콜 COLE

미장 투자함. AI랑 기술도 놓치기 싫은 사람. 관심사가 너무 많음.

[콜 COLE @JWMoonX]

@GyujinAAIG 현금도 충분하고, ATM도 있는데 또 전환사채까지는 너무 이른거 아닌가요..

[콜 COLE @JWMoonX]

@jackjasonz84 @SmallCapSnipa 그렇군요. 근데 이 전환사채 발행이 마이크로소프트와 엔비디아 다음 3번째 주요 계약을 확보했을 가능성으로 볼 수도 있을까요? 그건 너무 이른걸까요

[콜 COLE @JWMoonX]

@SmallCapSnipa 미란티스?? 처음 들어보는 기업인데 이것이 시사하는 바가 뭐죠.

[콜 COLE @JWMoonX]

$IREN 스위트워터 1 에너자이제이션!! 이제 계약만 남은건가요?

[IREN @IREN_Ltd]

Sweetwater 1 has been successfully energized – a key milestone in the development of the broader 2GW Sweetwater campus.

@danroberts0101, Co-Founder and Co-CEO of $IREN commented:

“Delivering Sweetwater 1 substation energization on schedule reflects our disciplined execution, the strength of our supply chain relationships and the efficiency of our vertically integrated development model. It is another example of our ability to design and construct large-scale infrastructure reliably and at speed to meet market demand.”

Learn more: https://t.co/Bo0tSv67jM

[콜 COLE @JWMoonX]

@supernovajunn 아 그렇군요 감사합니다

[콜 COLE @JWMoonX]

@supernovajunn 어떤 부분이요? 구체적으로 차이 설명해주실 수 있나요?

[콜 COLE @JWMoonX]

Claude Opus 4.7 나왔다

근데 솔직히 말하면, 커뮤니티에서 우스갯소리로 도는 말이 있다. "4.6 너프시켜놓고 원래 성능으로 돌려서 4.7이라고 발표했다"는 거다

황당하게 들리지만 근거가 없진 않다. Anthropic이 직접 인정한 내용만 봐도 출시 과정에서 기능을 의도적으로 단계적으로 낮추는 실험을 했고, 일부 안전성 항목에서는 4.7이 4.6보다 오히려 떨어진다. 그러면서 토큰 소모는 최대 1.35배 늘었다. 가격은 동일하다

[콜 COLE @JWMoonX]

@v81093933 알려주셔셔 감사합니다 ㅎㅎ

[콜 COLE @JWMoonX]

@v81093933 아 그럼 둘이 꽤 비슷한 편인가 보네요 메인으로 하나만 쓴다면 코덱스가 더 나을까요? 아니면 사람마다 좀 다른가요?

[콜 COLE @JWMoonX]

@v81093933 오 그렇군요 ㅎㅎ 근데 코덱스랑 클로드 코드 성능 비슷한가요? 뭐가 다른지 잘 몰라서요

[콜 COLE @JWMoonX]

@lucas_flatwhite AI 활용 과정을 설명하는 능력이 실제 업무 능력이랑 같은 건지는 모르겠는데 말 잘하는 사람이 뽑히는 구조가 AI 시대에도 반복되는 거 아닐까요

[콜 COLE @JWMoonX]

@supernovajunn 실수 AI ㅋㅋㅋ 근데 진짜 그게 답일 수도요 일부러 틀리게 설계된 AI 옆에서 배우는 인턴이라니

[콜 COLE @JWMoonX]

맥미니 3대, Claude로 날씨 데이터 읽어서 $101K 벌었다는 글.
항공 기상 데이터 공개돼 있고 prediction market 가격 비효율도 실제로 존재하니까 이론상 말이 되긴 하는데..
이게 진짜 재현 가능한 걸까요? 그냥 마케팅 글일까요?

[Carver @carverfomo]

A Chinese engineering student built a weather tracking station in his dorm. Three Mac Minis. Two monitors. Satellite maps on both screens. Labels on each box: UI/UX. DEV. ADMIN. Total cost under $2,000.

