Ari

Over-analyzing the inference economy.

Soft spot for great literature, grand historical schemes, and demanding cats.

[Ari @arithesis]

Nittobo ripping +20% in two days since the write-up

[Ari @arithesis]

I pointed Fable at the AI supply-chain trade. 12 hours, 5 agents, ~100 subagents each. CCL, substrates, PCBs, packaging, glass-core, the whole bottleneck map.

one name kept coming back as mispriced.

Nittobo. 3110.T.

it's the bottleneck under the bottleneck everyone's already trading.

it owns ~90% of high-end T-glass, the low-CTE cloth that keeps big AI substrates from warping and failing. bigger dies, denser boards, more layers means a lot more of it.

here's the part that's odd.

look who went vertical in this exact chain:
- Unimicron (ABF substrate): ~10x in a year, P/E ~125
- Nan Ya PCB: P/E ~300
- Elite Material (CCL): parabolic

every one of them buys Nittobo's glass. they're the tenants. Nittobo is the landlord.

the market bid the tenants to 125-300x earnings, and left the landlord down 37% off its highs.

the supplier with the harder moat and the pricing power is the cheap one. that's backwards.

why it wins either way:
CPO delayed? copper and organic substrates live longer, so more glass.
CPO ships? optics sit next to a hot, warping package, materials get harder, and it still needs the glass.

heads it works, tails it works.

and the shortage is mechanical, not a vibe: no new T-glass capacity, Nittobo's or anyone's, comes online before mid-2027. fixed supply against $725B of locked, supply-constrained 2026 hyperscaler capex. it can't break this year.

the numbers:
revenue +8%, operating profit +27%, and the AI glass segment did +40% profit on +20% revenue. that's pricing power. that segment is ~84% of operating profit, so the "boring conglomerate" take doesn't hold. forward guide: rev +16%, OP +25%, accelerating.

it's FCF-negative on purpose: ¥120B going into tripling capacity, METI-subsidized, with customers offering to help fund it. a monopoly spending into its own sold-out shortage is about as bullish a tell as you get.

valuation, straight:
not a meme steal. normalized it's ~mid-30s P/E. but that's cheap for a 90% monopoly growing OP 25% in a locked shortage, and a fraction of the 125-300x you'd pay for its own customers.

the target:
consensus is ¥27,144, which is +48% from here just to catch up to where the street already has it. retest the old high and it's +80%. you're buying it down 37%.

the fine print (both sides):
real risk is 2027, second-source qualification (Taiwan Glass). telegraphed and watchable. glass-core/CoPoS displacement is a 2029+ story.
and the August print will show net income "down 59%". that's just last year's one-time asset gain rolling off while OP grows 25%. if the tape sells it, that's the entry.

bottom line:
retail found the trade. they bought every layer except the one that taxes the whole stack.

Nittobo sells the glass to all of them. 90% share, spending into scarcity, down 37%, with +48% to consensus just to catch up.

curious if anyone sees a hole here.

3110.T / NBCLF

[Ari @arithesis]

lite-on looks like the cheapest credible way to play the inside-rack 800v pull-forward.

important distinction: this is not the same trade as $wolf.

$wolf is the layer 3 / grid-to-rack bet: higher-voltage sic, solid-state transformers, medium-voltage straight to 800vdc. bigger upside if the sockets land, more monopoly-like if it wins, but later and less proven. real ramp is more 2027+.

lite-on is the nearer-term rack layer.

power shelves. bbus. sidecars. rack power systems.

that part is already moving.

nvidia is also clearly multi-sourcing this stack: delta, lite-on, flex, megmeet, etc. they don’t want one supplier becoming the bottleneck or taking all the rent.

so this isn’t “who owns 800v?”

it’s “who is credible, cheap, and close enough to the ramp to get repriced?”

lite-on has the ingredients:

- 110kw power shelves already in mass production
800v rack volume late-2026 / q1’27
- two us csp orders
- bbu capacity tight, lines going from 8 to 12
- ai moving toward ~30% of sales

delta is cleaner and obv deserves the premium, but it also trades closer to 50x+.

lite-on is messier, but trades closer to 24x.

dilution check is clean too: share count basically flat for years, restricted shares cancelled, and the planned nt$8b convertible was revoked.

the bear case is simple: this is not a monopoly, margins are weaker than delta, consumer still drags, and if july doesn’t show margin stabilization the re-rate waits.

if ai power demand holds and margins stop leaking, the multiple probably changes.

right now lite-on is still priced like a messy electronics supplier.

the setup says it should start getting valued like one of the real rack-power names in the 800v buildout.

$2301.TW $LITEON

[Ari @arithesis]

if 800V HVDC is really pulling into Q3 2026, $FLEX is one of the benefactors.

it's one of only two credible grid-to-chip full stacks (per TrendForce), and it benefits two ways:
1. it builds the 800V power shelves
2. it owns the grey-space layer above them: switchgear, busways, distribution via Anord Mardix.

that Critical Power business already grew ~38% and is guided +65-75% standalone.

[Serenity @aleabitoreddit]

$NVDA and $GOOGL lead 800V DC ahead of schedule.

"Ahead of schedule", pulled up to Q3 2026 with small volume shipments starting .

