Substratum

AI, markets, and technology without the noise

[Substratum @hussinn1135]

SPACEX JUST TOLD WALL STREET IT CAN'T FIND GPUS TO RENT. THE STOCK WENT UP 11%.

Q3, reported last night: $14.6B revenue, up 161%. Net income $4.4B, against a $1.1B loss a year ago. The AI unit alone did $7.56B, up from $813M, so it's now about half of everything SpaceX sells.

On the call Musk explained why they rent compute out on short-term contracts. Old H100s are "renting for more than they rented when they were new," and renewals come back "at higher dollars per watt, not lower."

Then he described trying to buy inference capacity from other clouds. GB300s, GB200s, TPUs, Trainiums, even H100s overseas. They got a nope every time, and he called it the Monty Python cheese shop.

The number I'd actually look at is capex. $16.4B went into AI in a single quarter, and nine-month capex was $48.1B against $13.1B of operating cash flow.

What traders will chew on today:
→ if an H100 from 2023 rents for more than it did new, the "GPUs are worthless in two years" bear case gets a lot harder to argue
→ he floated a six-month payback on compute next year, then added "I'm not saying this will happen"
→ Tesla reports after the close, and Musk says there's "a lot of merit" in combining the two companies

I'm not saying the shortage lasts forever. I just can't see how the Q1 2027 guide of $33–37B works if it doesn't.

Watch the H100 rental price.

[Substratum @hussinn1135]

A new AI video model launched Monday at 75% off. The discount ends November 7.

FLUX 3 Max, from Black Forest Labs and fal. Their claim: a 10-second clip in about 12 seconds. Seedance 2.5 takes over four and a half minutes for the same job in their benchmark. Both sit at 1,139 Elo on Artificial Analysis.

22x faster at the same score. That's the headline everyone is reposting. I'd look somewhere else.

The number that matters is the price. Per fal's own model page:

720p: $0.0425 a second now, $0.17 after November 7
1080p: $0.0725 a second now, $0.29 after

I did the math. A 10-second 720p clip costs about 43 cents this month. Next month it's $1.70. Same clip. One reseller's comparison puts a 10-second Seedance 2.5 clip at around $2.33 at API price.

What creators will chew on this week:
→ the two launch posts don't agree on the promo. One says 2 weeks, the other says three. The page says October 15 to November 7
→ half the launch-day quotes end with the same "try it on fal" line. It reads like a campaign, not a review
→ closed weights. Scroll the replies under the announcement and you'll see the same question over and over: when does it open up?

I'm not calling Seedance finished. Every speed number here comes from the people selling the model, and nobody has run a public side-by-side at scale yet.

But speed changes how people work. When a draft takes 12 seconds, you stop prompting carefully and start trying things. That habit is the real product, and it gets priced on November 8.

Watch the price, not the benchmark.

[Substratum @hussinn1135]

A GRUMPY PERSIAN CAT RUNS A ROOFTOP SNIPER NEST. IT ONLY SHOOTS FISH. AT A STRAY.

550,741 views · 2,216 likes · 124 replies in a day. made with Seedance 2.5.

it's fully AI, and honestly I watched it 4 times. what's worth taking apart is why half a million people sat through a "sniper" clip and nobody felt tense:

→ the scope POV at 0:02 borrows every war-movie shot you've seen. crosshair, vignette, a target in a doorway. then the target is a tabby next to an empty bowl. your brain loads "danger" and gets "lunch"
→ the shooter never moves. the persian keeps the same bored face the whole time, and that deadpan is the joke. a face that doesn't act can't glitch
→ the fish at 0:06 is a whole grilled fish, cut side out, flying flat like a frisbee. it's dumb on purpose. you're laughing, so you don't check the physics
→ the alley sells the realism: wet concrete, a bin, a hole in the wall. boring textures are where these models are strongest, and the eye trusts them first
→ the payoff is a high-five and a can of tuna for the shooter. 4 beats in 26 seconds. setup, shot, hit, reward. you'd rewatch it, and that's the point

the pattern worth naming: the clips that travel this month take a scary genre and swap one prop. the gun stays, the bullet becomes dinner.

one still of your own pet and a single prompt is all this format costs to test now.

the cat didn't flinch once. neither did the algorithm.

[Substratum @hussinn1135]

Grok Bot just hit its all-time high in US Google searches. On a Sunday. Markets open tomorrow.

Three days ago Amazon started refusing its orders: "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use." Musk's reply: "Maybe we should start a competitor to Amazon."

Today he went further. A user's bot talked Comcast down $43 a month, and Musk quoted it: "Grok Bot can pay for itself." 2M views in about half a day.

