Accelerating Intelligence: Our Investment in Fractile

Every sentence, image, and piece of code generated by an AI model is the result of inference: data passed across billions or trillions of weights and parameters by powerful processors to create the outputs and applications that touch our everyday lives.

Producing those outputs is no simple task.

To continue scaling intelligence, inference must be blazingly fast, cheap, and concurrent. Models need to serve thousands of tokens per second to thousands of users in parallel, all at a power budget and cost that make sense.

Increasingly, the underlying chips that run these models are the binding constraint. Generation after generation of incremental improvement has failed to keep pace with the throughput and latency requirements of leading models.

Fractile is a chip-first frontier AI company building a new inference accelerator to remove this bottleneck and push the Pareto frontier of compute performance.

That’s why we’re excited to co-lead Fractile’s Series C.

The Exponential

Inference is growing exponentially. Frontier models are processing more than 10x more tokens every year, creating significant demand for the chips needed to serve them. Over the next five years, the market will reach hundreds of billions of dollars, with roughly 80% of that spend going towards inference.

This shift is being driven by a fundamental change in how frontier tokens are consumed. Long-horizon agentic tasks, heavy reasoning workloads, and subagent- and token-intensive use cases such as coding are now generating far more tokens per task than traditional chat-style interactions.

Serving this demand requires a massive buildout in the underlying compute infrastructure. But most chips today are selling into a compute supply shortage. Few are purpose-built to accelerate frontier workloads. Every token a model writes requires reading its weights and its context out of memory, so the speed a single user experiences is set by memory bandwidth, which has scaled far more slowly than raw compute.

During generation, a chip spends most of its time waiting on data to arrive. The architectures in the market today lack capacity to hold rich context, to move data between memory and compute quickly, or the ability to do this cost-effectively at scale (usually all three).

What Fractile Is Building

Fractile is designing a new generation of inference accelerators and rack-scale systems that are built on stacked dynamic random-access memory (DRAM). This approach takes the same memory that normally goes into high-bandwidth memory (HBM) and packages it directly on top of the logic. That tight coupling delivers far higher memory bandwidth than HBM, so models run faster, with significantly more capacity than static random-access memory (SRAM) designs. Speed and scale on the same chip.

The resulting accelerators can serve the most advanced AI models at over an order of magnitude faster speeds and at a fraction of the cost.

Delivering this step change in performance with a radically different chip architecture requires a truly special team. Fractile has been vertically integrated since day one. AI chip efforts often hand backend design to Broadcom or another outside partner; Fractile owns its entire stack, which lets the team reimagine the full system from first principles and better control its cost structure.

Founder and CEO, Walter Goodwin, started Fractile in 2022. Since then, we think he’s assembled one of the deepest benches of full-stack talent in semiconductors to execute on this ambitious mission. Fractile has recruited industry veterans who have shipped silicon at scale and worked across software, hardware, operations, and supply chain to push the boundaries of the chips running the most advanced models.

New and powerful AI workloads require ambitious designs if progress is going to hold its pace.

Fractile’s mission is simple to us: build the hardware that pushes AI’s frontier forward. We’re partnering with Walter and the entire Fractile team to help them do it.