This article is a part of Poland Unpacked. Weekly intelligence for decision-makers
Pathway has secured additional funding. The company has now raised a total of USD 30 million, bringing its valuation to USD 500 million. The money will go toward further development of technology that could enable the creation of “AI for one dollar,” Zuzanna Stamirowska, the company’s co-founder, tells XYZ.
Pathway is a startup that has been around for years. It began in 2020 as a company using machine-learning technology to improve efficiency in logistics. That is hardly surprising: Stamirowska is a scientist whose PhD focused on forecasting in maritime trade.
Over the years, the company changed direction. It really began attracting attention in late September 2025, when the Pathway team, with the support of Łukasz Kaiser – a Pole and one of the creators of the transformer architecture that now underpins every major GenAI model – published a research paper presenting BDH (Dragon Hatchling). It is an entirely new architecture for artificial intelligence, designed to bypass the limitations of existing approaches.
The team’s achievements have won over a broad group of international investors. They have just put several million dollars into the startup in a seed round. This comes on top of the USD 10 million investment raised in 2024. Pathway has therefore secured a total of USD 30 million. Investors believe that the startup founded by Poles can deliver extraordinary results in AI at low cost.
Big money for Pathway
The funding was provided by, among others, Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital and WS Investment Co., the investment arm of law firm Wilson Sonsini. Jonathan Frankle of Databricks has also joined the investor group as an angel investor. As a result, the startup’s valuation has reached USD 500 million.
TQ Ventures became a Pathway shareholder somewhat earlier. It was the fund’s 2024 investment that set the company’s change of direction in motion.
“In 2024, when we secured seed funding from TQ, which approached us with an investment offer before the formal round process had even begun (a so-called pre-emptive term sheet), established the company in the US and moved there ourselves, we effectively began an entirely new chapter that we kept secret for about a year—until the publication of the paper on BDH, which was covered by The Wall Street Journal. It was then, in the first half of 2024, that we redirected all our resources toward building a new architecture and AI models,” Zuzanna Stamirowska tells XYZ.
Stamirowska runs the company alongside Jan Chorowski, its CTO, and Adrian Kosowski, its CSO.
Pathway has also formalized its advisory group. Łukasz Kaiser is a co-creator of the transformer architecture and a co-author of TensorFlow. He was a researcher at Google Brain and later worked at OpenAI, where he co-created the o1 and o3 models and worked on ChatGPT, GPT-4 and GPT-5. At Pathway, he advises on the direction of the company’s R&D efforts.
The group also includes Jonathan Frankle, Chief AI Scientist at Databricks; Prof. Martín Farach-Colton of NYU Tandon; and economist Jacques Attali. They support the company in scaling its models, developing its scientific vision and funding strategy, and analyzing the economic and social implications of BDH.
Intelligence for a dollar
Beyond the funding round itself, Pathway is touting an achievement in AI efficiency, measured by cost relative to performance. In the ARC-AGI-1 benchmark, the BDH-CQ model developed by Pathway set a new record for “intelligence for a dollar.” It achieved a score close to that of GPT 5.6 Luna at more than 11 times lower cost.
How does the benchmark work? The system is given several examples of transformations involving small grids of colored cells. It must independently discover the underlying rule and apply it to a new case. An answer counts only if it matches exactly, cell by cell.
BDH-CQ scored 29.5% on the public set of 400 tasks. Pathway calculates the cost of each task at USD 0.0007 (about PLN 0.003, or EUR 0.0006), less than one-tenth of a cent. The company calls this a record for “intelligence for a dollar” because, according to the public leaderboard of the ARC Prize Foundation on which the comparison is based, no system with at least the same level of performance has operated at an equal or lower cost.
For comparison, the GPT 5.6 Luna Low variant cited by Pathway scores 34.2%. That is 4.7 percentage points higher. But the cost of running the task is about 11 times greater. Pathway’s small model is therefore not “better” than a large general-purpose model. What it does show, however, is that for a specific type of reasoning, a different architecture can change the economics of computing.
“What excited us most was the combination of linguistic and visual reasoning that BDH-CQ demonstrates on ARC-AGI-1. The fact that we achieve very high accuracy with tiny models that immediately pick up the rules with virtually no scaling, across a broad range of different problems – ARC-AGI is just one example – was quite shocking. It means we have enormous cost headroom for scaling,” Zuzanna Stamirowska says.
