guide. Upside 213%.

R/R is ~4.6x.

The bear case is not a conservative scenario. It requires management to miss the already-revised guide by 28%. H1 alone printed ¥1.79B. To hit the bear case, H2 operating profit would need to be ¥163M, a 91% sequential collapse from H1. That requires the business to break, not the cycle to weaken.

Roll-Forward EV Compression

Holding share price flat at ¥16,730, the roll-forward EV mechanics compound the valuation case rapidly.

At ¥16,730, today’s investor is paying ¥21.2B of enterprise value for the operating business and getting ¥5.7B of net cash and securities on top. By FY9/29 in the base case, accumulated free cash flow lifts the cash position to ¥20.4B. The roll-forward EV compresses to ¥6.5B. Cumulative EBIT generated over the four years reaches 130% of the starting EV.

The investor buys the entire enterprise value back in cumulative operating earnings inside four years while watching the EV they paid for compress by 69% on the balance sheet.

The peer discount is structural, not fundamental. Physics-anchored franchise names in Japan (Furuya Metal in PGM recycling, Nittobo in specialty glass, Mitsui Kinzoku in battery materials) trade at 30x+ forward earnings. The discount reflects the float, the absence of sell-side coverage, and the cyclical valve-and-actuator framing. As FY9/26 prints and the structural narrative becomes visible, both earnings and multiple should re-rate.

Regulatory monopoly. Trading at 19x company guidance EPS of ¥872/share. Beginning of generational nuclear restart cycle. Cash + Securities over 23% of Market Cap. More ways to win than to lose. Let me know your thoughts.

Leave a comment

Not investment advice. The author or affiliated funds may hold positions in the securities mentioned. Positions can change without notice. Do your own work.

The supply bottleneck is gone. The earnings inflection is here.

In October I published my initial thesis on Seikoh Giken, arguing that consensus implicitly priced in a sharp deceleration in AI-driven optical demand. The consensus was wrong. Q3 FY3/26 results confirmed it, and the company’s guidance revision makes the point emphatic.

New FY3/26 EPS guidance of ¥605 runs 15% above my bull case of ¥565. Read that again. The company is guiding above my most aggressive scenario with one quarter still to report.

What Happened?

Seikoh Giken revised FY3/26 consolidated guidance up across the board: net sales to ¥30.0B (+25% vs prior guidance of ¥24.0B), operating profit to ¥7.0B (+43% vs prior ¥4.9B). The revision came alongside Q3 results that demolished consensus on every line.

Q3 operating profit printed ¥2.09B against ¥1.45B consensus, a 44% beat. Sales hit ¥8.04B vs ¥6.50B expected (+24%). Nine-month cumulative operating profit reached ¥4.84B, up 184% year-over-year and already 69% of the new full-year target with Q4 remaining. The dividend was raised to ¥100 from ¥80. Consensus expected no change.

Subscribe

Optical Segment: The Engine

The upward revision stems almost entirely from Optical Products, where operating profit guidance jumped 50%, from ¥4.0B to ¥6.0B. Segment sales guidance rose 39% to ¥20.2B from ¥14.5B. Management attributes the step-change to two factors: accelerating data center demand and resolution of a critical parts bottleneck that had constrained production of optical connector polishing machines. Supply chain constraints on key components have lifted. Production can now meet accumulated backlog. Optics sales y/y growth accelerated from 55% to 156%!!!

This is the single most important development since I published the original note. The investment case was always about the size and durability of the demand wave. The constraint on realizing that demand was physical: Seikoh Giken could not build polishing machines fast enough because a critical input part was supply-limited. That bottleneck is now cleared.

Precision Machinery saw a modest uptick, with segment OP guidance rising 11% to ¥1.0B on automotive sensor and mold demand. It remains a secondary story.

Margin Expansion Confirms Operating Leverage

Profit growth is significantly outpacing revenue growth. Operating profit guidance rose 43% on a 25% sales revision. The mix shift toward higher-margin optical equipment is driving this wedge. Constrained capacity turning into fulfilled orders creates powerful operating leverage, exactly the scenario modeled in the original note with 55% incremental margins for the Optical segment.

The Q3 standalone numbers illustrate the point: ¥8.04B in sales produced ¥2.09B in operating profit, a 26% margin for the quarter. For a company that was running at 5.5% consolidated OP margin just two years ago, this is a structural transformation, not a cyclical blip.

From “Constrained Growth” to “Unleashed Capacity”

The supply unlock transforms the investment case. Prior to this quarter, the bull thesis rested on demand. The bear pushback centered on whether Seikoh Giken could actually convert backlog into revenue at scale. That objection is answered. Polishing equipment represents high-margin capex sales to hyperscalers and cabling suppliers. Backlog realization at these levels suggests the ¥6.0B optical OP figure establishes a new run-rate floor rather than a one-time spike.

If data center interconnect demand holds, and every hyperscaler capex datapoint says it will, removal of the production cap supports 15-20% sales CAGR over three years purely from fulfilling existing market demand. Margin expansion continues as the equipment mix stabilizes at higher levels.

What This Means for the Thesis

My original model assumed sustained fabric densification and continuous refresh cycles. Consensus assumed deceleration. The Q3 results and guidance revision validate the model and then some. Actual EPS guidance now exceeds my bull case by 15%, and we are still in the early innings of the 800G-to-1.6T transition that drives the superlinear uplift in polishing consumables, connectors, and ferrules per rack.

The structural thesis is unchanged: dominant market share, razor-and-blade economics, and an expanding addressable market driven by AI fabric density. What has changed is the constraint. The capacity ceiling that capped near-term realization is gone. Seikoh Giken can now convert its backlog and market position into earnings at a rate the market has not yet fully priced.

Share

The revision upcycle I projected is underway. Consensus needs to move substantially higher for FY3/27 and beyond. I will update my full model with revised estimates in a follow-up note. What do you think?

Comment

Hello friends long time no see. I’ll give an explanation for my recent silence on Substack in the fullness of time, but I’m in Tokyo now and Nittobo is reporting tomorrow.

DISCLAIMER: Funds and individuals who compensate me for advisory services may hold positions in securities discussed in these posts. I may also hold positions in securities mentioned. Nothing herein constitutes investment advice, and readers should conduct their own due diligence before making investment decisions. Past performance does not guarantee future results.​​​​​​​​​​​​​​​​

Shout out to Ray Wang at SemiAnalysis whose post inspired me to sharpen my pencils and take a look at this opportunity back in October ‘25. Stock is up a lot since then. A year from now, it is likely to be much higher than where it is now. Let me tell you why.

Why I own it: Nittobo is a mission-critical monopoly at the binding constraint of AI substrates, mispriced as a commodity materials business

Variant Perception: Market views this as a pecialty glass fiber enjoying a temporary supply/demand anomaly in a cyclical materials industry where Chinese low cost competition will over-expand, driving margins down over next 12-24 months so heavily discount the terminal value ≈ low multiple on out years.

My view: Nittobo behaves like a mission-critical tollbooth with physics-backed barriers, long qualification cycles, and demonstrated pricing power.

This is a structural monopoly reinforced by process know-how, yield physics, and roadmap control. It is not a cyclical demand spike that caught the entire semi supply chain unaware.

Summary

 Nittobo trades at roughly 8.5 times base-case FY3/2028 EPS, a valuation wholly inconsistent with its position as the dominant supplier of ultra-low-CTE and low-loss glass cloth that gates production of advanced AI substrates. (And as the news out of Apple/QCOM/MGC show- pretty much everything else too)

 Nittobo controls roughly 90% of global supply of T-glass, a mission-critical insulation layer whose performance determines warpage, yield, and signal integrity in high-end GPU and AI server packages. 

Despite being less than 1.5% of the BOM of a $25k+ accelerator, T-glass sits at the tightest bottleneck in the stack. Demand is driven by the deepest-pocketed customers in the world, and supply is constrained by physics, process control, and capital intensity, not by willingness to invest.

Share

The Core Moat:

New entrants do not fail because they cannot achieve low Dk / low Df or low CTE in a lab. They fail because they cannot simultaneously satisfy the joint constraint required for qualification-grade yield at scale:

      1.    Low dielectric loss (Df) for 112G+ signal integrity

      2.    Low surface roughness (Ra) to minimize conductor loss

      3.    High resin wet-out / adhesion to survive lamination and thermal cycling

      4.    Stable laminate yield at volume, with tight filament diameter and weave uniformity

Most entrants can hit two, sometimes three of these. Fixing one variable breaks another. This creates a technical whack-a-mole. Smoothing glass to lower Ra kills surface energy, which causes poor wet-out, which leads to delamination. Adding sizing agents to fix adhesion raises Df or degrades uniformity. Pilot-line success fails at scale due to filament and weave variance, which causes skew, impedance drift, and yield collapse.