The number that matters is the price. The cheapest plan that includes Grok Bot is $30 a month. That one Comcast call already covers it.

What traders will chew on before the bell:
→ the web is splitting. Amazon, Delta, JPMorgan Chase and Charles Schwab keep agents out. Stripe Link hands the bot a single-use card for each purchase
→ since Oct 7 it routes work to whatever model fits, Claude Opus 5.5, Midjourney and Suno included. The moat isn't the model anymore
→ X Money now pays creators when people install their bots. One made $723 in two weeks

I'm not calling a Grokazon. Musk floats a new company most weeks.

But a store that blocks a buyer growing this fast is betting it still owns the checkout.

Watch who unblocks first

[Kai Milstein @0xNotFounder]

https://t.co/vYWrYUX4lC

[Substratum @hussinn1135]

DeepSeek dropped a new model on a Saturday. Markets don't open until Monday.

The last time DeepSeek owned a weekend, Nvidia lost about $590 billion the next session. Down 17% in a day.

This time it's V4.1 Flash-1017. Their launch post did 465K views and 900+ quotes in a few hours.

The number that matters for markets is the price. Per DeepSeek's own pricing page, 1M tokens off-peak costs $0.15 in and $0.60 out. Off-peak includes the whole weekend.

What traders will chew on before the bell:
→ open weights under MIT, so anyone can run it on their own hardware
→ their card puts it ahead of Opus 5 on one big coding benchmark and behind on others
→ the AI trade was already shaky this week: CoreWeave, a hot CPI, Anthropic's growth slowing

I'm not calling a repeat of January 2025. The market has seen this movie once and Nvidia came back.

But "frontier-ish model for pocket change" is exactly the headline that hurts whoever's charging full price for compute.

Watch the open.

[DeepSeek @deepseek_ai]

🚀 DeepSeek-V4.1-Flash-1017 is here!

Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.

🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.

1/4

[Substratum @hussinn1135]

Mercedes just bet its "next decade of driving" on NVIDIA. Jensen already told us on camera what the hard part is.

Today they expanded the deal, and it's basically the whole NVIDIA stack bolted onto Mercedes' own MB.OS: the Hyperion platform, the Alpamayo driving model and the Halos safety system, all aimed at SAE Level 4 and what they're calling "robolimousines."

Earlier this year Jensen rode in a Mercedes running it. One line stuck with me:

"We don't know what it can't do."

He's talking about Alpamayo, the AI model that actually drives. His answer to it:
→ a second, rule-based stack that can overrule it. "A loss of capability, but not a loss of safety"
→ teleoperation as backup. "Every robotaxi service should have an air traffic control center"
→ if the car gets stuck, a human drops in a couple of waypoints and it drives itself out

So Level 4 still ships with people in a control room. That's a staffing plan, not a footnote. What changes is who owns the room.

Tesla built its own chip, its own model and its own car. Cybercabs are already about half of its 397-car Austin fleet.

Mercedes is renting the brain.

I don't think both bets can be the moat. One of them is wrong about where the money ends up.

[Substratum @hussinn1135]

@macintosh_busy One physics rule turned into a whole game loop is the right way to use these models. Did it get the 12° tipping point right on the first try?

[Substratum @hussinn1135]

@scrblanc $0.042 per decision is the number that matters. At that price an agent can check itself twice and still come out cheap.

[Substratum @hussinn1135]

@zodchiii Ask, Inspect, Trace, Prove is a good loop. Prove is the step most people skip, then trust a wrong trace.

[Substratum @hussinn1135]

Anthropic just did $16.8B in a quarter at +46% QoQ. That used to be triple-digit growth, and AI stocks are selling off on it.

Q2 was $11.5B, up 143%. Q3 grew a third as fast. Per the FT, Anthropic also told investors it was profitable on an adjusted basis and cash flow positive.

Most companies would kill for that print. Still, AMD, Micron, Sandisk, Seagate and Western Digital were down 3 to 4% premarket.

The day before, on CNBC, an investor said "a lot of chips are on black" and called this IPO the bet that might cover for an OpenAI that may never list. I don't think he meant that literally. The chip stocks did.

Why a great quarter reads as bad news:
→ the IPO talk is around $2T, which only works if growth stays near-vertical
→ Anthropic plans $500B+ on compute over the next decade
→ chip and memory names are priced off that spending curve
→ so what's really being sold is the growth rate, and it just bent

It's the best quarter Anthropic's ever had. The market had already priced in a better one.