His roommate thought it was a climate research project. His professors thought it was a thesis prototype. He let everyone keep thinking that.

Then someone noticed what the station was actually connected to.

A wallet. Making $101K. Betting on the temperature.

ColdMath. $101,042 profit. 5,252 predictions. Joined November 2025. Bio: Edge Compounds.

→ https://t.co/T1z9GFWKVT

The station does one thing. Claude pulls live pilot weather data. Real sensors. Real readings. Updated every 1-3 hours from stations worldwide. Compares it to prediction market prices. When they don't match the DEV box flags it.

Mismatch found. He places the trade. Green result.

$25 on Tokyo hitting 16C on March 20. Payout: $12,452. $24 on Chicago reaching 54F on March 11. Payout: $12,398. $13 on Lucknow hitting 39C on March 7. Payout: $6,850.

Twenty five dollar bets returning twelve thousand. On the weather.

A friend who flies commercial told him pilots get atmospheric data hours before any public forecast. Temperature to a tenth of a degree. This data is free. Aviation safety requires it. Nobody outside of aviation even looks at it.

He looked. Pointed Claude at the feeds. Said: find me every city where the forecast doesn't match the price.

Claude found dozens. Every single day.

His roommate saw the station running one morning and finally asked what it actually does. The student showed him the balance. The roommate didn't say anything. Just asked for a second monitor.

34K people watching. $96K still loaded in active positions. Three Mac Minis. Two screens. One quiet kid who realized the most predictable thing on Earth is the thing everyone ignores.

The weather.

[콜 COLE @JWMoonX]

카파시 LLM 위키 포스트가 1300만 조회를 넘겼음.

코드 없음. 앱 없음. 설치할 것도 없음.

AI한테 파일 던지고 요약 시키는 방식이 사실 비효율적이었다는 걸 다들 알고 있었기 때문임.

매번 새로 읽고, 매번 새로 연결하고, 쌓이는 게 없음.

카파시가 제안한 건 다름. AI가 읽을 때마다 그냥 답 뱉고 끝내는 게 아니라, 위키에 쌓아두는 것.

연결하고, 모순 찾고, 업데이트하고.

새 자료가 들어올수록 위키는 더 똑똑해짐.

근데 진짜 포인트는 따로 있음.

이 파일 자체가 코드가 아니라 아이디어임.

직접 실행하는 게 아니라 에이전트가 실행함.

오픈소스가 코드를 나눴다면, 다음은 아이디어를 나누는 거.

[Andrej Karpathy @karpathy]

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

[콜 COLE @JWMoonX]

@DollarCostAvg If $IREN gets into that range, I’m going all in lol

[콜 COLE @JWMoonX]

@FransBakker9812 Fair.
Either way, it feels like the “immaterial” label may be hiding more than the market has priced in.

[콜 COLE @JWMoonX]

Great post @FransBakker9812.

The multi-OEM angle is the really interesting part here.

To me, the key question is whether the Lenovo capacity sits inside the previously disclosed $IREN fleet, or outside it.

If it sits outside, the market may be missing undisclosed GPU purchases.

If it sits inside, the market may still be missing how much monetization is already building inside what management flags as immaterial.

Either way, the market may be systematically underpricing the revenue capacity that could show up from $IREN over the next few earnings reports.

[Frans Bakker @FransBakker9812]

$IREN x @Lenovo x $DELL

My new theory.
First:
1. Dan said on our space(1) 2 weeks ago: "you can certainly buy GPUs outside of $NVDA, and in fact we do, you buy them through the OEMs".
2. He also said that the new 10MW IT load data center in Prince George may be deployed as liquid-cooled for the "1,000" GB300s (1,200 actually bought, rounded down), or installed as air-cooled. "but it's immaterial either way".

