The Twenty Minute VCEiso Kant, CTO @Poolside: Raising $600M To Compete in the Race for AGI | E1211
CHAPTERS
- 0:00 – 0:53
Poolside’s $600M bet: earning a seat in the AGI race
Eiso frames AGI as a competitive race where timing matters, arguing Poolside’s latest raise is about securing the right to compete. He sets expectations: Poolside must execute on both capability progress and go-to-market without “stumbling.”
- •AGI framed as a once-in-a-generation inflection point (like internet/mobile)
- •Funding is positioned as an entry ticket to compete seriously
- •Race dynamics: no room for major missteps in capabilities or GTM
- •Sets the episode’s core theme: competition, scale, and urgency
- 0:53 – 3:09
What Poolside is building: AGI via software development capability
Poolside’s approach to AGI is to focus on building the most capable AI for software development. Eiso explains his view that we’ll first reach human-level capability in select, economically valuable domains before generalized parity across everything.
- •Poolside’s mission: most capable AI for coding as the path toward AGI
- •AGI definition: machines match/surpass humans across capabilities—still ‘quite a bit out’
- •Near-term future: uneven, domain-specific human-level capability
- •Foundation models as compression of web-scale data into neural nets
- 3:09 – 5:04
The missing dataset in coding: capturing intermediate reasoning and iteration
Even with trillions of tokens of code available, models lack the process data that shows how developers reach solutions. Eiso argues that the crucial missing piece is the chain of intermediate attempts, failures, debugging, and reasoning steps that lead to working software.
- •Scale of code data exists, but it’s mostly ‘final product’ output
- •Models struggle when the right kind of data is missing, not just when data is small
- •Key gap: intermediate actions—trial, error, debugging, reasoning, iteration
- •Poolside’s focus is building datasets that represent the full workflow to a solution
- 5:04 – 7:38
How to generate “uncharted” data: from AlphaGo to code execution feedback
Eiso uses AlphaGo as the archetype of synthetic data generation via reinforcement learning in a simulatable domain. He positions software as closer to deterministic environments, enabling large-scale learning through execution feedback and test results.
- •AlphaGo story: bootstrap with human games, then self-play via RL in simulation
- •Synthetic data works when the domain can be simulated and outcomes evaluated
- •Real-world domains (e.g., driving) require large-scale data capture (Tesla example)
- •Coding sits near deterministic: execution provides reliable feedback signals
- 7:38 – 8:58
Poolside’s core technique: RL from code execution in massive real codebases
Poolside trains models and then lets them explore solutions inside large environments of real repositories, using test pass/fail as feedback. This loop enables generating not only output code but also improved intermediate reasoning traces aligned with success criteria.
- •“Reinforcement learning from code execution feedback” as Poolside’s differentiator
- •Environment scale: 130,000 real-world codebases (claimed largest of its kind)
- •Models explore multiple solutions; tests act as deterministic validation
- •Synthetic generation includes intermediate reasoning, not just final answers
- 8:58 – 12:49
What’s the real bottleneck: compute, data, algorithms—and talent
Eiso argues algorithmic efficiency is table stakes across frontier labs; true differentiation comes from data advantages. Compute remains foundational because it powers both training and synthetic data generation, and he adds talent as a critical bottleneck alongside research and data.
- •Algorithms and hardware efficiency improvements are necessary but not differentiating
- •Data is the key source of model differentiation
- •Compute matters for generating/evaluating synthetic data at scale
- •Eiso expands the ‘three ingredients’ to include talent and proprietary applied research
- 12:49 – 15:46
When synthetic data works (and when it’s snake-eating-itself)
Synthetic data only improves models when paired with an ‘oracle’ that scores outputs as correct/incorrect or better/worse. Poolside’s oracle is code execution; other domains often require human feedback, which is slower and less deterministic to scale.
- •Core objection: model-generated data doesn’t help without a truth signal
- •Need an oracle: objective evaluation or preference ranking
- •Code execution is a strong oracle (tests, runtime correctness)
- •Human feedback remains essential where deterministic evaluation is unavailable
- 15:46 – 18:22
Scaling laws and the economics problem: train big, then distill to ship
Eiso believes scaling still has room—both in parameters and in data—especially via inference-time compute used for synthetic generation. But very large models are too expensive to serve directly, so the industry pattern is to distill big frontier models into smaller, deployable ones.
- •Scaling laws now include inference-time compute for data generation
- •Belief: still significant headroom in scaling parameters and data
- •Serving giant models is economically challenging for customers
- •Distillation: large model learns capabilities; smaller model inherits them for viable deployment
- 18:22 – 27:51
Model costs over 12–24 months: separating cost vs price in a price war
Eiso distinguishes the underlying cost stack (data centers, chips, energy, networking) from market pricing, which is being driven down by intense competition. Hyperscalers have structural cost advantages via vertical integration and increasingly through custom silicon.
- •Price ≠ cost: competition forces prices toward marginal cost
- •Cost stack: servers, networking, data centers, GPUs/chips, energy
- •Hyperscaler advantages: existing infra, scale, and custom silicon (Google TPUs, Amazon Trainium/Inferentia)
- •Model providers also reduce effective cost via distillation and efficiency at the intelligence layer
- 27:51 – 33:19
Compute reality check: is $600M enough, and can money buy GPUs?
