Lucas ✈️ CoRL

billions will thrive
robot inference @hf0, ex-quant 2x

[Lucas ✈️ CoRL @quantbagel]

I’m hiring for my robotics inference startup, we’re working on solving fast cheap long context batch invariant Inference for VLAs and WMs, billions will thrive!

[Lucas ✈️ CoRL @quantbagel]

@KyrieBlunders flamecharts😭

[Lucas ✈️ CoRL @quantbagel]

Me circa aug 2026 looking at our failed trajectories

[Daniel Yang (Toronto Nov 20-23) @punished_daniel]

Kind of fucked up how the moment an Asian guy experiences hypergamy society knocks him down

[Minh Nhat Nguyen 🦭 @menhguin]

one thing I noticed abt "bubble discourse" is that people really want to reuse comparisons. railroads, radio, telephone, real estate and dotcom all played out in meaningfully different ways, and you'd be wrong if you actually tried to reason about them exactly the same.

like for example, railroads do not actually have the same core economics as AI. railroads are closer to a fiber buildout bubble, bc they connect points and scale w and stimulate demand between points. two wealthy cities benefit from connecting railroads up to the point of their economic value/traffic, and saturate beyond that. but data centers serving one wealthy city are not similarly affected by the presence of surrounding cities. this is important bc the specific mechanism for railroad bubble popping (railroads saturated their competitors' routes with capex and drove margins to zero) is less strong for AI. you cannot pattern match the exact same way.

[Lucas ✈️ CoRL @quantbagel]

@Midnight_Captl @sur4js wtf is guppy level

[Lucas ✈️ CoRL @quantbagel]

@glennko @mhdempsey That’s what the data is for

[Pranjal @pranjalssh]

Megakernel blog is out now

We implement a full layer (MoE + Attention) megakernel, achieving 1100 toks/s for Muse Spark, 2.1x faster than baseline.

Nvidia's Instruction cache makes it really hard to write bigger kernels. We ended up tuning both instruction and data caches at all levels for extreme latency hiding!

With this, some warps of next operation can start running before previous operation has finished - redefining what GPU utilization means.

Full blog post: https://t.co/ubZbSQpkKl

[Lucas ✈️ CoRL @quantbagel]

@jehovahscript Git for the economy, lowk Palantir is like this

[Mike Taylor @hammer_mt]

Worst week of Jensen's life

[Baron H. @Dunetsk]

I wonder what the cartels are building with Claude code rn

[Stanley Lee @stanleylee21]

CoreWeave $CRWV is leaning harder on customer prepayments and returning to IG offtake for its next DDTL because public credit got flooded with supply and drew the line. That's as much a sign of compute's first credit cycle as it is a "broken" market.

In these cycles, the financing burden shifts to whoever can carry it. At the top of the mkt that's IG balance sheets, who have access to some of the world's cheapest capital. In the middle market it's actually venture rounds, the most expensive, carrying this through prepayments as lenders take a step back/have re-racked levels wider.

This actually makes inference company raises, not neocloud spreads, the signal to watch in the medium term. As long as those keep getting done, customers keep paying and lenders get comfortable funding more of each cluster again. Over time, prepays come back down and cheaper capital finds its way back to the good operators. Same path every financed asset class takes.

[Robert Lukoszko @Karmedge]

the worst nightmare of every robotics data lab is that simulations start to work

[Qiayuan Liao @qiayuanliao]

SuperMimic is open-sourced during my talk at CORL.

Have fun with it: https://t.co/VX3Z0cZlk7

[Carson Poole @CarsonPoole]

Of course that’s your contention.

You’re a first-year inference engineer.

You just read SemiAnalysis’s blog and saw megakernels coming to TPUs and AMD, so naturally you believe that’s the way to get to speed of light.

Next month, when you get your hands on an inference engineering cookbook—Baseten’s, probably—you’ll be singing from the treetops about the 500 nanoseconds of critical-path latency you save on every GEMM pipeline drain.

That’ll last until sometime next year, when you watch a pirated geohotz lecture livestream and become convinced that PDL gives you a strict superset of megakernel capabilities.