Key distinction: Competitors treat composition, sizing, weaving, and finishing as separable problems. Nittobo co-designs the entire system, including glass chemistry, draw conditions, sizing, and weave geometry, locking down narrow process windows, not just material specs.

This is why datasheet parity has not translated into supply relief, and why substrate makers’ yield dashboards still point to glass cloth as the binding constraint. Brochures do not tell the story.

If competitors were on track to scale, we would not see:

      ∙     Continued bottleneck language from top substrate makers like Ibiden

      ∙     Price hikes cascading through laminates and prepregs at companies like Resonac and MGC

      ∙     Taiwan PCB makers raising prices despite claims of second-source progress

None of this behavior is consistent with an easing constraint. It is consistent with qualification friction and yield failure at scale.

Message me

The Moat Is Widening, Not Narrowing

While competitors struggle to scale today’s spec, Nittobo is already moving the goalposts. T-glass 2.0 delivers roughly 30% improvement in thermal expansion, targeting around 2.0 ppm versus 2.8 ppm, aimed at larger, hotter AI packages with targeted use around 2028. NER / next-gen low-loss glass is ramping alongside T-glass. This raises switching costs for customers who demand low warpage and high signal integrity.

As new entrants struggle to scale production volumes, Nittobo's capacity is expanding. Fukushima expansion delivers roughly 3x T-glass output. Taiwan furnace investment is underway. Roughly 20% incremental supply comes via Nan Ya weaving, while Nittobo retains control of yarn chemistry and process-critical steps.

The market is not pricing the volume step-function post-2027, the pricing power reset as LTAs signed in 2023-24 roll off, or the fact that higher-performance glass increases pricing leverage as AI substrates become more unforgiving.

What Would Falsify the Thesis

These catalysts would be taken as strong  disconfirmations of my hypothesis:

      1.    A competitor demonstrates sustained, high-volume shipment of low-CTE / low-loss cloth into Tier-1 AI substrates without downstream yield or price tension.

      2.    Substrate and laminate price hikes reverse despite continued AI package complexity growth.

      3.    A true substitute like quartz cloth achieves mass adoption at scale, not announcements of pilot lines.

Bottom Line

My base case 3/28 EPS is ¥1,945 versus consensus at ¥503. The assumptions to get there aren’t crazy. Topline up on combination of 1) LTA price resets higher, 2) flow though of price hikes announced last year, and 3) repurposing lower grade glass furnace capacity to T/NE Glass. Incremental profit margins down from 92% to 80% q/q reflecting fixed costs added in as Fukushima and Taiwan plant upgrades and personnel expense stack in. I purposely spread the last 6 quarter with y/y comps including the 3/2024 quarters to highlight that, ya know, we are in the beginning of the up-cycle.

These assumptions yield consolidated annual numbers below (eps normalized to remove impact of the RE disposal reported in the current quarter.)

At roughly 8.5 times FY3/2028 EPS, we are paying a commodity multiple for a business that gates AI accelerator output, sells into customers who cannot substitute away, captures pricing power disproportionate to BOM share, and is actively widening its moat via technology and capacity.

That mismatch is the opportunity.​​​​​​​​​​​​​​​​

Double disclaimer:

Anything is on the table when it comes to tomorrow’s sendings. Could announce a huge capex program and drop 18% after hours like Ibiden (biggest customer of Nittobo’s biggest customers) did after they arnnounced ¥500bn ADVPKG substrate capacity expansion.

Or its limit up 20% three days in a row as market digests the fact that NO ONE IS COMING TO TSMC suppliers WHO HAVE NOT SECURED TGLASS ALLOCATION, and what this implies for their ability to raise price the next 12-24 months. Then, and only then, will we start to digest that whole Taiwan Glass is sold out and yet still struggles to ramp acceptable product, Nittobo is already moving on to T glass v2.0 and NE glass v2.0.

I skipped over other factors that have me sounding so convicted on the trajectory. Will lay those out in the post mortem, my jet lag is hitting me hard.

Subscribe

The axis of debate around the “AI Bubble” issue is almost always revolving around supply: Nvidia GPUs, TSMC fabs, Zuckerberg revealing $600 billion in capex for 2026-2028, Oracle and OpenAI signing mega deals.

But very few people are tracking demand. When labs share usage data, it’s always a different KPI at irregular intervals, in different units.

Each metric shows growth, but without consistent disclosure, it’s extremely difficult to see how compute demand is really evolving.

So I thought I’d take a shot at it. The numbers are shocking.

Methodology

I scraped data from labs and companies on DAU, MAU, prompts, tokens, and LLM usage surveys to estimate daily prompts for Q1 2023, Q1 2024, and each quarter of 2025.

I analyzed token intensity by prompt type (simple text, deep research, extended reasoning, code assist, multimodal) and combined it with prompt composition to calculate total tokens, assuming token intensity rises with expanding context windows.

**A note on v0.1:** This was my first attempt at these estimates, and my taxonomy of prompts had overlap and miscategorization between “extended reasoning” and “deep research.” GPT-4 Turbo and Claude 2.1 (both November 2023) had chain-of-thought (CoT) prompting, longer context windows for sustained logical processing, and improved problem-solving over multiple steps, so they went into the ER/DR buckets. After I put this out, I found at least 7 things that were wrong with my assumptions. I’ve been working on a new model and will be posting more about it in coming days.

Key findings:

1. Demand is accelerating exponentially, not linearly

Daily token generation: 42 billion (Q1 2023) → 94,847 billion (Q3 2025)

That’s a 2,258x increase in under three years.

But look at the quarterly incremental additions:

- Q1→Q2 2024: +775Bn tokens

- Q2→Q3 2024: +1,389Bn tokens

- Q3→Q4 2024: +3,419Bn tokens

- Q4 2024→Q1 2025: +6,329Bn tokens

- Q1→Q2 2025: +17,541Bn tokens

- Q2→Q3 2025: +39,086Bn tokens

We went from adding under 1 trillion tokens per quarter to nearly 40 trillion tokens per quarter in just five quarters.

2. Composition of demand

Complex prompts (Extended Reasoning, Deep Research, Multimodal) are only ~40% of total prompts sent.

However, they now represent 90%+ of all token demand, up from just 4% in early 2023.

This is the key insight: A small shift in what people are asking AI to do has created a massive shift in compute demand.

Token consumption per prompt:

- Simple text: 280 tokens

- Extended reasoning: 5,500 tokens (20x)

- Deep research: 210,000 tokens (750x)

- Multimodal: 24,000 tokens (86x)

- AI Coding: 52,000 tokens (186x)

3. We’re still in the early innings

Multimodal is only 23% of prompts, mostly static images. Video hasn’t scaled yet.

Consider what’s coming:

- A 10-minute video costs approximately 177,000 tokens just to load before any analysis begins

- Context windows keep expanding, heading toward millions of tokens

- Vision models are adding “perception tokens” for visual reasoning

- Deep research prompts now consume 210,000 tokens each, 20x higher than a year ago

- Agentic workflows are becoming the default for high-value tasks

What about total tokens in recent quarters?

We have passed the 1 quadrillion token milestone sometime between feb/march 2025. I estimate ~7 quadrillion tokens in 3Q25.

Number too big! I must be completely off the mark! No way!

And then I saw Sundar Pichai’s speech at the GCP dev day on October 9th.

Ok, maybe I’m in the right zip code.

Then I spread the Google level KPIs to get a sense of acceleration:

It appears to me that we are not even close to the knee of the curve.

What about the bubble?

CC Wei (TSMC) recently guided that 45%+ of AI GPU demand is sustainable. He also mentioned that Nvidia customers were telling him that token demand is currently doubling every three months, exactly what these figures show.

If token demand keeps accelerating at this pace, then 45% looks conservative.

The question isn’t whether the unprecedented fixed asset investment boom we’re experiencing will leave us with trillions of dollars in stranded assets.

The question is whether we’re building fast enough.

