Last month a tokenized pension prototype landed in my review queue.
ERC-4626 vault. T-bill collateral. A "target retirement date" NFT that transfers on secondary markets. And on page fourteen of the documentation, a mortality table dated 2019.
I ran the numbers twice.
The product assumed a cohort life expectancy of 78.4 years for a Western European male. It then priced a thirty-year payout stream off a discount rate that had been set by a governance vote in March — a vote taken while stablecoin utilization on a major lending market was spiking and the supply rate touched 9.1%.
Two weeks before that review, David Solomon told an audience that artificial intelligence would extend human lifespan, and that because people will live longer, they should expect to work longer, and retirement policy should adapt accordingly.
I read the two documents side by side. They describe the same future. They share none of the same mathematics.
That gap — between the story capital tells about longevity and the arithmetic actually embedded in the products being built — is the subject of this piece. It is not a small gap. In my experience auditing these things, it is the entire distance between a narrative and an instrument.
I have spent nine years translating lending and savings mechanisms for people who will never read a whitepaper. In 2020 I led a community project that turned a lending protocol's liquidation documentation into something five thousand non-technical savers in Eastern Europe could actually follow. I learned something on those weekly calls that I have never forgotten: the part of a financial product that people misunderstand is almost never the part the builders consider important.
Nobody asked me about the collateral ratio. Everyone asked what happens to their money when the rate changes.
Hold that thought. It comes back.
Context first, because the framing matters more than the quote.
Solomon runs a bank whose balance sheet, client base, and advisory franchise all sit downstream of the pension system. When a person in that seat says people will work longer, they are not making a demographic prediction. They are describing a liability management strategy. Longevity risk — the risk that people outlive the assets set aside for them — is the single largest unfunded exposure in the developed world, and it sits on the books of exactly the institutions that pay Solomon's salary.
The transmission chain he sketched is real, and it runs like this: AI improves health outcomes, healthspan and lifespan extend, working lives extend, pension solvency improves, human capital gets repriced.
Every link in that chain is plausible. The question is whether the chain is tight enough to underwrite anything.
Here is why the crypto industry should care, and why it should care carefully.
Over the last two years, the tokenized cash complex grew from a curiosity into a genuine market. Tokenized money market funds crossed into the tens of billions. On-chain savings products, no-loss prize pools, rotating credit circles modeled on ROSCAs — all of it now sits in the same liquidity bucket. The pitch writes itself: bring the pension on-chain, cut out the intermediary, let the beneficiary own the yield.
Longevity capital is moving in parallel. Altos Labs, Calico, Retro Biosciences, the AI-drug-discovery shops spun out of the large research labs. The money is real. The timelines are not short.
And that is the first thing worth saying plainly, before the analysis gets technical: this is a slow variable wearing a fast narrative. The distance between a model that predicts protein folding and a statistically significant drop in age-adjusted mortality is measured in years to decades, and it is mediated by regulators, clinicians, and reimbursement systems that move at their own pace.
A bull market wants to trade that. It cannot be traded. That mismatch is where people get hurt.
There is a media signal buried here that I have not seen anyone name. A macro flash about a bank CEO's longevity remark ran on a crypto outlet. That is not an accident. It is what content strategy looks like when an industry is searching for adjacent narratives in a soft cycle — AI plus macro is the direction of drift. Read the drift, not just the article.
Now the arithmetic.
Longevity risk is a duration problem, and duration is unforgiving.
A defined-benefit scheme with a mature membership carries a liability duration of roughly twelve to eighteen years. Duration governs sensitivity to the discount rate, but the length of the payout stream governs sensitivity to everything else. Add one year of expected life at retirement and you have added one year of payments to every surviving member, discounted back to today.
Rule of thumb from my own modeling work: a one-year increase in life expectancy at age 65 raises the present value of a mature pension liability by roughly three to five percent, depending on the discount curve.
Put numbers on it. A ten billion euro scheme. One year of added life. Four percent. Four hundred million euros of new liability appearing on a balance sheet that in most developed markets is already carrying a funding ratio below one hundred percent.
Now the discount rate itself, because this is where the pedagogy matters most. A single euro owed in thirty years is worth roughly thirty-seven cents today at a 4.5% discount rate. At 6%, it is worth seventeen cents. Same promise. Same beneficiary. The value of the obligation more than halves because of one parameter nobody outside the actuarial profession ever discusses.
That parameter is the whole ballgame. And on-chain, nobody is discussing it at all.
Compare the liability movement to what AI health tooling currently delivers at population scale. The regulatory record is the honest anchor here. By my count of the public authorization lists, the total number of AI-enabled medical devices cleared across the major jurisdictions sits in the low four figures — call it thirteen hundred — and the overwhelming majority are assistive: triage, image prioritization, clinical documentation, arrhythmia flags. Useful. Marginally productivity-enhancing. Not mortality-moving at population level.
The mortality that AI would have to bend is dominated by cardiovascular disease, cancer, metabolic conditions, and — critically — access to care. The last of those is not a model problem.