“Well, as a matter of fact, I won’t, because PDL introduces—”

“PDL introduces unnecessary stalls because its synchronization mechanism is very coarse.”

Yeah, I read that too. HazyResearch’s blog, “Look Ma, No Bubbles!” Just below Figure 2.

Do you have any thoughts of your own on the subject, or should I just point my agent at the SGLang repo and have it do it for me?

[Noah Ziems @NoahZiems]

Fun Fact: Basically none of the initial railroad companies made money. They overbuilt, took on way too much debt, and ended up in brutal pricing competition with each other, and nearly all went bankrupt.

[Lucas ✈️ CoRL @quantbagel]

@bill_xby do dflashv2

[Lucas ✈️ CoRL @quantbagel]

@_coopergadd @benjaminakar Tinder for horses is a great idea, ? marketplace for breeders or to get a ton of polygenic testing data

[Lihong | Freeport @Lihong062]

$CRWV reports tomorrow.

A few equity vol pods and prop desks believe this is one of the more consequential prints this quarter, with ripples into other Neoclouds and semis slightly underpriced, especially on the downside.

We now provide a brief note covering the operating model and key expectations, what matters in the print and on the call, our scenario analysis , potential read-throughs, and our longer-term views on industry economics.

The two competing narratives regarding Neoclouds are (1) compute scarcity makes everyone a cash machine vs (2) financing and depreciation eventually overwhelm the economics and they all go to 0.

On selective Neoclouds some specialized desks are selling CDS against long puts betting that debt is actually cheap to equity (this has done well in the past month).

Q3 Consensus: ~$3.5B rev, ~$240M adjusted EBIT, ~$12.5B capex, -$10.5B FCF, 1.7GW.

Q4 Consensus: ~$4.5B rev, ~$675M adjusted EBIT, ~$8.5B capex,

FY26 ~$13B rev, ~$1.1B adjusted EBIT, ~$38B capex, and ~2GW.

There are ~$130B backlogs so investors mostly care about execution e.g. revenue beat with delays could be worse than a miss with deployment accelerating.

In the print algos and traders will be watching: active MW delivered, margins & operating income, capex, and Q4/FY27 guidance.

The story being told the above numbers matter, e.g. more capex can be faster deployment or overruns; less can be efficiency or delays.

Algos / traders will be reading out the “story” in the minutes post print.

On the call we expect analysts to ask about how the next GW financed, at what spreads / terms? Are projects on schedule? What are marginal compute prices and unit economics?

Options imply ~20% moves. Historical implied moves averaged closer to 15% and realized moves were closer to 10% close to open.

We nonetheless think this is fair, since stock has been more volatile recently and earnings size is correlated with diffusive volatility.

We also think this earnings is modal
up (60%) tail down (40%) with more violent downward moves.

Word on the street is that the team has been executing well and we expect a good quarter, but the thinking goes that a good quarter can’t disprove financing and execution concerns but one bad quarter can validate them.

First-order read-throughs are to other Neoclouds $IREN (already reported), $NBIS (reports Nov. 11), and $GLXY $RIOT $MARA $CIFR $WULF $CLSK $HUT $CORZ $BTDR.

We think ripples on these are slightly underpriced especially on the downside, though options illiquidity makes it hard to trade.

The asymmetry is that with current sentiment, a durable bullish semis/macro read-through takes more than just one earnings. However, a bad surprise reprices peers’ cost of capital immediately.

We believe this potentially applies to the rest of the stack but to a lesser extent based on sentiment

compute $NVDA $AMD $AVGO

memory $MU $SKHY $SNDK

photonics $AAOI $COHR $LITE

semicap $AMAT $LRCX $KLAC

power $GEV $VRT $BE

Tomorrow will mostly trade on sentiment and positioning.

However, structurally, we’re long term bullish power, memory, interconnects, but bearish neocloud / inference margins,

since lower compute prices will drive token demand without protecting operator margins.

[Lucas ✈️ CoRL @quantbagel]

@jmdagdelen @aj_hugs @hugothomel This is how @taurobots should be delivered

[lindsay @puterfaction]

making claude talk like nick fuentes so i actually pay attention