Thinking Machines Is Betting Against the God Model. I Think They're Right.

Last week, on July 15, Thinking Machines Lab shipped its first model. Mira Murati's lab called it Inkling, and on paper it looks like another big release: 975 billion parameters in a Mixture-of-Experts design, 41 billion active per token, multimodal, Apache 2.0 open weights. But the interesting part is not the model. It's what the lab said about the model. They were explicit that Inkling is not a finished product. It's a starting point, raw material you're supposed to fine-tune through Tinker, their model-customization platform.

That framing is a bet against the entire strategy of OpenAI, Anthropic, and Google. And I think it's the right bet.

The business model is the story

Look at what Thinking Machines is not doing. No API metering. No model licensing. They gave away a 975B-parameter model under Apache 2.0, which means you can take it, modify it, and ship it commercially without paying them a dime. The thing they actually sell is Tinker, the infrastructure for turning that raw model into your model.

That only makes sense if you believe organizations want to own a customized model rather than rent a general one. If the future is everyone calling one giant API, giving away the weights is corporate suicide. If the future is thousands of companies fine-tuning open models on their own data, the weights are a customer acquisition channel and the fine-tuning platform is the business.

They shipped a proof point with the launch. In a project with Bridgewater Associates, a fine-tuned open-source model reportedly beat proprietary frontier models on financial reasoning tests, at a fraction of the cost. Bridgewater is not a company that tolerates worse answers to save money. If a hedge fund with that much on the line got better results from a model it customized and owns, that's not a demo. That's the template.

Even the model's design choices point the same direction. Inkling flags uncertainty instead of guessing, which matters enormously in regulated and high-stakes domains where a confident wrong answer is worse than no answer. It has adjustable compute effort, so you can dial speed against depth per task. Those are not consumer chatbot features. They're features for someone building a production system around a specific job.

The god model needs everything to go right

The big labs are all running the same play: build one enormous model that does everything for everyone, and keep scaling until generality wins every workload. That strategy requires effectively infinite capital, which is why we're watching raises like $122 billion happen. It also requires an assumption that rarely gets said out loud: that a general model beats a specialized one at every task that matters economically.

I don't think that assumption holds. Most of the economic value in AI sits in narrow, high-volume, domain-specific work. Claims processing. Contract review. Trade reconciliation. Clinical coding. Support triage. On those tasks, a customized model can win on accuracy, cost, latency, and data privacy at the same time. Not one of those dimensions. All four. The general model has to be so much better that it overcomes losing on all four, forever, and I've seen no evidence that's where things are headed.

The data backs this up. 46 percent of Fortune 500 companies now prefer open-source models for fine-tuning, up dramatically from where that number sat even a year ago. Yes, open-source trails the closed frontier by nine to twelve months on raw capability. But fine-tuning closes that gap entirely for specific applications. A fine-tuned small model frequently outperforms the frontier general model on the exact task it was trained for. That's not a benchmark curiosity. It's a measurable production outcome, and Bridgewater just demonstrated it in one of the least forgiving domains there is.

We have seen this movie

Here's why I'm confident rather than just intrigued. The specialization bet matches how every previous computing wave actually played out. Mainframes centralized computing until distributed systems pulled it apart. Monolithic enterprise software ruled until specialized SaaS carved it into a thousand focused products. The pattern is consistent: centralize first, while the technology is scarce and hard, then specialize and distribute once it commoditizes.

AI is hitting that transition right now. The frontier labs played the mainframe role, and they played it well. They proved what's possible and drove the science forward. But the moment capable open weights exist, and Inkling is very capable, the economics start pulling toward specialization, because that's where the margins and the moats are for everyone who isn't a frontier lab.

The market structure question is the one that should keep founders up at night. If the god-model thesis wins, value concentrates at two or three labs and everyone else pays rent forever. Your product is a thin wrapper, your costs are their prices, and your roadmap is their changelog. If the specialization thesis wins, value distributes across thousands of companies that own their fine-tuned models, and the infrastructure layer, meaning fine-tuning platforms, serving, and evals, becomes the real business. Thinking Machines is positioning to be that layer. It's the picks-and-shovels move, except the gold rush is every company's proprietary data.

What this means if you're building

I'll be direct about the practical takeaway, because this is the part that changes decisions this quarter.

Your domain data plus a fine-tuned open model is a real moat. Nobody else has your data, and after Inkling, the base model quality is no longer the bottleneck. A prompt on top of a god model is not a moat. It's a feature, one that the lab you're renting from can absorb in their next release, and one that any competitor can replicate in an afternoon.

The Bridgewater example is the template worth copying. Beat the frontier on your task, not every task. Do it at a fraction of the cost. Do it with a model you own, running where your data lives, with latency you control and no usage meter ticking against your margins.

Thinking Machines just made that path dramatically cheaper and more legitimate. The god model will keep getting bigger and keep getting funded. But I think the durable businesses of this era get built on the other side of the bet, and last week the other side of the bet shipped its opening move.