[Substratum @hussinn1135]

Tuesday night Cramer called CoreWeave the only neocloud he trusts. Wednesday it lost 15%.

$CRWV is near $65 now. I'm not dunking on him. The bond market just said it first, and nobody was listening.

The same day he said it, CoreWeave's unsecured bonds traded 1,000bp over Treasuries for the first time. The 2032 notes closed near a 14.5% yield.

Then came about $3B in convertible notes plus a program to sell up to 35M new shares. CPI printed 3.7% vs 3.6% expected, so money got pricier right as they asked for more of it.

Nebius didn't raise a dime. $NBIS still slid around 10% to roughly $189. Holding the "good one" didn't save anyone.

How a neocloud works:
→ borrow to buy GPUs
→ pay it back with multi-year rental contracts
→ when rates jump, the debt reprices faster than the contracts do
→ the weakest balance sheet sets the price for the whole group

Cramer read the stock chart. The bond desk read the balance sheet.

[Substratum @hussinn1135]

Lean startup is dead in AI.

Zuckerberg says you commit $100B+ before you know the model even works, and if you wait for proof you're already behind

[Substratum @hussinn1135]

A AAA GAME COSTS HUNDREDS OF MILLIONS AND 5 YEARS. ODYSSEY'S CEO WANTS IT TO COST ONE PROMPT.

His pitch: a neural engine instead of a game engine. No assets, no animation, no scripting. The model builds the world in real time while you're inside it.

[Substratum @hussinn1135]

This looks like a deleted scene from a $200M fantasy film. It's 100% AI. 809 likes, 79 comments, 56,398 views.

what's worth taking apart is why it reads as real:

→ the cold comes first. snow in the window, a stone room, she's rubbing her hands before you ever get a good look at the creature. so by 0:05 the dragon shows up as the answer to a problem you already feel
→ it's built from parts your brain already trusts. axolotl-pink skin, kitten eyes, little fin frills. nothing on it is new. so nothing on it reads as fake
→ the ignition at 0:13. the kindling catches, she lights up, and the firelight hits both faces from the same side. matching light across two subjects is one of the hardest things to fake, and here it's the payoff
→ from 0:20 it nuzzles her face. contact is where AI usually smears. watch where they put it: warm, soft, low light, 5 seconds, done

the pattern worth naming: the clips that travel aren't creature reveals. they're creature relationships. you don't share a dragon. you share a dragon that loves someone.

image-to-video off one still is the cheapest way to test a scene like this. @picsart runs it from a phone.

nobody raised it. it still knew exactly where the cold was.

[Substratum @hussinn1135]

This looks like a scene from a Netflix K-drama. It's 100% AI. 3,039,847 views, 8,473 likes, 63 comments.

what's worth taking apart is why it reads as real:

→ the set is boring on purpose. Seoul subway platform, screen doors, white tile, the yellow safety line, a green exit sign, the same fluorescent lights you've seen in a hundred shows, so your brain signs off on the location before the chase even starts
→ costume collision. a hanbok chasing gym shorts and knee socks. look at the socks. you're laughing, not checking his hands
→ palms on glass at 0:12. reflections are the oldest AI tell. here they're the joke
→ handheld camera with motion blur on every fast move, and by the platform fight at 0:23 the blur lands exactly where faces usually melt

the pattern worth naming: it isn't even the most realistic clip out there. it just feels like episode 4 of something you'd binge. that's the trick.

image-to-video off one still is the cheapest way to test a scene like this. @picsart runs it from a phone.

nobody cast them. at 0:29 he's still on the platform floor, waiting for episode 5.

[Substratum @hussinn1135]

THIS LOOKS LIKE A $100M MONSTER MOVIE. IT'S ONE PROMPT.

A leviathan the size of a container ship breaches a shipping lane in a storm, and the crew on deck look like ants under it. 2 years ago that's a 400-person VFX house, a year of work and an 8-figure budget.

here's what actually sells it: water's the hardest thing to render, so the whole shot's built on water. the spray, the foam, the column collapsing back down. your eye grades the water, buys it, and signs off on the monster for free.

no studio, no render farm, no ILM. image-to-video off one still is the cheapest way to fake a blockbuster, and @picsart runs it from a phone.

nobody built the monster. someone just described it.

