Tail Risk

perpetually curious...

its in the sitting, not doing

[Tail Risk @_tailrisk]

Daily Notes:

[Tail Risk @_tailrisk]

Two big employee departure? Time to short $GOOGL ?

[John Jumper @JohnJumperSci]

A bit of news: After nearly 9 years, I have decided to leave Google DeepMind and join Anthropic (after taking some time to recharge). I am incredibly grateful for my time at GDM. @demishassabis took a real chance letting me lead the AlphaFold team just six months after finishing my PhD, and the entire GDM team taught me so much about how to do great science. GDM is a special place, and I’ll still be excited to hear about what amazing things they discover next.

[북극성 @PolarisLog]

삼성 파운드리, 빅테크 AI 칩 수주 쌓인다.
(이형수)

- 3나노에서 아쉬움을 남겼던 삼성 파운드리가 2나노 선단 공정을 앞세워 빅테크 협업을 하나씩 따내는 분위기임

- 구글의 10세대 AI 칩 TPU v10 가능성이 거론되는데, 메인 연산 다이는 TSMC 1.4나노를 쓰고 입출력 인터페이스 칩에 삼성 2나노가 들어갈 것으로 보임

- 양산 목표가 2028년이라 시간은 남았지만 레퍼런스로서의 무게는 충분함

- 리사 수 CEO 방한 이후 AMD 에픽 서버용 CPU를 삼성 2나노로 만든다는 이야기가 흘러나옴

- 퀄컴도 스냅드래곤을 약 4년 만에 다시 삼성 파운드리에 맡길 가능성이 높아짐

- 구글 협력이 특히 값진 이유는 매출보다 설계 노하���를 배우고 핵심 IP를 함께 개발할 기회이기 때문임

- AI 워크로드가 텍스트에서 이미지, 동영상으로 빠르게 옮겨가다 보니 특정 연산만 잘하는 ASIC은 불리할 수 있음

- 그래도 구글처럼 인프라 전체를 쥐고 있는 기업은 워크로드 자체를 통제할 수 있어 상황이 다름

- 첨단 패키징은 TSMC의 CoWoS 해자가 워낙 단단하고 그 틈을 인텔 EMIB가 메우는 구도라 삼성이 당장 다 가져오긴 어려움

- 대신 삼성은 IDM답게 메모리부터 기판까지 묶어서 던지는 패키지 딜로 차별화를 만들고 있음

- 쇼티지 상태인 HBM, 소캠, LPU 캐시용 SRAM까지 커스텀으로 함께 제안할 수 있다는 점이 강점임

- 삼성전기와 엮으�� FC-BGA, 유리 기판에 AI 서버 한 대당 2만 8천 개나 들어가는 고마진 MLCC까지 한 번에 공급 가능함

- 실제로 브로드컴, 구글 관계자들이 샘플을 보려고 미팅을 진행 중인 상황임

- 직접 하기 어려운 독자 패키징은 앰코 같은 OSAT 2위 업체와 손잡거나 PLP 기반 대체재로 보완할 거란 추측도 나옴

- 공급망 강화와 가동률 상승 흐름에서 관심 갈 만한 국내 종목들도 함께 거론됨

- 원익IPS는 삼성 테일러 팹 초기부터 깊게 들어간 전공정 장비사임

- 케이씨텍은 본업인 CMP 장비에 더해 초임계 세정 장비가 SK하이닉스 퀄을 통과하며 주목받는 중임

- HPSP는 TSMC향으로 신고가를 찍은 고압 수소 어닐링 업체로 삼성 진입 여지도 충분함

- PSK는 PR 스트립과 세정, 식각 쪽에서 공정 노출도가 높은 편임

- 두산테스나와 하나마이크론은 물량이 넘칠 때 낙수 효과를 기대할 수 있는 OSAT, 테스트 업체임

- 삼성전기는 주가가 이미 많이 올랐지만 유리 기판, FC-BGA, 하이엔드 MLCC 수혜가 또렷함

[Bessemer @BessemerVP]

1/ We've been mapping the AI data center stack — and the opportunity is hiding in plain sight.

