Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Saturday, November 7, 2026

Back to the front page

From X

@convequity 1 click

Convequity on X

$NVDA, $CBRS, and Huawei's Kirin 9050 Pro are three different bets on the same turning point: what the chip industry does when shrinking transistors stops paying. In AI chips, most energy goes into moving data, not into the calculation itself. In our view, as transistors shrink, calculations get cheaper much faster than the wires around them, so the industry's centre of gravity is shifting from ever-smaller transistors to shorter paths for data. That shift is why Huawei's logic folding — designing a circuit vertically to cut the distance data travels — looks potentially transformative and disruptive for the semiconductor industry to us. Here is the comparison in one line: one 16-bit floating-point calculation is the baseline, 1×, about 0.52 picojoules in NVIDIA's published GH200 energy model. Every figure below is the energy spent moving data, measured against that one calculation. 1. Moving data to local SRAM next to the computing units: Cerebras uses about 2× the baseline. NVIDIA's highly optimized local-memory path uses about 5–15× the baseline. 2. Moving data across the chip: we estimate Cerebras's core-to-core network at about 12–37× the baseline. NVIDIA's optimized path through L2, its shared on-chip cache, is about 60–90× the baseline. 3. Moving data off the chip to external memory: Cerebras's MemoryX path is about 330–800× the baseline. NVIDIA's optimized path to HBM, its external high-bandwidth memory, is about 230–290× the baseline. The energy cost is set by how far the data travels: a short hop costs single-digit multiples of one calculation, a trip across the chip costs tens of times, and a trip to external memory costs hundreds of times. If moving data is the real constraint, then shortening paths, which is what Huawei is doing with logic folding, matters more than another process node — for NVIDIA and Cerebras as much as for Huawei. These are engineering estimates, workload-dependent, not measurements of complete systems. The optimized block transfers behind NVIDIA's figures are already counted above; what remains is the physical cost of the wires and memory themselves. In our view, Cerebras's commercial case depends on whether its compiler can keep data close to computation across enough workloads. NVIDIA's advantage is a broader range of software, not a way around the energy cost of moving data.

Investor Letters

All letters

Trading Post November 5, 2026

MercadoLibre’s logistics expansion cut margins by 60 basis points, while Mercado Pago grew credit 62% with 7% NPLs.

MELI $77B
Cassandra Unchained 2 clicks

Thursday Tailwinds #13: Nov 5, 2026

Florida marinas are 92% full as boat counts rise and permits constrain new supply, creating pricing power for existing operators such as…

SUI $14B
Trade Ideas Flow 6 clicks

New Ideas From The Road

Aurelion Research increased healthcare exposure and sees opportunities in banking and defense, alongside research on Alcoa, Nike, ONON and…

AA $10BNKE $60B
Aurelion Research 4 clicks

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 109

YouTube 4

Blogs 15

Seeking Alpha 74 of 275

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

91 accounts
Nominate an account

Investor letters we follow

359 firms

Sectors

Ordered by how often our editor reaches for them.