What Year Is It?

I. The most important sentence in finance this year

Larry Fink's 2026 letter contains a claim that almost nobody read carefully: "tokenization today may be roughly where the internet was in 1996… picture a bridge being built from both sides of a river… The task for policymakers is to help build that bridge as quickly and safely as possible." Set aside that the largest asset manager on earth has adopted the thesis crypto has argued for a decade. Look instead at what the sentence concedes — that policymakers decide the pace — and at the year.

1996 is a flattering year to claim. It is before the mania, before the collapse, before the shakeout. It means everything is early, nothing is decided, and any position is defensible. It is the year every builder wants to be in.

It is also checkable, and BlackRock's own balance sheet argues against it. Of roughly $150 billion in digital-asset AUM, about $65 billion is stablecoin reserve management and about $80 billion is ETPs. BlackRock did not enter this market by building anything native. It entered by selling custody, wrapping, and reserve management — three competences it already had. That is not 1996 behavior, when incumbents ignored the thing entirely. That is the behavior of an incumbent arriving after the infrastructure is built, which in the internet's case happened much later.

So: what year is it actually? The question sounds rhetorical. It is not — it can be fitted.

II. The fit

Take the Nasdaq Composite's post-peak path from March 2000 and the crypto drawdown from its own peak, index both to 100, and search for the time-scaling factor that minimizes the distance between them. If crypto is a compressed rerun of the dot-com arc, there is a stretch factor that makes the two curves lie on top of each other, and the position of "today" on the stretched clock is the answer.

The best fit is s = 1.83: one crypto month behaves like 1.83 dot-com months. The crypto cycle runs at roughly 1.8× speed — which is itself the most useful single number here, because it converts every dot-com interval into a crypto interval. On that clock, 27 July 2026 maps to late October 2008, with a confidence band running from March 2007 to July 2009. The obvious objection — that the fit is just matching crypto's crash to the 2008 financial crisis — was tested by refitting with that period excluded; the answer moves to December 2007, inside the same band. And the stretch is doing real work: forcing s = 1 makes the fit 72% worse.

Now the honest part, which matters more than the point estimate. The dot-com arc was a V; crypto is a W. Over 56 months since its peak, DeFi TVL posted seven major reversals and bitcoin five, against the Nasdaq's zero over the mapped window. DeFi recovered to 96% of its all-time high in October 2025 and then fell 60% in under nine months. A single-clock model is therefore structurally mis-specified, and I would not defend "October 2008" to the month. What survives is the ordinal claim, which is all the argument needs: we are past the shakeout, not before it. Not 1996.

One divergence deserves its own line, because it breaks the analogy in crypto's favor. The Nasdaq needed 15.1 years to reclaim its peak. Bitcoin's market cap reached 194% of its 2021 peak within four years. Price recovered; participation did not. That combination — capital returning while users leave — is not something the dot-com era ever did, and it is the tension the rest of this essay is about.

III. What the 2000s did to capital

If we are past the shakeout, the relevant question stops being "will this work" and becomes "who captures it." The dot-com era answered that question with unusual clarity, and the answer is brutal for infrastructure.

Total return from the March 2000 Nasdaq peak to today. Cisco, which built the internet's plumbing, returned 2.6× — below the index it helped create. Apple, which entered music in 2003 and phones in 2007, returned 358×. Provenance: daily histories re-fetched and recomputed independently by a verifier from a second provider; Cisco's peak close corroborated against its FY2000 10-K price table; revenues re-parsed from SEC XBRL.

Cisco is the cautionary tale nobody tells properly. It did not fail. It tripled revenue — $18.9B in FY2000 to $56.7B in FY2025 — and its shareholders still underperformed the index over the same twenty-six years. Building the essential infrastructure of the era, and executing on it, was not sufficient to capture the era's value. Meanwhile Apple, which was not an internet company in 2000 and arrived years after the crash, returned 358×.

The mechanism is visible in the physical layer. At the end of 1998, by the FCC's own survey, incumbents ran about half their fiber lit — while the overbuilders ran Qwest 2%, Williams 1%, GST 2%. Ninety-plus percent of new capacity was dark at the moment it was built. Internet transit collapsed from $675 per Mbps in 2000 to $5.00 in 2010, a 99.3% fall. The infrastructure became abundant and nearly free, and the surplus flowed to whoever sat closest to the user.

The modern rhyme is exact, and I sampled it live rather than assert it: across 293 consecutive Base blocks, the base fee sat at 0.005 gwei — the protocol floor — in every single one. Not "cheap." Pinned at the minimum the protocol will accept. Blockspace is the dark fiber of this cycle, and the same conclusion follows: the surplus will not stay with the people selling it.

One more number, because it reframes the whole period. At the March 2000 peak, world internet penetration was around six percent. The greatest infrastructure buildout in modern history was undertaken for a market that essentially did not exist yet — and it was still, in the end, correct. Being early is not the same as being wrong. It is only the same as not being paid.

IV. What this technology actually makes cheaper

Before forecasting anything, it is worth doing what Musk does to rockets: strip the thing to physics and ask what the irreducible costs actually are. Finance has four. Verifying trust in a counterparty or issuer. Achieving settlement finality. Holding capital against risk. Distributing to human beings.

Blockchains lower one and a half of the four. They genuinely collapse reconciliation — the messaging and matching half of settlement, which is most of what settlement costs in practice — and they genuinely lower the cost of collateral mobility and composability. That is real and it is not small.