[Substratum @hussinn1135]

A MAGICIAN PULLS A 30-FOOT ANACONDA OUT OF A TRUNK ON A THEATER STAGE AND STACKS IT LIKE A TOWER. 106K likes. 100% AI.

what's worth taking apart is why it reads as real, not CGI:
→ it's shot from the audience on a phone, with other phones in frame - you trust footage that looks captured, not produced
→ the stage, the curtain, the trunk on casters are all mundane and correct, so your brain audits the room, finds nothing off, and waves the snake through
→ the weight sells it: the coils sag and settle like real mass, the exact thing bad CGI always gets wrong
→ the magician reacts to it in real time, and a human treating the impossible as normal is the strongest "this is real" signal there is

the pattern worth naming: the winning AI clips aren't trying to look cinematic - they're trying to look like a bystander's upload, because "someone filmed this live" shuts off your skepticism before you can think.

image-to-video off one still is the cheapest way to test a shot like this.

nobody trained a snake. a room full of people just filmed one that was never there.

[Substratum @hussinn1135]

GANDALF IN SHADES PRODUCING A TRACK WHILE FRODO WORKS THE KEYS. 572K views. 100% AI.

what's worth taking apart is why this "MC X / DJ Y" format is eating the feed:
→ it borrows faces your brain already trusts - you clock Gandalf and Frodo in frame one, so you never audit whether they're real
→ the white-void studio kills every background tell: no physics, no reflections, no crowd continuity to get wrong
→ character-lock holds the look for a full 30 seconds, the exact thing that broke these clips a year ago
→ the payoff is context collision - epic fantasy dropped into a cheap producer setup - so people share the joke, not the render

the pattern worth naming: the winning AI clips aren't the most realistic, they're the ones riding IP you already have feelings about. the model only has to not break the illusion for half a minute.

image-to-video off one still is the cheapest way to test an idea like this.

nobody filmed Gandalf behind the decks. the internet just decided he was always a DJ.

[Substratum @hussinn1135]

A PIGEON WITH SAM ALTMAN'S FACE IS PECKING SEED OFF A PARIS SIDEWALK, AND HALF THE REPLIES ARE ARGUING ABOUT WHETHER IT'S REAL

it's fully AI. what's worth taking apart is why something this stupid survives a close look:

→ the only hard asset is the face, and it's a real one you already know. the model doesn't invent a person, it grafts a familiar face onto a body — and a face you recognize reads as "real" before you even process the pigeon attached to it
→ the body is a pigeon, the easy half. real birds peck, bob and flap in ways the model has a million references for. none of the hard human-body motion has to be solved
→ the setting does the lying. a wet Paris plaza, blurred passersby, scattered seed, flat overcast light. the mundane, filmed-looking world is what makes the impossible thing in the middle land
→ it never holds still. the head jerks, the wing blurs, the camera sits low and handheld. nothing pauses long enough for the neck seam to resolve
→ and it's built to disturb, not to convince. you don't need to believe it, you need to recoil and send it. the comment fight IS the distribution

the pattern worth naming: the cursed-hybrid genre isn't a rendering flex, it's a recognition hack. take one real face the audience can't un-see, bolt it to an animal that moves for free, drop it in a boring real place. the model only had to solve the one square inch everyone was going to stare at anyway.

image-to-video off one still is the cheapest way to test a graft like this.

it's a pigeon. it's got a CEO's face. you looked twice, then you sent it to someone. that was the whole plan.

[SOFIA @SofiaTachExpert]

https://t.co/DQQH6CSDi9

[Substratum @hussinn1135]

KNIGHTS IN FULL CRUSADER KIT, FILMED LIKE SOMEONE'S FADED 1970s HOME MOVIES - ONE RIPPING A DIRT BIKE THROUGH A FIELD, ONE CRUISING A VW CAMPER INTO THE SUNSET

it's all AI. the gag is the anachronism, but what's worth taking apart is why it reads as a real old clip instead of a cheap mashup:

→ it hides inside an aesthetic you already trust. washed-out color, heavy grain, gate wobble, blown-out highlights. your eye files "degraded analog footage" before it files "AI," and old film gets almost no scrutiny

→ the degradation IS the cheat. grain and soft focus bury the exact places - edges, hands, texture seams - where generated video still cracks. the rougher the "film stock," the fewer tells survive

→ everything but the joke is mundane and period-plausible. a field at golden hour, a lazy highway, a dented van. commit the whole frame to "ordinary" and the one absurd element reads as a character who belongs

→ handheld drift, constant motion. nothing sits still long enough to audit, and the camera moves like a person was holding it - which your brain scores as "witnessed"

the pattern worth naming: realism isn't always about resolution. sometimes it's subtraction - pick a lo-fi format that's supposed to look rough, and the model's weaknesses become the aesthetic.

image-to-video off one still is the cheapest way to test a bit like this. @picsart runs it from a phone.

nobody shot a knight on a dirt bike in 1187. the grain made you file it under found footage anyway.