The Flywheel and the Meter

A day of heavy work inside an agent harness throws off corrections the whole time: I take this, I rewrite that, I throw the third one out. Each of those judgments lands on disk as a byproduct of the work, and a pile of them is training data. Feed it back and it routes calls between expensive and cheap models, adapts open weights to what they keep getting wrong, and cuts what goes to a metered frontier API. I have been reading my own usage for exactly that. I think one person can now do it, and that whoever does it for teams, enterprises, and air-gapped nation-states ends up mediating the market of AI use. Three releases this week are why I am writing it down now.

On August 10, Meta released Muse Glimmer, a 30B model under Apache 2.0, quantizable to under 20GB and onto a consumer GPU. On August 11, NVIDIA released Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters, pitched as the execution layer for long-running agents, with an open-source router, NeMo Switchyard beside it. On August 12, Alibaba published the weights of Qwen3.8-Max, a 2.4T-parameter flagship under a license I will come back to.

The bills came due the same season. Uber reportedly burned its entire 2026 AI budget in about four months on coding agents, then capped staff spend. Anthropic added weekly usage caps to its flat-rate subscriptions, picked up a class action over it, and trimmed its margin guidance after inference costs ran over plan. Neither of those is a story about a bad deal. Uber’s number was unforecastable, and the flat rate Anthropic offered instead did not cover the inference, so the meter stays.

What I want to settle is narrower: are documented open models and pricing pain actually arriving together, or have I lined up two news cycles because they rhyme?

Where the Words Come From

Flywheel, data flywheel, harness: three borrowed words, so I looked up what each one arrived carrying.

“Flywheel” as a business metaphor is Jim Collins, Good to Great, 2001: the flywheel effect is momentum from many consistent pushes rather than one dramatic shove. What it asserts is compounding: each turn makes the next one cheaper.

“Data flywheel” is the AI-specific extension, and I could not find who coined it. The best-documented early instance is Tesla’s Autopilot data engine: ship it, mine the failures, label the long tail, retrain, ship again. NVIDIA has since adopted the term as marketing architecture.

“Harness” is older than it sounds: a test harness is the stubs and drivers needed to execute a test, formalized in IEEE 610.12 in 1990. It reached AI agents two ways, as evaluation scaffolding (SWE-bench’s eval code is a module named harness) and the environment wrapped around a model, the sense Steve Yegge’s “Welcome to Gas Town” runs with and Anthropic’s agent documentation settled into. harness-explorer exists because I wanted to measure my own harness rather than describe it.

The Stack I Keep Seeing

Different vendors, different words on the boxes. The same four layers underneath.

A harness people work inside. Under it a flywheel, recording what happened to each piece of output: accepted, corrected, thrown away, per task. A router spending that record on every call, sending the work to the cheapest model that clears its bar. Then three ways to answer a call, on the same judgment: a frontier API used selectively, an open model you host and adapt, or a week of rented accelerators, quarterly-ish, aimed at one business unit’s task.

Routing has the most behind it. RouteLLM showed in 2024 that a learned router could hold 95% of GPT-4 quality while sending only 26% of calls to it. OpenAI then built a real-time router into GPT-5, which reads to me as the frontier vendors taking the pattern in-house before their customers can capture the savings. Switchyard shipping beside Lightning, with NVIDIA claiming task costs near one third of a frontier-only baseline (vendor-reported), makes routing something you download rather than implement.

Training moved most recently, and it moved on price. Thinking Machines’ on-policy distillation work reports matching reinforcement-learning results on far less compute: one published math-reasoning run reached its target in 1,800 GPU-hours against 17,920 for the RL baseline. With H100s renting at two to three-and-a-half dollars an hour, a week of adaptation compute for a bounded task prices in the tens of thousands, not the millions, and a LoRA-scale adaptation is already an afternoon. I will not project that price two years out. Rental rates and distillation efficiency have moved one way for two years running, which makes it a line item a business unit can plan for.

Building a Flywheel to Traverse the Ladder

I started the one-person version yesterday, an instrument rather than a product. A small reverse-proxy shim tees my agent traffic, request and response, into a local corpus. A DuckDB store unions that capture with backfilled transcripts, and a read-only dashboard sits on top. The unit of work is a study: a query over the corpus ending in a dated verdict. The code is meant to be cheap to throw away; what I keep is the corpus, the datasets mined out of it, and the verdicts.

The part I expect to outlive the code is an ordering out of the project’s charter. Steering a model you did not train is a ladder, cheapest and most reversible at the bottom: context, system prompts, skills, tools, deterministic hooks, model routing, adapter training, full training. The bottom rungs change in minutes and roll back for free. Each step up buys durability and costs more to undo. The second rung from the top is the adaptation priced above; full training sits past it at a cost I did not price. The rule is the lowest rung that durably fixes the behavior, picked on evidence from the corpus rather than instinct.