- Delta Electronics (2308), $VRT
- Song Chuan Precision (7788)
- Schneider Electric, Eaton, Siemens.

All flagged as beneficiaries.

"Market sources indicate that Nvidia’s Vera Rubin platform and Google’s next-generation AI data centers will be the first to adopt the technology"

Source: Commercial Times

The power semi trade should be happy to hear this.

[Ari @arithesis]

Yeah I dug into this. Management not hiking up prices but fully committing to expansion is a smart move and obviously defensive against other players. I forecasted assuming the 2027 risk of a second source qualification, and even accounting for that, nittobo is still currently cheap

[Ari @arithesis]

I pointed Fable at the AI supply-chain trade. 12 hours, 5 agents, ~100 subagents each. CCL, substrates, PCBs, packaging, glass-core, the whole bottleneck map.

one name kept coming back as mispriced.

Nittobo. 3110.T.

it's the bottleneck under the bottleneck everyone's already trading.

it owns ~90% of high-end T-glass, the low-CTE cloth that keeps big AI substrates from warping and failing. bigger dies, denser boards, more layers means a lot more of it.

here's the part that's odd.

look who went vertical in this exact chain:
- Unimicron (ABF substrate): ~10x in a year, P/E ~125
- Nan Ya PCB: P/E ~300
- Elite Material (CCL): parabolic

every one of them buys Nittobo's glass. they're the tenants. Nittobo is the landlord.

the market bid the tenants to 125-300x earnings, and left the landlord down 37% off its highs.

the supplier with the harder moat and the pricing power is the cheap one. that's backwards.

why it wins either way:
CPO delayed? copper and organic substrates live longer, so more glass.
CPO ships? optics sit next to a hot, warping package, materials get harder, and it still needs the glass.

heads it works, tails it works.

and the shortage is mechanical, not a vibe: no new T-glass capacity, Nittobo's or anyone's, comes online before mid-2027. fixed supply against $725B of locked, supply-constrained 2026 hyperscaler capex. it can't break this year.

the numbers:
revenue +8%, operating profit +27%, and the AI glass segment did +40% profit on +20% revenue. that's pricing power. that segment is ~84% of operating profit, so the "boring conglomerate" take doesn't hold. forward guide: rev +16%, OP +25%, accelerating.

it's FCF-negative on purpose: ¥120B going into tripling capacity, METI-subsidized, with customers offering to help fund it. a monopoly spending into its own sold-out shortage is about as bullish a tell as you get.

valuation, straight:
not a meme steal. normalized it's ~mid-30s P/E. but that's cheap for a 90% monopoly growing OP 25% in a locked shortage, and a fraction of the 125-300x you'd pay for its own customers.

the target:
consensus is ¥27,144, which is +48% from here just to catch up to where the street already has it. retest the old high and it's +80%. you're buying it down 37%.

the fine print (both sides):
real risk is 2027, second-source qualification (Taiwan Glass). telegraphed and watchable. glass-core/CoPoS displacement is a 2029+ story.
and the August print will show net income "down 59%". that's just last year's one-time asset gain rolling off while OP grows 25%. if the tape sells it, that's the entry.

bottom line:
retail found the trade. they bought every layer except the one that taxes the whole stack.

Nittobo sells the glass to all of them. 90% share, spending into scarcity, down 37%, with +48% to consensus just to catch up.

curious if anyone sees a hole here.

3110.T / NBCLF

[Ari @arithesis]

Oh i use AI in everything I do, from analysis to writing to back testing, its ability to supercharge my workflow is why I'm so bullish that we are only in the beginning of the AI infra super cycle

If you are calling it slop because of the writing style, i can't care less.

If you've got actual angles on aaoi that's worth a look, feel free to share

[Ari @arithesis]

AAOI ripped 10× from $15 to $233 on the perfect AI optics shortage story.

But its own earnings is calling the entire narrative bullshit.

Longs need flawless execution quarter after quarter plus a world that stays friendly. Bears only need one thing: hope to finally run dry.

And hope is always loudest right before it dies.

AAOI is screaming its head off.

The 10x that the company's own numbers don't believe
$15 to $233 on the hottest story in hardware: an 800G/1.6T shortage 40-60% undersupplied through 2027, datacenter revenue +154%, $1.1B FY26 guide, and “China InP ban equals US fab jackpot.”

Then you check the one number that actually matters. At the absolute peak of the scarcity every bull is high on:

- Lumentum gross margin: +1,270 bps
- Coherent: +105 bps
- AAOI: -150 bps, down to 29.1%

The grown-ups raised prices and printed fatter margins in the best market they will ever see. AAOI’s margin went backwards. They are not capturing the bottleneck. They are the sweatiest guy inside it — selling everything they can make at prices other people set, while the market pays them like they own the choke point.

A real bottleneck raises price into scarcity. AAOI couldn’t lift a single basis point during the greatest optics squeeze in history. That is not a moat. That is a tell.

The "bullish" China ban is loading the other guy's gun
The InP export control is supposed to bless US laser fabs. Except AAOI’s lasers come off an outdated small-wafer line in Texas, a full generation behind Coherent, so domestic Chinese substrate access does nothing for them.

