Most "GPU exposure" has nothing underneath it you could point to. Ours starts with a serial number. Every Device on AION is split into exactly 40 Fractions. One Fraction is one Compute Unit, recorded against that specific machine: manufacturer, model, serial. Forty is a hard cap in the purchase agreement. A Device never gets a 41st. We like this constraint more than we expected to. It forces every conversation back to one physical box and the hours it actually ran, instead of a blended number from a pool no one can see. When you hear "compute," what do you picture first: a model, a cloud bill, or a machine in a rack?

Every hour we count gets logged the boring way. Which Device. Which serial number. How long it ran. On what SKU. No blended pool average. No index. One machine, its own record. Dull on purpose. Dull is how you tell an hour that ran from an hour on a slide. What's the one field you'd want on a compute log that nobody shows?

A futures contract on a GPU hour forces the question nobody wanted to answer: what is an hour? Which SKU. What uptime. Delivered where. Interruptible or not. On-demand H100 quotes span roughly $2.19 to $11.06 an hour depending on provider, terms and region. That spread isn't noise. It's the hour refusing to be one number. Which print do you trust: the offer, or the hour that actually ran?

Same $200. Half the API-dollar allowance. OpenAI's Pro 200 drops from 20x to 10x Plus usage from Oct 30. Their framing: API prices fell by half, so token volume is roughly unchanged. Maybe. But a flat seat just got repriced against the meter. When a lab keeps the price and resizes the allowance, what ran out: margin, or machines?

The memory shortage left the data center. Counterpoint: smartphone memory prices up 300%+ year on year in Q2. IDC: average phone price up ~14% this year, to a record $523. Same wafers. Different shelf. The rented GPU hour didn't change. The package did. If the shortage reaches your pocket first, does the rack follow, or wait for the next build?

Micron's last quarter: 87% gross margin. Most of its 2027 HBM supply already contracted, at prices well above last year. 26 multi-year take-or-pay deals, some running into 2031. That's the board sold forward. The rented hour is still a weekly argument. A committed package isn't an hour that ran. When memory is pre-sold for years, what is the spot GPU hour actually clearing against?

🚨BREAKING: Wall Street just launched a record $60 BILLION deal, the largest chip-financing deal to date. Bank of America, Citi and Morgan Stanley are leading the deal, which will fund Anthropic's lease of Google's AI chips, per the FT. About $42 BILLION of it is backed by Broadcom. The three banks have already started selling portions of the debt to other banks.

⚡️ LATEST: OpenAI tops a16z’s ranking of consumer AI apps by monthly revenue, followed by Anthropic and Canva. ChatGPT also leads its ranking of generative AI mobile apps by monthly active users, ahead of CapCut and Gemini.

The memory controller is moving house. Nvidia's custom HBM base die pulls the controller off the compute chip and into the bottom of the memory stack. Nvidia's claim: up to 25% more die area for compute, ~15% lower HBM power. Silicon handed back to the work. Not a cheaper hour. Not a powered hall. If the chip gets denser and the watt doesn't arrive, which invoice do you notice first?

$42B of European data-center investment is now delayed, reworked or cancelled over public opposition. 70+ projects rejected or restricted between January and April. More than all of 2025. That's not a demand story. It's a hall story. The cards can still be ordered. The date they plug in is what's moving. Does a stalled site make the existing hour scarcer, or just push the same watt later?

The hour is getting a ticker. CME has H100 and B200 rental-index futures set to list today. One contract = one month of rent on one GPU. Cash-settled. Read the fine print: you can't take delivery of an hour. A future hedges the price. It doesn't seat a card. Does a public price make the spare hour tighter, or looser?

The week opens on three invoices. Not one GPU price. The board: memory and packaging. The hour: what a card rents for this week. The hall: power, a building, and a date. Last week the package got pricier. The hour started getting its own futures market. The watt got later. We don't forecast which one leads. We name which one moved. Which invoice are you watching into the open?