Everyone's talking about GPUs. The real bottleneck is physical infrastructure: power, permitting, cooling, and the grid. 🧵

[Moatless Capital @MoatlessCapital]

No opinion on $META, but its precisely this behavior that allowed $GOOGL to win out buying Deepmind

[Ricardo @Ric_RTP]

JD Vance just admitted the White House plan is to take ownership of every major AI company in America.

Steven Bartlett brought up Bernie Sanders' proposal that workers should own 50% of the major AI companies.

Vance's response: "The president by the way likes that idea too. He likes that idea."

Trump's preferred mechanism, Vance said, is a sovereign wealth fund where the US government takes equity stakes in private AI companies.

The Vice President literally just confirmed that an administration is planning the most radical economic policy proposed in modern American history. Partial nationalization of the MOST valuable private companies on earth. And the idea originally came from Bernie Sanders, who Vance said Trump agrees with on this point.

This is not a small thing:

The US has spent 80 years selling the world on the model where private companies stay private and the government stays off the cap table.

The countries that did the opposite, with sovereign wealth funds owning slices of their biggest firms, are Norway, Saudi Arabia, China, and Singapore. And the Trump administration told you on a podcast it wants to do the same to Silicon Valley.

But the reasoning Vance gave for it is where it gets really interesting...

He said the historical analogy that scares him is the original Industrial Revolution. His own words:

"Rich people got way richer. And that led to in Europe fascism and communism."

He believes AI will not cause mass unemployment but mass inequality, and that mass inequality is what breaks societies. His fix is that workers need a seat at the bargaining table before the wealth gets created, not a redistribution check after.

"I think labor unions are a very important model here."

And the other thing about AI that scares him is surveillance. His exact phrase was that AI is "fundamentally a communist technology" because it lets governments and corporations watch and score people in ways NOTHING else can.

He said he doesn't want a social credit system, doesn't want a tech CEO deciding whether you can buy a beer based on an algorithm nobody understands, and is afraid of exactly that outcome.

So here is the full picture:

The sitting Republican administration believes AI will make the rich dramatically richer, that this will radicalize the country the way the Industrial Revolution radicalized Europe, that the answer is government equity stakes plus stronger labor unions, and that the second-biggest threat is the surveillance state these companies are building.

That is not a Republican worldview. That is not even a Democratic worldview.

This is a worldview that has no political home in the United States right now.

Most people are still arguing about whether ChatGPT will take their jobs. But the people with the actual power are already past that argument.

They are quietly designing the framework for owning the companies that will.

The craziest part is how casually Vance dropped it as a sidenote on a podcast millions will half-listen to in the background.

If you have money in OpenAI, Anthropic, or anything like that, you should be watching the full thing yourself.

What do you think?

[Dan Nystedt @dnystedt]

Nvidia Vera Rubin-based AI data centers cost US$47 billion per 1GW (gigawatt) data center holding 3,557 server racks – $9.1M each – with an annual $1.3B electric bill and hardware depreciation 6x higher than power costs, said Young Liu, chairman of Foxconn, the world’s biggest AI server maker, in a speech. 1/3 $NVDA #Foxconn $TSM $ASX $DELL $HPE $SMCI

[郭明錤｜Ming-Chi Kuo @mingchikuo]

Breaking down TSMC's glass core substrate slide

On June 11, at JPCA Show 2026 in Japan, TSMC gave a roughly 40-slide presentation titled "Advanced Packaging Technology Essential to the Evolution of AI" (AIの進化に不可欠な先端パッケージング技術). One slide from the deck, titled "Glass Substrate Development for CoWoS," has since leaked online and widespread attention.

Here's a closer read of that slide (see attached image). I'll skip the technical background that is already widely available. One thing to flag: the "COP" on the slide does not stand for Chip-on-Package. It means Coplanarity.

