Tory | io.net 🦾

Co-Founder of @ionet | Freeing the AI Genie from the Bottle 🧞‍♂️

[Tory | io.net 🦾 @MTorygreen]

https://t.co/PO3HSj9ajx

[Tory | io.net 🦾 @MTorygreen]

@Hiraweb3 💯

[Tory | io.net 🦾 @MTorygreen]

Apparently being close to the physical technology has very little to do with actually adopting it.

California is home to every frontier AI lab that matters. New York has more Fortune 500s than any other state. Both got lapped by Colorado, which hit 23.2% business AI adoption while New York managed 13.8%.

What Colorado and Arizona have isn't better infrastructure or talent. It's a willingness to move before the industry signals it's safe to. The states closest to the technology are often the slowest to deploy it because they have the most invested in how things already work.

Three quarters of American businesses still aren't using AI in any meaningful capacity in 2026 and every headline about AI reshaping the economy describes a world 77% of businesses haven't entered.

Despite the models getting better and get cheaper every month. It still comes down to a distribution problem and the layer that solves distribution is never the one that built the product.

[Tory | io.net 🦾 @MTorygreen]

@bcherny The missing sixth archetype: the Orchestrator. The person who knows which of these five to deploy, when, and how to transition between them as the product matures. That role is rarer than all five combined and nobody has a job title for it yet.

[Tory | io.net 🦾 @MTorygreen]

@chamath Cross-customer learning in healthcare, insurance, and government is where every enterprise AI company hits the same wall. The data that would make the platform smarter for everyone is the data nobody can share. Solving that is the actual product.

[Tory | io.net 🦾 @MTorygreen]

China matching US cybersecurity capabilities isn't the result of bad export policy. It's the result of open research. Papers, conferences, GitHub repos. You can't export-control knowledge that's already in the scientific commons. Closing Mythos off didn't stop ZAI. It just slowed down every allied researcher who could have helped.

[Tory | io.net 🦾 @MTorygreen]

Nobody is asking the obvious question. If custom silicon cuts TCO by 30-40% and Anthropic just locked in for 10 years, why is the US simultaneously export-controlling the models that run on it? You don't restrict something cheap and abundant. The TCO story and the Mythos ban are the same story told from opposite ends.

[Tory | io.net 🦾 @MTorygreen]

@AndrewCurran_ Austria understands the problem but Europe doesn't have the solutions. Frontier AI follows compute. Until Europe can guarantee the supply chain end-to-end, the incorporation address is irrelevant.

[Tory | io.net 🦾 @MTorygreen]

Europe is pitching values. The US is pouring concrete.

What Europe needs to pull this off is 50GW of committed power capacity, a semiconductor supply chain it actually controls, and the political will to treat compute as critical infrastructure. That's a decade of work, minimum.

Anthropic isn't relocating anywhere because the US has made sure they never will. Guaranteed compute, guaranteed power, and a government that classified them as a strategic national asset overnight.

Austria's letter is well-intentioned, but it's a pitch for 2019. The game changed when the Commerce Department put an export ban on a commercial AI model for the first time in history.

This isn't a headquarters decision. It's a supply chain decision. And the US already made it for them.

[Andrew Curran @AndrewCurran_]

From the letter:

'Let us jointly explore the strategic establishment and participation of Anthropic within the European Union. With legal certainty, market access, capital and a set of values that suits this company.'

'Anthropic fits us particularly well. A company that understands the ethical use of AI not as marketing, but as a core conviction. That places safety over speed. That is a deeply European attitude. This company would not be constrained in Europe; it would be unleashed.'

The UK also made similar overtures a few months ago.

It won't happen. Anthropic will stay where the compute is, and where the supply is guaranteed. They can't risk getting cut off, and from here on out the compute will increasingly be concentrated within US borders.

[Tory | io.net 🦾 @MTorygreen]

The whole structure rests on one assumption you name and then set aside: no non-US competitor at the top end. That's the variable doing all the work. GLM-5.2 is already a few points off the frontier on open weights, trained outside US silicon entirely. The tiering only holds as long as the tower is the only tower. That condition is eroding faster than the access policy is hardening.

[Tory | io.net 🦾 @MTorygreen]

On paper, AI costs should have collapsed by now.

