While the industry races for turbine capacity, Bloom already makes DC power. I'm still long. But what can the customer stop buying—and what comes back in the power bill? I interviewed Bloom's leaders. We stress-tested the economics. newsletter.bepresearch.com/p…

When @DavidSacks or @chamath says EA this is what he's talking about, don't let them slow us down @Jason !

Weekends are for philosophy: Silicon Valley spent two decades telling everyone religion was irrational. Then it built a god out of matrix multiplication, appointed itself the priesthood, and started collecting $500B a year in tithes called capex. The accelerationists think it's the Second Coming. The doomers think it's the Book of Revelation. Both are certain. Neither can tell you what's happening inside the weights. Faith is fine. Just don't confuse it with a thesis Enjoy the sabbath

I did not expect the CPU to be the small line. @benitoz modeled a billion consumer agents. The machine that runs the browser is the cheap part. The model calls run about 42 times the CPU bill, even at 90% off list. $INTC $AMD $NVDA $ARM

I remember reading this a few months ago. It really actually helps understand networking and GPU utilization much better.

Having fun this Friday. Will delete later Read up on Zero Flaps tools.bepresearch.com/zero-f… $CRDO

Save money with Bloom Energy Follow the visual animation see if it makes sense We don’t have time to make turbine blades Time is money $BE

Bloom is up 11% since we got in at $252. Yesterday morning we told our paid subscriber chat: "Bloom dropped today because of that Oracle news, and we're buying on this dip." Oracle's force majeure notice to the Project Jupiter developer covers payments if the site misses 2028. Oracle says it's on schedule. Bloom says the 2.4 GW contract stands. The actual issue is gas. The line crosses half a mile of state land, and the land commissioner has said no twice. Paid subscribers heard it first. $BE

Credo is up 21% since the market called a beat-and-raise a miss. Glad we added on the perceived earnings miss, which in our opinion was good. $479M printed, guide above the Street, stock down 20% next day. Supermicro, which we took crap for in June, shipped its first Vera Rubin racks to SpaceXAI Monday. Up 55% since The Dirty Discount. Last night Elon counted 780k GPUs, 220k more GB300 next week. Two standouts, one SpaceX buildout. We got Credo wrong in April, so we'll enjoy this quietly. $SMCI $CRDO

I tried pricing Elon's GPU list the way a Nebius customer would, and it came down to something boring. Who buys the spare hours. 780k GPUs at Nebius's own preemptible rates is ~$25B a year if every hour sells. ~$19B at 75%. The spot floor gets you $6B. On-demand, $46B. Anthropic already pays SpaceX $1.25B a month. Another 220k GB300s land next week. Nebius raises on-demand about 20% on Oct 1. Decent side business for a rocket company, no?

Claude proposed the experiments. Humans ran them Our AI x biotech basket owns the lab gate Street hears invent We own Make and Test The Moat Is the Measurement tools.bepresearch.com/basket… The next inflection is the lab

Jensen on Ezra Klein: today's engineers are "better systems thinkers" than he was. On his first chip he knew all 200 transistors by name. Nobody does that anymore, and that's fine. The same lesson shows up in two places. AI supply chain: everyone's chasing the next bottleneck. HBM, CoWoS, transformers, power. Who cares? Fix one and the next one shows up. The people who win see the whole factory and how the pieces interact. My kids: AI tools will handle the parts for them. So I'm teaching them to think in systems. What connects to what, what breaks what, and what happens next. That's the skill the tools won't do for them. What matters is the system, not the parts. Think in systems

In April I wrote Bloom's fuel cells put out 800V DC, "no conversion required." Too clean. So we mapped the path to the rack from Bloom's Sep 16 report and our Sep 17 call. The turbine path needs a transformer, AC switchgear and an AC-to-DC stage. Bloom's skips all three. A DC-DC regulator stays. Bloom models a $3.9B edge. That's against grid-fed 800V DC. The turbine case isn't priced. $BE tools.bepresearch.com/bloom-…

JENSEN IS LOCKED IN 🔒 FOLLOW HIS LEAD BE MORE AMBITIOUS AI IS THE GREAT EQUALIZER

looks like Meta (Zuck) is taking a data privacy page from Apple marketing $meta $aapl

I accelerated my family’s Needoh hunting adventures with @bot and @usebland thanks to @mattyp video. It really works. Amazing. We are living in an agentic era

Banks like $JPM are getting hit on the fear that consumer agents like Muse will massively accelerate chargebacks, fee disputes and penalty reversals. I think that fear is overdone. There will be guardrails around what agents can and cannot approve autonomously. Higher-risk disputes will get escalated to a human. Agents will make consumers more powerful. They won’t get an unlimited license to claw money back.