Executive Summary

NHK Spring trades at 12.4× consensus EPS despite operating as a duopoly supplier of critical components for AI data center storage infrastructure. The company manufactures precision disk drive suspensions—the mechanical assemblies that position read/write heads in hard disk drives. As AI models evolve from text to multimodal systems generating video, audio, and 3D content, data creation is growing by orders of magnitude. Storage demand now scales faster than compute demand itself, with hyperscalers projecting 30-40% annual HDD exabyte growth through 2030.

The HDD suspension market is poised for explosive TAM expansion through 2030. In my bear case, the market grows 4.4× driven by modest volume increases and modest pricing power. In the base case, the market expands 6.8× as exabyte growth compounds with rising suspensions-per-drive and duopoly pricing dynamics. The bull case projects 9.2× expansion if HAMR adoption accelerates and hyperscaler buildouts exceed current forecasts. NHK Spring, controlling 45-50% market share in a newly formed duopoly (following Suncall’s September 2024 exit), is positioned to capture the majority of this growth.

The market still values NHK as a cyclical auto-parts manufacturer. In reality, its disk-drive suspension segment has grown roughly 60% year-over-year on average over the past five quarters, now drives over half of EBIT, and is transforming into a high-margin, secular growth engine. With 52-week lead times on advanced drives and incremental margins above 40%, NHK’s earnings can compound at more than 40% annually through FY2030—yet the business trades at a mature cyclical multiple.

At 12.4× consensus EPS of ¥195/share in Mar-2026 earnings, a rerating to just 15x on my base case ¥270 EPS implies over 60% upside, with further optionality if AI data creation continues to outpace compute growth. This is a classic case of secular growth trapped in a cyclical multiple, a type of asymmetry that doesn’t last for long.

The Opportunity: AI’s Exponential Storage Problem

The Lost Decade (2014–2023): Secular Decline in HDD Demand

To understand the magnitude of the current inflection, you must first understand what preceded it.

Unit Collapse Despite Exabyte Growth

Global HDD unit shipments plummeted from 564 million in 2014 to approximately 127 million in 2023—a roughly 77% decline—driven primarily by the migration of client devices to SSDs, which erased over 400 million annual drives from PCs, notebooks, and consumer electronics like DVRs and game consoles.

Three structural headwinds drove this collapse:

Flash cost curve (3D NAND scaling): SSDs achieved cost parity for client PCs and mid-tier enterprise tiers, accelerating HDD displacement.

Hyperscaler focus on efficiency: Cloud providers postponed capacity expansions and extended refresh cycles amid economic pressures.

COVID-related inventory whiplash: Over-ordering in 2020–2022 triggered a 2023 digestion phase, slashing demand.

Slower content growth: Streaming platforms enhanced compression algorithms, curbing video storage expansion rates.

Data Center Offset, But Not Enough

Enterprise nearline drives for cloud and data centers emerged as the sole growth segment, with shipped exabytes surging from ~35 EB in 2014 to 350–400 EB by 2023—a tenfold rise fueled by capacity per drive escalating from ~4 TB to over 20 TB, rather than increased unit volumes.

By 2022–2023, the industry hit a cyclical trough: utilization rates and ASPs plummeted, prompting Seagate, Western Digital, and Toshiba to slash production amid a 33% unit drop in 2022 and continued declines into 2023.

The AI Inflection: Surge from ChatGPT Day (Nov 2022) to Today

The release of ChatGPT on November 30, 2022, ignited an AI boom that profoundly altered HDD dynamics. Hyperscaler demand for cost-effective, high-density nearline storage in AI training and inference pipelines reversed prior declines in unit volumes, sparked hyper-growth in capacity shipments, and led to an OEM revenue rebound.

Here is the HDD rebound story in figures, using STX and WDC data:

Explosive Growth in Exabyte Shipments

Exabyte shipments recovered post-ChatGPT: from an estimated ~600 EB in 2022 to ~897 EB in 2023, then surging 49% to 1,337 EB in 2024 amid AI-fueled hyperscaler builds. Full-year forecasts are pointing to 1,600+ EB as 30TB+ HAMR drives ramp, dwarfing the ~10× decade-long growth to 2023. Seagate now projects mass-capacity exabyte shipments doubling roughly every two years, a pace last seen before 2017.

OEM Revenues: Robust Recovery

HDD industry revenues, squeezed to ~$13.4 billion in 2023 amid inventory digestion, roared back 48% to $19.8 billion in 2024 on premium nearline pricing. Seagate’s HDD-focused FY2025 (ending July 1, 2025) hit $9.1 billion (up ~42% from FY2024’s ~$6.4 billion), Western Digital’s disk revenues climbed ~51% to ~$4.8 billion in FY2025.

This starkly contrasts pre-AI trends of unit erosion despite incremental EB gains: AI has recast HDDs as indispensable for petabyte-scale datasets, driving value over volume in a $20B+ market poised for continued expansion.

Impact on NHK Spring

Here is a summary of NHK Spring’s last 5 years financials, showing the impact of the rebound in nearline HDD on its DDS business:

These results reflect a fundamental shift in storage economics driven by AI. To understand why this growth will continue accelerating, we need to examine what makes AI workloads uniquely storage-intensive.

Why AI Creates Exponentially Growing Storage Demand

AI data centers produce massive volumes of warm and cold storage tiers, including intermediate model checkpoints, fine-tuning corpora, and video frames, which demand petabyte-scale capacity for infrequent access.

Demand Trajectory

AI data centers such as Microsoft Fairwater, Google Jupiter AI, and AWS Project Atlas now integrate hundreds of thousands of HDDs per campus. These drives store model checkpoints, logs, embeddings, and video intermediates. Aggregate exabyte demand now scales with model-training throughput, not with consumer devices.

For a decade, HDD demand shrank as the world shifted to solid-state storage. AI has reversed that trend by generating exponentially growing cold and warm data that is too large and too slow for flash. HDDs are once again essential as the bulk storage foundation of the AI era.

The Solution: Why HDDs & NHK Win

Why HDD Suspensions Are Critical: An Engineering Marvel

HDD suspensions hold the read/write head, allowing it to fly just above the surface of the spinning disk, with typical clearance of about 3nm. For comparison, a standard sheet of paper is approximately 100,000 nanometers thick.

This is an engineering marvel: imagine maintaining a 3-nanometer gap while the platter spins at 7,200 RPM and the actuator arm swings the head across tracks thousands of times per second. The precision spring mechanism of the disk drive suspension (DDS) must absorb vibrations, maintain exact positioning, and ensure that the head doesn’t crash into the platter.

The Next Generation: HAMR Technology Increases Complexity

The next generation of high-capacity HDD depends on even more sophisticated suspension technology. Seagate’s Heat Assisted Magnetic Recording (HAMR) technology uses a laser integrated into the read/write head assembly to heat the recording media to approximately 450°C, enabling 30TB+ drives with 3TB per platter. Seagate envisions 50TB drives within a few years, using 4-5TB platters.

These HAMR heads are far more complex than traditional heads:

• They incorporate nanophotonic lasers

• They require precise thermal management

• The suspension must accommodate additional components while maintaining nanometer-level precision

NHK Spring’s Flagship “Triple Stage Actuator HDD Suspension”

NHK Spring produces advanced HDD suspensions designed for ultra-high-capacity nearline HDDs. These devices position electrical conductors directly onto the suspension structure at three pivot points: the Voice Coil Motor (VCM), the milliactuator, and the microactuator. This enables more precise control over the head’s position, which enhances areal density and allows for larger drive capacities.

NHK’s Unique Position: From Auto Springs to Digital Infrastructure

NHK Spring’s entry into HDD suspensions emerged from its core expertise in precision spring engineering. In the 1980s, as hard drives shrank and data densities soared, HDD makers needed suspension assemblies with micron-level stiffness, damping, and fatigue performance. NHK had already mastered these mechanical challenges in automotive coil and valve springs.

Leveraging its metallurgy, heat treatment, and precision forming capabilities, NHK developed ultra-thin stainless spring components capable of maintaining nanometer fly-height tolerances. By investing early in clean-room manufacturing and automated inspection, it built trust with HDD OEMs like Seagate (STX) and Western Digital (WDC).

Over time, as barriers to entry rose and the supplier base consolidated, NHK’s manufacturing discipline and reliability reputation transformed an auto spring maker into a critical player in the digital infrastructure supply chain.