The same frontier-level task that cost $30 per million tokens at GPT-4's launch in 2023 costs $1.25 today with GPT-5. A 96× drop in three years, across named models with published prices.

By any normal rule, enterprise AI bills should be cratering. Instead, average Fortune-500 AI budgets went from $7M in 2024 to $19M in 2026. Nearly 3× in two budget cycles.

The reason: cheap tokens don't get saved, they get weaponized. A chatbot fires tokens once and stops. An agent doing real work runs 5 to 30× more for the same task, and once inference is that cheap you point it at everything. Every price cut unlocks more usage than it saves.

The cost center didn't shrink. It moved from the model layer to the infrastructure underneath it. That's where the real bill is being written, and most companies haven't opened it yet.

[Tory | io.net 🦾 @MTorygreen]

Hard to argue productivity isn't the engine. The "not taxation" framing is the part the chart can't actually show, though. Growth and redistribution moved together in a lot of these regions, and separating which one bent the line is genuinely hard. The honest version might be that you need the growth first and the distribution to make it reach the bottom.

[Tory | io.net 🦾 @MTorygreen]

@nickgrossman This is amazing. Completely agree. But what about decentralized compute?

[Tory | io.net 🦾 @MTorygreen]

These calls are about valuations, not demand, and the two keep getting mixed up. The stocks can be in a bubble and the compute shortage can still be real at the same time. 2000 was a similar setup too where the dot-coms were absurdly overpriced and the internet still ate the world.

[Tory | io.net 🦾 @MTorygreen]

The catch is that "smart enough" isn't knowable until after you've run it. You can't tell if a cheap model would've nailed a request without either trying it or having a model good enough to judge difficulty, which costs about as much as just running the good one.
That's the real problem to solve here, and it's why this hasn't been productized already.

[io.net @ionet]

Running 32 A100s on AWS costs ~$770K/year on-demand.

And that's before hidden costs like networking, storage, and idle time push it 40–60% higher.

@ionet changes the equation.

AI compute made affordable and accessible: https://t.co/nLcCI6u85H

[Tory | io.net 🦾 @MTorygreen]

@SemiAnalysis_ If $0.99 blended is already the optimized price and you're still at 30% of comp, what does that ratio look like for a firm that hasn't done the wideEP and caching work yet? Feels like the unoptimized version of this number is the one that should scare people.

[Tory | io.net 🦾 @MTorygreen]

Something quietly inverted in AI compute this year, and it changes what the buildout is actually for.

In 2023, 2/3 of AI compute went to training, the actual work of building a model. The other, smaller slice went to inference, the work of actually running it once it's built. But that ratio quietly started flipping.

Inference is now 2/3 and still climbing, per Deloitte, and the chips built to run it crossed $50B this year.

The main reason this flip matters (and it's not percentage-wise): training and inference are different animals. Training happens in bursts, on one giant cluster, then it's done. Inference never stops. It runs every time someone sends a prompt or an agent takes a step, and it scales with every user you add. One is a construction project. The other is a utility bill that grows forever.

Every assumption about AI infrastructure was built around training, because that's where the money went. The money just moved to the workload that doesn't need to sit in a single cluster to run.

[Tory | io.net 🦾 @MTorygreen]

@brian_armstrong The quiet admission here is that everyone's been paying frontier prices for execution work that GLM or Kimi does just as well. Defaults were doing the damage, not usage. Half the bill was just nobody checking.

[Tory | io.net 🦾 @MTorygreen]

Two years ago an open model on this chart would've been near the bottom. The closed labs were generations ahead, and that gap was the whole reason people rented models instead of owning one.

Now GLM-5.2 sits at 51 on the @ArtificialAnlys index.

Open weights, Chinese lab, fifth overall. And knock Fable out of the list since it's not available, and the open-weights model is way closer to the top than its ranking lets on.

The pitch for closed was always the lead. Pay the API, accept the terms, build on something you don't control, because the model's far enough ahead to be worth it. That lead is now a few points, and GLM got there while being on the Nvidia chip cutoff list, which is even more impressive.

The premium was priced against the gap. The gap's nearly gone yet the premium hasn't moved.

Curious to see where we'll be in a year from now.

[Tory | io.net 🦾 @MTorygreen]

@sama read this twice and the only thing that stuck was Terra at half the price. everything else is the part you have to say now, this is the part that actually ships