Georgian has invested in Fractile's $720M Series C round at a $7B post-money valuation, led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, alongside new investors Dragoneer, DST Global, BlackRock, T. Rowe Price and CapitalG and returning investors Accel, Conviction and Felicis.
Fractile is a chip-first frontier AI company, building AI inference chips and systems designed to run AI models faster and at lower cost.
Walter Goodwin founded Fractile in 2022 as a PhD student at Oxford's Robotics Institute, on the belief that the most capable AI systems would eventually be limited by the time taken to produce useful outputs. Fractile has worked full-stack ever since to build chips and systems that tackle this problem: how to efficiently generate output at thousands of tokens per second, while handling the complexity and capacity challenges of running large models at very long contexts. Fractile aims to serve frontier models at more than an order of magnitude faster speeds and at a fraction of the cost.
The Market Opportunity
As AI becomes more widely used and capable of handling increasingly complex tasks, the demand for computing power continues to grow. Using an AI model is a two-step process. First, the model is trained, a one-time effort in which it learns from large amounts of data. Then, every time someone uses the model to ask a question or run a task, the model applies what it learned to produce an answer. That second step is called inference. Unlike training, inference isn’t a one-time cost. Every question or task a user submits requires computing power, so the cost scales with both how many people use the model and how demanding each task is.
As adoption grows, inference is becoming the dominant AI workload, accounting for ~2/3 of AI compute in 2026, up from ~1/3 in 2023. Gartner forecasts that spending on inference in AI-optimized infrastructure as a service will reach $23.3B in 2026, above the $19B it projects for training. This shift is also driving significant increases in energy requirements, with McKinsey projecting that power demand for AI inference will more than quadruple by 2030.
The tasks AI models are performing are also getting more complicated. Long-horizon agentic tasks and heavy reasoning workloads now generate far more tokens per task than chat-style interactions, and serving them means delivering thousands of tokens per second to thousands of users at once. One of the difficulties existing AI chips face as demand continues to grow is the ability to process the amount of information required for more complex tasks. On the toughest problems, LLMs are already generating up to 100M tokens, the units of text a model reads and writes. Fractile asserts that on conventional chips, models typically run at about 40 tokens per second, so a single output of 100M tokens would take up to a month to complete. Compressing that into one day would require ~1,200 tokens per second. This illustrates how much faster AI hardware needs to become to support the increasingly complex work these models are being asked to perform.
One of the main factors that limit hardware speed is how AI chips store and access information. When a model generates a response, its processor needs to repeatedly retrieve information stored in memory. On conventional chips, moving this information back and forth takes time, often leaving the processor waiting for data instead of computing a response. This slows the model down and raises the cost of running it. The gap between processing power and memory speed has also widened considerably. Over the past 20 years, computing power has grown roughly a million-fold, while memory bandwidth, which measures how quickly information can move between memory and the processor, has grown about 40-fold. As a result, even as processors become more powerful, their performance can still be limited by how quickly they can access the information they need.
Fractile’s Approach
A core challenge in AI hardware is that processors can’t pull information out of memory fast enough.
Picture a world-class surgeon whose instruments are kept in a supply room down the hall. The surgeon is fast and skilled, but for every cut, a runner has to fetch the right instrument, hand it over and take it back. Most of the operation is spent waiting on the runner, not performing the surgery. Today’s AI chips work the same way. To produce each word of a response, the chip has to pull a large amount of learned information across from memory, leaving the processor idle and burning power while data travels. Making the processor faster doesn’t fix this. What’s needed is a way to keep large amounts of information right next to where the work happens, so it’s available the moment it’s needed.
Existing approaches each solve only half of that problem. Chips built on high-bandwidth memory (HBM) can store the vast amount of information modern models rely on, but HBM limits memory bandwidth, so they struggle to deliver it fast enough for demanding workloads. Other approaches, such as on-chip SRAM, a fast type of memory built into the chip, deliver data faster but can hold only a small fraction of the information large models need.