Economics: HDD’s Unassailable Advantage over SSD for AI Use Cases

HDDs maintain a ~83% lower storage cost per GB (6× advantage over enterprise SSDs), enabling economical scaling for high-volume storage tiers. While SSDs offer lower idle power per TB (~0.16W/TB vs. HDD’s ~0.26W/TB), HDDs deliver 19–94% superior overall power efficiency (TB per watt) in read- and write-intensive AI workloads, balancing cost and energy demands at scale.

These datasets’ sheer size and low access velocity exceed SSDs’ cost-effective thresholds, leaving HDDs as the only scalable option for bulk AI storage.

Scale-Up + Scale-Out = Superlinear Suspension Volume Growth

Most investors understand that AI drives more data and therefore more storage. What they miss is that storage demand scales non-linearly with capacity and technology shifts. As the below table shows, when data-center scale rises from a 100-megawatt pod to a 1-gigawatt campus, storage capacity grows 15× but total suspensions rise nearly 14× even though drive units rise only 12×. Higher-density drives and more heads per drive offset unit efficiency gains, creating convex growth in suspension demand. In the bull case, this accelerates further as multi-actuator and higher-platter designs push suspensions per drive from 18 to 22. The market models storage linearly by exabytes, but NHK’s exposure is to this second-order, structural nonlinearity.

Oligopolistic HDD suspension market structure has become more oligopolistic

Pre-2025 HDD Suspension market structure:

• NHK Spring: ~45% share

• TDK: ~45% share

• Suncall: ~10% share

Suncall announced permanent exit from HDD suspension business in September 2024

This volume shift is permanent - HDD OEMs won’t re-qualify a third supplier given:

• 6-12 month qualification timelines

• Yield ramp takes another 6 months

• Any line disruption costs millions per day in lost production

Pricing power dynamics: TDK and NHK have been fined by regulators in the USA, Japan, Taiwan, and China for “cartel-like” price-fixing over the past few years. With Suncall gone and demand surging (52-week lead times on HAMR drives), pricing environment has never been better.

Nearline HDD BOM analysis suggests room for NHK to raise pricing

Suspensions cost $22.5 for 20 units in both models, representing just 6-7.5% of total COGS, dwarfed by heads (22-30%) and media platters (23-26%). This low relative share highlights suspensions as a stable, low-impact component amid rising costs for advanced recording tech like HAMR, which boosts capacity but inflates head and media expenses. As nearline HDD shipments grow 7-10% annually through 2030, driven by data center and AI demand, absolute suspension volumes per drive (one per head) will scale, amplifying supplier leverage without proportionally burdening OEMs.

This cost structure suggests ample room for NHK Spring, a dominant suspension supplier with 40-50% market share, to raise prices by 10-20%—adding only $2.50-4.50 to unit COGS (under 1.5% impact)—while capitalizing on recovering HDD production and firming average selling prices (up ~5% YoY in 2025). In an oligopolistic market, NHK’s precision engineering for HAMR/MAMR compatibility justifies premiums, especially as material costs like stainless steel rise modestly and can be passed through. With nearline ASPs strengthening amid enterprise storage growth (HDDs holding 36% share), targeted hikes could enhance margins sustainably, as suspensions enable high-capacity designs without equivalent cost escalation.

Valuation & Modeling

What Does This Mean for Modeling Assumptions?

Based on the structural dynamics outlined above—content-per-drive increases, scale-up effects, and scale-out multiplication—a straightforward bull case would project 8-15× unit volume growth through 2030. These aren’t fantasy numbers: they simply reflect the combination of exabyte growth (2-3x every two years), rising suspensions per drive (from 18 to 22 units), and the non-linear effects of data center scaling.

However, to build in a substantial margin of safety, I discount these projections by 50% across the board to arrive at base case assumptions of 4.0/5.8/7.5× volume growth through 2030. Combined with 15/25/50% ASP uplift over 5 years (reflecting duopoly pricing power and HAMR complexity premiums), this gets me to 4.4/6.8/9.2× TAM expansion in bear/base/bull scenarios. Even in the bear case, NHK’s disk-drive business more than doubles. In the base case, it grows nearly 7× over five years.

Leave a comment

Here is my base case for the next three quarters:

And next 5 years:

Conclusion

NHK Spring’s transformation from cyclical auto-parts supplier to secular AI infrastructure play is already underway—the disk-drive suspension segment now drives more than half of EBIT and has compounded at 60% annually over the past five quarters. Yet the market still prices the business at 12.4× consensus earnings, as if the AI storage build-out doesn’t exist.

The investment case rests on three pillars:

Structural volume growth: With exabyte shipments doubling every two years, suspensions-per-drive rising from 18 to 22 units, and data centers scaling non-linearly, NHK’s addressable market is expanding by 5-10× over the next five years.

Pricing power: Suncall’s exit created a duopoly. With 52-week lead times and suspensions representing just 6-8% of HDD bill-of-materials, NHK and TDK can raise prices by 10-20% without meaningfully impacting OEM economics.

Operating leverage: With fixed costs largely behind them and incremental margins running at 40%, every dollar of additional revenue drops 40 cents to the bottom line. The result: earnings can compound at 40%+ annually through FY2030, yet the stock trades at a 12-13× multiple typical of mature auto-parts suppliers. A rerating to just 15× on my base case FY2027 EPS of ¥270 implies 60%+ upside from current levels, with further optionality if AI data creation continues to outpace even bullish forecasts.

This is textbook asymmetry: structural growth trapped in a cyclical multiple. These mispricings rarely persist once the market recognizes the transition. The catalysts outlined above should accelerate that recognition over the next 12-18 months.

Catalysts (Next 12-18 Months)

Near-term:

1. FY25 Q2 earnings (Nov. 12, 2025): Should show sequential revenue and margin re-acceleration as inventory normalization completes. Key metric: DDS margin expanding from 20.8% (1Q25) toward ~28%.

2. STX/WDC 1Q26 earnings (Oct. 22 and Oct. 30, 2025): Further nearline guidance raises validate AI storage thesis. Watch for HAMR mix commentary (target: 25-30% of nearline).

Medium-term:

4. HAMR volume ramp: STX/WDC disclosing 32TB+ HAMR at 30-40% of nearline mix = ASP tailwind confirmation for NHK.

5. Hyperscaler capex guidance: Further MSFT/GOOG/META 2026 capex guidance = TAM expansion.

6. NHK management “share gains” disclosure: Incremental color on Suncall volume = Street wakes up to duopoly value.

7. Video AI usage data: Sora 2 hitting 15M+ users or Videogen 150K+ enterprise customers = storage thesis confirmed.

8. DDS margin re-acceleration: FY26 margins breaking back above 24-25% (vs. FY25 23.8%) = incremental leverage visible.

9. M&A optionality: At current valuation, DDS could be acquisition target for TDK (duopoly consolidation) or PE (carve-out).

10. Share buybacks: NHK Spring bought back around 10% of the current market cap in shares over the past five years, if I am anywhere close to reality in my modeling, even in a bear case they will have lots of excess cash to deploy into buybacks.

    ---

Key Risks

• Erosion of HDD cost advantage over SSD:

  • Faster-than-expected SSD price compression to ≤$0.03/GB at enterprise scale, or advancements in QLC SSDs with sustained write endurance optimized for object storage, could narrow the $/GB cost gap between SSDs and HDDs

  • Mitigants:

    • Current enterprise SSD pricing remains well above $0.03/GB, with high-capacity models like the 61.44TB Solidigm D5-P5336 at around $0.06–0.12/GB and street prices for 60TB+ NVMe drives exceeding $0.09/GB, maintaining a 4~6x gap over HDDs;

    • Hyperscaler tenders show contract pricing rising 10% QoQ due to AI-driven demand, while QLC endurance improvements are promising but not yet transformative enough to fully displace HDDs in AI DC.

• Shift from MAMR to HAMR reducing # of platters per drive:

  • HAMR technology boosts areal density, enabling 30TB+ capacities with fewer platters per unit, which directly reduces the number of suspensions needed per drive; this could lower overall suspension demand as HDD manufacturers optimize bill-of-materials costs.