With this in mind, and looking at the Lenovo case, I am considering the following:
1. $IREN rounds down 1,200 GB300s to 1,000 GPUs, this makes 200 GB300s worth "nothing", and 1,200 or 1,000 GB300s "immaterial".
2. Material or immaterial is being compared to either the $MSFT contract (76k), or to the fleet (currently ~150k disclosed as purchased).
3. IREN never cared to announce that they would be delivering the first NVL72 GB300 cluster in Canada with Dell(2), it was probably immaterial as well.
4. Dell delivered(?) these systems as pre-configured and fully integrated Dell Integrated Rack Scalable Systems, making an overlap with Lenovo's ThinkSystem nodes extremely unlikely.

Now we have to ask ourselves:
1. If 648 nodes of Lenovo's ThinkSystem SR680a V3 holds 8 units of B200 GPU each, we get to 5,184 GPUs.
2. These are already deployed per today (see quoted post below), and as Dan said on the same space: they don't have GPUs energized in data centers that are idle.
3. Is this in any way MNPI to disclose?

Based on all of the above, I think we can conclude that:
1. IREN didn't disclose a separate purchase of 5,184 B200s related to the Lenovo deployment.
2. 5,184 GPUs are probably immaterial for IREN compared to a fleet of 150k+ GPUs.
3. The purchase price of 648 complete SR680a nodes with B200 GPUs is probably around $350m, this is most likely immaterial for IREN with almost $10B of financing behind them.

So:
1. IREN most likely didn't include these GPUs in their ARR guidance, as they were all based on Dell purchases that were disclosed.
2. So, these new GPUs could very well be deployed in their immaterial 10MW data center, and we would never know.
3. If so, these are not sitting idle, but are probably rented out for 3,4,5 years to AI natives and/or enterprises, forcing no MNPI disclosure either.
4. Even so, 5,184 B200s at $3.5 per GPU hour, would bring in ~$158m in ARR.
5. The additional 600x8 = 4,800 B300s at $4 per GPU hour, could bring in another ~$168m of ARR.

I think it's completely plausible that $IREN has bought multiple smaller batches of <10k GPUs from different OEMs and other cloud providers, without telling the market.

This Lenovo "discovery", was probably by request from, or a contractual obligation to Lenovo.

IREN is setting up for a surprise to the upside in the coming 2,3 earnings reports.

I have satellite imagery that confirms there is construction happening in Mackenzie, which could mean another "immaterial" data center is being contructed.

I think we know now why we have a $6B ATM.
I think they will tap this to pay for small <$500m purchases of GPUs to deal with the memory constraints, without having to access the debt markets.

By the same token, they can tap the ATM to pay for the construction of these immaterial data centers, with possibly incremental immaterial megawatts.

Everything is immaterial if it's small by comparison to the disclosures.

IREN is here to fuck spiders, but they won't tell us about it.
Not because they are shy, but because it's not going to move the needle.
What it will move, is the revenue by aggregate, and the earnings surprise to the upside.

The only thing that bothers me though, is the fact that Dell said in November 2025 that they will deploy Canada's first GB300 cluster with IREN, but Dan Roberts said 2 weeks they aren't sure what they will do with those "1000" GB300s they bought.

Would Dell leave this on their website, if it never actually happened?

I still lean to the "incremental OEM supplier" theory, and not to the "substitute OEM supplier".
Typically IREN will disclose GPU purchases in their quarterly earnings reports, or their management discussion of the period.
If not, I think this would make for a good analyst question.

I am always very critical of lacking data points of the companies I invest in. For this reason (amongst others), $NBIS is uninvestible for me.

$IREN on the other hand, claims they disclose much more than their competitors, and in a way that's true. But since we as retail investors are used to having the tools to calculate the future earnings, it's becoming quite hard to be precise enough, as the degree of immateriality is increasing, and there is basically an entire shadow operation going on outside of the disclosed projects and guidance.

Many people like to blame IR.
But I think this is a company culture "issue", if anything.

But where we see a future of high margin earnings, the market sees a glass that is half empty.

To be invested in $IREN, you need to have conviction, patience, and understanding.
We do our best on here, to cover what we can find.
But at the end of the day, the EPS will tell the story, and if IREN has indeed utilized their different OEM partners to deal with the time-to-GPU issue, then we will be rewarded to the upside.

I'm long IREN, and I think 2026 will be our inflection year.