*Part 1 of two. Part 2, "Intelligence Authoritarianism," takes up the technical premise this essay leaves open.

The prefix “deep-” marks the places where a cheap process runs at a new depth: deep learning, deep fakes. Deep-waste belongs in the same register. The familiar kind of waste sits at the visible edges of an economy, in the landfills, the failed startups, the burned venture capital that everyone already knows how to price. Deep-waste runs further into the substrate. It is the quiet sequestration of the scarcest inputs a civilization has for producing knowledge, its best researchers, its frontier compute, and the capital that could fund either, into closed organizations that return almost nothing verifiable, on the strength of a promise built so it cannot be checked.

This essay concerns a specific class of institution where that sequestration is easiest to see: the second wave of private AI laboratories. The revenue-generating leaders are not the subject. The subject is the stealth megaraise “neo-labs” that have raised billions at double-digit-billion valuations inside eighteen to twenty-four months, with little or nothing shipped. The argument moves in four steps: the ledger, the belief the ledger runs on, the pattern that belief produces, and what the whole structure finally rests upon.

1. The Ledger

Count the cost. — Luke 14:28.

On one side stands the population of public scientists who hold almost the entire heavy tail of human scientific talent: the faculty, postdocs, and research staff in computing, engineering, mathematics, and the physical and life sciences at the 187 R1 universities in the United States, together with the scientists of the national laboratories beside them. Counted conservatively, this is on the order of 150,000 people. They are the engine that has produced essentially the entire verified corpus of modern science, along with the foundational technology the private sector later commercializes.

On the other side sits the average researcher inside a neo-lab.

The public side’s funding is thin by any measure. The National Science Foundation operates on roughly $9 billion a year, about $8.75 billion in the FY2026 budget, and that sum spreads across some 2,000 institutions and the hundreds of thousands of researchers they employ. Narrow it to the directorates closest to what the neo-labs claim to do and it shrinks further. The Computer and Information Science and Engineering directorate runs around $1.05 billion, Engineering around $0.95 billion, and the physics-bearing part of Mathematical and Physical Sciences a comparable figure, so the computing, engineering, and physics envelope most relevant to AI work comes to roughly $2 to $2.5 billion a year, divided among more than a thousand laboratories. Counting every federal source that reaches these scientists, including the Department of Energy’s Office of Science and the relevant slices of other agencies, the total runs to perhaps $20 to $25 billion a year. Per researcher, the directly relevant federal flow comes to something like $130,000 to $170,000 a year. Add institutional support, state funds, and indirect recovery and the all-in figure reaches perhaps $300,000 to $400,000. Over the eighteen-month window that matters here, that is somewhere between $200,000 and $600,000 per scientist, the shaded band in Figure 1.

Now the other side. Take a basket of roughly twenty second-wave labs founded since 2023, the ones in Figure 1. Together they have raised on the order of $25 to $30 billion against a combined headcount of about 2,500 people. The arithmetic mean is something like $10 to $12 million of fresh capital standing behind each researcher. Weighted toward the purest cases, the labs that have shipped nothing public, the figure climbs past $40 million a head. At the far end of the distribution a single lab has raised on the order of $150 million per employee and carries a paper valuation near $1.6 billion per employee, with no product and no revenue. None of these is chosen as a freak. They are the right tail of a distribution whose body already sits one to two orders of magnitude above the public band.

Step back to the whole sector and the figures stop reading as startup news and start reading as macroeconomics. Roughly $258 billion flowed into AI in 2025, a record, with something like $80 billion of it going to foundation-model labs. The twenty-odd labs in the basket raised, inside eighteen months, more than the National Science Foundation and the Department of Energy’s Office of Science deploy together in a full year.

One fairness objection arrives immediately. Am I setting a one-time cumulative raise against an annual public budget, a stock against a flow? The objection does not hold here, which is the point of choosing these labs. Their raises occurred inside eighteen to twenty-four months and they are still raising. Measured over the same window, the comparison is flow against flow.