  • Mitigants:

    • Despite fewer platters per drive, HAMR is driving explosive growth in total HDD shipments and exabyte capacity for AI DC, with Seagate’s 30TB HAMR roadmaps targeting 100TB by decade’s end, offsetting per-drive reductions through higher unit volume;

    • NHK is scaling production for these denser designs to meet demand

• Meaningful improvements in storage efficiency:

  • Architectural shifts that cut persistent storage per model, such as improved checkpoint sparsification, de-duplication (“dedup”) at scale, and data retention policy changes that trim log requirements.

  • Mitigant:

    • As AI datasets and models grow from petabyte to exabyte scale ensures continued need for nearline HDDs, while dedup and sparsification may only optimize existing data rather than eliminate future storage requirements

    • Emerging multimodal/spatial AI workloads are OOM step-ups versus current workloads

• Cold-storage demand spike:

  • A spike in cold-storage deployment could cannibalize nearline HDD demand by shifting infrequently accessed AI data (e.g., training datasets or model outputs) to tape, which is lower-cost and energy-efficient

  • Mitigants:

    • Tape primarily targets colder, archival data while nearline HDDs handle high R/W frequency, sequential-access workloads essential for AI inference and training

    • Tape’s slower access speeds limit its substitution for HDDs in active data centers

• New entrants:

  • Emergence of new entrants outside the STX/WDCl duopoly could flood the market with additional suspension capacity or alternative micro-actuator technologies, disrupting supply chains and eroding pricing power

  • Mitigants:

    • The HDD duopoly remains unchallenged in 2025, with no significant new entrants reported;

    • NHK and TDK are scaling within the ecosystem to meet AI-driven growth, projecting the market to reach $2.5 billion by year-end

    • Micro-actuator advancements stay incremental without disruptive substitutes

NHK Spring Business Description

Founded in 1939, NHK Spring began as a manufacturer of steel springs for automobiles and railways, supplying suspension and engine components to Japan’s early automotive industry. Through the postwar boom, it became a key supplier to Nissan, Toyota, and other domestic automakers, expanding into coil, leaf, and torsion springs as Japan’s vehicle exports surged. By the 1970s, NHK had built a global reputation for metallurgy, fatigue resistance, and precision forming, establishing overseas plants to support automakers abroad.

NHK Spring today operates in 5 segments

1. Automotive Suspension Springs
   NHK Spring produces coil, leaf, and stabilizer springs for cars and commercial vehicles, emphasizing light weight, high strength, and fatigue resistance. It holds a global top share and supplies most major automakers worldwide.

2. Automotive Seating
   Through NHK Seating and joint ventures with Faurecia, NHK designs and assembles complete seat systems and frames. It focuses on comfort, safety, and modular lightweight structures, with strong OEM integration and record recent profits.

3. Precision Springs & Components
   This unit makes small, high-precision springs and related components for engines, transmissions, and industrial systems. It leverages NHK’s core expertise in metallurgy and fatigue control for automotive and electronics applications.

4. Disk Drive Suspensions: Was included within “Precision Springs & Components” until 3/31/25, when they provided historical quarterly data for 3/24 to 3/25.

5. Industrial Machinery & Equipment
   NHK supplies precision parts and systems for semiconductor tools, infrastructure, and machinery, plus specialty products like pipe supports and parking systems. The segment extends NHK’s mechanical engineering capabilities beyond automotive markets.

   Subscribe now

You’ve probably heard the bear case by now: “AI data centers are a cash-burning bubble. Hardware becomes obsolete every year, real profits are inflated by understating depreciation, and old GPUs are worthless as collateral.”

It sounds scary. People have strong opinions. But most opinions I see are vibe-based, not grounded in reality.

When I ran the numbers, the bear case seems like a hysteric fantasy. Let’s walk through the analysis.

The Power Economics Are Straightforward

Start with simple math that anyone can verify.

Each A100 GPU, launched back in 2020, uses approximately 0.4 kW. Add in cooling overhead (a power usage effectiveness ratio of 1.2), and you’re at 0.48 kWh per hour of operation.

At a typical industrial electricity rate of $0.20/kWh, that’s 10c/GPU-hour in power costs.

Even at a low spot market rental rate of just $0.50/hour, each GPU earns $0.40/hour in gross margin. That’s roughly 4 times the power cost.

But let’s make this more rigorous. Here’s how the economics change across different electricity costs and rental rates:

A100 GPU Gross Margin per Hour (Rental Rate vs. Electricity Cost)

Notice what this table shows: even at a realistic bear scenario (with $0.50/hour rental rates and $0.20/kWh electricity, nearly triple typical industrial rates), each GPU still generates ~$0.40/hour in gross margin.

Electricity costs would need to exceed $1.04/kWh before a GPU renting at $0.50/hour would lose money. For context, the average U.S. industrial electricity rate is around $0.07/kWh, and even expensive states like California average around $0.15-0.20/kWh.

The point is simple: margin drives cash flow, not depreciation schedules.

OK, how does it look at scale?

• Assuming a 25MW cluster at 1.2x PUE we can fit ~52,083 A100 GPUs

• At 75% utilization we get ~342M GPU-hours/yr = 164.3kWh per year

• Assuming $0.50/hr rental rate and $0.20/kWh electricity cost, we get ~$138mm gross profit

• Taking out $12mm in fixed costs nets ~$126mm EBITDA (74% margin)

You will notice that depreciation does not appear in the calculation: it is an accounting entry, not a cash cost. The data center either generates cash or it doesn’t.

What are realistic GPU rental rate assumptions?

For A100 rentals, I’ve cobbled together some quotes:

• Latest quoted prices for Spot/Batch rentals is $0.5-$1.5/hr

• On-Demand rates are $1.3-$1.8/hr

• Hyperscaler rates are $3-$5/hr

Depending on your customer base composition, the realized rental pricing right now is likely well above $1.00/hr.

If truly worthless, A100 rates would drop the cost of electricity per GPU-hour, ie $0.096. The reality is that they are 5-15x higher.

Sources (I AM NOT MAKING THIS $H!T UP)

• Lambda Cloud pricing (A100, H100 on‑demand detail). Lambda

• RunPod model pages (A100 PCIe $1.64; A100 SXM $1.74; H100 PCIe $2.39; H200 $3.59). Runpod+3Runpod+3Runpod+3

• AWS Capacity Blocks for ML (per‑accelerator effective hourly rates: A100, H100, H200). Amazon Web Services, Inc.

• GCP Spot VMs GPU pricing (A100 80GB $1.5712; H100 $2.253–$2.3791; H200 $3.7247). Google Cloud

• Azure ND96isr H100 v5 pricing (on‑demand $98.32/instance; spot $23.488/instance). Vantage

• AWS on‑demand instance totals used to derive per‑GPU (p4de, p5.48xl). Vantage+2Cost Calculator+2

• Oracle Cloud (OCI) GPU pricing (A100 80GB $4.00; H200 $10.00 per GPU‑hr). Oracle+2Oracle+2

• Vast.ai pricing (A100 SXM4 ranges; P25 reference). Vast AI

What are realistic power price assumptions?

Lets look at the wholesale power price forecasts from the U.S. Energy Information Administration’s Short-Term Energy Outlook (October 7, 2025)

The USA WAVG wholesale power price (weighted by projected power consumption by regions in ‘25 and ‘26 respectively) is 4.4c/kWh in 2025 growing to 5c/kWh in 2026. There is variation by region, and because of the high margins, DC operators are willing to pay well above average to secure power. I’ve seen ranges of $0.07/kWh to as much as $0.27/kWh. $0.20/kWh looks to me like a realistic base case.

Stress-Testing the Downside

Lets make it really, really bad:

• The lowest quoted A100 rental rates fall 50% to 25c/hr

• Power prices surge from 20c/kWh to 60c/kWh

When the lowest quoted A100 rental pricing crashes 50% to $0.25/hour, and power prices triple, the 25MW cluster burns cash.

We need to see a catastrophic rental price collapse AND an unprecedented electricity crunch for the “obsolete” cluster to burn cash. That’s not a realistic downside case IMHO.

The “Rapid Obsolescence” Myth

Now let’s address the claim that GPUs lose value immediately and become worthless within a year or two.

The book value of the GPUs goes down 20%/year assuming 5-year depreciation.

Based on that, I estimate the BV of the A100 (2020), H100 (2022), and H200 (2024)

• A100: $3,000

• H100: $10,000

• H200: $24,000

Here’s what those ‘obsolete after a year’ GPUs actually sell for (median prices, Oct 2025):

• A100: $12,000 (4x BV)

• H100: $25,000 (2.5x BV)

• H200: $33,000 (1.4x BV)

Are the buyers of these GPUs stupid? No: they understand that these assets produce cash flow, and price their purchases accordingly.