Meanwhile that same loophole exempts the actual killers. InnoLight and Eoptolink — 60% of global 800G between them, $3B+ scale, 20%+ net margins — get cheap permit-free substrate at home and quietly reload for the 2027 price war. InnoLight ships more in a month than AAOI ships in a year. Beijing didn’t build AAOI a wall. It handed its competitors ammunition and a head start.

A trapdoor with a rug on it
- 98% of revenue from the top 10 customers. Microsoft alone ~29%.
- Purchase orders, not contracts. No demand floor. Amazon even pays partly in dilution.
- We have seen this exact movie. 2018: the top customer blinked, revenue went $330M to $224M, stock went $100 to under $10.

No pricing power and no committed demand. That is the financial equivalent of standing on a rug over an open elevator shaft and admiring the view.

So is it about to pop?
Honestly, probably not tomorrow. This is the part people get wrong about shorting bubbles: being right early is just being wrong with extra steps. The shortage has real legs into 2027. AAOI keeps beating and raising, the headlines keep landing ("800G is now our biggest segment"), and a stock like this can stay irrational longer than the impatient can stay solvent. Fighting the melt-up while the squeeze is still on is how shorts become the bagholders.

So near-term, the air is still in it. The hype has fuel left.

So when does it actually pop?
When the scarcity that's holding the whole thing up goes away. That's not a vibe, it's a schedule. The flood of new supply, including AAOI's own 930k-modules-a-month of 2027 capacity and the substrate-free Chinese giants ramping into the same window, all lands in 2027. The moment 800G stops being short, a company with no pricing power and no contracts has nothing left to stand on.

The tells that the air is leaving, in order:
- Chinese 1.6T price cuts from InnoLight or Eoptolink. That's the war starting.
- AAOI's gross margin cracking below 28% while datacenter mix rises. That's the price-taker getting confirmed in the numbers.
- 800G spot premiums collapsing under 10%. That's the shortage ending in real time.

And the things that would mean the bears are wrong, worth watching honestly: a real equity or long-term-agreement deal from a hyperscaler, the way NVIDIA put $2B each into Lumentum and Coherent, or margins that somehow sustain above 33%. Either of those rewrites the story, and you'd have to respect it.

This isn't a bet that the company is a fraud. It's a read that the story sprinted miles ahead of the cash flows, and gravity tends to show up the moment supply normalizes.

The bull case, steelmanned
To be fair, the bulls aren't delusional. Demand genuinely exceeds what AAOI can build, the Texas ramp is executing faster than skeptics expected, and management is guiding gross margin back to ~35% by end of 2026 and 40%+ in 2027 as 800G/1.6T takes over the mix. If they hit that, my single best line, that they can't raise price, gets a lot weaker. And the US-fab angle is real: hyperscalers want to diversify off Chinese supply, and if AAOI lands an equity or long-term-agreement deal like the ones NVIDIA gave Lumentum and Coherent, the concentration and no-contract risks get fixed overnight.

I take both seriously enough to make them my exit triggers, not footnotes. A sustained gross margin above 33%, or a real hyperscaler LTA, and I'm wrong and I'm out. The bear case here isn't "the company is bad." It's "one quarter of margins falling at peak scarcity is the tell, and they have to disprove it." Until they do, the burden of proof is on the story, not the skeptic.

Bottom line
AAOI is the rare setup where all three legs of a perfect short line up at once:

1. No pricing power. At the peak of the greatest optics shortage in history, its margin went down while Lumentum's and Coherent's went up. If you can't raise price in this market, you never will.
2. The "tailwind" is actually a headwind. The China InP control everyone cites as bullish does nothing for AAOI's outdated Texas fab and quietly arms its substrate-free Chinese competitors for the 2027 price war. There's no cavalry coming. The thing they're cheering is the thing that kills them.
3. No floor under the demand. 98% of revenue in ten customers, ~29% in one, all on purchase orders with no commitments. We've seen exactly how this ends: 2018, top customer blinked, $100 to under $10.

AAOI is a price-taker with no moat, no savior, and no demand floor — trading at a choke-point multiple while flying on vibes. It is not a choke point. It is just busy. And busy is not a multiple.

It's not a chokepoint. It's just busy. And busy is not a multiple.

[Ari @arithesis]

@NotJustAnyWhale If you've actually read it, you'd realize I've no plans to short it any time soon

[Ari @arithesis]

the cheapest name in AI infrastructure is also the one that just built the first optical rack. 15x, 62% growth, and the future demo'd back in March.

can you guess the company?

[Ari @arithesis]

@KawzInvests the anthropic share is estimated to be ~0.5%, not 0.25% like some have quoted, so you're basically getting a $5B stake of anthropic for free

https://t.co/r2mDEsN9QS

[Ari @arithesis]

$SKM: "the cleanest pre-IPO Anthropic play." I traced the actual share count. might be nearly double what everyone quotes.