There is a second distinction that almost nobody in this conversation makes, and it changes the labor market conclusion entirely.
Lifespan and healthspan are different products. If AI extends both, you get the outcome Solomon describes: more years of capable work, and a plausible case for a longer career. If AI extends lifespan but not healthspan, you get the opposite: more years of disability, higher care costs, and no additional labor supply at all. The compression-of-morbidity hypothesis — that illness gets squeezed into the final years rather than stretched across decades — is still a hypothesis. The pension math depends entirely on which version arrives, and the industry is pricing the flattering one.
So the chain holds at the front. AI genuinely compresses drug discovery cycles, genuinely improves diagnostic throughput, genuinely reduces clinical documentation burden. Those are measurable productivity gains and I have seen them in the data.
The chain snaps at the back. From tool efficiency to population mortality is a decade-scale causal distance, and the confounders — lifestyle, screening uptake, reimbursement, public health infrastructure — are larger than the AI effect for most of that window.
Goldman is not wrong about the direction. It is early about the timeline, and it is silent about the distribution.
Now the part that keeps me up, and the part that makes me unpopular at conferences.
The interest rate models that the entire on-chain lending market runs on are arbitrary. They have nothing to do with the real supply and demand for capital.
Let me explain what I mean, because this is not a rhetorical jab. It is a structural observation.
Take the dominant lending protocols. Their rates are set by a kinked utilization curve. Below an optimal utilization point — usually set between eighty and ninety percent — the borrow rate rises gently from a governance-set base. Above that point it rises steeply, to incentivize repayment and deposit inflow. Three parameters, plus the optimal point. Four numbers.
Who sets them? Governance. Who votes? Token holders. On what basis? A forum thread, a risk team's recommendation, and a snapshot vote.
There is no term structure. There is no yield curve. There is no connection whatsoever to the price at which a thirty-year obligation should be discounted.
And these parameters are not stable. In 2022 I watched the stablecoin supply rate on a major market travel from roughly half a percent to above nine percent and back inside a single quarter, driven entirely by utilization spikes during stress. That is an eight-hundred-basis-point move in ninety days on the instrument that on-chain products use as their risk-free reference.
You cannot discount a thirty-year liability at a rate that moves eight hundred basis points in a quarter. The variance alone destroys the product. The expected value might be fine. The path is uninvestable.
A pension is a long-duration promise. The on-chain rate complex is a short-duration reflex. Every on-chain annuity I have reviewed is a front-end instrument cosplaying as a long-duration liability.
This is not a fixable bug in a specific protocol. It is a missing market. To price a real pension on-chain you need a deep, liquid, observable long-duration discount curve. Tokenized T-bills give you the zero-to-twelve-month segment and nothing beyond it. The thirty-year point simply does not exist in this market. Until it does, the DeFi pension is a marketing phrase, not a financial structure.
I say this as someone who spent a year translating a lending protocol's risk documentation for non-technical savers because I believed — and still believe — that access to these mechanisms is a form of empowerment. Belief in the mission does not require credulity about the math. Build for humans, not just nodes. And humans retire on thirty-year horizons.
The second structural problem is governance, and it compounds the first.
On-chain governance turnout sits persistently below five percent. That is not a bug in a particular protocol. It is the steady state. Delegation concentrates it further. In practice, a handful of wallets and a few venture funds determine parameter changes that affect every depositor in the system.
Now put a pension inside that structure.
The people whose retirement is at stake are beneficiaries. The people who vote are token holders. In a mature system those two sets overlap almost completely, because the members are the governance. In a bull-market DAO they overlap almost not at all, because the token was distributed to airdrop farmers, early depositors, and funds.
Consider a concrete decision. A scheme is running an optimistic discount assumption — say six percent — that flatters its funding ratio. The honest actuarial view says four and a half. Lowering the assumption raises required contributions by roughly twenty percent. Who votes for that? Nobody holding a yield-bearing token whose headline APY depends on the assumption staying high.
So the incentive inside a token-governed pension is to keep the assumptions optimistic — which is precisely the failure mode that destroyed every underfunded defined-benefit scheme of the last two decades, now with a governance token attached.
That is the whole trick, and it is not even a clever one. It is the same accounting optimism that ran through the pension system for forty years, wrapped in a new interface.
There is a deeper incompatibility underneath, and I want to state it as sharply as I can.
A pension is a promise from the young to the old. A DAO is a promise you can revoke by selling your tokens. These are structurally incompatible.
Pensions work because they are illiquid by design. You cannot exit. Contributions are locked, cross-subsidized across generations, and enforced by law. That illiquidity is not a limitation. It is the mechanism. It is what allows a scheme to hold long-duration assets and smooth returns across decades instead of across quarters.
A DAO's defining feature is exit. Rage-quit. Redeem. Sell on the secondary market. That is the feature people love most about the structure, and it is exactly the feature that makes the structure unusable for retirement.