The A100 launched five years ago. It’s “obsolete” by the standards of the bear case. On the books of a hyperscaler using 5-year depreciation, it would be worth essentially zero today.

Yet it still commands $12,000 on the secondary market, 4 times its depreciated book value.

Ask yourself a question: Why does TSMC continue to produce chips on fab assets that are fully depreciated (zero book value)? Aren’t those assets worthless?

Obviously, the answer is no, those assets are not worthless. They can continue to fab chips and generate cash far longer than their depreciation schedule suggests.

To be fair, basic physics tells us thatGPUs running hot 24/7 for 75% of the days over the past 5 years can’t be run as as hard for as long as a semiconductor fab.

The fact of the matter is that 5 year old A100 GPUs continue to be run hot and rented out for over $1.00/hr today. Time will tell how long their useful lives actually are, but claiming a 2-3 year useable life for an NVIDIA GPU is demonstrably false.

The Depreciation Red Herring

“But depreciation is understated! Profits are inflated!”

Let’s be clear about what’s actually happening.

Most GPU operators took advantage of 100% bonus depreciation and fully expensed their hardware in Year 1 for tax purposes. The large hyperscalers (AWS, Microsoft, Google) now use 5-6 year accounting lives based on actual usage data and historical experience, not arbitrary guesses.

Could they adjust these schedules? Yes. Would it matter?

Not to cash flow. Changing depreciation schedules moves numbers between accounting lines. It affects reported GAAP earnings. It doesn’t change the amount of cash the data center generates or the money available to service debt.

Depreciation provides a rule of thumb for when CAPEX needs to be spent to replace assets that have reached the end of their useful lives and cease generating cash flows. Fully depreciated assets are, in both theory and practice, fully capable of generating cash flows.

The Collateral Question

“Old GPUs make terrible collateral for loans!”

This argument misunderstands how asset-based lending works.

Lenders underwrite based on cash generation, not book values.

• Lets say you buy 52,083 A100s on the used market today. The median price is ~$12,000 = $625mm cost. Under previous assumptions the cluster generates $126mm EBITDA and $107mm cash net income = 6 year payback at rock-bottom rental rates

• Assuming a realistic rental rate of $1.00/hr, the cluster generates $383mm of EBITDA, $326mm of Cash Net Income = 1.9 years payback

Would a lender prefer brand-new hardware? YES.

Does the second-hand A100 GPU hardware using 5 year old tech at mid-range rental pricing generate observable, predictable cash flows? YES.

Observable, predictable cash flows ARE THE SAME THING AS STRONG COLLATERAL.

The Bottom Line

The “GPU depreciation bubble” argument confuses accounting presentation with the actual cash economics of the productive assets.

Here’s what the actual data shows:

• Operating costs are less than $0.20/GPU-hour

• Spot pricing runs 10-30× operating costs, even for 5-year old hardware

• Resale values for 5-year-old GPUs are 3-5× their depreciated book values

• Data centers produce substantial cash flows even in extremely pessimistic scenarios

Book depreciation is a method for allocating capital costs across time periods for financial reporting purposes.

Cash flow is economic reality. It’s the money that pays employees, covers electricity bills, services debt, and returns capital to investors.

When people warn about a “GPU bubble,” ask them which number they’re worried about: the inflated GAAP EPS, or the underlying cash flows.

Follow me on X: Superhoogie

Follow me on Linkedin

What do you think? Are concerns about GPU data center economics overblown, or are there risks this analysis doesn’t capture? Are my assumptions off by a country mile? Is my analysis sloppy? Let me know in the comments. Share your thoughts in the comments.

Especially you, Jim. Especially you.

The flywheel is spinning

Two days ago I published my research on Seikoh Giken, a Japanese fiber-optic equipment manufacturer with a dominant position in connector polishing machinery. The thesis centered on AI data center build-out driving exponential demand for fiber-optic interconnects—and Seikoh sitting at the critical chokepoint with ~45% global share of the connector polishing machinery market as the de facto global standard supplier. Key competitors are Domaille (20%), NTT Advanced Technologies (~12%), and Krell Technologies (~9%).

I’ve done deeper analysis on their Optics subsegments, which disaggregates manufacturing machinery (razors + blades), connectors/ferrules (end products), and others (consumables). What I found materially strengthens the investment case and clarifies the multi-year earnings trajectory in a way that consensus estimates appear to completely miss.

This post focuses on three things:

1. The razor-blade business model that creates predictable, high-margin cash flows

2. Why the semiannual disclosure structure gives us unusual visibility into the installed base expansion

The headline: 2H24 manufacturing machinery revenue hit ¥3,142 million (+107% YoY)—obliterating the previous all-time high by 30% and locking in subtantial stream of future consumables revenue.

Razor-Blade Economics

Seikoh Giken’s optics business operates on a classic razor-blade model, though the market treats it like a cyclical component supplier. Here’s the actual structure:

The Razor (Manufacturing Machinery):

• Fiber-optic connector polishing equipment (estimated at $250-350mm total market in 2025, CAGR-ing at 6.5% through 2030)

• Precision grinding machines

• Automated inspection systems

• Upfront capital sale: ¥50-200 million per major equipment installation

• 7-10 year useful life

• One-time revenue recognition

The Blades (Consumables & Connectors):

• Ferrules (ceramic pins that hold optical fibers)

• Polishing films (consumable abrasives for end-face finishing, estimated ~$180mm global market, growing at 7% through 2030)

• Inspection equipment upgrades

• Connector assemblies

• Recurring purchases over equipment lifetime

• 50-60% gross margins vs. 35-40% on machinery

The Economics:

For every ¥1 million of machinery placed, Seikoh generates an estimated ¥5-7 million in consumables revenue over the equipment’s 7-10 year operational life. Here’s the math assuming 7 years:

Example: Hyperscaler Data Center Equipment Placement

• Initial machinery sale: ¥150 million (polishing line + inspection systems)

• Annual consumables spend: ¥80 million

  • Ferrules: ¥40 million (continuous replacement as fiber connections scale)

  • Polishing films: ¥20 million (consumable abrasives, monthly replenishment)

  • Maintenance/upgrades: ¥20 million (software, inspection equipment, spare parts)

• Equipment life: 7years

• Total lifetime value: ¥800 million (¥150mm machinery + ¥650mm consumables)

The ¥150 million machinery placement generates ¥650 million in future consumables at 50-60% gross margins.

What the Optics subsegment data shows

FY2024 manufacturing machinery revenue: ¥5,598 million (+86.7% YoY)

For context:

• Previous peak: ¥3,504 million (FY2023)

• 11-year average (FY2014-2024): ¥2,637 million

• FY2024 is 2.1× the historical average and 60% above the previous peak

But what matters more is the composition of total revenue:

Machinery as % of Total Revenue:

• FY2014-2024 average: 18.2%

• FY2024: 28.02% (2H24: 28.8%)

• Previous high: 21.52% (FY2022)

When machinery represents 28% of revenue vs. an 18% historical norm, it signals a structural shift in the business mix—not a cyclical spike. The company is placing equipment at unprecedented rates, which means the installed base is expanding faster than at any point in the last decade

Growth Inflecting Across 1/2/3 Year Time Horizons

Here’s what separates structural demand from cyclical noise: when growth accelerates across multiple time horizons simultaneously.

FY2024 Machinery Growth Rates:

• 1-year: +86.7% (previous decade high: +36.0% in FY2021)

• 2-year: +59.8% (previous decade high: +49.1% in FY2017)

• 3-year: +74.6% (previous decade high: +68.8% in FY2018)

Furthermore, growth accelerated into 2H24, with machinery up 107% on 2H23 versus a high of ~50% in 1H21.

Semi-annual Machinery Growth Rates:

All three metrics hitting decade highs at once suggest demand driven by AI data center fiber interconnect build-out hitting Seikoh’s order book:

• Data rate transitions: 400G/800G/1.6T/3.2T deployments are accelerating

• Rack-to-rack fiber density: GPU clusters requiring 10-40× more fiber connections than traditional servers

• MPO/MTP evolution: Shift from 12-fiber to 24-fiber to 32-fiber connectors increasing ferrules per connection

• Single-mode adoption: Longer-reach interconnects requiring tighter polishing tolerances (premium equipment/consumables)

Orders Are Outpacing Revenue

FY2024 orders growth outpacing sales growth implies supply tightness and pricing power:

• Orders received: ¥21,380 million (+33.7% YoY)

• Revenue recognized: ¥19,982 million (+26.6% YoY)

• Order backlog: ¥4,384 million (rebuilt from ¥2,778mm in FY2023)

]

Orders growing faster than revenue (+34% vs. +27%) while backlog rebuilds means demand exceeds supply capacity. Despite Seikoh converting record revenue, they’re still accumulating backlog.