Anthropic IPOs within months. betting markets give 59% odds it opens above $1.8 trillion. you can get in via korea's biggest phone company, trading at $39.

the issue: no one agrees how much of Anthropic it actually owns. english press says 0.3%. korean brokers say 0.58%. X says up to 1%. the company itself won't say (NDA with Anthropic).

so I pulled the korean filings. the debate resolves up, not down. and at $39, the stake is priced at zero anyway.

the filed number
korean annual reports include a mandatory investments table that the 16-page english summary omits. pages 441-442 of the 442-page korean original:

3,703,141 shares. 0.7% of Anthropic as of year-end 2024. carrying value $132M.

and the math checks itself: $100M / 3,703,141 = exactly $27 a share. at the $5B valuation that's 185.2M total shares, making SKT's stake exactly 2.0%, matching their original "2%" language to the decimal. you can't fake a number that lines up three ways.

the dilution, traced round by round
SKT's shares never shrink. the percentage dilutes as Anthropic issues new stock:
- aug 2023: 2.0% ($100M at $5B)
- YE 2024: 0.7% (the filed number. the amazon and google mega-rounds did the damage)
- sep 2025, series F ($13B at $183B): ~0.65%
- feb 2026, series G ($30B at $380B): ~0.60%
- spring 2026, employee tender + amazon $5B + google $10B equity: ~0.54%
- series H ($50B new at $965B post): ~0.51% basic today

korean sell-side (Korea Investment, Meritz) independently lands at ~0.58% pre-series-H

so where did 0.3% come from?
a phantom denominator. PitchBook carries 1.47 billion "fully diluted" shares, a count that includes option pool that was authorized but never granted. shares that don't exist can't dilute.

3.7M / 1.47B = 0.25%, the english press picked it up after the auditor switched disclosure conventions mid-year, and it stuck around. the share count never changed. the presentation did.

the honest range today is 0.40% to 0.51% -- nearly double the consensus figure. the august DART filing and Anthropic's public S-1 will both print the real denominator, both to arrive before the IPO.

there's another detail the 0.3% camp missed entirely: on june 10, SKT confirmed it made an additional investment in Anthropic's latest round, size undisclosed.

what you pay for the stake today: $0
SKM's market cap is $15.0B at $39. the business excluding Anthropic is worth $16-17.7B on conservative broker math, and closer to $20B on a DCF.

This is korea's #1 carrier, with 31.8M subscribers, $4B+ of annual EBITDA, and a finished 5G capex cycle. plus an AI datacenter unit growing 89% year over year at full capacity, + a venture book holding Rebellions, Perplexity, and Persona.

basically, everything except Anthropic already covers the entire market cap. the 3.7 million Anthropic shares are priced at less than zero.

and this isn't a distressed company being valued for parts. the 2025 data breach that made the stock cheap is healing, with the first operating-profit quarter above ₩500B in a year, 210K subscribers returning, the dividend reinstated at roughly 3.5%, and KKR buying 49% of the datacenter unit at a ₩2-2.5T valuation, expected to sign around june 30.

the model, and please read this part before the numbers
every target below applies a 30% haircut to the Anthropic stake. that's deliberate. korean conglomerates historically get credited for only 50-70% of the assets they hold (the "korea holdco discount"), so the skepticism is built in up front. it also means that if you recompute these numbers without the haircut, you'll get bigger targets than mine, not smaller.

is 30% even the right haircut? sister company SK Square trades at a 45% discount today, so the bears have a real reference point. on the other side, the discount regime is being dismantled in real time (the fiduciary-duty amendment, the Value-Up program, a KOSPI up 80% this year), and the moment Anthropic becomes a publicly quoted stock, the opacity that justifies half the discount disappears. I use 30% to err on the harsh side, on purpose.

the ladder, on the central ~0.45% share basis. punitive 45% haircut | my 30% base | gross:
$1.8T: $55 | $58 | $64
$2.0T: $57 | $60 | $67
$2.2T: $58 | $61 | $69
$2.5T: $60 | $64 | $72

(on the filed 0.51% basis, add $2-3 to every cell; under the phantom 0.25% math, subtract $6-9. that spread is exactly why the august filing matters: it collapses it.)

the stock trades at $39. the worst cell on that ladder, a punitive discount at the lower side of the betting markets' range, still reads $55. the pessimist's view of this model is still +42%.

honest read
- the korea discount could stay stubborn longer than the reform headlines suggest. that's why every number above already absorbs 30%
- management says it won't sell the stake. the NAV re-rates anyway once Anthropic trades daily, but how value reaches shareholders remains the long-game question
- the street is either asleep or hostile: goldman carries a sell rating and models none of this, and the consensus target of $31 sits below the market price. either the telco analysts are right, or they haven't read page 441 of a korean filing
- the stacked worst case (IPO pulled, phantom share math, a KOSPI unwind) is roughly $22-26, about -40%. it takes four independent failures to get there, and the dividend would yield over 6% at those prices
- the stock has already doubled on anticipation, so the IPO itself could trade as a sell-the-news event even in the winning scenario

what's next
- in two days: the SpaceX IPO - is it going to run hot?
- ~june 30: the KKR datacenter deal signs, hardening the cash floor
- ~august: the H1 DART filing, with an updated share count that includes the june add-on. the most important document in this thesis
- Q3: Anthropic's public S-1, will disclose official total shares
- oct-dec: the IPO itself. the markets say 81% by october, with 59% odds of opening above $1.8T

bottom line
if Anthropic opens at $1.8-2T, where the betting markets have it, SKM is worth $58-60 after a harsh 30% discount. that's +50%, and it's the boring outcome. every 10% Anthropic runs above $2T stacks another ~5% on top, with a 3.5% dividend while you wait.

august prints the real number. until then: this might be a free stake in the fastest-growing company in history.