Eiso says the raise is enough to be an entrant now (10,000 H200s online), but not indefinitely. He emphasizes real constraints—networking, interconnect limits, and physical buildout—meaning money doesn’t instantly translate into limitless training clusters.
- •$600M enables entry now; long-run needs will grow
- •Poolside brought 10,000 GPUs online using the new capital
- •Cluster scaling constraints: interconnect and physical limitations
- •Cash-to-compute is imperfect due to supply shortages and long lead-time decisions
- 33:19 – 36:50
CapEx vs OpEx and why data centers must look different for training vs inference
They discuss the massive infrastructure buildout required to run AI at global scale and the distinction between training and inference requirements. Eiso explains why training needs tightly coupled, co-located clusters (high bandwidth/low latency synchronization), while inference can be geographically distributed near users.
- •Spend to create models (CapEx) differs from the cost to run them (OpEx)
- •Scaling AI adoption requires data centers near users due to latency
- •Training requires tight synchronization across many machines in one place
- •Inference can be replicated/distributed; training cannot (at scale)
- 36:50 – 45:01
Nvidia’s dominance, Blackwell delays, and the next compute landscape
Eiso credits Nvidia’s early conviction and outlines the emerging chip ecosystem led by Nvidia, Google, and Amazon (with Microsoft catching up). He notes Blackwell’s delay can be strategically beneficial to incumbents on current-gen chips and that the biggest potential leap may be on inference efficiency more than training.
- •Nvidia’s moat built from early AI conviction and iteration cadence
- •Three major volume players: Nvidia + Google + Amazon (with Microsoft developing silicon)
- •Blackwell delay: helps teams currently training on H200s; training gains typically ~1.5–2x per generation
- •Blackwell’s more dramatic promise may be on inference economics
- 45:01 – 51:00
Consolidation and ‘which lab to buy’: OpenAI vs Anthropic vs xAI
Harry presses on M&A and hypothetical acquisitions; Eiso says few serious frontier labs remain independent. On choosing among OpenAI, Anthropic, and xAI, he highlights each one’s distinctive advantage—compute execution speed, product/revenue momentum, and scientific rigor—while refusing to make a simplistic pick.
- •Few remaining frontier candidates: Cohere, Reka, Mistral (xAI unlikely)
- •xAI’s advantage: rapid physical infrastructure buildout (100k GPU cluster)
- •OpenAI’s advantage: ChatGPT momentum and business traction
- •Anthropic’s advantage: rigorous research culture and methodology
- 51:00 – 55:23
Crypto vs AI: why AI trends toward centralization (for now)
Eiso contrasts crypto’s decentralization ideals with AI’s resource intensity, arguing that scarcity in talent, research, and compute concentrates power in fewer organizations. He also critiques crypto’s incentive problems (‘bad money drives out good money’) while expecting AI to support both hyperscalers and breakout independent labs.
- •Crypto’s decentralization ideals vs incentive failures and bad actors
- •AI centralizes due to scarce inputs: talent, research, compute (capital least scarce)
- •Historical analogies: many entrants, few survivors (autos)
- •Prediction: a handful of hyperscalers plus a few ‘escape velocity’ AI companies
- 55:23 – 59:01
Why stay Europe-connected: recruiting strategy and where talent clusters formed
Eiso clarifies Poolside is an American company but chose a Europe-anchored talent strategy after mapping ~3,300 potential hires globally. He argues Europe and Israel contain high-end talent that often prefers not to relocate to the Bay Area, with London strengthened by DeepMind’s legacy and additional diaspora effects (e.g., ex-Yandex).
- •Founders planned Bay Area, then built a global talent map (3,300 people)
- •Europe/Israel had deep talent but fewer ‘massively ambitious’ frontier startups
- •London talent concentration tied to DeepMind; Meta contributed across London/Paris
- •Yandex diaspora as a significant European talent pool
- 59:01 – 1:04:53
Work ethic, the ‘race’ mindset, and what it costs personally
They discuss Europe’s work-life stereotypes and the reality of building frontier AI under competitive pressure. Eiso frames joining Poolside as opting into a race that requires sacrifices, and he pushes back that highly driven people exist everywhere; you just have to find them.
- •AGI race demands sustained intensity; sacrifices are explicit upfront
- •Hiring filter: do candidates want to ‘join a race’ and compete for gold?
- •Stereotypes about Europe vs reality: high-intensity talent exists globally
- •Execution risk: must win both capability development and go-to-market
- 1:04:53 – 1:19:00
China’s AI position and the closing quick-fire: regrets, regulation, and motivation
Eiso rejects the idea that China is ‘two years behind,’ citing research output and strategic openness to attract talent. In quick-fire, he covers personal lessons (data scale), a near-sale to GitHub, concerns about chip supply disruptions, preferences on regulation (focus on applications), and deeper personal motivation and values.
- •China is not far behind; strong research output and talent strategy
- •Biggest misconception: progress halting—main risk is chip supply disruption from conflict
- •Regulation: risk of bureaucratic overhead that hurts startups; regulate end-use applications
- •Personal reflections: ‘stuff’ doesn’t matter; motivation comes from pursuing the hardest meaningful problems