The result was independently replicated by Łukasz Kaiser and Richard Zhong of New York University. Remigiusz Kinas of Bielik AI is also a co-author of the technical paper.
Expert's perspective
Pathway has taken on one of the toughest challenges in AI
They have taken on one of the toughest challenges in technology today: building a new engine for AI in a field where competitors are spending billions. And they have just shown that this is not fantasy. Their model reasons without generating a costly “scratchpad,” making it more than 10 times cheaper and less energy-intensive than today’s approach. Cost and energy are now real ceilings on AI’s development, so whoever raises that ceiling holds the cards.
I am delighted that, alongside ElevenLabs and Viktor, we have another Polish project playing for global stakes – and as early investors, we want to see more of them.
How Pathway is developing
Pathway positions itself as a “neolab” – a laboratory focused on researching a new kind of artificial intelligence and building its foundational architecture.
“In most cases, company valuations in this category are driven largely by the founders’ names. In our case, it is somewhat different, because a large part of the R&D work is already behind us, and our differentiation from the transformer architecture is becoming visible. The round we are discussing now was raised from investors who believed in us early on, when the risk was even higher, with the conviction that BDH could become a new path for AI,” says Pathway’s CEO.
The newly raised capital will go primarily toward computing power, including Nvidia GB300 systems, and training larger models. Adam Kurzrok, who was previously responsible for Gemini at Google DeepMind, has been appointed head of product. Among the company’s advisers is Łukasz Kaiser, a co-creator of the transformer architecture.
Chart showing Pathway’s model performance compared with other models. Photo: press materials
What Pathway has actually built
The technological core of the company is BDH. Pathway describes it as a post-transformer architecture in which memory, adaptation and reasoning are part of a single model. This is a different operating mechanism from the popular transformer architectures, which require larger contexts or new memory mechanisms.
The architecture draws inspiration from biology. The system is designed to retain information, learn through successive interactions and reason without turning every step into text.
This is where the concept of “reasoning in latent space” becomes important. Today’s most popular models using chain-of-thought reasoning develop their reasoning sequentially, generating successive tokens that effectively serve as a scratchpad. This increases the number of operations, consumes context-window capacity and drives up costs. BDH-CQ is designed to work differently.
“Chain-of-thought means that a model develops its reasoning token by token, in real time, in natural language. Imagine solving a crossword puzzle and having to describe everything you see and read in words on a piece of paper. BDH-CQ, by contrast, thinks in an abstract space, more like a chess player analyzing several possibilities at once without having to go through the inefficient step of translating that reasoning into language,” explains Zuzanna Stamirowska.
The model can still respond in natural language. According to Pathway, however, part of the reasoning takes place in parallel within a latent state, without generating and then reprocessing a long sequence of intermediate tokens.
“Right now, intelligence for a dollar is the most important metric for AI development. Ultimately, however, what matters to corporations and users is the value created by solving a problem and safety. The number of parameters is merely a means to an end—and one that generates costs. I would not say that maximizing it is in the user’s interest,” says Zuzanna Stamirowska.
What the startup is planning next
Pathway has also scaled BDH across models ranging from 1 billion to 600 billion parameters.
“We ran experiments with 600 billion parameters to test how the architecture behaves, but we did not train it further on a large volume of data because that is not our goal at this stage,” says Zuzanna Stamirowska.
The company says its initial experiments show scaling laws similar to those observed in transformers. For now, however, it wants to iterate on smaller models. Pathway has demonstrated the mechanism under specific conditions; the next step is to see how it performs in a large, more general-purpose system.
Expert's perspective
Why AI must evolve
Accessing AI models through APIs is a bit like shopping at big-box home-improvement stores. At first, they entered the market promising, “We won’t let anyone beat our prices,” shaping our habits in the process. Once we became dependent on them, we go in for six screws and some wall plugs and leave PLN 40 (about EUR 9) at the checkout, paying a premium for convenience and a closed ecosystem.
The race to optimize AI costs is now a matter of survival for many startups. If a company builds its value solely as a typical AI wrapper around OpenAI’s or Anthropic’s API, rising token costs will eat up its entire margin, because products of this kind either become cheaper or get swallowed up by the growing capabilities of new models.