[Ari @arithesis]

@saso_capital's whole argument is that the anthropic value can transfer entirely between chaebol companies without a dime reaching shareholders. Let's unpack that.

Could the group move the stake internally? Technically yes. But SKM is NYSE-listed. Related-party transfers require SEC disclosure, independent board approval, and KFTC review. Korea's 2024 fiduciary reforms added personal liability for directors who sign off on value extraction.

Moving a $5B+ asset between listed affiliates would be front-page news. Samsung C&T is the worst-case example anyone can cite. It was messy and litigated, and the discount still settled at 40-50%, not 100%.

At $39, the Anthropic stake is priced below zero. Even if only half the value passes through (via dividends, NAV re-rating, or monetization), that's $10-15/share on a $1.8T IPO

[Ari @arithesis]

$SKM: "the cleanest pre-IPO Anthropic play." I traced the actual share count. might be nearly double what everyone quotes.

Anthropic IPOs within months. betting markets give 59% odds it opens above $1.8 trillion. you can get in via korea's biggest phone company, trading at $39.

the issue: no one agrees how much of Anthropic it actually owns. english press says 0.3%. korean brokers say 0.58%. X says up to 1%. the company itself won't say (NDA with Anthropic).

so I pulled the korean filings. the debate resolves up, not down. and at $39, the stake is priced at zero anyway.

the filed number
korean annual reports include a mandatory investments table that the 16-page english summary omits. pages 441-442 of the 442-page korean original:

3,703,141 shares. 0.7% of Anthropic as of year-end 2024. carrying value $132M.

and the math checks itself: $100M / 3,703,141 = exactly $27 a share. at the $5B valuation that's 185.2M total shares, making SKT's stake exactly 2.0%, matching their original "2%" language to the decimal. you can't fake a number that lines up three ways.

the dilution, traced round by round
SKT's shares never shrink. the percentage dilutes as Anthropic issues new stock:
- aug 2023: 2.0% ($100M at $5B)
- YE 2024: 0.7% (the filed number. the amazon and google mega-rounds did the damage)
- sep 2025, series F ($13B at $183B): ~0.65%
- feb 2026, series G ($30B at $380B): ~0.60%
- spring 2026, employee tender + amazon $5B + google $10B equity: ~0.54%
- series H ($50B new at $965B post): ~0.51% basic today

korean sell-side (Korea Investment, Meritz) independently lands at ~0.58% pre-series-H

so where did 0.3% come from?
a phantom denominator. PitchBook carries 1.47 billion "fully diluted" shares, a count that includes option pool that was authorized but never granted. shares that don't exist can't dilute.

3.7M / 1.47B = 0.25%, the english press picked it up after the auditor switched disclosure conventions mid-year, and it stuck around. the share count never changed. the presentation did.

the honest range today is 0.40% to 0.51% -- nearly double the consensus figure. the august DART filing and Anthropic's public S-1 will both print the real denominator, both to arrive before the IPO.

there's another detail the 0.3% camp missed entirely: on june 10, SKT confirmed it made an additional investment in Anthropic's latest round, size undisclosed.

what you pay for the stake today: $0
SKM's market cap is $15.0B at $39. the business excluding Anthropic is worth $16-17.7B on conservative broker math, and closer to $20B on a DCF.

This is korea's #1 carrier, with 31.8M subscribers, $4B+ of annual EBITDA, and a finished 5G capex cycle. plus an AI datacenter unit growing 89% year over year at full capacity, + a venture book holding Rebellions, Perplexity, and Persona.

basically, everything except Anthropic already covers the entire market cap. the 3.7 million Anthropic shares are priced at less than zero.

and this isn't a distressed company being valued for parts. the 2025 data breach that made the stock cheap is healing, with the first operating-profit quarter above ₩500B in a year, 210K subscribers returning, the dividend reinstated at roughly 3.5%, and KKR buying 49% of the datacenter unit at a ₩2-2.5T valuation, expected to sign around june 30.

the model, and please read this part before the numbers
every target below applies a 30% haircut to the Anthropic stake. that's deliberate. korean conglomerates historically get credited for only 50-70% of the assets they hold (the "korea holdco discount"), so the skepticism is built in up front. it also means that if you recompute these numbers without the haircut, you'll get bigger targets than mine, not smaller.

is 30% even the right haircut? sister company SK Square trades at a 45% discount today, so the bears have a real reference point. on the other side, the discount regime is being dismantled in real time (the fiduciary-duty amendment, the Value-Up program, a KOSPI up 80% this year), and the moment Anthropic becomes a publicly quoted stock, the opacity that justifies half the discount disappears. I use 30% to err on the harsh side, on purpose.

the ladder, on the central ~0.45% share basis. punitive 45% haircut | my 30% base | gross:
$1.8T: $55 | $58 | $64
$2.0T: $57 | $60 | $67
$2.2T: $58 | $61 | $69
$2.5T: $60 | $64 | $72