In a market where Seikoh holds ~40-50% share, there is a limited pool of alternative credible suppliers, customers facing lead times for high-speed optical systems, supply tightness = sustainable pricing power. This isn’t a temporary allocation squeeze—it’s structural undersupply relative to AI infrastructure demand.

The Optics vs. Precision Mix Shift

The margin story gets even better when you look at segment mix:

The high-margin segment growing 4× faster than the low-margin segment drives group-level margin expansion even before operating leverage kicks in. As optics revenue scales to 60%+ of the mix, consolidated EBIT margins expand 500bps in my base case.

Why the Semiannual Disclosures Matter

Most Japanese component manufacturers provide minimal segment disclosure—making it nearly impossible to track installed base dynamics in real-time. Seikoh is different.

What Gets Disclosed (Every Six Months)

At 1H and full-year earnings, Seikoh breaks out:

• Manufacturing machinery sales

• Connectors and ferrules sales (consumables proxy)

• Other optics revenue

• Orders received and order backlogs (consolidated, not broken down by segment)

This granularity is unusual for Japanese industrials and gives investors a real-time window into:

1. How fast the installed base is expanding (machinery sales & orders)

2. Whether consumables are inflecting (connectors/ferrules growth)

3. Pricing power dynamics (margins + backlog trends)

4. Forward visibility through order backlog

November 13 Earnings

The 1H FY2026 earnings release (scheduled for November 13, 2025) is the next major inflection point, as we will get more granular details on the Optics segment and the orerbook. Here’s what I’ll be watching:

1. Orders received ¥13Bn+ for 1H25 would imply full-year FY2025 could hit ¥26Bn+ confirming structural demand expansion at 30%+ growth rather than one-time surge.

2. Consumables growing 20%+ would confirm that FY24 machinery placements are maturing and driving attach.

3. Optics margins expanding to 32%+ would validate my assumptions about operating leverage

4. Backlog expansion to ¥5B+ would confirm demand still exceeds supply despite record conversion, increasing revenue visibility and supporting my hypothesis on pricing power.

Valuation

At the current share price, I see 200%+ upside in the bull case, and 100%+ in a base case, versus ~3% downside in my bear case. Asymetric enough for me.

I’d be keen to hear your thoughts on Seikoh Giken, especially if it is to skewer my thesis. Let me know what I’m missing here.

Subscribe now

Executive Summary

Founded in 1972, Seikoh Giken (“SG”) is the world’s leading supplier of fiber-optic connector manufacturing and testing equipment. The top five producers together hold roughly two-thirds of global market share, and SG ranks first. Its customers include Corning and Fujikura, among others, who depend on SG’s precision equipment and consumables to maintain yield, throughput, and reliability in their connector production lines.

SG began as a precision engineering firm during Japan’s 1980s–1990s electronics boom and developed the first mass-production optical-fiber polishing machine in 1987. This innovation became the global standard for connector finishing and established SG’s reputation for quality and precision. Over time, the company expanded downstream, selling consumables, ferrules, and connector accessories into the large installed base of its own machinery. This created a recurring, high-margin revenue flywheel that strengthens with each new generation of optical production tools.

SG’s competitive moat is built on three structural advantages:

• Decades-long vendor integration with every major connector manufacturer.

• High switching costs and qualification barriers within telecom and datacom production.

• Deep engineering know-how and proprietary IP accumulated through decades of precision manufacturing.

The company is entering a period of accelerating earnings growth as the AI data center build-out expands its addressable market. Two secular forces are driving this inflection:

• The rapid increase in network-fabric density within hyperscale data centers.

• The faster cadence of data-rate transitions from 400G to 800G and 1.6T, which refresh both equipment sales and accelerate the pace of consumable purchases.

The operating leverage of the Optical segment is now visible in the financials. EBIT margins declined from 14.8% in FY22 to 5.5% in FY24, but demand from AI-driven connector and testing refresh cycles lifted consolidated margins to 14.1% in FY25. In the latest quarter, sales rose 35% year over year, and segment EBIT margin reached 21% in 1Q FY26, highlighting the strong incremental margins in the Optical business.

In summary, Seikoh Giken combines dominant market share, pricing power, and an expanding total addressable market, positioning it for supernormal earnings growth over the next several years. Recent earnings performance confirms that the fundamentals are accelerating far ahead of the current consensus.

Valuation and Consensus Comparison

Consensus forecasts imply a sharp and unrealistic deceleration in AI-driven optical growth and data-rate adoption. Our model instead captures sustained fabric densification and refresh cycles, which support materially higher EPS and a likely revision upcycle over FY26–FY27. With all leading indicators of fiber optic demand inflecting higher, we see a conservative base case IRR of ~73% over the next 18 months with limited downside.

Seikoh Giken’s business

Optical Products (1Q26: 57% of sales, 83% of EBIT)
Seikoh Giken’s core growth engine is its Optical Products segment, which supplies ultra-precision ceramic ferrules, multi-fiber connectors (MPO/MTP, LC/SC), polishing systems, and consumables. The segment is deeply embedded in customer production flows because Seikoh’s proprietary ceramic processing and polishing platforms require qualification and generate recurring consumable demand. Customers include telecom carriers, data centers, and optical OEMs, so revenue is directly exposed to the expansion of hyperscale data centers, cloud infrastructure, and high-speed broadband. The result is high operating leverage and predictable attach-rate economics as connector volumes and data-rate upgrades accelerate.

Precision Parts (1Q26: 43% of sales, 17% of EBIT)
The Precision Parts segment produces high-precision molds and molded or pressed components such as insert-molded housings, thin-wall resin parts, and precision metal pieces for automotive sensors, electronics, and medical devices. The business has moved from tooling-only to turnkey manufacturing that covers design, molding, assembly, and quality control, which increases stickiness through qualification-led programs. Margins are highest in complex insert-molding for EV sensors and regulated medical production where switching costs and validation barriers are steep. Key risks are exposure to cyclical smartphone and auto demand and commoditization in lower-spec parts.

Investment Highlights

1. Dominant market leader with entrenched market position operating a razor & blade business model enabling AI data center networking connectivity.

   SG faces limited global competition at the top-end where it operates due to its reputation for reliability, quality, and innovation. By providing an integrated system of machines, fixtures, and consumables that is known to produce consistent, high-yield results, making customers willing to pay a premium and discourages them from experimenting with unproven, lower-cost alternatives. This is especially important in a period of unprecedented demand, where no manufacturer can risk operational disruption.

2. Forward deployed production model makes switching costs prohibitively high.

   When a customer deploys SG production/testing equipment, they also need to buy an array of proprietary polishing fixtures specific to different connector types (LC, SC, MPO) as well as polish types (PC, APC). Switching to a competitor’s consumables or fixtures invalidates the customer’s production process, and after-service agreement. Finally, production technicians are trained and certified on the SG product ecosystem - besides retooling the production line, the customer would have to retrain the entire workforce, and then requalify the product with end-customers. That is a 3-6 month process for commodity products, 6-12 months for leading edge data center products.

3. Powerful economic flywheel: Largest install base > integration into customer processes > unparalleled process knowledge > superior R&D > expand technological edge > grow install base.

   SG is deeply integrated into the production processes closely with leading global fiber optic cable manufacturers SG has unique insight in the evolving requirements and manufacturing challenges in the sector long before its competitors do, which combined with high margin, recurring consumables sales to those customers enables the funding of R&D into solving customer pain points well ahead of its competitors. To wit, SG has been first to market with solutions in recent years that directly address AI data center customer needs: control solutions for multi-fiber connectors (MFC), fixtures for Very Small Form Factor (VSFF) connectors, and One-Lever MT fixture. This entrenches SG’s technological advantage and grows the install base.