Scaling a business forces evolution. The winners will be those that build a smart architecture: using the most expensive models only for the most complex tasks, while shifting 90% of repetitive work to smaller, cheaper open-source models that they fully control on their own hardware.
Where BDH could appear in companies
Pathway does not plan to start by replacing the entire AI stack at enterprises. The company is working with AWS and so-called design partners on small reasoning models, while awaiting a larger general-purpose model. Its first applications are expected to focus on narrowly defined tasks that are costly and require multiple stages of reasoning.
“I would not ask a CIO to replace their entire transformer-based stack. I would, however, ask them to identify the most difficult process in which the company is currently incurring high costs because of architectural limitations, or where transformers simply do not work well,” says Pathway’s CEO.
These are processes in which a system needs to maintain state over long periods, respond to changing information or draw on experience from previous interactions. Initially, BDH is expected to operate alongside today’s models.
According to Zuzanna Stamirowska, the first thing BDH could displace may not be the LLM itself. It is more likely to replace the infrastructure currently built around models – for example, repeated context retrieval, additional memory mechanisms, repeated fine-tuning and specialized optimization software.
Expert's perspective
Intelligence is needed by people, too
AI labs want us to think about those prestigious tasks, and by default many businesses simply use the latest models for everything because we are still in the phase of discovering use cases. In slang, we call this “tokenmaxxing”: broadly applying AI to everything in order to collect data.
Once businesses – sooner or later – demand hard ROI data and a clear cost per task, this will force the use of smart routers that can dynamically select a model based on the complexity of the task. Small, dedicated models will be created for specific tasks based on logs showing what worked, rather than by fine-tuning small models.
Pathway wants to build its own models
Zuzanna Stamirowska also points to Pathway’s longer-term direction.
“We want to build our own models,” Stamirowska says.
The long-term vision is for a model not to start practically from scratch every time it is run, rereading the entire context of a company or user. Instead, it would accumulate experience. For an enterprise, this could mean learning the terminology, exceptions, patterns and decision-making rules of a particular organization.
For an ordinary user, the first difference would be simpler.
“I think the first element visible to the average user will be its memory. A model that understands you,” adds Zuzanna Stamirowska.
Pathway is currently planning further scaling of the architecture, as well as tests in mathematics and on ARC-AGI-2 and ARC-AGI-3. In parallel, it is developing a large language model that will also reason in latent space.
The capital raised in the round will fund greater computing capacity, further experiments and the transition from an architecture demonstration to models with broader applications. Adam Kurzrok is among those responsible for the product side. Stamirowska also highlights the partnership with AWS, which is intended to facilitate access to enterprise customers and cooperation with Nvidia.
“We raised the current funding to secure the computing power needed to scale, prove and deepen BDH’s advantage. The benchmark we are presenting today, in which we set a record for intelligence for a dollar, is one of the first results of this round,” says Pathway’s CEO.
“We will demonstrate more of BDH’s capabilities, increase its accuracy and work closely with partners such as AWS and Nvidia to deliver business value to enterprises,” Stamirowska says.
Key Takeaways
- Pathway has raised USD 20 million in a seed round, bringing the company’s valuation to USD 500 million. The capital will primarily finance the acquisition of greater computing capacity and further scaling of the BDH architecture. Led by Zuzanna Stamirowska, Jan Chorowski and Adrian Kosowski, the startup has undergone a major transformation in recent years. It has evolved from a company developing logistics solutions into a laboratory working on a new architecture for artificial intelligence. Its investors include Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital and WS Investment Co.
- Pathway has raised USD 20 million in a seed round, bringing the company’s valuation to USD 500 million. The capital will primarily finance the acquisition of greater computing capacity and further scaling of the BDH architecture. Led by Zuzanna Stamirowska, Jan Chorowski and Adrian Kosowski, the startup has undergone a major transformation in recent years. It has evolved from a company developing logistics solutions into a laboratory working on a new architecture for artificial intelligence. Its investors include Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital and WS Investment Co.
- Pathway wants to develop its own models and introduce BDH into companies gradually, starting with specific, costly processes that require memory and multi-step reasoning. The startup does not envisage immediately replacing transformer-based infrastructure. Initial deployments are intended to operate alongside existing models and address problems where maintaining state, responding to changing information and drawing on experience from previous interactions are important. The company is working with AWS and Nvidia, among others, while continuing to test BDH and developing a larger language model that uses reasoning in latent space.