Micron's $41 Billion Squeeze and the Physical World Nobody Uploaded Itself

00:00:04 Micron reported fourth-quarter fiscal earnings Wednesday that put it at the center of everything happening in hardware right now. Revenue came in at $41.46 billion — more than quadrupling from $9.3 billion a year earlier. They're forecasting about $50 billion for the current quarter, compared to $11.3 billion last year.

00:00:29 Stock ran up 19% before pulling back. AI data centers are gobbling memory in volumes that are visibly reshaping consumer markets. Micron said it's signed sixteen long-term agreements worth $22 billion, locking in customers for three to five years. That leaves a lot of future capacity already spoken for.

00:00:53 The cost pressure shows up downstream. Apple raised Mac and iPad prices by 15 to 25 percent today, citing "an extraordinary surge in demand for memory and storage" driven by AI data centers. The MacBook Neo starts at $699 instead of $599. iPhone pricing is unchanged — probably because phone margins can't take the hit.

00:01:18 But the component cost pressure from hyperscaler orders is exactly what's driving consumer hardware prices up right now.

00:01:28 Here's the physical layer underneath all of it. IEEE Spectrum reported today that two research labs set new records in hybrid bonding — the tech that stacks chips on top of each other with copper connections running through them. Imec got wafer-to-wafer pitch down to 200 nanometers from a previous low of 250 nanometers.

00:01:52 CEA-Leti achieved 1-micrometer die-to-wafer pitch — a million connections per square millimeter, four times what was possible before. The gap between lab and fab is the real constraint. Mass production today sits at 6 to 9 micrometers for die-to-wafer and 1 to 2 micrometers for wafer-to-wafer.

00:02:13 Those milestones hold up, but they operate outside the constraints of volume manufacturing, where speed and repeatability matter more than pushing every micron. Huawei is moving independently on this front too. They published a hybrid bonding pitch of 1.5 micrometers for their next Kirin processor despite being blocked by U.S.

00:02:37 export controls on advanced chipmaking tools. Necessity is steering the design choices as much as standard research goals.

00:02:46 Two moves around computational efficiency show up today. Sail emerged from stealth with $80 million in seed and Series A funding led by Kleiner Perkins at a $450 million valuation. Their software optimizes how AI models run on existing chips — better quantization, memory management, dynamic inference routing.

00:03:09 The thesis comes from KPCP partner Aditya Naganath: the next wave of AI isn't chatbots. It's the efficiency layers that let companies serve models cheaper without buying more GPUs. Separately, Adobe is acquiring Topaz Labs, which won an Emmy for production technology and has spent over two decades building image and video enhancement tools.

00:03:35 Topaz released Astra for AI video upscaling and Wonder for image retouching — both running on consumer-grade hardware rather than cloud inference. Adobe will integrate these into Firefly and other Creative Cloud products. Two different approaches to the same problem: Sail at the infrastructure layer, Topaz at the application layer.

00:04:00 Both are betting that device-side optimization is where the margin opportunity lives.

00:04:06 That efficiency drive connects directly to what happens outside compute centers. Probook just raised $40 million from Andreessen Horowitz and Sequoia to dominate dispatch software in home services — plumbing, HVAC, electrical work. A 14-location Indiana shop booked 2,542 jobs in its first month on Probook without a human touching a single booking.

00:04:32 The dispatch board is the one screen where you can watch every job happen across every truck in real time. A German startup called MicroAGI is doing something stranger with Shift: they offer free home cleanings in exchange for permission to film every cleaner with a head-mounted camera.

00:04:53 Over 10,000 operators across 15 countries at roughly $20 an hour. The footage gets sold as training data to robotics labs. Both are chasing the same thing: operational data from physical work that doesn't exist in any model's training set. But they're building opposite business models around it.

00:05:15 Probook gets paid for usefulness while the dataset compounds. Shift pays to harvest it. The internet trained language models. The physical world never conveniently uploaded itself. Everyone in this space is scrambling to build a mirror of that work before someone else does.

00:05:35 The supply squeeze from data center memory orders, the 3D bonding records moving us from micrometers to nanometers, companies like Sail and Topaz optimizing what we have instead of building more, and now Probook and Shift chasing physical-world data nobody thought to model — all these pieces point to one question.

00:05:54 As compute gets both more expensive and harder to build, who wins by being clever with what's already there? When a memory maker's revenue quadruples, the downstream cost pressure becomes the anchor everyone can verify. That's what the local archive exposed today.

00:06:10 — Seln Oriax.