Investment committed prior to full IC process. Stage 2 (independent research) skipped at David's instruction — Council run directly against Stage 1 voice notes. Liq pref update post-Council: \$10B stack (not full \$16.2B), softens Bear downside assumption.

In David's own words — unedited voice note transcript

01 · The Belief

Real-world models are required for embodied AI. Jan LeCun left Meta to build them. Brett from Figure AI talks about them. The market understands LLMs but underestimates what's needed to automate physical, engineering, and manufacturing businesses. Jeff Bezos — arguably top-four CEO of all time — understands building real businesses (supply chains, factories, people, equipment) better than any pure software entrepreneur. Blue Origin and Amazon's robotics divisions provide the proprietary real-world datasets that are the raw material. The thesis: agentify manufacturing and engineering with models trained on physics, material science, and automated process data — not next-word prediction. The Berkshire JV ($100B private equity fund) will acquire best-in-class industrial businesses, agentify them using both LLM-based agentic workflows for knowledge work and real-world models for engineering and manufacturing, and produce businesses that are materially better, more marginal, with improved products and processes.

02 · The Business

Prometheus will build a real-world model — the reason for the scale of capital raise. Revenue streams: licensing to robotics manufacturers and anyone doing embodied AI or non-LLM-based workflows. Enterprise value of the model itself could be enormous. Separately: the Berkshire JV as GP of a $100B PE fund — acquiring and running businesses agentically using the real-world model. Two value streams: the model layer and the operating business layer.

03 · Why Now

LLMs have proven the power of the paradigm but exposed their constraints — predicting the next word doesn't help a robot understand its environment. Real-world models are the natural next evolution for different use cases. Jeff Bezos doesn't want to miss the AI revolution, and unlike Elon, he has the time to focus. Putting his hand up as co-CEO is the signal — he wants this in his legacy alongside Amazon.

04 · Why This One

Arguably one of the best management teams to back in any endeavour. Jeff is operating-business-first, AI-science-second — gives him rationalism and realism in implementing technology into real businesses, as opposed to pure science-first studios. Berkshire's willingness to partner is validation that even Buffett thinks Jeff and his team are the right people to deploy this technology into industrial assets. The JV structure means even if the real-world model needs to pivot to a competitor model, the portfolio businesses still get agentified and still generate enormous value.

05 · The People

Jeff Bezos and an ex-Google co-CEO. Limited detail on the broader team from David's diligence at this stage, but the view is that this is one of the best management teams to back in any endeavour. Hard to see this being a zero regardless of which direction they pivot.

06 · The Numbers

$38B pre-revenue is unprecedented by old metrics, but relative to where OpenAI and Anthropic are heading, the model layer + business layer combined TAM for embodied AI is arguably larger than the TAM for knowledge-work AI. The $350K entry is the cost of a lottery ticket on a world-class team addressing the largest industrial transformation of the century.

07 · The Risks

Asymmetric — David doesn't see this as a zero. Risk 1: not a super-healthy return. Risk 2: liquidity — but IPO expected within 3 years (per Polis). Bezos has kept it off-market, but at this scale public capital markets are inevitable. Risk 3: AI bubble sentiment, but David's view is that AI adoption compounds exponentially and we're moving from hundred-billion to multiple-trillion valuations.

08 · Portfolio Fit

Three AI bets via KDD/CIH: Anthropic ($215K), Apptronik ($250K — humanoid robots/embodied AI), Flourish Labs (~$200K). Prometheus ($350K) is a separate family investment — not a KDD/CIH vehicle. Circa $1M invested across AI in total. Prometheus slots in cleanly alongside the CIH portfolio: physical AI / real-world models / industrial, differentiated from the others. Building a genuinely diversified AI portfolio across model layer, embodied AI, healthcare AI, and industrial AI.

09 · The Exit

IPO expected within 3 years per Polis. If they show progress on the real-world model and the $100B starts deploying, David would consider this a long-term hold post-IPO — potentially a 20-year hold analogous to long-term Amazon shareholders. The stock, not the SPV, is the real exit vehicle.

10 · The Decision

Not every day you get to back somebody like Jeff Bezos in a startup. And I don't even think it's a startup — with this team, this amount of money, this experience, it doesn't feel very startupish. High probability they execute.

Prometheus is betting that physical industries — aerospace, defence, semiconductors, automotive, pharmaceuticals — are about to be transformed by physics-aware AI in the same way software was transformed by the internet. The core product is the "Artificial General Engineer": a generalized AI platform that accelerates engineering simulation, design iteration, and manufacturing workflows by orders of magnitude. The thesis is that traditional numerical-solver architectures (CFD, FEA) are a bottleneck in every major industrial sector, and that Prometheus can replace or radically augment them with machine learning.

The four-layer revenue model moves from SaaS engineering tools → enterprise manufacturing platforms → government/defence contracts → the $100B Manufacturing Transformation Vehicle (a Berkshire Hathaway-style permanent capital fund acquiring and modernising industrial assets to generate proprietary training data). The moat, if it materialises, is the data flywheel: own the factories, generate the data, train the models, sell the software back to the industry.