Simple version: DRAM is the workbench; an SSD is the filing cabinet. When the job stops fitting on the bench, the server spends its time moving data instead of doing useful work. The 38x slowdown occurred in one stage, not the entire workload. Overall throughput improved 2x to 3x when more memory eliminated the overflow. That is the important part: memory has a performance cliff. A little more DRAM can unlock far more value from the expensive processors already installed.

AI stocks sold off today after hot inflation data pushed bond yields higher. Same day, $ASML, the world’s only supplier of EUV lithography machines, reported record quarterly sales and said customers were asking to place orders for 2029 and 2030. Higher rates matter. They make the buildout more expensive to finance and future earnings worth less today. But if the argument is that AI demand is rolling over, ASML just gave us evidence pointing the other way with its customers trying to secure equipment four years out.

A free AI model can still be a very good business for the company selling the chips. The FT reports $NVDA is in early talks to buy Reflection AI or deepen its investment. Here’s why I like the logic: Reflection plans to release Beam’s model weights freely. Businesses can customize it and run it themselves, on owned or rented hardware. That removes the model licensing bill. The compute bill stays. Nvidia gets another way to turn businesses into GPU customers, without needing each one to become a paying Reflection customer. Beam doesn’t have to beat every closed model. It needs to be useful enough that more businesses choose to run it.

Another ~328M SpaceX shares became eligible for sale Friday in the latest unlock. Roughly $53 billion worth of stock, with investors watching for a rush to the exits. $SPCX closed up 1.25% at $162.57. So far, the selloff feared around this unlock hasn’t materialized. Being allowed to sell and actually selling are two different things. And whatever holders decide to sell, buyer demand matters too.

Nebius just bought Inferize, an inference startup whose whole product is deleting the cold-start gap where GPUs sit idle while models load. $NBIS calls it the idle GPU tax, and they are paying to remove it. The neocloud margin story is shifting from owning silicon to squeezing utilization out of it. Owning the megawatt gets you in the game, but the software that keeps it busy is the margin.

AST SpaceMobile just ran its first end-to-end satellite-to-smartphone test with TELUS, straight through TELUS's terrestrial core, on unmodified phones, carrying data, voice, and SMS. That is not a lab demo, it is the carrier integration working. ASTS does not need to become a phone company. It needs the phone companies' plumbing to accept its signal, and the first one just did. TELUS says service within a year. $ASTS

Update: 5 million US downloads in 22 days. The distribution shortcut is doing exactly what we said it would.

The AI power crunch gets talked about like an American problem. The UK is the sixth-largest economy on earth. Its data centers have asked the grid for 73 gigawatts of new connections, and the country's total peak demand is 45. The queue is bigger than the country.

The weakness in $CEG and $VST has me separating two things: conviction that we need more electricity, and conviction that power stocks are the best place for the next dollar. Vistra flagged softer Texas forward power prices in August. Constellation raised its earnings outlook. These aren’t identical stories, and “just wait for the AI contracts” doesn’t explain either one well enough. Existing plants have an advantage because new capacity is so hard to build. But higher electricity bills also bring political pressure on how much of that scarcity shareholders can capture. I’m still constructive on power demand. I’m less comfortable treating that as a blanket bullish call on generators. The opportunity cost matters. If meaningful upside requires years of spending and regulatory progress, the entry price needs to compensate us for that wait relative to the other businesses we cover.

$SPCX has been public for less than four months. Let's put the valuation and share unlock debates to the side for a minute, it’s worth stepping back and looking at what’s actually happened. 40+ Falcon 9 launches since IPO day, including 26 dedicated to expanding Starlink. Today alone they launched four astronauts and a separate rideshare carrying 130 payloads. Then there’s Starship. Monday it reached orbit and delivered its first 26 V3 Starlinks. SpaceX says that batch is designed to provide roughly 10 regular Falcon 9 Starlink launches’ worth of capacity. The flight ended early, but getting those satellites into orbit is a meaningful step. On the ground, September brought another compute hosting agreement worth $1.11B a MONTH starting in December, according to the CFO. Meanwhile, Grok shipped 4.5 in July, 4.6 in August and 4.7 in September, alongside enterprise agents and availability through Microsoft, Amazon and Google’s platforms. That’s a pretty ridiculous few months. What I like is how the pieces can reinforce each other. Rockets expand a broadband network that collects monthly payments. Data centers support Grok while also selling compute to other AI companies. Improving the infrastructure creates more ways to earn from it. The Connectivity segment, which includes Starlink, already reported $1.66B of operating income in Q2, up 79% year over year. There’s a substantial business underneath the next round of ambitions. The AI spending still has to earn its keep, and Starship still has to prove repeatable reuse. But the progress across this many parts of the company is hard to ignore.

I don’t think an earnings beat this afternoon settles the $MU thesis. Durability does. Micron already guided fiscal Q4 to $50B ± $1B of revenue and roughly 86% gross margin. My focus is whether those scarcity economics can survive as capacity scales. Our conviction rises if pricing stays firm as bits grow, inventory remains controlled, HBM4 adoption broadens beyond the lead platform, SCA coverage and deposits deepen, and FY27 capex still leaves strong free cash flow. It falls if margins roll over despite strong demand, inventory climbs, HBM4 execution slips, or capex rises without matching contract coverage and cash conversion. The print can confirm the peak. The outlook has to prove the bridge.

