If you need AI to be more than just a chatbot, you need a lot of CPUs and memory to get real work done $INTC $SKHY

My takeaway AI demand is extremely strong and continues to exceed supply across most of the ecosystem. What is the biggest bottleneck holding back AI growth? Infrastructure, not demand, mainly: Power and data center capacity Advanced packaging and foundry capacity Optical networking Memory (HBM, DRAM, NAND) Is NVIDIA still the whole AI story? No. Several companies highlighted the increasing importance of custom AI chips. Is AI software making money yet? Not really. AI revenue remains a relatively small portion of overall revenue for many firms.

中秋節快樂

Naga Chandrasekaran who is the EVP, CTO, and GM, Intel Foundry. In a meeting with Keybanc told them: Intel, "Expects 14A performance to be within 5% of TSM's A14". Intel is not afraid of CoPoS. "Sees sustained leadership in advanced packaging with EMIB-T even with TSM introducing CoPoS". I must say that the news around 14A PPA is truly extraordinary given 14A yields and ramp are extremely aggressive. Intel is finally sticking its neck out to say how their nodes compare with a % range to TSMC. Note goes on to say that Intel has a server part on 14A in 2028. Coral Rapids might not be on 18A-U if that's the case. Coral could be 14A, but I have capacity concerns (2029 is the big ramp, pointings towards 2029 Ohio volume) $INTC

AI memory demand remains extremely strong. Supply is still constrained. UBS thinks shortages extend into 2027. Buybacks could become enormous. Wall Street earnings estimates are still too low. Investors should focus less on the next quarter and more on the multi-year earnings power of Micron. $MU

Slowdown? Not today! Nomura’s Chip Shortage Index clock shows semis have been in a Boom phase for 27 consecutive months, surpassing the 26-month pandemic-era cycle and setting the longest boom since 2012.

I guess most of the people actually don’t need state of the art models We need a good enough, just work, and get things done app, and that’s $META Muse

JPM on China memory (Page 8, Q7) Is China a threat? Yes, structurally. Will they crash DRAM/NAND prices soon? Unlikely, given lag + tool bans. What actually constrains them? EUV (and possibly more DUV), yields, customer qualification. Is Chinese HBM bad for the $MU, $SKHY and Samsung? Not immediately — it can tighten conventional DRAM. When do you flip bearish? Yields jump, server/HBM sockets open, or export volumes appear.

The 70% can't generate a single token without the 20%. Compute + networking account for roughly 70% of data center capex. Electrical systems + cooling account for nearly 20%, already more than the cost of the building itself. More powerful AI racks require more electricity and generate more heat. Even if chips get cheaper over time, every new chip still demands more power and cooling. We need more power!!!

No HBM oversupply Even after lowering assumptions, supply still cannot fully meet demand. $SKHY $MU

Hyperscaler capex keeps rising Memory's share of that capex rises even faster The chart is created by combining two UBS tables to visualize the relationship. Page 17 and 18

Shortage through 2027 is basically confirmed by all major banks But I think UBS overstates how cleanly prices can keep compounding into 2028 without a demand or capacity response

Takeaways from Bernstein management meetings in Silicon Valley $INTC 1. Margins beat because of mix, volume, price — not because the turnaround is done. 2. Old plants are very profitable. New 18A plants are not, until they get huge. 3. Boss doesn’t want 60% margins. He wants growth + profit to add to ~45. 4. The cost problem is partly the factories — and a lot the oversized chip designs. 5. Servers: they can sell every chip they can make into 2027. 6. Packaging and custom chips will grow fast and look worse on margins at first. 7. PCs still print cash, but expensive memory is slowing buyers. 8. Real test is 2028, when the industry’s new capacity arrives.

Intel is hunting packaging help outside CoWoS. The latest whisper is AUO and Micro LED glass. $INTC Economic Daily News says it is talking to AUO on Micro LED substrates for co-packaged optics and dense compute chips. Intel would not comment. AUO chair Paul Peng has said the firm is already working with partners on advanced packaging and glass, and would not name them. The reported overlap is glass, through-glass vias, redistribution layers, and light next to the chip. Intel holds a US patent on an IC package with Micro LEDs. The idea is to embed the chip and a glass via in the same glass base so the package can light up, show status, and run optical tests without extra parts. Peng has said AUO is adding trial lines for TGV, RDL, and CPO, plus more Micro LED products this year. Glass antennas for low-earth-orbit satellites are also in the mix. This is still a rumor plus a patent. TSMC remains the volume CoWoS house. Intel is also pushing EMIB for overflow. A panel maker with 30 years of glass is useful. It is not a qualified AI package. If Intel needs glass and light in the same stack, does AUO become a real CPO partner, or just another name on the evaluation list? $TSM

Wafer prices have to rise a lot within two years, or new supply does not get built. China adds capacity. Not enough qualified high-end material to stop 12" utilization running into the high 90s.

One of Russia’s key oil industry facilities and the aggressor’s logistics facility were hit. Ukrainian drones have struck ~90% of Russia’s refining capacity in total and ~73% of that capacity just in August 2026 I will leave the data here without further comment

A chart is worth more than a thousand words Flow into equity funds, US funds drove the net inflows after several weeks of driving net outflows

We are 70% of the way through the year, It's time to start looking at 2027 The most important takeaway isn't how much earnings are expected to grow next year. It's that analysts are still raising their forecasts. Normally, estimates start high and only to get cut over time. This year, they've been moving higher across many different sectors, not just semis. Growth will probably slow down in 2027, but the data so far suggests growth is broadening, not fading. Again, the market doesn't need bad earnings to crash. Earnings can be good, but not good enough for today's standards.