DIGITAL CASTE — THE SORTING HAS ALREADY STARTED

In 2025, AI minted more than 50 new billionaires.

Not over a decade. Not through a slow accumulation of market share. In a single year.

Founders in their twenties. Executives who left one AI company to start another and were worth a billion dollars before the ink dried. According to the Financial Times, citing Bloomberg data, the top ten wealthiest US tech founders and executives added more than $550 billion to their combined net worth — taking their total to nearly $2.5 trillion. Elon Musk became the first person in history to surpass $500 billion in personal wealth. At least 20 existing billionaires gained a combined $460 billion through AI investments alone. Forbes reported more than 50 new AI-linked billionaires across foundation model firms, SaaS companies, and businesses explicitly built to replace human workers with software.

One year. One technology wave. More new fortunes than most revolutions produce in a generation.

AI captured 61% of all global venture capital in 2025 — $258.7 billion out of a total $427.1 billion — according to an OECD policy brief published in February 2026. That's double AI's share from just three years earlier. The two largest foundation model companies, OpenAI and Anthropic, together captured 14% of all AI venture investment. Mega-deals above $1 billion now represent roughly half of all AI investment value. The market is concentrating at the top even within AI itself.

The San Francisco Bay Area has more billionaires than New York. More homes sold above $20 million in San Francisco last year than in any recorded year in the city's history.

This is not a technology story. It is a geography story. A concentration story. A story about who is inside the wall and who is being sorted by it.

THE MECHANISM.

The digital caste isn't announced. It doesn't have a name. It doesn't send a letter.

It operates through systems you already use. Your credit score isn't just a number — it's an input into algorithmic decisions about what insurance you qualify for, what interest rate you pay, whether your rental application clears. Behavioral profiling from data brokers — your browsing patterns, your purchase history, your location data — feeds into models that determine access to financial products, healthcare plans, and employment screening. These systems don't call it sorting. They call it personalization. They call it risk assessment. They call it efficiency.

Thomas Piketty's central argument in Capital was that when the rate of return on capital exceeds the rate of economic growth — when owning things earns more than doing things — wealth concentrates until something breaks. The AI moment is Piketty accelerated. Not because the math changed. Because the speed changed. What took a generation now takes a funding round.

A landmark study published in Science found that a widely used health management algorithm exhibited significant racial bias — at identical risk scores, Black patients were considerably sicker than white patients. Fixing the disparity would have increased the percentage of Black patients receiving additional care from 17.7% to 46.5%. The algorithm wasn't malicious. It was trained on healthcare cost data, which reflects historical inequity in access, and then it scaled that inequity across millions of patients.

A 2025 study published in npj Digital Medicine tested four major AI systems — including ChatGPT and Gemini — across psychiatric patient cases. When patient race was explicitly or implicitly indicated, the models proposed inferior treatments. A Cedars-Sinai study found that multiple AI platforms omitted medication recommendations for ADHD cases when race was stated, but included them when it wasn't. The University of Michigan documented that medical testing rates for white patients run up to 4.5% higher than for Black patients with identical age, symptoms, and triage scores — and that bias gets baked directly into any AI trained on that data.

The federal government took notice. The Agency for Healthcare Research and Quality and the National Institute on Minority Health and Health Disparities convened an expert panel and published guidance on eliminating algorithmic bias in clinical care. When federal agencies stand up a dedicated task force, that is not a research finding. That is an acknowledgment that the problem exists at scale.

The sorting algorithm isn't being built in the future. It is running right now, inside financial institutions, insurance underwriting models, the hiring pipelines of the same companies that just replaced their junior workforce with AI. The people designing those algorithms are inside the wall. The people being sorted by them are not.

WHAT ACCESS ACTUALLY MEANS.

In late 2024, National Taiwan University Hospital deployed the world's first clinically operational AI system for early pancreatic cancer detection. PANCREASaver — developed in collaboration with Taiwan's national mathematics institute — analyzes CT scans using deep learning, flags suspicious lesions invisible to the human eye, and integrates directly into radiology workflows. National clinical validation trials showed 80% sensitivity for early-stage tumors under 2 centimeters and overall diagnostic accuracy exceeding 90%. It is not a pilot. It is not a research project. It is a live diagnostic service, integrated into multidisciplinary care, available now.

PANCREASaver holds FDA Breakthrough Device Designation and patents in both Taiwan and the United States. It is on the path to US clinical deployment.