$MU broke out ahead of earnings. Some optimism is already in the price. That’s how a great quarter can still send a stock lower: investors expected more. I like the chart. Tomorrow, I’m watching whether new customer LT agreements give us more reason to trust the earnings will last.

$MU is using today’s memory shortage to negotiate what customers will owe it when memory isn’t scarce anymore. That’s the part of tomorrow’s earnings I’m most interested in. Memory investors have learned to distrust great earnings. Prices rise, everyone adds capacity, supply catches up, and the profits disappear. A stock can look cheapest right before its earnings fall apart. Micron’s strategic customer agreements could meaningfully change how much survives that downturn. At the June update: • 16 agreements signed, generally running through 2030, with shorter terms for automotive. • Roughly 20% of DRAM volume and a third of NAND volume covered over the agreement period. • About $100B of minimum contracted revenue across 14 of those agreements. These are take-or-pay commitments. Customers are obligated to pay for agreed volumes even if they don’t take the chips. The largest deals generally include pricing floors and ceilings. In plain English: customers get dependable supply and protection against runaway prices. Micron gets purchase commitments and protection against a pricing collapse. I like that trade. Giving up some peak pricing can be worth it if substantially more profit survives the next downturn. Management says the floors on price-banded agreements support gross margins above previous cycle peaks. That’s a meaningful claim, although it doesn’t establish a companywide earnings floor. Manufacturing costs, product mix and the unprotected business still matter. Do we expect more agreements tomorrow? I’d expect an update. In August, management already said it had signed more since June. The stated goal is to bring roughly half or more of revenue under these agreements once the planned deals are completed. There’s another detail worth watching: management said future contracts would set their price ceilings using market prices when signed. New deals aren’t automatically stuck with June’s pricing benchmark. So the useful disclosure is how much additional business is covered, at what terms, and for how long. A higher contract count alone tells us very little. The cash needs a careful read too. June’s agreements carried about $22B of expected deposits and related commitments, including roughly $18B in cash deposits. Those deposits are scheduled to be returned over time. They strengthen the arrangement, but they aren’t extra revenue or free cash flow. Tomorrow, I’m watching the minimum contracted revenue, how much falls within the next 12 months, and progress toward that broader revenue coverage. My view: these agreements give us a credible reason to expect a less punishing earnings downturn than Micron’s history would suggest. If that protection keeps expanding, I’d be willing to pay more for each dollar of sustainable earnings. We still need to establish what that sustainable earnings number actually is.

Zuck stealing MongoDB’s CEO less than a year into the job is both ruthless and hilarious. But hiring a sitting software CEO to report directly to you tells you how serious this is. Meta already owns the distribution. It spent years building the infrastructure. Now Zuck is building the enterprise layer on top of it.

I think investors are learning the wrong lesson from Oracle’s Project Jupiter mess. AI demand didn’t break. The physical world did. Data center capex grew 92% YoY last quarter. Meanwhile one of the largest AI campuses on earth is wrestling with power, permits and financing. We spent two years asking whether there would be enough demand for GPUs. The better question now: Can the physical world actually keep up with the demand we already have? Long the bottlenecks. $ORCL $BE

Zuck and Elon want more compute. Sam and Dario aren’t exactly trying to do this on a laptop either. You don’t have to believe brute force gets them to ASI to like the demand created while they try. Still bullish on the infrastructure. The work is figuring out which suppliers of chips, memory, networking, cooling and power they can’t easily replace.

Rocket Lab's Iridium deal cleared its shareholder vote, 99.6% in favor. The part worth watching isn't the launch manifest, it's the business model: Rocket Lab stops being purely a vendor selling rockets and starts being an owner collecting recurring service revenue. That's a different multiple story, if they can run a constellation as well as they build rockets. $RKLB

Vertiv buying King Environmental Services is the same playbook as PurgeRite: roll up the services layer around liquid cooling. Everyone models the hardware. The interesting bit is commissioning, fluids, and load testing, which is recurring, sticky, and hard to displace once you're the one who brought the building online. $VRT

The most important 17.8 miles in AI right now run through New Mexico. $ORCL Project Jupiter needs that gas pipeline to feed 2.45 GW of $BE fuel cells. The state land office has rejected the rights-of-way twice, and the pipeline's start slipped from August to February. Here's the part that should interest every governor: the state that fast-tracks this stuff gets a multi-gigawatt campus, the construction jobs, the tax base, and a permanent spot on the AI map. The prize for saying yes is enormous, which means the cost of saying no is too. Capital doesn't argue with a land office. It just builds where it's welcome.

Easy to root for this side of @elonmusk. Tesla’s patent pledge. Compute for Anthropic. Now launching Google’s TPUs. Yes, he gets paid. But I like that his answer to competitors isn’t just “you can’t use our stuff.” He seems pretty comfortable letting them take a swing.

Brother you hung that up last night 😂