(on the filed 0.51% basis, add $2-3 to every cell; under the phantom 0.25% math, subtract $6-9. that spread is exactly why the august filing matters: it collapses it.)

the stock trades at $39. the worst cell on that ladder, a punitive discount at the lower side of the betting markets' range, still reads $55. the pessimist's view of this model is still +42%.

honest read
- the korea discount could stay stubborn longer than the reform headlines suggest. that's why every number above already absorbs 30%
- management says it won't sell the stake. the NAV re-rates anyway once Anthropic trades daily, but how value reaches shareholders remains the long-game question
- the street is either asleep or hostile: goldman carries a sell rating and models none of this, and the consensus target of $31 sits below the market price. either the telco analysts are right, or they haven't read page 441 of a korean filing
- the stacked worst case (IPO pulled, phantom share math, a KOSPI unwind) is roughly $22-26, about -40%. it takes four independent failures to get there, and the dividend would yield over 6% at those prices
- the stock has already doubled on anticipation, so the IPO itself could trade as a sell-the-news event even in the winning scenario

what's next
- in two days: the SpaceX IPO - is it going to run hot?
- ~june 30: the KKR datacenter deal signs, hardening the cash floor
- ~august: the H1 DART filing, with an updated share count that includes the june add-on. the most important document in this thesis
- Q3: Anthropic's public S-1, will disclose official total shares
- oct-dec: the IPO itself. the markets say 81% by october, with 59% odds of opening above $1.8T

bottom line
if Anthropic opens at $1.8-2T, where the betting markets have it, SKM is worth $58-60 after a harsh 30% discount. that's +50%, and it's the boring outcome. every 10% Anthropic runs above $2T stacks another ~5% on top, with a 3.5% dividend while you wait.

august prints the real number. until then: this might be a free stake in the fastest-growing company in history.

[Ari @arithesis]

@saso_capital Fair concerns, but the regulations make this much harder than the old chaebol playbook.

SKT is NYSE-listed. Every related-party deal has to be disclosed to the SEC in its annual filings, and anything misleading opens it to US securities litigation. You can't quietly move a $4-5B asset through that.

Korea changed too. The KFTC polices intra-group transfers, and it has fined SK before for exactly the inflated-contract scheme you describe (2012, SK C&C). Since July 2025, directors are also personally liable to shareholders under the amended Commercial Act. Signing off on value extraction is now a career-ending legal risk, not a favor to the chairman.

And the structure itself protects you. SK Inc owns ~30% of SKT, minorities own ~70%. The only legal way to move cash up is dividends, which pay minorities $2 for every $1 the family's side receives.

So the leakage routes are either pro-rata (you get paid too) or illegal, disclosed, and litigated.

At $39 you're paying for the telecom and getting the Anthropic stake for free. For the bear case to work, shareholders have to see none of it. That's never happened in Korea, not even in Samsung C&T.

[Ari @arithesis]

@stefabrudan depends on how your brain handles dopamine

adhd people are 3x more likely to be entrepreneurs because they NEED that risk in their lives

[Ari @arithesis]

What the...

How does Elon talk to Grok to get Grok to talk back like that?

[Elon Musk @elonmusk]

Based Grok 🤣🤣
https://t.co/GBmdo0rmPu

[Ari @arithesis]

@HyperTechInvest yeah, 0.25% is a misprint, and $SKM just added more shares in the last round

https://t.co/r2mDEsN9QS

[Ari @arithesis]

$SKM: "the cleanest pre-IPO Anthropic play." I traced the actual share count. might be nearly double what everyone quotes.

Anthropic IPOs within months. betting markets give 59% odds it opens above $1.8 trillion. you can get in via korea's biggest phone company, trading at $39.

the issue: no one agrees how much of Anthropic it actually owns. english press says 0.3%. korean brokers say 0.58%. X says up to 1%. the company itself won't say (NDA with Anthropic).

so I pulled the korean filings. the debate resolves up, not down. and at $39, the stake is priced at zero anyway.

the filed number
korean annual reports include a mandatory investments table that the 16-page english summary omits. pages 441-442 of the 442-page korean original:

3,703,141 shares. 0.7% of Anthropic as of year-end 2024. carrying value $132M.

and the math checks itself: $100M / 3,703,141 = exactly $27 a share. at the $5B valuation that's 185.2M total shares, making SKT's stake exactly 2.0%, matching their original "2%" language to the decimal. you can't fake a number that lines up three ways.

the dilution, traced round by round
SKT's shares never shrink. the percentage dilutes as Anthropic issues new stock:
- aug 2023: 2.0% ($100M at $5B)
- YE 2024: 0.7% (the filed number. the amazon and google mega-rounds did the damage)
- sep 2025, series F ($13B at $183B): ~0.65%
- feb 2026, series G ($30B at $380B): ~0.60%
- spring 2026, employee tender + amazon $5B + google $10B equity: ~0.54%
- series H ($50B new at $965B post): ~0.51% basic today

korean sell-side (Korea Investment, Meritz) independently lands at ~0.58% pre-series-H

so where did 0.3% come from?
a phantom denominator. PitchBook carries 1.47 billion "fully diluted" shares, a count that includes option pool that was authorized but never granted. shares that don't exist can't dilute.

