Three stories landed within a day of each other, from different parts of the industry, with nothing obvious connecting them.
A workflow automation company shut down and its founder took a job at Google. A chip company gave up on its chips and became a cloud provider at half its old valuation. And a Forbes analysis pointed out that most enterprise deployments of open models run on artifacts nobody can trace.
Read together, they describe the same movement. The independent middle of this stack is being absorbed, and the parts that resist absorption are drifting out of anyone’s control.
The feature that became a browser
Relay is shutting down. TechCrunch reports free access ended on 15 August and paying customers lose access on 14 September. Founder Jacob Bank is rejoining Google as VP of Product for Chrome.
Relay launched in 2021 to take on Zapier with AI-assisted workflow automation. Drafting documents, moving work between tools, the connective tissue of an operating business.
Bank has said he has plans to help people work with AI inside Chrome to get things done. That is the entire Relay thesis, relocated into a browser with a few billion users, at a moment when Gemini has crossed a billion users of its own.
The lesson for anyone building in this space is old and keeps getting re-learned. If your product is a capability rather than a system of record, the platform that owns the surface will eventually ship it. Not because they copied you, but because it belongs there.
The commercial version of that lesson: never build a business whose moat is being early to a feature the platform obviously wants. The buyer’s version is not letting one vendor own your agents.
The chip company that became a landlord
Groq raised 350 million dollars led by Disruptive, with Nvidia participating, at a 3.5 billion valuation. In September 2025 that valuation was 6.9 billion.
Half the value, and a completely different business. Groq built custom LPU inference chips as an alternative to Nvidia. It now runs Nvidia GPUs as a neocloud, after founder Jonathan Ross and much of the technical talent left in a 20 billion dollar licensing arrangement with Nvidia.
The company plans to scale from 54 megawatts to over 200 by 2027, across 13 data centres, serving more than six million users.
That is not a failure, it is a retreat to a defensible position. But look at what the defensible position is. Renting capacity from the incumbent’s ecosystem, competing on operations and price, in a business whose primary risks are capital intensity and hardware depreciation.
The challenger to Nvidia is now a customer of Nvidia. When the alternative to the standard becomes a reseller of the standard, the standard has won that round.
The supply chain nobody is tracking
The third signal is the quietest and probably the most operationally dangerous.
A Forbes analysis by Janakiram MSV looked at how open models actually reach production, and the numbers are striking. Repository activity declaring GGUF support grew 464 percent in the first seven months of 2026, against 21.5 percent growth in model repositories themselves.
The packaging layer is growing twenty times faster than the models. And it is almost entirely community-produced. Hugging Face carries 28,531 community GGUF conversions of Alibaba’s Qwen models against 54 official ones from the lab.
So when a company says it has deployed a particular open model, that statement is usually wrong in a specific way. It deployed a conversion of that model, quantized by an unnamed party, with a recipe nobody recorded.
Signing does not solve it. OpenSSF released model signing v1.0 in spring 2025, but a signature confirms bytes and identity. It cannot confirm that a quantization faithfully represents the original weights.
The recommendation in that piece is the right one and nobody will enjoy implementing it. Track a bill of materials for every model you run: upstream repository, conversion recipe, artifact hash, tokenizer version, runtime, hardware target, evaluation results. Names are not identifiers. And if you cannot say what you deployed, you cannot test it either, which is why evals are the new product spec.
What connects them
Each of these is a story about the middle disappearing.
The application layer gets absorbed upward into platforms, because platforms own distribution. The hardware layer collapses downward into infrastructure economics, because capital wins there. And the packaging layer in between is expanding fast in the hands of anonymous volunteers, because nobody with a budget claimed responsibility for it.
If you are building, that geography tells you where not to stand. If you are buying, it tells you which questions your vendor cannot currently answer.
The stack is settling into a shape. Owning a layer of it is a strategy. Sitting between two of them is a timeline.