torygreen

.@ionet is at ~165B tokens through @OpenRouter in the first 13 days of October.
That’s already ~1.8x everything we did in July.
If the current pace holds, October lands somewhere around 4x July’s total volume.
I care more about that than raw GPU count tbh.
The useful output per deployed GPU keeps moving as sparse models get better and the serving stack gets tighter. So the network can grow in two directions at once: more hardware coming online, and more work getting squeezed out of what’s already there.
Tokens served per GPU is a great sign of whether the network itself is actually getting bet

By 2032, the current AI buildout math wants roughly 183 GW of capacity earning ~$6.90 per billed GPU-hour at 80% utilization.
That is a huge amount of infrastructure that has to stay busy.
Van Nieuwerburgh’s central case assumes a ~$9.6T capital base, 10% unlevered returns, 50% operating cash-flow margins and a six-year IT equipment life.
Under that setup, the system needs $3.725T in annual revenue, around 9.2% of projected US GDP.
Move the assumptions around and that hurdle runs from $2.3T to $6.7T.
Meanwhile models keep getting cheaper to run, inference stacks keep improving and more capacit

.@ionet processed almost as many tokens through @OpenRouter in the first week of October as it did in all of July.
~92B tokens in July
~86B in Oct 1–7
20.9B on Oct 7 alone
Same machines.
What I like most is how the vol is showing up.
A lot of the recent vol is flowing through sparse models like MiMo, Gemma and Nemotron, where only a small fraction of the model fires on each token. Pair that with better batching, caching, routing + serving infra and the economics of the same GPU fleet keep changing underneath you.
For @ionet both sides of the equation keep improving:
inference demand is growin

Jev might be the perfect example of how fast a good AI idea gets commoditized now.
It launched Sep 15. Within two weeks, OpenAI, Databricks + Cloudflare were already shipping around the same primitive. Cloudflare’s Clef-flash is even Jev-compatible and claims ~13x lower median latency.
So demand wasn’t the problem. The idea was just easy to absorb once everyone saw it work.
Feels like the moat here has to be the feedback loop, proprietary data + distribution.
Jev found the category. Everyone else showed up almost immediately.

Same Deepseek harness, same short instruction and:
Deepseek-v4-flash scores 0.52 on BIABench.
Claude Opus 5 gets 0.65.
So ~50x more spend buys 0.13 more score across 16 real bioimage-analysis tasks.
And the ugly tail barely moves. Some 3D + time-series tasks stay below 0.19 across every agent, while rerunning the same model can vary more than switching models entirely.
I’d probably burn the cheap model on routine volume, then escalate the uncertain cases to humans or specialized tools.
And in science, a $4.53 answer you still need to verify isn’t really a $4.53 answer.

Google has a pretty unfair loop with Argon compared to the rest of frontier labs
they own the TPUs underneath Argon and can drop the model straight into APIs already processing around 22B tokens/min.
Thats a pretty impressive setup.
Build the chips -> train the model -> ship it into absurd distribution -> feed the gains back to the next run.
And Argon is already improving the stack around itself too, freeing up 300+ TiB of datacenter memory by migrating +800k lines from C/C++ to Rust and pushing a quantum-computing routine 40% past the published baseline.
Every Gemini generation can increa

.@AnthropicAI S-1 is kinda fascinating because the business is moving almost too fast for the filing itself.
2025 revenue was ~$4.6B. Q2 2026 alone hit ~$11.5B, and by late July ARR was reportedly above $65B.
Then you get ~80 pages of risk factors vs 48 on the actual business, a disclosure that ~6% of AI research compute went to safety work in one sampled week, and a pretty blunt admission that continuous + overlapping model releases are basically required to stay at the frontier.
Feels like public markets could force a level of AI transparency we haven’t really had yet.
Revenue growth, comput

AI adoption barely grew this year. Consumer spend still went from ~$12B → $40B.
That’s a way more interesting signal than another user-count milestone.
U.S. adoption only moved 61% → 64%. Global users grew ~11%.
Existing users just started handing AI way more of their lives.
41% have tried an agent. 32% have already let AI act without final approval. And 92% of agent users are paying for AI.
So the next leg of consumer AI probably doesn’t need another billion people discovering chatbots.
It can come from the people already here going from:
“answer this”
to
“handle this.”
That’s a completely di