4. SG’s end markets have superlinear leverage to AI data center growth.

   We see a structural, superlinear uplift in Seikoh Giken’s addressable volumes as AI campuses move to 1 GW scale. The combination of (1) fabric-driven link growth exceeding rack growth, (2) MPO densification at higher speeds, and (3) recurring polishing & MAC work creates a volume flywheel in connector polishers, consumables, and ferrules that should compound ahead of simple AI DC GW connections or rack count growth curves.

   

Key Implications for revenue outlook

Network fabric complexity increases volumes: Denser optics mix of AI data centers pushes roughly 20x more polishing consumables and 20-40x more connectors/ferrules. “Scaling up” by adding more GPUs into a rack increase per-rack optical density. “Scale out” from 100MW pods to 1GW clusters doesn’t just add more racks, but creates an exponential rise in east-west traffic across pods/zones, and the high density cabling, MPO connectors, and polishing cycles.

Strict tech requirements drive volume, ASP, and mix: Higher data transfer rates (ie transition from 400G>800G) demand much more precise connector and endface specs (lower signal loss, backreflection, and dust particle contamination), which translates into more polishing cycles per physical connection (consumables sales), and more high end (precise geometry) multi-fiber ferrules (product sales). Multi-row MPO (e.g., 16/24/32-fiber) and “low-loss” ferrules command higher ASPs; fixtures and polishing recipes become SKU-richer.

Faster pace of network reconfigurations drives volumes further: AI data centers engage in significantly more network moves/adds/changes (MAC) than traditional data centers, which also drives demand for polishing consumables and connectors on a recurring basis

Data rate transitions are a long term secular tailwind: Faster, overlapping transitions (400G → 800G → 1.6T → 3.2T) make links, terminations, and MAC activity grow faster than rack or MW adds, and they tighten optical tolerances. For Seikoh Giken (polishers, fixtures/consumables, connectors/ferrules), that means more units, more consumables per unit, and better mix. The below chart is from Coherent’s investor day, based on internal data and consultancy LightCounting.

Implications for SG

1. Rapid cadence of leading edge tech adoption: The adoption waves for the leading edge data rates overlap substantially in 2025 an onwards, so we should not expect any digestion lull between stages. 800G has been ramping while 400G is still large, and and the 1.6T rollout has already started (obviously led by AI data center demand) before 800G has even peaked! Meanwhile, 3.2T adoption is expected to commence in CY26. It is reasonable to assume that this overlap in adoption cycles implies a sustained cadence of continuous install + retrofit + MAC work, which drives steady polishing/connector churn. Compare this to the stop-start cycles observed in other industries, like SPE/WFE.

2. Step-up in network fabric density: Higher data transfer rates themselves drive demand more parallel fiber & denser connectors. Inside the data hall, short-reach links favor parallel optics (MPO/MTP) and multi-fiber ferrules (MT-type). Port-for-port, more ferrules and more endfaces need polishing (1µm aluminum oxide film: ~$18 for 10 sheets) and finishing (0.01µm “final polish” films: ~$47 for 5 sheets).

Key Risks

• Hyperscaler Demand Timing / Slippage

  • Risk: The thesis assumes the pace of AI-related capex continues apace, driving disproportionate optical interconnect volume. If spending slows due to macro, supply chain bottlenecks, or internal reprioritization, unit growth could undershoot forecasts for a year or more.

  • Mitigant: Hyperscaler network investment is not discretionary in the long run. Data traffic, AI workload scaling, and latency requirements mean optical interconnect upgrades must happen eventually — delays affect timing, not direction.

  • Historical precedent: Hyperscaler pauses in 2018–19 and 2022–23 were followed by catch-up capex bursts that overcompensated, which can be seen in SG’s Optical segment performance in FY23-FY25 in the table above. Seikoh’s broad product coverage (connectors, ferrules, polishers) positions it to capture pent-up demand when cycles resume, limiting the long-term impact of a slip.

• Technology Substitution

  • Risk: New architectures (e.g., silicon photonics, active optical cables, on-board optics) could bypass traditional ferrules/connectors. This undermines the thesis that Seikoh’s high-density interconnects deliver structural margin lift.

  • Mitigant: In practice, even advanced optical technologies still require connectorized interfaces for interoperability, field repair, and standards compliance. The inertia of global standards bodies and multi-vendor qualification processes makes wholesale displacement improbable within a 5–7 year horizon. Moreover, the industry shift toward high-density MPO/MT arrays aligns directly with Seikoh’s product portfolio. Substitution risk is more about gradual mix shifts than sudden obsolescence, as illustrated in the following chart from Coherent’s most recent investor day presentation. CPO is indeed coming, but the ramp will be gradual while pluggable connectivity continues to grow rapidly.

• Commoditization

  • Risk: Price pressure from low-cost Asian competitors could erode Seikoh’s share and compress ASPs.

  • Mitigant: In hyperscale data centers, the cost of optical connectivity is trivial relative to the cost of servers and downtime risk. A single rack can carry millions of dollars of GPUs, so project managers prioritize proven reliability, quality assurance, and service capability over marginal savings. Switching to an untested supplier for a few percent lower ASP is economically irrational when even minor connector failures can cause outages, troubleshooting costs, and reputational risk. This makes Seikoh’s qualification history, installed base, and field-service credibility a far stronger moat than headline pricing.

• Manufacturing & Execution (Ramp, Yields, Geopolitics)

  • Risk: Rapid scale-up in Dalian, Hangzhou, and Thailand raises risk of yield losses, capacity coming online too slowly, or geopolitical disruptions (China tariffs, logistics shocks). Poor execution could turn top-line leverage into margin erosion if unit costs rise faster than volumes.

  • Mitigant: Optical components are qualification-intensive — once a customer validates a ferrule or connector line, they rarely requalify alternative suppliers unless forced. That makes demand sticky once Seikoh clears ramp curves. The company has also built geographic redundancy, reducing single-point failure risk. Finally, its automation investments structurally lower execution risk by improving yields and throughput. This makes the probability of execution failures persisting long enough to derail the investment case relatively low.

Substantial Insider Ownership

Masatoshi Ueno (former CEO/Chairman), Jun Ueno (current CEO), and Tamotsu Kimura (director) together hold 9.74% + 3.89% + 6.49% = 20.12%. Adding the founder family vehicles (7.28% + 6.49% = 13.77%) brings insiders + founding family to ≈33.89% of outstanding shares.

Conclusion

Seikoh Giken is uniquely positioned at the intersection of three powerful, durable tailwinds reshaping global optical demand. First, AI data centers are driving unprecedented fiber-optic density as rack-to-rack and pod-to-pod interconnect complexity compounds faster than physical rack growth. Second, the cadence of data-rate transitions—from 400G to 800G, 1.6T, and soon 3.2T—is accelerating, forcing continuous equipment refresh and higher-spec connector demand.

Together, these structural forces create a multi-year, non-cyclical expansion in Seikoh Giken’s revenue and margin opportunity. The firm’s installed base, razor-and-blade economics, and entrenched vendor relationships make it the de-facto standard in fiber-connector precision tooling.

Consensus estimates implicitly assume a rapid and unrealistic deceleration in AI data center fiber demand and a slowdown in the cadence of data transfer technology adoption. That assumption is inconsistent with observed hyperscaler capex trends and the overlapping rollouts of 400G, 800G, 1.6T, and early 3.2T deployments. My model instead assumes sustained fabric densification and continuous refresh cycles that drive materially higher connector, ferrule, and consumable volumes. The result is a clear EPS revision upcycle rather than the sharp slowdown implied by consensus.

The path to upside combines two compounding effects. First, higher than consensus EPS as attach rates and consumables intensity normalize above the street. Second, a multiple rerating as the market recognizes durable, superlinear exposure to AI fabric growth and upgrades. Even my bull case is likely conservative given the recent pace of hyperscaler announcements. U.S. telco fiber capex is an incremental upside that we have not baked into base case estimates.

Key Modeling Assumptions

Though modeling 35% and 55% incremental margins for Precision and Optics may feel like aggressive assumptions, if one looks at the last 5 reported quarters in the table below, the hefty and stable incremental margins of both business units stands out. The quarterly revenue growth assumptions through end of current fiscal year assume an 800bp deceleration in the base case, and only a 200bp acceleration in the bull case. Going forward I project a tapering off in growth rates through 2030, but if we continue to see AI Hyperscaler capex estimates rise, these growth rates will certainly become unrealisiticaly conservative.

Base Case Output

Bull Case Output