None of the mining is built yet: corrections pulled out of the corpus, whatever the low rungs cannot fix fed into a LoRA experiment. It is one operator’s traffic, so every finding is a single data point, and I want to know whether the loop closes at n of 1 before arguing it closes for anyone else.

Two Moats

Push the opening belief far enough and it says adapted small models eventually retire the frontier API. I went looking for that retirement and did not find it: there are two moats here, and they are not moving together.

The reproduction moat is shrinking. AI2’s OLMo 3 ships weights, the full 6T-token dataset, training code, and checkpoints; Hugging Face’s SmolLM3 publishes its data mixtures and post-training recipe; Lightning ships weights, data, and recipes under the Linux Foundation’s OpenMDW license. A team can start from documented parts now, and the head-to-head evidence, most of it vendor-published, has adapted small models matching or beating frontier ones on bounded work: classification, extraction, tool calls (an academic check agrees on the classification slice). The gap re-opens on long-horizon reasoning, which is why Lightning’s own positioning keeps a frontier model planning.

The other moat has not moved, or nobody has measured it moving. Epoch’s capability index puts the best open-weight models about four months behind the closed frontier since January 2026, flat to slightly wider than the prior three years’ average. Glimmer is Apache 2.0 and ships weights plus a methodology narrative: no data, no token counts, no training code. Qwen3.8-Max backed away from its ancestors’ Apache 2.0 to a custom revenue-gated license. The reinforcement-learning infrastructure, reward models, and task distributions behind every frontier release stay proprietary even where weights do not. One countersignal I cannot read yet: OpenAI is shutting down self-serve fine-tuning by early 2027, saying newer base models make it unnecessary. Either they are handing the adaptation market to the open-weight ecosystem, or they know something about prompting-sufficiency the published case studies do not show. I do not know which.

One more row landed while I was finishing the piece; the closing tells that part.

So, my question. The two curves did arrive together, and I cannot fully rule out the rhyme: the meter would hurt this year with or without OLMo’s dataset, and the open models would keep improving with or without Uber’s bill. What ties them is the stack, which needs both: routing and adaptation want documented parts and a meter painful enough to justify building. The frontier keeps its four months either way. What shrinks is the set of jobs paying frontier prices, and that shrinking is the part a bill can measure.

The Air-Gapped Case

One set of users never had the economics question, and I think they run ahead of everyone else. Classified networks under ICD 503 and the DoD cloud security guides do not meter the external connection. They forbid it, so egress lands as a policy violation rather than a line item.

Scale’s Defense Llama, built on open Llama 3 weights, was reportedly the first LLM on a U.S. classified network. Palantir’s AIP fine-tunes open-weight Nemotron models inside the customer’s perimeter, the customer keeping the weights, and France’s armed forces ministry signed a three-year framework with Mistral for on-premises deployment. The part I did not expect: the frontier vendors came to them. Claude Gov runs at Impact Level 6 through accredited regions, as does Azure OpenAI. The air-gapped remainder runs open weights on owned hardware, adapted in-perimeter by cleared staff, updates carried in by accredited one-way transfer.

That remainder already lives in the end state this argument predicts, compute on a capital budget and no meter anywhere. My own controlled-egress work (egress) sits next to it, and from close up the architecture is unexotic: the same four layers with the API option struck out, a natural experiment for whether the other three carry the load. Published recipes and distillation-grade small models keep making it cheaper to run.

What I Want to Watch

The measured trend runs against me. Open-model share of enterprise workloads fell from 19% to 13% across the first half of 2025 while total spend tripled, and the cloud-repatriation precedent people reach for stayed niche (Gartner found about 4% ever moved back). For any of this to be right, that number turns around on the mechanism above: judgment captured for free, recipes published, a meter that pays for the upkeep. The upkeep is the part I suspect everyone underestimates, me included. An adapted model drifts against a teacher that keeps moving, and the eval rig that catches the drift costs money every month and assumes a frontier-quality judge inside it.

The silicon vendors are betting the same way at the edge. AMD ships up to 128GB of unified memory in machines it brands agent computers, with a 192GB part announced claiming 300B-parameter local models, and NVIDIA’s DGX Spark puts 128GB on a desk. IDC’s framing is that always-on agents cannot keep paying cloud rates for every token. Then there is the memory market. Apple spent 2026 cutting the Mac Studio’s configurable ceiling from 512GB to 96GB as the DRAM shortage bit, and every one of those bets buys from the same constrained supply.

The license question answered itself while I was finishing this. Alibaba shipped Qwen3.8-27B on August 14, the week’s fourth release and the commodity-sized member of the generation whose flagship just went revenue-gated, and its LICENSE file reads plain Apache 2.0. No revenue test. AMD had day-0 support running on the same 128GB machines from the last section. Training data and recipes did not ship, so its row in the table above fills weights and license and nothing else. I did not expect the answer the same week I flagged the question. That is the size of thing I can check, and I would rather collect a season of them than argue the abstraction. The next two: what the meter costs, and whether anyone ships recipes beside their weights.