▌ Key conclusions:

1. TSMC has officially announced a partnership with Ibiden and Innolux to develop a glass core substrate. The structure is a three-layer design, a glass core sandwiched between two ABF build-up layers. This is the "oS" in CoPoS.

2. The market underestimates how important the glass core substrate is. It's a must-have capability for TSMC. In other words, within CoPoS the "oS" matters more than the "CoP", which is also why, when it was tested, it was paired with the existing CoW rather than with CoP.

3. The glass core substrate costs several times more per unit than existing ABF substrates. The glass processed by Innolux is very expensive per unit and is the single most critical material. Besides Nvidia, two US-based customers have also expressed strong interest.

▌ Industry checks tied to this slide:

1. The glass core substrate shown on the slide is cut from a full-size 250×250mm one. The ABF build-up layers mainly use Ajinomoto's GL107, mixed with ABF-GCP, and were tested at 24–28 layers, which is the mainstream ABF spec for AI chips in 2027–2028.

2. The CoW used in TSMC's experiment is a test vehicle. It is sufficient to validate the most challenging mechanical-structure issues that arise when working with composite materials. Good results mean TSMC, Ibiden, and Innolux have together broken through the critical technical bottleneck.

3. Ibiden currently handles cutting the 250×250mm glass core substrate. When the 510×515mm format is used for pre-mass-production simulation in 2H27, if Ibiden still wants to reduce production complexity to protect its ultra-high gross margins, it may hand the cutting over to Innolux, which is more familiar with the properties of glass.

▌ The leaked slide shows the validation results of pairing CoW with the "oS" in CoPoS, i.e., the glass core substrate (labeled "glass-SBT" on the slide). This addresses the "Substrate mechanical and electrical Dilemma" raised on the previous slide, and it strongly underscores how important the "oS" is within CoPoS.

1. Within CoPoS, what CoP solves is production efficiency / cutting economics, which ties to cost and price. What the oS solves is warpage and durability, which determines whether the chip can be made at all, and whether it can work.

2. CoP and oS complement each other well when integrated, but looking out over the next few years their technical roles still differ. CoP is a very-nice-to-have optimization, and going without it simply means a more expensive chip. But the oS is a must-have. Without it, even being able to make a usable chip is in doubt.

3. Comparing their roles isn't about elevating oS at the expense of CoP. It comes down to the practical question of which technical piece customers are willing to pay for. Details below.

▌ The real gold here is the power integrity (PI) improvement shown on the slide. This matters a great deal to customers, and it means that once glass core substrate production stabilizes, TSMC's profitability and competitive edge should rise in tandem.

1. How it works: the glass core substrate is thin → the vertical conduction path through TGV (through-glass vias) is short → conduction-path resistance (R) and loop inductance (L) both drop → PI improves.

2. Why it matters to customers: better PI → more stable power delivery → frees up power headroom → room to integrate more transistors, or to push clock speeds higher → more AI compute.

3. For customers, production efficiency is TSMC's basic responsibility, so they won't pay extra for it. But gains in AI compute translate directly into the customer's own competitiveness and profit, so customers are willing to pay for that. This is why Nvidia is so positive on the glass core substrate.

4. For TSMC, the glass core substrate raises yield and lowers cost while also boosting both the compute and the selling price of AI chips. It's both a cost-cutting tool and a pricing lever, a plus for profitability and competitiveness alike.

5. Substrate cost currently accounts for a low single-digit percentage of an AI chip's BOM, while losses from packaging yield run roughly 5–10× the substrate cost. So even if the glass core substrate ends up costing several times more than today's, its share of the BOM stays low, and it can cut the losses from packaging yield. The high unit price is therefore not expected to dampen customers' willingness to adopt it.

▌ In the Q&A after the presentation, an audience member asked about TGV details for the glass core substrate. TSMC declined to answer on the spot, because TGV is the key technology behind the glass core substrate, and the core know-how currently sits with TSMC and Innolux. By contrast, when another attendee asked about integrating IVR, eDTC, and LSI, TSMC answered at length.

