DanielRomero

$GOOG Head of Infra:
“A Gemini prompt now consumes the equivalent of 7 seconds of TV.”

$GOOG showing off its 1.09 PUE
Very impressive, actually

$GOOG Head of Infrastructure:
“Energy is perhaps the biggest challenge.”

$GOOG Amin Vahdat is such an impressive speaker
The difference in presentation quality between American and Asian companies is a whole different world
Even when Asian companies present in their native language, they often don’t communicate information as American companies do

First official day of SEMICON Taiwan
This place is packed!
The booths are already full of people
$GOOG Amin Vahdat will be the first to present

$AMD says hybrid-bonded 3D DRAM is 17× more energy-efficient than HBM using microbumps

$SKHY claims up to 5.15x LLM inference improvement with Custom HBM
Looking further ahead, the company is exploring a new type of memory:
- Compute-using DRAM
This would allow DRAM to perform certain computations directly inside the memory itself, reducing the need to move data back and forth between memory and the processor

I'd say the themes capturing the most attention at SEMICON this year are:
1. Advanced packaging
2. Memory
3. Photonics
Glass substrates are getting a lot of attention as well. I didn't expect that, but people were really pumped for Unimicron
Big vibe shift with $AMD too. People were really interested in them

$AAPL being the second-largest company in the world by market cap in 2026, despite having barely any direct AI exposure, is genuinely impressive
No data centers, no frontier AI models, just good ol’ smartphones and laptops

Unimicron says TGV yields are still poor and is targeting 2029/2030 for mass production of glass-core substrates

Glass will be used alongside ABF build-up layers and PID, an insulating polymer layer
This substrate material transition is crucial for FOPLP

$LITE likes VCSELs for AI scale-up interconnects
Their main limitation is reach

Samsung’s zNAND-O looks amazing
10× DRAM bit density, with 7× higher read bandwidth and 7× better read power efficiency than conventional flash
3D NAND could offload a lot memory capacity from HBM, particularly for inference workloads where not all data needs HBM-class latency
First samples targeted for 2028

The US manufactured around 450 humanoid robots in 2025
China manufactured 20,000
If US humanoid production increases 1,000x from here,
Which stocks stand to capture the most value from that growth?