Mayo Clinic's Dr. Ajit Goenka, a radiologist and nuclear medicine specialist, stated plainly in 2025: AI can help diagnose pancreatic cancer almost a year before its clinical presentation. When detected early and confined to the pancreas, the five-year survival rate rises from 13% to 44%. MIT's PRISM model, validated on large-scale US clinical data from a federated network spanning multiple institutions, identified 3.5 times more high-risk patients than current screening criteria. A Nature Medicine study trained on 6 million patients from two countries predicted pancreatic cancer risk up to three years in advance from clinical histories alone.

The science exists. It is not theoretical. The question is who gets access to it when it scales — and on whose terms.

PANCREASaver launched in Taiwan as a self-pay service. No national health insurance reimbursement. The patients who can afford out-of-pocket AI diagnostics get early detection. The patients who cannot wait for symptoms. And the model itself was trained predominantly on Taiwan's national health database — a Han Chinese population. The PMC literature is explicit about this limitation: reproducibility across diverse populations requires validation studies that do not yet exist for this tool.

Meanwhile, the Trump administration cut approximately $2.7 billion in NIH funding in the first three months of 2025 — including a 31% decrease in cancer research funding compared to the same period the previous year. A study published in JAMA Internal Medicine found that 383 clinical trials lost grant funding between February and August 2025. More than 74,000 patients enrolled in those trials were affected. Of the terminated trials, 181 specifically studied cancer — per the National Cancer Institute's own grant termination records.

Dr. Anupam Jena of Harvard Medical School put it directly: clinical trials are not light switches. Cutting off funding mid-trial wastes research dollars and puts patients at risk. The patients most dependent on publicly funded trials — those with rare cancers, pediatric diagnoses, conditions underserved by pharmaceutical investment — are precisely the patients the market does not reach on its own.

The breakthrough is real. The distribution is a different question entirely.

AI CURES IT. HUMANS DECIDE WHO GETS IT.

That is not a prediction. That is a description of the current architecture.

The decisions about who gets access to precision oncology, to AI-matched clinical trials, to the diagnostic tools that catch cancer before it becomes late-stage — those decisions run through insurance systems, hospital networks, the same financial infrastructure quietly being restructured around new ownership models. The efficiency gains do not distribute themselves. They flow through ownership structures.

City of Hope's CEO said in January 2026 that too many patients are still getting lost in a system not built for the realities of modern cancer care. The system isn't broken. It's working as designed. The question is who designed it and what they were optimizing for.

WHAT THE CASTE ACTUALLY IS.

It isn't slavery. It isn't feudalism. It is something new enough that we don't have clean language for it yet.

It is a tiered access system built on data, capital, and algorithmic inference — where your position in the tier is determined before you know the tier exists, by systems you didn't consent to, trained on data you didn't know was collected, optimized for outcomes you weren't part of defining.

The people at the top of the tier are the 50 new billionaires. The companies capturing 61% of global venture capital. The institutions writing the models. The sovereign wealth funds and family offices moving into new asset classes before the public knows what those asset classes are.

The people at the bottom are everyone else. Not because they failed. Because the sorting happened before they arrived.

THE WINDOW.

The window to shape how this concentrates is not permanently open.

Infrastructure is being laid before the governance conversation has started. In AI healthcare, the tools exist in clinical deployment in other countries while validation for diverse US populations remains unfunded. The same federal agency cutting cancer research funding is being warned by its own commissioner that the US is losing clinical trial ground to China. The FDA chief said this month that the US fell behind China in Phase 1 clinical trials conducted in 2024 and called existing oversight processes non-competitive.

The pattern is consistent across every domain this series has documented. The infrastructure moves first. The public conversation about who it serves comes later — usually after the people who built it are already inside it and everyone else is outside.

That's been the pattern with every previous concentration of power. The railroad. The broadcast spectrum. The internet. Each created enormous wealth. Each sorted. Each left behind a class of people who were told the tide would lift all boats.

The tide is rising. The boats are not distributed equally.

AI cures it. Humans decide who gets it.

That sentence is the whole piece. Everything above is the evidence.

The cure is coming. The question is whether it arrives as infrastructure — something everyone can access — or as a product. Something you qualify for. Something the algorithm decides you deserve.

That decision is being made right now. Not in a government hearing. Not in a public referendum. In a series of very quiet, very consequential choices about who gets to train the models, who owns the data, who funds the validation studies, and who writes the terms of service.

The caste doesn't announce itself.

It just becomes the system you were born into.

This is the ninth piece in The Rubicon. One piece remains.