Eiso Kant on Poolside, Compute and AI for Software Development

Poolside co-founder and CTO Eiso Kant joins Harry Stebbings to discuss AI for software development, compute, funding and the company's ambitions. Kant has a direct commercial interest in Poolside, and the technical claims and forecasts are his account rather than independent evaluations.

Poolside is pursuing more capable AI for software development as a path toward broader machine intelligence. Kant distinguishes a distant vision of general human-level capability from earlier gains in particular economically valuable tasks. Both the path and timing are hypotheses.

He describes foundation models as learning generalizations from large datasets. In his view, scarce data can limit reasoning, planning and domain-specific capability. For coding, he estimates about three trillion usable code tokens and hundreds of millions of public codebases; these are his estimates and depend on definitions and data rights.

Public code typically shows the final artifact, not all the intermediate attempts, tests and reasoning involved in producing it. Kant says Poolside aims to create data about that process as well as the final output, so a model can learn from attempts that fail and are revised.

He uses DeepMind's AlphaGo as an analogy: a model can explore moves and use feedback about game outcomes to improve. Later systems could learn in simulated environments without initial human games. His account is simplified and does not establish that software engineering is equivalent to the game of Go.

Kant contrasts domains that can be simulated with the messy physical world. He cites vehicle data collection as an example and predicts Tesla will lead full self-driving. That is his competitive prediction, not a verified conclusion about safety, capability or future market leadership.

Poolside calls its approach reinforcement learning from code execution feedback. Kant says the company lets models try solutions in an environment containing approximately 130,000 real-world codebases, using execution and tests as feedback. The scale and superiority claims are company statements; passing available tests alone does not prove code is correct, secure or meets every requirement.

He separates compute, data and algorithms, adding that talent is essential. In his view, research that improves compute efficiency is necessary to remain competitive, while distinctive data can differentiate models. Compute also matters for generating and evaluating synthetic examples.

Kant uses compression as a mental model for explaining why larger models may retain and generalize more information than smaller ones. He expects continued improvements from scaling and better algorithms, while acknowledging that size has limits. This analogy is explanatory, not a complete theory of model behavior.

Synthetic data is not automatically useful merely because a model generates it. Kant describes producing many candidate solutions and then using an evaluator to distinguish better from worse—tests in code, or human feedback in less directly verifiable domains. The quality and coverage of the evaluator remain important.

He believes software development is a promising early domain because execution feedback can be scaled with compute, while collecting expert human judgments is harder. He nevertheless describes models at the time as useful assistants with a substantial remaining gap to human developers. His claim that software will close that gap first is a company thesis.

On scaling laws, Kant expects room to increase data and model size, and views synthetic-data generation as another use of compute. He does not claim a known endpoint. Better algorithms and hardware could also improve efficiency, though the pace and results are uncertain.

One commercial approach is to train a large, capable model and distill some of its behavior into a smaller model that is cheaper to serve. Kant says Poolside follows this broad path. Distillation can trade off capability and cost; the interview does not give performance or unit-economics measurements.

He distinguishes model price from provider cost. Hyperscalers with data centers, networking, energy and their own chips can have different cost structures from a company buying hardware or cloud capacity. He expects competition to put pressure on margins, while recognizing that model quality and efficiency also affect cost.

Kant discusses Google's TPUs, Amazon's chips and Microsoft's silicon efforts as examples of vertical integration, alongside Nvidia's hardware. His comparisons of manufacturing and cost advantages are industry assessments made at the interview, not audited cost data.

In the longer term, he expects demand for machine intelligence across many problems and believes hardware and software efficiency will improve. He presents the broader economic and social benefits as a vision, not a demonstrated outcome or a precise forecast.

To assess an opportunity, Kant considers three questions: how valuable the work is, how large the gap is between current models and human capability, and whether enough useful data can be obtained to narrow it. He applies that framework to coding, speech and driving; his assessments of those gaps are subjective and time-specific.

Stebbings asks whether GitHub's code gives it an unbeatable data advantage. Kant says public code is broadly available to developers of models, while access to private code requires appropriate rights. He emphasizes intermediate execution data and research as potential differentiators; his broad statement about private-code training should not be read as a universal legal determination.

Poolside had raised $600 million in total, including a $500 million round. Kant says this made it an entrant in the race, not fully funded indefinitely, and reports bringing 10,000 GPUs online. These are company-reported financing and infrastructure figures, not evidence of independent model performance.

He says building larger training clusters faces practical limits in networking, algorithms, hardware supply and physical infrastructure. Money matters, but it does not instantly buy unlimited connected compute. He does not give a definitive amount needed to reach AGI.

Compute procurement also requires relationships, contingency plans and decisions well ahead of deployment. Kant says he sees demand outstripping supply but acknowledges hyperscalers have a better overall view. His assessment reflects the market at the time of the interview.

Asked whether entry requires $100 billion, he distinguishes building a global hyperscale infrastructure footprint from training increasingly capable models. He thinks the first may require sums of that magnitude, but says nobody knows the exact cost of closing the capability gap.