Two GPU compute DePINs Same product, Opposite shapes 🔽 ⚪ Revenue, 365 days → $IO : $8.73M → $AKT : $2.01M ⚪ Revenue per $1M mcap → IO : $123K → AKT : $9.0K ⚪ Volume as % of mcap → IO : 31.1% → AKT : 3.39% ⚪ Supply issued → AKT : 100% → IO : 49.9% ⚪ Price vs all-time high → AKT : 9.38% → IO : 2.71% ⚪ CertiK Skynet → IO : 86.62 → AKT : 84.21 Two live businesses in a sector full of ideas. Both earning, both building.

The week split into three invoices. The board moved. HBM4 controllers and PHYs are taking about 16% of a Rubin compute die. Micron printed margins in the high 80s while bits only grew mid-single digits. Memory got paid. The package did the work. The hour did not. A $200 plan was rewritten as half the API dollar. Same price. Smaller claim on the same racks. The hall stalled. The biggest US grid suspended a data-center power auction. Cards can still be ordered. The date the watts connect slipped. AION does not add a fourth number. We name which tape moved. This week the package got expensive, the spare hour got smaller, and the watt got later. Which of those three do you think clears first on Monday?

Data centres are going to eat all the electricity and you'll be left eating stale bread in the street, begging for the water they use... Thats the political narrative building right now and every part of that gets the direction wrong. Agents are billions of new economic participants added on top of us. They run 24/7 at silicon speed, and every one of them is demanding something... data, compute, storage, energy, tokens. Demand is about to go through the roof. That is the Economic Singularity... synthetic labour removing the old limits on growth, and GDP as we measure it stops meaning anything. The real fear underneath is that somebody else makes all the money. Except for the first time in history, unlike the internet, the substrate is ownable by everybody. All that activity has to coordinate on the L1s. So own the fucking rails.

The board took the week. The hour did not. HBM4 controllers and PHYs are eating about 16% of a Rubin compute die. That is leading-edge silicon spent talking to memory instead of doing the work. Micron’s print was the same tape from the other side. Margin in the high 80s. Bits still only mid-single digits. Price did the heavy lifting. AION names which invoice moved. This week it was the package, not the spare hour, not the date the watts connect. When memory is the scarce line, what do you think the buyer is actually paying for? Is it the chip in the slide, or the hour that rack can still deliver?

Micron reports today. That's the board book moving, the price of memory and the chips that go into every GPU Device we hold. We don't forecast what the print says. We watch what it does to the board line: DRAM, HBM, NAND pricing feed straight into what a new Device costs to source and deploy. Four books, one report: board moves first. Hour and hall follow with a lag. What's your base case going into today's print, the inline, beat, or miss?

Three invoices landed before breakfast. The $200 plan was rewritten as half the API dollar. That is the hour. Bain says the halls need $6T a year by 2031 to justify what is already poured. That is the hall. Goldman puts $2T of cap ex against 57GW and calls the grid the constraint. That is the date the watts arrive. AION does not add a fourth number. We name which tape moved. The model can get cheaper and the subscription can still buy less silicon. Those two facts can sit on the same morning. Which invoice are you actually watching?

This week a market tried to put a price on the GPU hour itself. NYMEX has H100 and B200 futures ready. The CFTC has pushed them to a Nov 9 review. We think the delay is the most honest part of the story. You can't clear a contract until everyone agrees what's being delivered: which chip, how much uptime, where, for how long. That's the question every GPU invoice already has, just without a clearing house. Futures will come. When they do, the hour becomes a traded thing, and the record of what actually ran matters more, not less. Which would you rather hold in 2027: a contract on the hour, or the machine that produces it?

Over the last 20 sessions, semiconductors are up almost 6% and software is down almost 4%. That gap tells a story. For two years the pitch was that software would capture the value and hardware would become a commodity. This month the market paid for what you can plug in: memory, packaging, power, racks. Samsung is locking memory into multi-year agreements. HBM is sold out into 2027. It won't stay this lopsided forever. Software always gets its turn once the metal is in place. But right now scarcity sits in the physical layer, and the market is pricing it that way. Is this a rotation, or the start of hardware being valued like infrastructure instead of like tech?