3.7M / 1.47B = 0.25%, the english press picked it up after the auditor switched disclosure conventions mid-year, and it stuck around. the share count never changed. the presentation did.

the honest range today is 0.40% to 0.51% -- nearly double the consensus figure. the august DART filing and Anthropic's public S-1 will both print the real denominator, both to arrive before the IPO.

there's another detail the 0.3% camp missed entirely: on june 10, SKT confirmed it made an additional investment in Anthropic's latest round, size undisclosed.

what you pay for the stake today: $0
SKM's market cap is $15.0B at $39. the business excluding Anthropic is worth $16-17.7B on conservative broker math, and closer to $20B on a DCF.

This is korea's #1 carrier, with 31.8M subscribers, $4B+ of annual EBITDA, and a finished 5G capex cycle. plus an AI datacenter unit growing 89% year over year at full capacity, + a venture book holding Rebellions, Perplexity, and Persona.

basically, everything except Anthropic already covers the entire market cap. the 3.7 million Anthropic shares are priced at less than zero.

and this isn't a distressed company being valued for parts. the 2025 data breach that made the stock cheap is healing, with the first operating-profit quarter above ₩500B in a year, 210K subscribers returning, the dividend reinstated at roughly 3.5%, and KKR buying 49% of the datacenter unit at a ₩2-2.5T valuation, expected to sign around june 30.

the model, and please read this part before the numbers
every target below applies a 30% haircut to the Anthropic stake. that's deliberate. korean conglomerates historically get credited for only 50-70% of the assets they hold (the "korea holdco discount"), so the skepticism is built in up front. it also means that if you recompute these numbers without the haircut, you'll get bigger targets than mine, not smaller.

is 30% even the right haircut? sister company SK Square trades at a 45% discount today, so the bears have a real reference point. on the other side, the discount regime is being dismantled in real time (the fiduciary-duty amendment, the Value-Up program, a KOSPI up 80% this year), and the moment Anthropic becomes a publicly quoted stock, the opacity that justifies half the discount disappears. I use 30% to err on the harsh side, on purpose.

the ladder, on the central ~0.45% share basis. punitive 45% haircut | my 30% base | gross:
$1.8T: $55 | $58 | $64
$2.0T: $57 | $60 | $67
$2.2T: $58 | $61 | $69
$2.5T: $60 | $64 | $72

(on the filed 0.51% basis, add $2-3 to every cell; under the phantom 0.25% math, subtract $6-9. that spread is exactly why the august filing matters: it collapses it.)

the stock trades at $39. the worst cell on that ladder, a punitive discount at the lower side of the betting markets' range, still reads $55. the pessimist's view of this model is still +42%.

honest read
- the korea discount could stay stubborn longer than the reform headlines suggest. that's why every number above already absorbs 30%
- management says it won't sell the stake. the NAV re-rates anyway once Anthropic trades daily, but how value reaches shareholders remains the long-game question
- the street is either asleep or hostile: goldman carries a sell rating and models none of this, and the consensus target of $31 sits below the market price. either the telco analysts are right, or they haven't read page 441 of a korean filing
- the stacked worst case (IPO pulled, phantom share math, a KOSPI unwind) is roughly $22-26, about -40%. it takes four independent failures to get there, and the dividend would yield over 6% at those prices
- the stock has already doubled on anticipation, so the IPO itself could trade as a sell-the-news event even in the winning scenario

what's next
- in two days: the SpaceX IPO - is it going to run hot?
- ~june 30: the KKR datacenter deal signs, hardening the cash floor
- ~august: the H1 DART filing, with an updated share count that includes the june add-on. the most important document in this thesis
- Q3: Anthropic's public S-1, will disclose official total shares
- oct-dec: the IPO itself. the markets say 81% by october, with 59% odds of opening above $1.8T

bottom line
if Anthropic opens at $1.8-2T, where the betting markets have it, SKM is worth $58-60 after a harsh 30% discount. that's +50%, and it's the boring outcome. every 10% Anthropic runs above $2T stacks another ~5% on top, with a 3.5% dividend while you wait.

august prints the real number. until then: this might be a free stake in the fastest-growing company in history.

[Ari @arithesis]

@gulVasikova $SKM is so overlooked, wrote a detailed post on why:
https://t.co/r2mDEsN9QS

Anthropic is also most likely to IPO ~$2T, $1T is wayy too low

[Ari @arithesis]

$SKM: "the cleanest pre-IPO Anthropic play." I traced the actual share count. might be nearly double what everyone quotes.