Fractile’s approach is to bring the supply room right up to the operating table, so the runner is no longer needed. Its chips and systems use stacked DRAM, the same type of memory found in HBM, packaged directly on top of the logic. That tight coupling gives far higher memory bandwidth than HBM and significantly more capacity than SRAM, so models run faster at scale. By rethinking how memory and compute are brought together, Fractile aims to combine the capacity of large memory with fast access to it, so AI models can run faster at lower cost. The goal is to remove the dead time that makes every operation slower and more expensive than it needs to be.
Why We Invested in Fractile
1. Product and Technology Moat: A Memory-First Architecture
Fractile was built on the conviction that inference speed would be the defining constraint for AI. Fractile’s team has continued to sharpen the company’s architecture as the demands on hardware have evolved. By late 2023, it recognized that growing context lengths would make memory capacity a ceiling for any design built around speed alone. Over the next two years, Fractile worked closely with memory suppliers and manufacturing partners to develop an approach that delivers capacity and speed together, without sacrificing affordability. The result is a stacked DRAM architecture with advanced packaging. Because Fractile is full-stack, it has been able to co-design the DRAM and logic far more closely than conventional approaches allow.
That memory bandwidth matters most for the largest models as we outlined above. As Walter puts it, inference economics at data center scale come down to the cost per gigabyte of the memory employed, so capacity cannot be sacrificed for speed. Fractile’s approach of pairing high memory capacity with far higher bandwidth is intended to address the memory speed, capacity and cost implications that are important to running AI efficiently as demand scales.
2. Traction and Scale: Demand Is Forming Before First Silicon
Although Fractile has not yet shipped production chips, it is gaining momentum as demand for faster and more efficient AI inference continues to grow. The company is also earning recognition as governments look to strengthen their domestic AI infrastructure. In June 2026, the UK government identified Fractile among British companies advancing next-generation AI hardware as part of its AI Hardware Plan, which includes £150M to purchase inference chips from British firms. Fractile has also committed £100M over three years to its UK operations, including expanding its London presence and establishing a new hardware engineering and testing facility in Bristol.
This progress comes amid significant investment in AI infrastructure. Microsoft, Amazon, Alphabet and Meta plan more than $725B in combined capital expenditures in 2026, up 77% from 2025, with much of that spending directed toward AI infrastructure. While Fractile remains in the early stages of commercialization, we believe its technical progress, government recognition and investment in operational capacity position it to pursue opportunities in this expanding market.
3. Team and Founder Pedigree: Early Conviction Backed by Deep Technical Expertise
Walter's background is in robotics, where his PhD focused on using early large language models to build general-purpose robots. Towards the end of his studies, he recognized that existing hardware would struggle to keep up as compute demands shifted from training to inference. He founded Fractile in 2022, roughly two years before reasoning models began demonstrating just how much computing power advanced AI applications would require.
Fractile's leadership team brings together experienced executives and engineers with full-stack expertise spanning AI research, chip design, hardware, software and supply chain management, including talent drawn from established chip companies. Fractile has been vertically integrated since day one, designing the full system from first principles, and has recruited industry veterans who have shipped silicon at scale. This breadth of experience enables Fractile to manage nearly every stage of chip development internally, from initial architecture and design through physical layout and advanced packaging of the finished chip. The company runs its own systems build-out, has its own software stacks and does core ML research. With a full-stack team, Fractile can iterate on architectural decisions more quickly and optimize its hardware to meet the evolving demands of increasingly complex AI models.
What’s Next
Fractile’s mission is to build the hardware that pushes AI’s frontier forward. The financing is intended to help Fractile bring its first chips and systems to customers with a platform ramp targeted for the second half of 2027.
Throughout our engagement with Fractile, Georgian’s AI Lab has valued the opportunity to exchange technical insights with the company’s talented software team. We look forward to building on those conversations and exploring opportunities for collaboration as Fractile enters its next phase of growth.
We are pleased to partner with Walter and the entire Fractile team.