▌ According to industry checks, if all goes well, TSMC is aiming to start mass production of the glass core substrate in 4Q28–1Q29, to match the cadence of Nvidia's AI chip iterations. As a side note: the Ibiden earnings presentation slide that many people have been circulating lists the glass core substrate timeline as CY30. My read is this: Ibiden, which has always been conservative and cautious in public, has now formally put the glass core substrate on its roadmap, which further confirms the long-term trend for this technology. That said, some other details on Ibiden's slide don't fully line up with what's known in the market. For example, its reticle timeline is off from TSMC's public claims by about a generation, and the Rubin Ultra substrate size is clearly larger than the 90×90 it marked for CY26–27. It's a reminder to always cross-check across multiple sources when forecasting the future.

[Dan Nystedt @dnystedt]

Elon Musk’s TeraFab is a major project and new opportunity that will test ASML’s supply chain, said ASML CEO Christophe Fouquet, comparing the size to the millions-of-wafers DRAM fab projects going on in South Korea, in an interview with Bloomberg. 1/3 $SPCX $ASML $SSNLF $HXSCL $AMAT $TOELY $LRCX #terafab #semiconductors

[Tail Risk @_tailrisk]

do you think perception matters? like how others perceive the fundamentals of META. the biggest worry I have is the stock can remain absurdly cheap for a long enough time frame that the opportunity cost is just too high which makes the expected return lower. (ie, waiting 2 years for it to get back to your expected multiple vs 5 years)

[Podcast Alpha @PodcastAlphaX]

If you hold $NET, this one changes an input.

Gavin Baker @GavinSBaker of Atreides Management told @jordihays on @tbpn: CDNs like Cloudflare and Akamai deliver under 1% of global tokens today. Possibly under 10 basis points.

The bulk of token generation happens inside hyperscaler and neocloud data centers. CDNs handle the final delivery hop - real, but small. Cloudflare is trading at all-time highs on AI narrative. Baker's framework says the stock is pricing in infrastructure upside that isn't flowing through token throughput yet.

Token factory = yes. Token deliverer = maybe. The gap between those two labels is a valuation gap.

Who passes Baker's token path test and who doesn't:
https://t.co/C9hHmizK85
Source: TBPN - https://t.co/waLjhPoyWH

[북극성 @PolarisLog]

삼성전자
- 현재 캐파: 65만장/월
- 26년 예상: P4 증설 반영 시 76만장/월
(현재 대비 +11만장, +17%)
- 27년 예상: P4 램프업 반영 구간, 76만장/월 수준 유지 가정
- 28년 예상: P5 증설 반영 시 106만장/월
(현재 대비 +41만장, +63%)
- 29년 예상: P6 증설 반영 시 136만장/월
(현재 대비 +71만장, +109%)

SK하이닉스
- 현재 캐파: 55만장/월
- 26~27년 예상: M15X 증설 반영 시 63만장/월
(현재 대비 +8만장, +15%)
- 28~29년 예상: 용인 1차 램프업 반영 시 81만장/월
(현재 대비 +26만장, +47%)
- 30년 예상: 용인 Fab 1 풀 반영 시 99만장/월
(현재 대비 +44만장, +80%)

[Tail Risk @_tailrisk]

in the long run 90% of the usage don't require frontier intelligence. that means there is a long tail of inference revenue that can be monetized...

[Dan Nystedt @dnystedt]

https://t.co/OA9mnHgBAB

[Jukan @jukan05]

Korean Media: SK Hynix Preparing Massive $66.4 Billion Shareholder Return Program Following ADR Listing

- Korean media report that SK Hynix is expected to announce a large-scale shareholder return program worth approximately $66.4 billion after its ADR listing.
- To minimize dilution concerns and potential backlash from existing shareholders, the company has reportedly reduced the ADR offering size from 2.4% of outstanding shares to the low-2% range.
- SK Hynix is expected to raise approximately KRW 40 trillion (about $26.5 billion) through the issuance of new shares as part of the ADR offering.

