Who Pays for the Science?

Interest disclosure: T-OpenAI is a tokenized loan participation right providing economic exposure linked to OpenAI's pre-IPO valuation, issued through a dedicated issuing subsidiary, as are T-SpaceX and T-Kalshi. Following the SpaceX listing, the investment underlying T-SpaceX is being divested. Redemption opens only once a Redemption Start Date is announced, and as of 26 August 2026 none has been. Tessera's founder, and entities he controls, hold related exposure. Read what follows with that in mind.

First of three. This week's pieces work down the capital stack of the AI build (the research, the collateral, and the revenue) and ask, at each layer, who is holding a risk they were not structured to hold.

This argument has surfaced in this space twice in passing and never properly. It deserves more, because it is arguably the most underpriced structural question in the sector, and it has almost nothing to do with whether the models are good.

The question is this: basic scientific research is a long-duration, largely unappropriable risk. Who is capable of holding it?

The Tranche Nobody Wants

Technological development has a capital structure. At the bottom sits basic science: long duration, high failure rate, and, critically, returns that leak. The firm that funds a discovery captures a fraction of the value it creates; the rest accrues to imitators, adjacent industries, and the public. This is Arrow's 1962 argument and Nelson's from 1959, and it has held up. The US Congressional Research Service summarized the consensus last year: social returns to R&D typically run two to four times the private return to the firm doing the work.

That is an equity tranche. It absorbs first loss, it has no fixed maturity, and its upside is systematically expropriated. Above it sit commercialization, scale-up, and distribution: shorter duration, appropriable, financeable. Mezzanine and senior.

Failures happen when an entity holds the wrong tranche. Governments holding the commercialization tranche produce Concorde and Minitel, technically superb and commercially directionless. Private balance sheets holding the science tranche are supposed to produce underinvestment. What they are producing at the moment is the opposite, which is a problem for the theory, and this piece comes back to it.

Bell Labs Was a Containment Machine

Bell Labs is invoked as proof that private industry can fund basic science. That reading has the causation backwards.

AT&T could fund the transistor, information theory, and the laser because it was a regulated monopoly earning a guaranteed rate of return. Research costs entered the rate base. Spillovers did not matter, because the return was underwritten by regulation rather than by exclusivity. AT&T did not need to contain the knowledge; it needed only to keep the network. Bell Labs was possible because the appropriability problem had been solved by statute.

And when the statute changed, the model went with it, twice.

The 1956 consent decree, settling a 1949 antitrust suit, obliged AT&T to license its entire patent portfolio royalty-free, the transistor included. Watzinger, Fackler, Nagler, and Schnitzer studied the result in the American Economic Journal: Economic Policy in 2020: compulsory licensing raised follow-on innovation building on Bell patents by roughly 17% over five years, disproportionately from young and small firms, but only in markets outside telecommunications. That last clause cuts slightly against the neat version of the argument here and is worth stating: inside its own industry, Bell's containment held even after the decree, because it retained other means of foreclosure. What was socialized was the part Bell was never going to monetize directly.

The 1984 breakup finished the job. Divestiture cut AT&T's book value by around 70% and severed the Labs from the operating companies that funded them. Watzinger and Schnitzer's follow-on work argues plainly that the breakup reduced the Bell System's ability to appropriate its own innovation, and with it the incentive to invest. Headcount went from a peak around 25,000 toward roughly a thousand at Nokia Bell Labs today.

Everyone Who Tried It Without the Statute

The pattern repeats with unusual consistency.

Xerox PARC produced the graphical interface, the mouse, Ethernet, and object-oriented programming. Xerox was a copier company and commercialized almost none of it. Its $1 million stake in Apple was worth around $17 million at the IPO, a rounding error against what Apple and Microsoft extracted from the interface work. PARC alumni founded 3Com and Adobe; Microsoft hired Charles Simonyi in 1981 to build Word.

IBM Research implemented Codd's relational model as System R and invented SQL, then hesitated to ship it because it would cannibalize IMS. Oracle shipped the first commercial relational database in 1979. IBM followed with SQL/DS in 1981 and DB2 in 1983. The 801 RISC program seeded an industry while IBM's own RISC workstation flopped.

DuPont Central Research, which gave the world nylon, published more JACS papers in the 1960s than MIT and Caltech combined. By 2011 that number was zero. Under activist pressure the lab was dissolved in 2015-16 ahead of the Dow merger, with more than half of some 270 researchers let go.

The Obvious Objection, Which Improves the Argument

At this point a reader should be reaching for pharmaceuticals. Merck, Pfizer, the whole industry: decades of enormous research spend, sustained on competitive balance sheets, with no regulated monopoly anywhere in sight. If the tranche logic were right, that industry should not exist in its present form.

It exists because it has a statute of its own. Patents give pharmaceutical research the containment that rate-of-return regulation gave Bell: a legally enforced period of exclusivity long enough to earn back the science. Different mechanism, same function. The state solves the appropriability problem so that private capital can hold a tranche it otherwise could not.

There is a second half that makes the parallel stronger still. A great deal of the genuinely basic work upstream of a drug is not done by the manufacturer at all; it is publicly funded, in universities and national institutes, and licensed in. So the industry does not really hold the science tranche. It holds the development tranche, with statutory containment on top and a publicly funded science tranche beneath. That is a well-designed capital structure, and it took decades of policy to build.