The contrast resolves into a single comparison. The capital standing behind one neo-lab researcher exceeds the funding behind a comparable public scientist by something between twenty-five and forty-five times across the basket, and by five hundred to a thousand times at the extreme. Figure 1 draws this on a logarithmic axis, the only axis on which it fits. Set that capital intensity against output and the second measure points the other way. The verified, field-advancing results the neo-labs produced over this window come nowhere near what the public institutions and national laboratories produced beside them. Figure 2 plots the two together, capital per researcher climbing steadily to the right while verified results refuse to follow it. The gap is sustained, and it is widening. This is the first sense of deep-waste, a maximal concentration of inputs returning a minimal yield of anything the world can check.

2. The Belief the Ledger Runs On

A gap like that cannot rest on fundamentals, because there are no fundamentals: no revenue, no product, in several cases no published result. It rests on something else, and naming that something precisely matters, because the obvious accusation misses.

The obvious accusation is fraud, that the principals know there is nothing there and use their reputations as deliberate cover. That charge cannot be made from the outside, and it does not need to be. Almost no one here is lying, and lying is not required for the structure to work. Sincere belief is the better fuel.

Here is why. A claim such as “we will reach safe superintelligence on a straight shot, with no interim products” is, for now, unfalsifiable. There is no intermediate milestone, no output, no way to mark it wrong. An unfalsifiable claim cannot accrue evidential credibility, because no evidence could move it. Whatever credibility it commands in the market therefore has to come from somewhere outside the evidence, and the only thing on offer is the reputation of the people asserting it. Reputation is not decorating the bet here. It is the entire collateral. A sincere believer with a marquee track record transmits more credibility than a cynic could, and leaves nothing falsifiable to prosecute.

The reputation is also being spent across domains. A track record earned in scaling deep networks is real, demonstrated competence. The straight-shot-to-superintelligence bet is a different proposition with its own track record, which is empty. Credibility transfers from the domain where it was earned to the one where it has not been, and the transfer stays invisible precisely because the name is the same.

The structure is reflexive at the reputational layer, not only the price layer. A valuation that climbs sixfold in under a year with zero observable output is rising largely because the prior valuation rose and the next round cleared higher. The marquee number itself becomes part of the reputation, the lab everyone is watching, and the reputation then re-justifies the number. Reputation underwrites price, and price feeds back into reputation. Nowhere in the loop is there an external referent.

What turns this from an odd private bet into a circulating financial object is the secondary machinery, and it wraps the whole private-AI ecosystem, leaders and neo-labs alike. A valuation set by a tiny circle of inside-diligence investors becomes a public reference price. Special purpose vehicles pool outside money into single concentrated stakes. Layered SPVs, tokenized wrappers, and platform listings then push exposure down to retail buyers who are told outright that they cannot evaluate the company. The fragility shows most plainly where the numbers are largest. In May 2026 the tokenized shares of the two best-known private AI labs, listed on a Solana-based platform, fell from about $1,400 to about $900 in a single day once the companies declared such transfers void. That is a vivid measure of how loosely the circulating prices are tied to anything underneath. Opacity joined to a marquee number and aimed at buyers who cannot perform diligence is the exact substrate these structures monetize, and the neo-labs, with the least underneath, are the cleanest substrate of all.

The structure perpetuates itself, and it needs no architect. The issuers often fight the gray market themselves. The field’s two largest labs have declared unauthorized SPV and tokenized transfers void, one of them naming the platforms involved. That the most powerful issuers cannot suppress the secondary market cuts against any single-actor conspiracy and shows the dynamic for what it is, emergent and ecosystem-wide and sturdier for being emergent. Nobody has to intend it. The ecosystem selects for whatever narratives sustain the belief, because those are the narratives that clear the next round.

3. The Pattern: Choosing the Unfalsifiable Problem

Watch which problems these labs choose, and a pattern resolves.

They tend to avoid the grand claims that carry hard, near-term verification. For the most part they do not promise to cure a named cancer by a named date, because that promise gets checked in a wet lab, on a timeline, against a control arm, by people who can say no. They gravitate instead to problems where intermediate progress stays vaguely grounded, where “we are making progress toward superintelligence” admits no decisive disconfirmation, and where the absence of milestones can be rebranded as discipline, the straight shot with no interim products.

This is the verification asymmetry. Some domains have cheap oracles, where a result checks itself for free: the proof verifies or it does not, the program runs or it does not, the agent wins the game or loses it. Other domains have expensive oracles. Wet-lab biology, materials, anything confirmed against physical reality, all answer slowly, serially, at cost, and without mercy. The neo-lab pattern is to work in the cheap zone, the deferred zone, or the wholly abstract one, and to keep the terminal claim in a register no near-term experiment can touch.