If your claim is transferable, it has a price. If it has a price, that price is set by the same short-horizon market that cannot price thirty years. And the moment the market gets scared, your pension is worth whatever the bid is that afternoon.
I have watched this movie. In the 2022 drawdown I ran a peer-support network for burned-out builders in Prague — two hundred people, weekly sessions, most of them watching years of savings reprice in real time. The lesson those sessions taught me was not that volatility is bad. It was that a structure that lets you exit at the worst moment is a structure that will make you exit at the worst moment.
Pensions exist to prevent that. DAOs exist to enable it. The intersection of the two is mostly a landing page.
A ledger can record a promise. Only a community can keep it.
Which brings us to the part Solomon's framing quietly steps around: distribution.
His word was opportunity. For whom?
Extend the working life by three years and you have given a knowledge worker three more years of compounding, seniority, and remote-friendly income. You have given a roofer three more years of joint damage. Life expectancy at 65 is not uniform — it tracks income, occupation, and geography with brutal consistency. A uniform retirement age increase is therefore regressive by construction: the people with the shortest expected retirement contribute the most years and collect the fewest.
On-chain governance does not fix this. By default it makes it worse, because token-weighted voting is the purest available expression of one-dollar-one-vote, and the people with the shortest life expectancy also hold the fewest tokens.
I have spent enough time in policy rooms — I advised a European regulatory task force last year on decentralized governance standards — to know the fix here is not technical. It is differentiated policy design, portable benefits, and a safety net for people who cannot or will not work longer. That is political work. Code can support it. Code cannot substitute for it.
So what can the technology honestly claim?
Three things, none of them glamorous.
Verifiable credentials and proof-of-personhood for benefit delivery and cross-border portability, so a worker who moves between jurisdictions does not lose accrued entitlement in the transfer.
Data provenance for the clinical datasets that AI health models train on, so that when a model claims an outcome, the record of what it was trained on is auditable. That is a real contribution to the longevity thesis, and it is not a token.
Escrow and portable credentialing for re-skilling, so the transition costs of a longer working life are financed at the point of transition rather than deferred to a cohort that never gets asked.
None of these produce a yield. All of them address the actual problem.
Now the contrarian turn, because I have been building to it.
The same institution that now says AI will extend your life published, two years ago, a research note estimating that generative AI could expose three hundred million full-time jobs to automation.
Sit with that. Same shop. Both stories. Both probably true in direction.
The synthesis is not that you will live longer and work longer. The synthesis is that you will live longer and be less employable for more of it. Those are the same sentence with the optimism removed, and only one of them gets a keynote slot.
Here is the second contrarian point, and it is the one I would stake the article on.
Longevity is not the binding constraint on pension solvency. Fertility is.
The old-age dependency ratio is retirees over workers. Longevity grows the numerator slowly, one cohort year at a time. Fertility shrinks the denominator fast, because a sustained decline compounds through every working-age cohort simultaneously. A one-year gain in life expectancy at 65 moves the dependency ratio by a few percentage points over a generation. A ten percent decline in completed fertility moves it by more, sooner, and permanently.
Demographers know this. Actuaries know this. Solomon, who employs actuaries, knows this.
So why is the spotlight on longevity?
Because longevity is the flattering variable. It is the one where the story is technology giving you more years, not your society stopped having children and the arithmetic no longer closes. One of those is a product narrative. The other is a political crisis.
I am not accusing anyone of deception. I am pointing at a framing choice, and framing choices made by people with balance-sheet exposure deserve to be named.
And here is the prediction that will make me tiresome at the next conference.
Within ninety days of a remark like this, someone launches a Longevity Index token. Healthspan yield vault. Biological age staking. Something with a countdown clock in the interface.
It will raise money. It will have a governance forum. It will set its own parameters, on its own kinked curve, with turnout under five percent.
And that will be the tell that the industry learned nothing from the last cycle — that our reflex when confronted with a structural problem is to financialize the headline and ship the interface.
The correct response to Solomon's remark is not a product. It is a refusal, followed by the boring work: long-duration yield curves, honest actuarial assumptions, and governance designed for the people who will actually retire rather than the people who bought the token.
What I would track, if I were building anything adjacent to this: authorization counts for AI-enabled medical devices as a proxy for real clinical penetration, not press releases. Clinical endpoints on the drug-discovery side, not valuations. And the dependency ratio releases from national statistics offices, which will tell you the truth about pension solvency roughly five years before any CEO does.
So where does that leave us.
The question is not how long you will live. The actuaries will argue about that for another decade and the answer will arrive too late to change your savings rate.
The question is who sets the discount rate on those extra years, and whether you are in the room when the vote happens.
Right now you are not in the room. The rate is set by a curve someone tuned in a governance forum, by a turnout that would embarrass a student union, on assumptions inherited from a table dated 2019.
That is fixable. It is just not fixable by shipping faster.
Build for humans, not just nodes. And remember what those long calls taught me: education is the ultimate yield — the only one that compounds whether the market is up or down, and the only one nobody can vote away from you.