▶ YouTube Summaries

• https://www.youtube.com/watch?v=nGqGrrYeYGc
• Original title: The Craziest AI Pivot yet

Midjourney, the image-generation company, has announced a pivot into medicine: a tube you step into that vibrates water and uses ultrasonic computational tomography to image your body's soft tissue. Prime isn't interested in mocking the medical efficacy or the spa-themed rollout — instead he runs the numbers in the company's own blog and technical video and finds the data-throughput, processing, and storage figures wildly implausible. His verdict: the goal of a billion scans a month from 50,000 machines is arithmetically impossible at the stated data volumes, and he suspects the whole thing may be vaporware — while still respecting that Midjourney is at least attempting something concrete in medicine.

What Midjourney announced (the source material)

• Midjourney is entering the medical arena with a device Prime describes as a tube you enter that vibrates water; measuring the resulting vibrations produces a reasonably accurate image of the body's soft tissue (ultrasonic scanning, pitched as an alternative to MRI).
• The official name: "full body ultra-sonic computational tomography" — which Prime mocks for the absurdity of its acronym.
• Accompanying materials: a slick, futuristic promo video ("they look so futuristic, I've seen this on aliens"), a blog post on the rollout strategy, and a technical breakdown video on how it works.
• Rollout plan from the blog: Midjourney becomes a spa. First location opens in San Francisco in 2027 with hot tubs, saunas, cold plunges, and "cozy rooms with pools of golden light" where you scan your body. (Prime jokes the golden room is the construct from Halo.)
• Stated ambition: by 2031, a fleet of over 50,000 scanners worldwide with capacity of one billion scans per month.
• Stated mission: monthly self-scanning lets you catch organ shape/density changes early, enabling faster root-cause analysis and potentially saving or extending lives.
• Prime explicitly sets aside the medical questions (ultrasonic vs MRI efficacy) and the privacy joke (getting your MRI results on a Discord server). His focus is purely the data/throughput numbers.

Prime's first reality check: scans per machine

• 50,000 machines, 1 billion scans/month, 30 days × 24 h × 60 min × 60 s → each machine must scan someone roughly every 129.6 seconds, 24/7, continuously.
• If a single scan takes ~60 seconds, the machine must scan-out and reset constantly with almost no slack — Prime calls it already implausible.
• Even doubling to 100,000 machines only gets you to ~one person every 4 minutes, around the clock — still feels undoable.
• He notes this is the least crazy number in the video; everything after gets more absurd.

The technical-video data numbers

• At minute 3:32 of Midjourney's technical breakdown: 306 TB of raw data generated, reconstructed across 21 servers.
• Each body "slice" takes ~40 GB to move through the system. Slices are very thin — claimed ~0.1–0.5 mm (others estimate ~2 mm).
• System captures ~17 GB per second. Goal is to capture "several hundred" slices per body (Prime insists "several" means at least five hundred-ish, i.e. many).
• Prime's reconstruction: 17 GB/s × 60 s ≈ 1 TB per scan → ~500 images at 1–2 GB each (some estimates say 300). High-resolution, medical-grade interior images, so large files make sense.

The data-transfer problem

• Architecture: the tube machine on one side, 21 servers on the other, with a connection between for image processing (raw analog sensor data can't become images without processing).
• The blog/video mentions ~800 TB of raw data. Prime calls this number "just crazy."
• Even with the world's fastest switch at 1.6 TB/s, moving 800 TB takes ~66 minutes; splitting across 21 connections still ~3 minutes — assuming zero latency, zero frame drops, 100% utilization. Impossible against a 129 s budget.
• His charitable interpretation: the 800 TB / 40 GB-per-image must get reduced on the machine before transfer (maybe 40 MB, or 40 GB per image fanned to the 21 machines). Even ~20 TB total transfer eats heavily into the 129 s window.

The processing problem

• Pretend transfer is free; consider only frame production. Reference point: the "1 Billion Row Challenge" (memory-map a ~12 GB file, integer-parse, compute a statistic) — fastest times ~1.5 s, some sub-1 s.
• Raw image data is ~40 GB and likely needs more processing than the 1BRC.
• Being generous: 500 images across 21 servers at ~3 s each ≈ 71 seconds just to process — already blowing past the 129 s budget before transfer, read/write, or RAM constraints.

The storage and internet-bandwidth problem

• 1 billion scans/month × ~500 × 1 GB images ≈ 16 exabytes/month. Even gzipped down to ~4 exabytes.
• That ~4 EB represents roughly 12.5–25% of all internet transfer today — just to move medical images.
• For scale: ~4 exabytes is about what Netflix transfers every single day. Midjourney would become "Netflix" purely off medical images.
• Plus the storage: massive data warehouses, "a dedicated nuclear reactor just to power the hard drives." And none of this even accounts for RAM read/disk write speeds or spillover when memory fills.
• Conclusion: you won't hit a billion scans/month with 50,000 machines — you'd need maybe 500,000.

Prime's overall verdict

• Cost objection: a spa holds ~10 units, so 50,000 machines ≈ 5,000 spas worldwide, each needing a ~$10M server farm. (He jokingly says nobody can multiply 5,000 × $10M — the numbers are too big — landing on ~$100 billion, "as much as an AI laboratory.")
• The hard, unanswered question: how do you make this affordable to a billion people at these data volumes?
• Tone: not a hit piece. He's genuinely impressed that Midjourney — an image company — is attempting more in medicine than OpenAI, Anthropic, or others he's seen. The early-detection mission is admirable.
• But he flatly doesn't believe the numbers: "I'm glad to be proved wrong." His closing plea: "Please don't let this be vaporware. Please don't let this be the greatest rug pull of all time."