Kant distinguishes spending to train a model from the operating cost and physical footprint required to serve it. A model must eventually create enough value to justify those investments, he argues. His expectation of a very large inference buildout is a prediction about adoption and demand.

Training large models requires extensive communication among machines as they update during learning; serving separate requests can often be distributed differently. Kant uses this distinction to explain why training clusters have demanding networking and power needs, without claiming every training setup must use one physical room.

He sees Nvidia, Google and Amazon as major chip participants and considers AMD and Microsoft's own silicon important possibilities. His ranking reflects his perspective at the time, not a guarantee about future market share.

Kant says a delay in Nvidia's Blackwell generation gave Poolside more time with its existing H200 cluster. He estimates successive generations have historically improved training performance roughly 1.5–2 times, while expecting a potentially larger inference gain. The figures and expectations are his estimates and depend on workload.

He says a more efficient chip does not necessarily make an existing cluster unusable or force an immediate upgrade. It can change the cost and competitive balance; the calculations depend on hardware, workload and economics.

Rather than focus only on what the next GPT release delivers, Kant expects the coming decade to reveal broader effects of AI. He identifies intelligence, energy and space as major long-term challenges. These are ambitions and forecasts, not established outcomes.

When comparing funding with larger rivals, he argues that dollars buy compute but do not translate directly into talent, original research or data advantage. Physical constraints can give smaller companies time to compete. That is his strategic argument as a founder raising capital in the same market.

Poolside deliberately did not include Google, Microsoft or Amazon in the cited round, Kant says, because it saw an independent path. Nvidia did participate in a smaller strategic role, and he describes close collaboration on hardware and software. This relationship is relevant to his comments about Nvidia.

He sees only a limited number of independent frontier-model companies sufficiently far along, naming several while acknowledging he is observing them from outside. Industry consolidation and each company's future are uncertain.

Asked to choose among hypothetical OpenAI, Anthropic and xAI investments, he declines to pick without meeting their leadership teams. He identifies strengths in infrastructure execution, distribution and research at each. His descriptions and the hypothetical valuations are time-specific opinions, not investment recommendations.

Kant describes building Poolside as an obsession requiring sacrifice, and reflects on the pressure of hard decisions and disrupted sleep. He admires Elon Musk's willingness to take risks across difficult projects; that is his personal assessment, not evidence that any particular strategy will succeed.

He contrasts decentralization ideals in crypto with the concentration that scarce compute, talent and research can produce in AI. He argues that a small number of independent companies may nonetheless join hyperscalers as major players, explicitly including Poolside in the future he hopes for.

Enterprise AI adoption varies by use case, he says: some projects remain experiments, while developer assistance is already useful. He expects speech recognition and image generation to face increasing commoditization. These are market judgments rather than measured outcomes for every deployment.

Kant clarifies that Poolside is an American company with employees across the United States, Europe and Israel. The founders mapped roughly 3,300 potential hires and saw strong AI talent outside the Bay Area; they built a European presence partly to reach people who preferred not to relocate.

He credits DeepMind, Google, Meta and Yandex with developing talent pools, including researchers who left Russia. He rejects a simple stereotype about European work ethic and says Poolside is explicit with candidates that it intends to compete intensely. These are hiring and management preferences, not evidence that long hours guarantee progress.

Stebbings argues that much internet value was created long after the first wave. Kant agrees later applications may create more value than model builders alone, but expects some companies formed now to become important long-term players. The outcome remains uncertain.

Poolside aims to build both models and an end-to-end software-development business, Kant says, while expecting others to create additional value on top. This is its strategy for avoiding a situation where expensive breakthroughs mainly benefit downstream companies.

He rejects the assumption that China is necessarily years behind in AI, citing published research and talent. He urges making Western countries attractive to researchers from China and cautions technology leaders against assuming their technical expertise settles geopolitical questions. His comparative assessment is an opinion, not a comprehensive measurement.

In quick fire, he says he has become more convinced of the importance of data scale. He regrets the financial opportunity of not selling a prior company, Source, to GitHub, but values the journey and relationship with co-founder Jason Warner that followed. The hypothetical proceeds are not realized returns.

He identifies disruption of chip supply chains by global conflict as a possible brake on progress. He would choose Mark Zuckerberg as a board member for his long-term conviction, an assessment that reflects Kant's own leadership preferences.

On regulation, Kant favors accountability for AI's end uses over restrictions on training compute, arguing that administrative costs can hurt small startups. This is a policy preference from a frontier-model founder; it does not resolve debates about risks from models or their deployment.

He describes living on a sailboat with his partner and dog as a way to keep life simple and focus on experiences rather than possessions. Extensive travel for Poolside has meant spending little time at home since starting the company.

In a premortem, he says Poolside could fall behind if it stumbles either on model capabilities or on bringing a product to market. He sees both as essential to the company's chances, not guaranteed successes.

Kant wishes people asked more often why founders pursue their work. He says he feels most at peace when working on an ambitious, difficult problem with people he values, even though the effort is intense. The conversation closes with thanks.