Anthropic IPOs within months. betting markets give 59% odds it opens above $1.8 trillion. you can get in via korea's biggest phone company, trading at $39.

the issue: no one agrees how much of Anthropic it actually owns. english press says 0.3%. korean brokers say 0.58%. X says up to 1%. the company itself won't say (NDA with Anthropic).

so I pulled the korean filings. the debate resolves up, not down. and at $39, the stake is priced at zero anyway.

the filed number
korean annual reports include a mandatory investments table that the 16-page english summary omits. pages 441-442 of the 442-page korean original:

3,703,141 shares. 0.7% of Anthropic as of year-end 2024. carrying value $132M.

and the math checks itself: $100M / 3,703,141 = exactly $27 a share. at the $5B valuation that's 185.2M total shares, making SKT's stake exactly 2.0%, matching their original "2%" language to the decimal. you can't fake a number that lines up three ways.

the dilution, traced round by round
SKT's shares never shrink. the percentage dilutes as Anthropic issues new stock:
- aug 2023: 2.0% ($100M at $5B)
- YE 2024: 0.7% (the filed number. the amazon and google mega-rounds did the damage)
- sep 2025, series F ($13B at $183B): ~0.65%
- feb 2026, series G ($30B at $380B): ~0.60%
- spring 2026, employee tender + amazon $5B + google $10B equity: ~0.54%
- series H ($50B new at $965B post): ~0.51% basic today

korean sell-side (Korea Investment, Meritz) independently lands at ~0.58% pre-series-H

so where did 0.3% come from?
a phantom denominator. PitchBook carries 1.47 billion "fully diluted" shares, a count that includes option pool that was authorized but never granted. shares that don't exist can't dilute.

3.7M / 1.47B = 0.25%, the english press picked it up after the auditor switched disclosure conventions mid-year, and it stuck around. the share count never changed. the presentation did.

the honest range today is 0.40% to 0.51% -- nearly double the consensus figure. the august DART filing and Anthropic's public S-1 will both print the real denominator, both to arrive before the IPO.

there's another detail the 0.3% camp missed entirely: on june 10, SKT confirmed it made an additional investment in Anthropic's latest round, size undisclosed.

what you pay for the stake today: $0
SKM's market cap is $15.0B at $39. the business excluding Anthropic is worth $16-17.7B on conservative broker math, and closer to $20B on a DCF.

This is korea's #1 carrier, with 31.8M subscribers, $4B+ of annual EBITDA, and a finished 5G capex cycle. plus an AI datacenter unit growing 89% year over year at full capacity, + a venture book holding Rebellions, Perplexity, and Persona.

basically, everything except Anthropic already covers the entire market cap. the 3.7 million Anthropic shares are priced at less than zero.

and this isn't a distressed company being valued for parts. the 2025 data breach that made the stock cheap is healing, with the first operating-profit quarter above ₩500B in a year, 210K subscribers returning, the dividend reinstated at roughly 3.5%, and KKR buying 49% of the datacenter unit at a ₩2-2.5T valuation, expected to sign around june 30.

the model, and please read this part before the numbers
every target below applies a 30% haircut to the Anthropic stake. that's deliberate. korean conglomerates historically get credited for only 50-70% of the assets they hold (the "korea holdco discount"), so the skepticism is built in up front. it also means that if you recompute these numbers without the haircut, you'll get bigger targets than mine, not smaller.

is 30% even the right haircut? sister company SK Square trades at a 45% discount today, so the bears have a real reference point. on the other side, the discount regime is being dismantled in real time (the fiduciary-duty amendment, the Value-Up program, a KOSPI up 80% this year), and the moment Anthropic becomes a publicly quoted stock, the opacity that justifies half the discount disappears. I use 30% to err on the harsh side, on purpose.

the ladder, on the central ~0.45% share basis. punitive 45% haircut | my 30% base | gross:
$1.8T: $55 | $58 | $64
$2.0T: $57 | $60 | $67
$2.2T: $58 | $61 | $69
$2.5T: $60 | $64 | $72

(on the filed 0.51% basis, add $2-3 to every cell; under the phantom 0.25% math, subtract $6-9. that spread is exactly why the august filing matters: it collapses it.)

the stock trades at $39. the worst cell on that ladder, a punitive discount at the lower side of the betting markets' range, still reads $55. the pessimist's view of this model is still +42%.

honest read
- the korea discount could stay stubborn longer than the reform headlines suggest. that's why every number above already absorbs 30%
- management says it won't sell the stake. the NAV re-rates anyway once Anthropic trades daily, but how value reaches shareholders remains the long-game question
- the street is either asleep or hostile: goldman carries a sell rating and models none of this, and the consensus target of $31 sits below the market price. either the telco analysts are right, or they haven't read page 441 of a korean filing
- the stacked worst case (IPO pulled, phantom share math, a KOSPI unwind) is roughly $22-26, about -40%. it takes four independent failures to get there, and the dividend would yield over 6% at those prices
- the stock has already doubled on anticipation, so the IPO itself could trade as a sell-the-news event even in the winning scenario

what's next
- in two days: the SpaceX IPO - is it going to run hot?
- ~june 30: the KKR datacenter deal signs, hardening the cash floor
- ~august: the H1 DART filing, with an updated share count that includes the june add-on. the most important document in this thesis
- Q3: Anthropic's public S-1, will disclose official total shares
- oct-dec: the IPO itself. the markets say 81% by october, with 59% odds of opening above $1.8T

bottom line
if Anthropic opens at $1.8-2T, where the betting markets have it, SKM is worth $58-60 after a harsh 30% discount. that's +50%, and it's the boring outcome. every 10% Anthropic runs above $2T stacks another ~5% on top, with a 3.5% dividend while you wait.

august prints the real number. until then: this might be a free stake in the fastest-growing company in history.