Thesis

Intelligence is still extraordinarily early. Frontier labs claim AGI, yet economic growth remains ordinary. The models exist; the products and infrastructure that turn them into useful work mostly do not. The engine may be running, but the car is not built.

Most AI software still waits for a person to ask a question. That will change. Software is becoming increasingly proactive: it can research, plan, use tools, and keep working in the cloud for hours. This is logical because proactive software is objectively better than reactive, once users can trust it. As more real work moves into these systems, demand for inference will be close to insatiable.

Electricity followed a similar path. Factories first replaced steam engines with electric motors to do the same work more cheaply. Once power became cheap and widely available, it reorganized industry and made entirely new products possible. Intelligence will do the same, but only when it becomes practical to use far more of it.

Inference should fit the work

Inference is usually sold as if every request has the same requirements. It does not. A person waiting for an answer cares about latency. A cloud task running for hours may care much more about accuracy, reliability, and cost. Today developers compare providers, benchmarks, prices, and latency, then commit to a compromise they cannot properly control.

Gradiated provides inference designed for the workload. We turn its traces into an RL environment, build the evals, and post-train an open model automatically. The developer just changes the base URL. The result is a model shaped around the work it actually performs, with direct control over the tradeoff between latency and cost.

Open models are taking the market

Open-weight models are not merely growing with inference. They are taking share from closed models. Between September 2025 and August 2026, their share of OpenRouter tokens rose from 26% to 60% for requests originating in the US and from 16% to 65% in the EU. Open models overtook closed models in both regions in early 2026, according to OpenRouter's regional usage data.

That change is happening inside a market that is itself compounding.OpenRouter's weekly volume grew fivefold in six months, from 5 trillion to 25 trillion tokens. By September 2026, nine of its ten most used models by 30-day token volume were open-weight, according to its usage data.

OpenRouter is one marketplace, not the whole market. But its direction is hard to ignore. Open weights can be post-trained, measured, and served in different ways. No single model will be best for every task simply because using a general purpose, many parameter model is less efficient to serve, and post-trained models frequently have better task-specific performance.

Go all the way down

Assume inference becomes a commodity. Reliability and trust are table stakes, in the same way a gas station would go out of business near instantly if its fuel wasn't good enough for cars to use. Beyond that, the provider that can deliver the most useful intelligence with the least wasted computation wins.

At the extreme, that requires control of the whole stack. A model's architecture determines its memory access. The runtime and kernels determine how much of the hardware is used. Scheduling and infrastructure determine whether expensive capacity sits idle. Eventually, the workload should determine the silicon itself where possible.

General-purpose accelerators must support almost anything. We do not need to. Sparse mixture-of-experts models such as Kimi K3 activate only part of the model for each forward pass. Their ideal memory layout, interconnect, and scheduling are different from those of a dense model. A chip designed around a small number of known model families can be simpler and much more efficient than one designed around every possible model.

We are building an end-to-end simulation environment so that software and hardware can be designed together. It lets us test compilers, runtimes, memory hierarchies, and silicon before tape-out, then change the design as models and workloads change.

There is a boring way to do this

We start with a reliable inference service on existing GPUs. Real usage tells us which models matter, how long requests run, what latency they need, and where computation is wasted. We optimize those model families, post-train them for customer workloads, and sell unused capacity through marketplaces.

We do not need to guess at a chip on day one. The API collects its requirements first. Once the workload is clear and the volume is there, we can build a straightforward chip around the parts that matter. Then we repeat the same process at the data-centre level.

The sequence is:

• Use traces to build evals and RL environments, then post-train models for the workload automatically. Provide trusted open-model inference with latency configurability as a by-product.
• Optimize the model, runtime, kernels, scheduling, and infrastructure as one system.
• Turn the requirements learned from real demand into silicon and get to the scale where we can buy specialist chips if needed for workloads our own silicon cannot service.
• Repeat, using the efficiency to drive more scale, to the point where we can get to data center scale and beyond. We will lose eventually if we do not go this deep.

Why now

The model landscape is still moving, open weights have crossed from the minority to the majority on a major marketplace, and no inference provider has won the market. Distribution matters because scale lowers purchasing costs, makes specialization worthwhile, and funds the next layer of the stack. The market is open now. It will not remain open forever.

If Gradiated succeeds, every workload gets the intelligence it actually needs: accurate, fast, and cheap, in the combination that matters for that work. Ultimately, driving down tokens per joule we believe is the greatest impact we can have on mankind as it is the best way to empower every person and every system to research, build, understand and do more.