[Tobias Reisner @reisnertobias]

Trust me: Nobody is ready for that Coinbase bid.

People assume 500k in daily buyback pressure because they do the back of the napkin math on $5B TVL.

What if I tell you USDC on HL goes to $50B within 24 months

That’s $5m in daily buyback pressure.

Crazier things have happened.

[Tail Risk @_tailrisk]

People freaking about Fable storing your prompts and context for 30 days…. But forgetting that Apple listens to every conversation you have…

Nobody cares about privacy. We are the product…

[bubble boi @bubbleboi]

It’s only going to get worse.

I know of several features that new accelerators are adding that have no analogous operation on GPU, TPU, etc. most of these “special tricks” are on the inference domain and not only lower the TCO but also increase the model quality & capabilities.

Hardware-algo codesign is the last frontier left and it’s going to be the area that picks the winner in the end.

[Gonzalo Nuñez @gonzalo__nunez]

It’s telling that MSFT put this out before one of the labs did.

I’ve never seen anything from the labs frame the future in this way, or sympathize with firms and consumers like this.

The labs are still too confused and unclear about where they slot in, or they’re fully cognizant and simply trying to dominate.

[Susan Zhang @suchenzang]

current LLMs fundamentally consist of four main components:

- input layer: where input "words" (prompt) get mapped to "latents" aka some-model-representation-you-don't-understand-unless-you-start-reading-tea-leaves-of-spurious-correlations (some quite compelling à la word2vec style; latents is also unnecessary lingo so i will refer to these as "inputs" with quotes from now on)

- mixing layers: where you jumble all your "inputs" together to see if any correlations between "inputs" can become useful (commonly used to compress or expand dims; predicting a single classification target == compress to a single dim, etc)

- attention layers: where you learn how "inputs" relate to each other (aka discern what's important to remember vs fluff)

- residuals: where you short-circuit a mixing/attention layer because it's probably adding too much confusion (aka avoid overthinking for simple things)

-----

a "big" LLM simply scales two things:

- width == how many dimensions you give to your "inputs" (the more dims, in theory the more unique/discerning/precise/complex your knowledge can become)

- depth == how many mixing/attention/residual layers you can stack/loop between (aka "reason" over, where more of these ~= more "reasoning" abilities)

"capabilities" that seem impressive to humans usually arise from taking advantage of both depth & width: where a model seemingly makes connections between disparate ideas, beyond what an average human can hold in working memory.

this requires models to "completely light up" when responding to a "hard prompt", where effectively no param/layer goes unused.

-----

the anatomy of a "model capability" is precisely the same mechanism that can be co-opted for a jailbreaking exploit:

your goal is simply to "light up" as much of the model as possible, dodging any shallow input-classifiers at the beginning by triggering as many disparate "input ideologies" as possible, and subsequently have these "inputs" relate to each other in seemingly unrelated-yet-related ways that ideally have similar "complexity" as your jailbreak goal (to make it past enough layers of the model).

think of the attack-vector as bundling your goal in a series of schizo-nerd-snipes:

a sufficiently capable model will try to reason through everything all at once, eliminate the dead-ends, and successfully deliver the one jailbreak use-case you bubble-wrapped for.

of course, there's an art to the above, and some are already extraordinarily proficient at the trojan-horse-packaging, but at some point there's no difference between "a capability" and "a jailbreak", though i'll be happy to be proven otherwise.

-----

tl;dr ant flew too close to the sun, better kiss the ring or get buried.

[Tail Risk @_tailrisk]

What’s the likelihood of someone in Anthropic actually going rogue for the sake of humanity and open sources fable 5 weights? 🤞

[Piotr Pomorski @PtrPomorski]

Someone put Fable 5 on the pirate bay, 3.4TB 😂