The sleight lives in the gap between what is promised and how the work is arranged. The deliverables these labs describe, systems that transform medicine and materials and the economy itself, would each be a physical-world event of the first magnitude if they were real. Their promised impact is exactly the impact of confirmed discovery. The work, though, is structured to avoid the falsifiable checkpoints that confirmed discovery must pass. They promise the destination of the most demanding and most checkable science while taking the one route that has no mile-markers.

Where the cheap-oracle successes are real, a system that solves olympiad geometry or writes competitive code, the credibility from them gets transferred quietly onto the expensive-oracle promises. It solved the proof, therefore it will cure the disease. The transfer is unearned. It is the cross-domain reputation laundering of the valuation game again, performed this time by the system’s own record rather than its founder’s. Competence verified in a checkable domain is capitalized as authority over the uncheckable ones.

In the vocabulary of the philosophy of science, this is a degenerating research programme dressed as a disciplined one, a programme that retrofits and defers instead of predicting and corroborating novel facts, with the missing checkpoints reframed as a virtue. The pattern is the operational signature of deep-waste rather than an accident of it. The waste is protected by the choice of unfalsifiable ground.

4. Conclusion, and the Question This Essay Leaves Open

Pull the three moves together. Scarce capital, scarce talent, and scarce compute concentrate into closed loops that admit little verification. Those loops are priced on reputation, since reputation is the only collateral an unfalsifiable claim allows. The pricing is reflexive and self-perpetuating, insulated by a deliberate preference for problems that cannot be checked. The distributed public research base, the R1 institutions and the national laboratories that hold the real heavy tail and produce the real corpus, operates on a rounding error beside it. This is what deep-waste names. The waste is not the money itself, which an economy knows how to value and discard. It is the quiet strip-mining of the substrate of distributed progress to feed a machine that returns little anyone can verify.

Honesty requires naming the one place this verdict could be wrong, because the whole edifice rests on it. The deepest defense of these labs does not reduce to “geniuses beat fields.” It is a claim about discontinuity. The claim holds that we stand at a point where the slow, field-paced, distributed human process of discovery is about to be compressed or replaced by the very systems under construction, so that the historical base rate, the one in which progress is cacophonous and cumulative, is about to stop applying. If that is true, concentrating everything into a single closed node is not waste at all. It is the rational response to a phase change.

So the charge is not yet complete. It hangs on a single load-bearing premise, the discontinuity. If the discontinuity is real, the concentration is justified and this essay misreads a revolution. If it is unreal, whether unsupported or finally incoherent, then the concentration is the strip-mining described here, and the reputational pricing is collateralized against nothing.

Anticipation of Part 2

Part 2, Intelligence Authoritarianism, takes up that premise directly. Beneath the discontinuity claim lies an older conviction with a long and uneasy lineage: that a few exceptional minds can stand in for the distributed, error-correcting process through which knowledge is actually made. Part 2 traces that conviction to its root, then asks whether it can survive contact with the constraints that govern energy and information. What those constraints permit, and what they rule out, is where the question left open here is settled.

Addendum — Sources and Output Benchmarks

*Supporting material for “The Era of Deep-Waste,” Part 1 (The Ledger), and for Figures 1 and 2.

List 1 — Sources used to compile the ledger

The capital, valuation, headcount, and public-funding figures behind Figure 1 and the per-capita comparison were drawn from the following sources. Where private-company headcount or valuation is contested, profile aggregators (Tracxn, Crunchbase) were cross-checked against primary press reporting.

Macro AI venture capital

Crunchbase News — Global venture funding in 2025 set records

Neo-lab capital, valuations, and headcount

Public research — funding and headcount

Public-side field-advancing results referenced in the report

Additional named primary sources referenced in the prose but not directly linked above (canonical and easily located): OECD policy brief on AI venture capital (2026), the Carnegie Classification 2025 update, NSF NCSES survey Graduate Students and Postdoctorates in S&E (NSF 26-308), the DOE Office of Science FY2024 budget, and reporting from Bloomberg, TechCrunch, CNBC, Reuters, the Financial Times, and The Information on individual neo-lab rounds.