So the pharmaceutical case is no second example of private capital holding the science tranche. It is an example of a stack in which nobody holds it unprotected: the public sector carries the science, patents contain the development, and the manufacturer takes a risk it can actually price. Bell was the other solution, one entity holding everything, with regulation guaranteeing the return.

Which means the honest count is that no industry has sustained the science tranche on a competitive private balance sheet without the state solving appropriability first, by one route or the other. That is what makes frontier AI unusual. There is no upstream public science tranche of any size beneath it, and the work itself is largely unpatentable in practice: training methods, architectural choices, and empirical know-how, protected by secrecy and speed rather than by exclusivity.

Trade secret law is a real legal protection and should not be waved away. But it protects against misappropriation, not against independent rediscovery, and independent rediscovery is precisely what a well-resourced competitor with published benchmarks and a target to aim at is able to do. A patent stops a rival who works it out honestly. A trade secret does not. On that distinction, frontier AI may be the first science-intensive sector at this scale with no containment mechanism that binds a competitor who simply does the work again.

The Current Position, and a Theory That Predicts the Opposite

Frontier AI labs are now holding that tranche at a scale none of the historical labs approached, on venture and corporate balance sheets structurally less patient than a regulated utility's.

The clearest single figure, on audited-document reporting via The Information and the Financial Times: OpenAI's R&D spend rose from $7.81 billion in 2024 to $19.18 billion in 2025, more than the company's entire $13.07 billion of revenue that year. Reported net loss for 2025 was $38.53 billion, though that figure includes substantial non-cash items and should not be read as cash burn. Anthropic burned roughly $5.6 billion in 2024, though its trajectory since has improved markedly enough that Friday's piece returns to it. Whatever the precise accounting, the shape is unmistakable: this is equity-tranche spending, funded by investors who expect appropriable returns on a venture timetable.

Here the framework has to confront something that cuts against it. Arrow predicts underinvestment when returns leak. What is observable is the opposite: an investment boom in exactly the conditions the theory says should suppress one.

Two things explain it, and both are worth holding onto. The first is race dynamics: when the perceived prize is a winner-take-most position, the private calculus stops being about expected spillover-adjusted returns and starts being about not being excluded. That produces overinvestment relative to any individual firm's appropriable return, which is a description of a great many capital cycles. The second is a genuine belief that this science is different, that scale converts research into something ownable in a way ordinary science does not.

That second belief deserves its own hearing, because it is the strongest counterargument to everything above.

The Counterargument, Which Is Real

Scale may be a containment mechanism pure knowledge never had. No one distills their way to a gigawatt of interconnect. The four largest US technology firms have guided, on the running tallies compiled after second-quarter earnings this year, to roughly $725 billion of capital expenditure for 2026, up around 77% on the ~$410 billion spent in 2025, with Alphabet raising its ceiling again in July and Meta having raised twice. Capital of that order is a genuine barrier, and no such barrier existed for PARC.

The point stands, with one qualification. Capex is a moat around deployment, the senior tranche. Around ideas it is no moat at all. On the ideas, diffusion runs at months: on Epoch AI's tracking, open-weight models have been trailing the closed frontier by an average of roughly four months, with DeepSeek the clearest case.

The honest rejoinder is that four months may be plenty. If a lab can monetize a lead before it evaporates, the leak is a cost of doing business rather than a structural flaw, and the revenue now being booked across the sector is some evidence that it can. That deserves to be taken seriously. The narrower reply is that a four-month lead is a commercial asset rather than a scientific one: it is worth something to the deployment business and close to nothing against the R&D line that produced it. The containment holds precisely where returns were already appropriable, and fails precisely where they were not.

That is the Bell Labs problem inverted. Bell had containment on the science and monetized through the network. AI labs have containment on the network and are exposed on the science.

What Usually Happens Next

When private appropriability fails, the science tranche historically migrates to a balance sheet that can hold it. Something like that migration has started: the UK's £500 million Sovereign AI Unit launched in April alongside state-funded compute, France's €109 billion commitment framed around sovereignty, Canada's compute strategy, the CHIPS and Science Act's NSF authorizations (authorized is not appropriated), and Gulf and Asian sovereign funds anchoring rounds across the labs.

The amounts are not yet material against private capex, and no claim otherwise is made here: £500 million does not move a $700 billion build. What is worth noting is the direction, and the fact that the justification offered for each of these programs is precisely the one the tranche framework predicts: that the returns to this work will not be captured by whoever funds it.

So the question for anyone holding paper anywhere in this sector is not whether the technology works. Bell Labs' science was extraordinary and largely unappropriated by the firm that paid for it; both were true simultaneously, and the shareholders still had a view. The question is which tranche is being held, and whether the capital holding it was built for the job.

Wednesday's piece takes the moat the labs are relying on, the capital expenditure itself, and asks what it is worth as security.

Sources are as cited in the text: Congressional Research Service on returns to R&D; Watzinger, Fackler, Nagler and Schnitzer, American Economic Journal: Economic Policy, 2020, and follow-on work on the AT&T breakup; OpenAI figures per audited-document reporting via The Information and the Financial Times; Anthropic figure as reported; hyperscaler capital-expenditure guidance per running tallies following second-quarter 2026 earnings; open-weight lag per Epoch AI. Figures are as of the dates and reports stated.

This is market commentary, not investment advice. It is not a recommendation to acquire, hold or redeem any Tessera product, or to take or avoid exposure to any company mentioned.

T-Tokens are tokenized loan participation rights, not equity. High risk. DYOR. Not financial advice. Not available in the US or other restricted territories. tessera.pe/terms