The "flippening" we didn't expect to arrive this early.
Open weights are starting to look a lot like baseload power for AI.
They're not beating the best closed model outright atm. But they're close enough and good enough for the massive pile of work that doesn’t justify frontier pricing.
Then Astra, Opus, Fable etc become more like peaker plants: expensive capacity you pull in only when the task actually earns it.
If this keeps up, open weights could end up carrying most routine inference while closed labs make their money on the ugly tail of problems where failure is expensive.
Way more inte

We spent decades making software less physical and AI is reversing that in a hurry.
The current estimates are that the US AI buildout will absorb about 3.63% of GDP every year from 2025–32, roughly $10.3T in total.
> Railroads averaged 2.24%.
> Highways 1.13%.
> Telecom + fiber 1.1%.
That’s a ridiculous amount of concrete, copper, power, GPUs + cooling being built for intelligence.
The comparison to railroads is useful, but there’s one nasty difference: a railroad built in 1880 could stay useful for decades. A rack full of accelerators can look old a few model generations later.
So this migh

Frontier models barely get time to become the default anymore before their own cheaper siblings show up and start eating the workload.
36 major model-family launches by Sep 23 this year, up from 27 across the same seven labs last year.
The price ladders are widening just as fast. @AnthropicAI has taken Opus-class pricing from $15/$75 per million tokens to $4/$20.
@OpenAI now stretches from Luna at $0.10/$0.50 all the way to Astra at $10/$50.
That does something pretty weird to model economics.
The expensive model starts looking more like burst capacity. You call it for the small slice of task

There’s another open-model flywheel that barely gets talked about.
DeepSeek V4.1 Flash did ~18T tokens on OpenRouter last week, and its outputs can explicitly be used for distillation.
So the model can do two jobs at once:
run production today, then become teacher data for whatever someone trains tomorrow.
OpenAI + Anthropic understandably wall that off for competing models.
But that means every widely used permissive model is also spreading some of its capability into the next generation of smaller/specialized models.
Cheap inference is already a distribution advantage. Turning that usage int

Xiaomi just published something labs usually keep buried: the actual RL bill.
MiMo-V2.6 Pro and Flash both trained on 753k samples and stopped at step 30.
> Pro cost $2.62M.
> Flash cost $854k.
That extra 3.1x spend bought 6.89 points on DeepSWE.
Then Flash still edged Pro on AutomationBench.
You can already see the same product shape showing up everywhere now: one model for the absolute hard stuff, another sitting just underneath it that’s cheap enough to run all day.
ZAI, Deepseek, Qwen and Xiaomi now all have this in common.
Once the smaller model clears the reliability bar for most work,

GPU-hours might be turning into something you hedge.
The CFTC is already exploring compute derivatives, and Liquid Compute just raised $15M to build a regulated market for GPU futures, options + swaps.
Makes sense if AI compute really grows from ~$360B in 2025 to ~$2.3T by 2030.
The weird part is standardization.
An H100-hour in one region isn’t the same product as a B200-hour somewhere else with different networking, power costs and cluster size.
So before compute gets a real futures market, someone has to invent the benchmark contract everyone agrees to trade around.
Once that exists, labs c

.@OpenAI spent $6.5B building the device team and now it may be spending another $300M+ on the part that decides what the model gets to see in the first place (Glass imaging).
Glass is working below the camera app directly on raw sensor data. Its proposed mobile system uses a sensor ~2x larger than today’s biggest smartphone sensors and claims up to 10x more light capture.
For an always-on assistant, that matters a lot.
Anything the sensor misses is gone. The model has to guess around it later, which means more uncertainty, more compute, and worse understanding of whatever is happening around

Query counts are about to become a pretty bad proxy for AI demand.
Over 8 weeks, one Claude Code user typed 1,138 prompts. Those turned into ~14k model calls and 3.2B processed tokens. Estimated energy use came out to ~150 Wh per typed prompt, around 625x a median Gemini text prompt.
Most of the compute happens after the user stops typing. The agent rereads context, calls tools, retries, branches, and keeps working in the background.
So one human request is starting to behave more like a small batch of jobs than a single query.
For infra, the better unit might end up being agent-hours, not pro