List 2 — How the y-axis in Figure 2 was qualified

Figure 2 plots, on its vertical axis, the number of verified field-advancing results an entity produced in the window (≈ mid-2024 to mid-2026). To avoid the quantity trap, this is not a paper or patent count. A result qualified only if it was a discrete, externally verifiable advance that moved the technical frontier of a field — a measured physical first, a demonstrated new capability, a deployed standard, or a designed-and-experimentally-validated artifact — and not a funding round, a product launch, a benchmark score, or an award. (Awards were excluded deliberately: the prize is an event; the underlying advance it honors often happened years earlier.) Each result is counted once.

The list below is illustrative, not exhaustive — it is meant to convey the bar, so a reader can fill in the rest. It is weighted toward AI, robotics, computation, materials science, and physics. The first group (public research and national labs) is why the public points sit high on the y-axis; the second group shows the same bar applied to the neo-labs.

Public research and national laboratories

Most precise measurement of the muon’s anomalous magnetic moment(127 ppb) — Fermilab Muon g-2 collaboration, final result, June 2025. news.fnal.gov
Repeated fusion ignition with record energy gain (8.6 MJ out from 2.08 MJ in) — Lawrence Livermore National Laboratory, National Ignition Facility, 2024–2025. lasers.llnl.gov
Ambient-pressure superconductivity in bilayer-nickelate thin films — Stanford / SLAC (Hwang group), Nature 638, 935–940, Feb 2025. nature.com
First standardized post-quantum cryptography algorithms (FIPS 203 / 204 / 205) — NIST, Aug 13, 2024. nist.gov
De novo enzymes with complex multistep active sites, crystal structures within 1 Å of design — Baker Lab / UW Institute for Protein Design, Feb 2025. bakerlab.org
RFdiffusion3 — open-sourced model designing DNA binders and advanced enzymes — Baker Lab / UW IPD, Dec 2025. genengnews.com
First de novo-designed proteins with anti-cancer activity (IL-2 mimetics) — UW Institute for Protein Design, Nature. ipd.uw.edu
LigandMPNN — atom-context protein sequence design, 100+ experimentally validated binders — Baker Lab, Nature Methods, 2025. bakerlab.org/publications
AERIS AI Earth-system model trained across 60,000+ GPUs, and Aurora exascale system opened to science — Argonne National Laboratory / ALCF, 2025 (Gordon Bell finalist). anl.gov
Trillion-particle cosmological simulation on Frontier — Oak Ridge National Laboratory, 2025 (Gordon Bell finalist). anl.gov (finalist list)
Casgevy — first FDA-approved CRISPR/Cas9 gene therapy, Phase-3 results in NEJM, Apr 2024 — academic-origin CRISPR commercialized by Vertex / CRISPR Therapeutics. crisprtx.com
Mobile ALOHA / ALOHA 2 — low-cost bimanual mobile manipulationvia imitation learning — Stanford University, CoRL 2024. mobile-aloha.github.io

Neo-labs (the same bar applied)

ESM3 / esmGFP — generative protein language model that designed a novel fluorescent protein ~500M years of evolution from any natural one — EvolutionaryScale, Science (doi 10.1126/science.ads0018), 2024–2025. science.org
π0 (pi-zero) — first generalist vision-language-action robot foundation model, open-sourced Feb 2025 — Physical Intelligence, Oct 2024. pi.website/blog/pi0
The AI Scientist — fully automated research pipeline (ideation → experiments → manuscript), later peer-reviewed in Nature (2026) — Sakana AI, Aug 2024. sakana.ai/ai-scientist
Marble — generative 3D world-model API — World Labs, late 2025. worldlabs.ai
Open-weight frontier model family (Mistral Large, Magistral reasoning, Devstral) — Mistral AI, 2025. mistral.ai

Scoring notes: Thinking Machines Lab’s Tinker (Oct 2025) was scored as a product, not a field-advancing result, so its y-value is 0. EvolutionaryScale’s ESM3 is genuine, but the company was effectively absorbed into Chan Zuckerberg Biohub in 2025 — the standalone neo-lab did not persist. SSI, Reflection AI, Skild AI, Periodic Labs, Magic AI, and Liquid AI had no verified field-advancing result in the window at the time of writing.