Lunar Think Trade

Trading, AI, and geopolitics. NOT FINANCIAL ADVICE.
BS Math MS Quant/CS

[Lunar Think Trade @LunarThinkTrade]

https://t.co/niEj1CS6C0

[Jukan @jukan05]

SK HYNIX IS REPORTEDLY IN TALKS TO ACQUIRE INTEL’S OHIO FAB, AIMING TO PRODUCE MEMORY CHIPS IN THE U.S., ACCORDING TO SOUTH KOREAN MEDIA REPORTS.

$INTC

[Lunar Think Trade @LunarThinkTrade]

@AndrewCurran_ Struggling to align right now makes me concerned about superalignment

[Ehrmantraut Capital @EhrmantrautCap_]

$AMD doubling down on HBM?

AMD's MI450 consuming 432GB of HBM4 per GPU would be significantly more than $NVDA's Vera Rubin GPU, which is expected to carry 288GB of HBM4.

Very remarkable to see how limited demand destruction is in the AI sector as opposed to the consumer electronics sector, hyperscalers are not going to stop using HBM any time soon despite skyhigh prices (and we can expect that trend to continue into 2027).

[Lunar Think Trade @LunarThinkTrade]

@jukan05 EMIB is a bet on yield ceiling and for 9.5+. CoWoS is better in volume, especially in 2027-28. I would not doubt TSMC as they are known for overcoming yield barriers while maintaining volume. The bet between the 2 typically favors TSMC over Intel.

[Lunar Think Trade @LunarThinkTrade]

@jukan05 We need to pay careful attention to who solves large-area, high-yield packaging. Everything, including memory, depends on who can deliver 9.5-12x reticle scaling. Are the 2 biggest contenders EMIB-T and CoWoS-L? I still think TSMC has the best chance.

[Lunar Think Trade @LunarThinkTrade]

@jukan05 Demand and production are not scaling at unprecedented rates in NAND. Unless HBF works, NAND market should be rather quiet and much more predictable.

[Lunar Think Trade @LunarThinkTrade]

@oguzerkan The way you capture enterprise value is not just about the cheapness of the model. It is about uptime after 50 steps into a project. Kimi is rendered useless if it runs out of capacity 3 steps into a hype phase.

[Lunar Think Trade @LunarThinkTrade]

It is easy to perform well on benchmarks or a few questions. It is another thing to continuously perform and stay up during a long-term project that requires large context windows and 100+ steps. This is why I am still bullish on closed-source frontier models.

[Lunar Think Trade @LunarThinkTrade]

People could not fathom $50B in AI investments in a single year 3 years ago. People could not fathom $1T in AI investments in a single year 2 years ago. Why do we keep using historical data to try and predict something that is revolutionary? Natural human instinct I guess.

[Teo @Teo_Sinamin]

China's heavy rare earth tap stays closed for Japan in June

China's exports of strategically important, controlled rare earths and minerals to Japan were tiny again in June, extending the raw material ​fallout of a diplomatic spat between the two nations.

China sent no ‌gallium, dysprosium, terbium or yttrium to Japan last month, Chinese customs data released on Monday showed. Exports have been tightly throttled for several months.

Japan has the world's biggest rare earth ​magnet sector outside of China but, like other countries, is reliant on ​China for its sourcing of some rare earth inputs.

Exports of controlled ⁠metals plunged after comments made by Japanese Prime Minister Sanae Takaichi on Taiwan ​in November caused a diplomatic rift.

Beijing introduced export controls on types of heavy rare ​earths and the magnets that contain them in April 2025. It publicly tightened controls on exports to Japan in January, and then twice again the following month, targeting major conglomerates.

Companies in the ​island nation have sounded the alarm over access to critical minerals, warning it is ​beginning to affect the broader Japanese economy.

China's export controls allow it to leverage its dominant ‌position ⁠in critical mineral supply chains and have become one of Beijing's most powerful diplomatic levers, also discussed in recent negotiations with U.S. President Donald Trump.

Exports of the chip and magnet-making metal gallium to Japan fell back to zero in June, after ​consumers there got ​some respite with ⁠a large shipment in May.

Export control-related shortages of yttrium — a rare earth used to protect turbine blades in aircraft engines or ​power plants from extreme heat — have wracked the market.

China exported ​no yttrium ⁠to Japan in June. Meanwhile, the U.S., a major destination for Chinese exports of the rare earth before the restrictions came into effect, had none sent there ⁠for the ​second month running.
Elsewhere, China's overall exports of rare ​earth magnets remained strong in June. It exported some 5,649 metric tons of magnets in June, ​up from 4,730 tons a month earlier.

[Jukan @jukan05]

Morgan Stanley’s Joseph Moore said that, after speaking with several purchasing contacts in the data center space last week, the intensity of the memory shortages shows no signs of abating. He added that prices appear to be up at least 25% on a like-for-like basis from 2Q to 3Q.

This is above both Morgan Stanley’s and third-party estimates.

Moore also noted that longer-term concerns that the memory shortage will intensify in 2027 and again in 2028 remain as strong as ever. Morgan Stanley added that there is not enough memory relative to AI requirements and that it does not see this situation changing.

Notable quotes:

“Cloud customers are paying premiums to the expected 2Q price for six-week expedites; do we think those customers are paying those premia to stockpile memory in a warehouse?”

“AI is consuming so much DRAM that there isn’t enough left over for other sectors, and everywhere we look, we see indications that it is a true bottleneck. It’s holding back PC builds and smartphone builds.”

“Memory is not just constrained by AI demand—memory is increasingly one of the major primary constraints on AI demand, along with space and power.”

$MU $SKHY $SNDK

[Lunar Think Trade @LunarThinkTrade]

@The_AI_Investor It baffles my mind. $100k/month for early access to market manipulation. Please tell me this is not real.

[Lunar Think Trade @LunarThinkTrade]

@IbrahimDagher20 Twitter algorithm favors over-reaction and bandwagoning the latest headline. Time to find the alpha.

[Lunar Think Trade @LunarThinkTrade]

@EhrmantrautCap_ Jane St got paid this week. free alpha

[Lunar Think Trade @LunarThinkTrade]

@aleabitoreddit I am bullish on Jane Street's quarterly numbers.

[Lunar Think Trade @LunarThinkTrade]

From a financial POV, SOTA models are already too big to fail with regards to spending and investments.

With the release of Kimi being on par with Opus 4.8 and ChatGPT 5.5, Jevons paradox only gets larger.

As capex spending begins to go online in 27-28, if Chinese models are able to keep up, investments in memory and compute only become stronger.

[Lunar Think Trade @LunarThinkTrade]

@oguzerkan Most of capex spending has yet to come online. Build outs take at least 2 years. We will see a larger disparity in pre-training sizes in 2027-28 between open source models and frontier models.

[Lunar Think Trade @LunarThinkTrade]

@YingYan74551198 @oguzerkan Most of capex spending has yet to come online. Build outs take at least 2 years.

[sensho @sensho]

i love kimi a lot but even they admit in their blog in daily use both sol and fable are still better

the $650B in capex that you mention is mostly yet to come online

the bitter lesson pills had always won out in ai (eg as it did with mythos/fable many months ago) and while i think it’s amazing they caught up i still think they will need OOMs more compute to stay competitive long term

[Lunar Think Trade @LunarThinkTrade]

@GavinSBaker Kimi is great for the world and only encourages more efficiency, more private-backed subsidies for closed-source models. It is a panic moment for national security in the short-term, but a positive net utility for everyone imo.