But they do not lower the cost of trusting an issuer: a tokenized claim on a dollar is exactly as good as the entity holding the dollar, which is why stablecoins re-concentrated. They do not create finality — finality is a legal and balance-sheet fact, not a cryptographic one; a chain can be certain about its own state and still leave you with an unenforceable claim. They do not reduce required capital by a single dollar, because capital requirements attach to risk, not to record-keeping. And they actively raise the cost of distribution: self-custody, key loss, no chargebacks, no recourse.

That inventory is the most useful paragraph a builder can carry, because it makes the strategy deterministic. This technology must win on reconciliation and collateral mobility, where its advantage is physical. Everywhere else it can only win on politics — which is to say, on rules, licenses, and permission, which are slow and can be revoked. Any product whose pitch rests on the other two and a half costs is fighting physics with a narrative.

It is worth noting that Ethereum's founder has effectively conceded this. Vitalik Buterin's own framing now centers on "low-risk defi" — payments, savings, fully-collateralized lending — with the flat admission that "crypto does not have magic secret sauce for sustainably creating much higher yields." The differentiator he claims is access, not return. That is the same conclusion the cost inventory forces.

V. Will banks hold bitcoin? The rulebook already answered

This is the question every "institutions are coming" thesis rests on, and it is usually argued with sentiment. It does not need to be. It is arithmetic, and the arithmetic is decisive.

Basel's cryptoasset standard assigns Group 2 assets — bitcoin among them — a 1250% risk weight, and caps a bank's total Group 2 exposure at 2% of Tier 1 capital, with the expectation it stays under 1%. Critically, the cap explicitly includes ETF and ETN holdings, so a bank cannot route around it by buying IBIT. Then do the sum. The eight US G-SIBs held $1,116.5 billion of Tier 1 capital at 31 March 2026 — parsed from all eight 10-Qs, not estimated. Two percent of that is $22.3 billion: 1.7% of bitcoin's market cap.

What the rulebook permits, against what already exists. Every large bank on earth, simultaneously at its maximum permitted crypto exposure, reaches roughly $92.8B [INFERRED — global Tier 1 estimated from BIS Group 1 aggregates, range $4.2–5.1T] — barely more than US spot bitcoin ETFs already hold. Provenance: Basel SCO60 as amended (July 2024) read from the BIS PDF; US Tier 1 from eight Q1-2026 10-Qs; BTC market cap live.

Extend it to every large bank on the planet and the ceiling is about $92.8 billion — roughly 1.06× what eleven US spot bitcoin ETFs were already holding at the end of Q1. The entire global banking system, at its regulatory maximum, could not be bitcoin's marginal buyer. The "banks are coming to buy" thesis is not merely optimistic; it is arithmetically foreclosed.

The capital math explains why nobody will even approach the ceiling. At a 1250% risk weight, a bank posts roughly $100 of capital for every $100 of bitcoin — about $187.50 at JPMorgan's actual capital requirement. Owning a zero-yield asset dollar-for-dollar with equity implies a breakeven of roughly 16% per year just to clear cost of capital. A subtlety most commentary misses: Group 2a and Group 2b cost the same for an unhedged long position, so the "but bitcoin is liquid" argument buys hedge recognition, not relief. For the risk weight to make holding merely plausible it would have to fall to around 400% — equity-like treatment — which is a political decision, not a market one.

Custody is the opposite. It consumes only operational-risk capital and earns fees on assets the bank never owns. The rulebook does not merely discourage banks from buying bitcoin — it assigns them the servant role. They will hold, custody, and issue claims; they will not hold the assets.

Who decides, precisely? Basel writes the ceiling, but Basel is a committee, not a legislature. Its standard is not self-executing: the UK's PRA said in May 2026 that its expectations remain interim and it expects to consult "in 2028 at the earliest," and a Federal Register search this session returns zero US banking-agency documents implementing a cryptoasset capital rule. So the honest answer is that the constraint binds today by anticipation rather than by law — banks are behaving as though the rule is real because their capital planning must assume it will be. Which means the decision chain runs: Basel proposes → national regulators dispose → bank boards act years ahead of both.

The behavioral tell is beautiful. JPMorgan's Q1 2026 10-Q contains zero instances of "blockchain," "crypto," "digital asset," "stablecoin," "tokenization," or "bitcoin" — across a document where "deposits" appears 152 times. The same bank launched a deposit token on Coinbase's Base network in November 2025. Strategy live; materiality nil. That gap is the entire current state of bank crypto adoption, in one filing.

VI. Who serves whom

If banks are servants and infrastructure doesn't capture, the question becomes who owns the customer. Here the asymmetry is not close.

Reported users, Q1 2026, from primary filings. Five custodial front-ends alone account for roughly 692 million accounts (not deduplicated). Coinbase's monthly transacting users: 8.2 million, down 15% year over year. Morpho, DeFi's most successful distribution story, reports 1.4 million users cumulative since inception — so the true gap against period-active comparators is wider than shown.

PayPal alone has 53× Coinbase's monthly transacting user base. Coinbase's own 10-K has stopped disclosing "verified users" entirely. Self-custody is smaller than either: the widely-cited 181 million monthly active addresses correspond to an estimated 40–70 million actual people [INFERRED — a16z's own range], and daily active addresses across 27 EVM chains average 2.3 million.

Then the detail that settles the argument about what retail actually wants. The most successful crypto retail product ever built is the spot bitcoin ETF — and 79% of its assets sit outside 13F reporting, which is to say with retail brokerage accounts and smaller holders. Every one of them holds a share, not a key. Given a free choice between self-custody and a ticker in an account they already had, retail chose the ticker, overwhelmingly.

The economics compound the point. Circle pays out 58% of its revenue in distribution costs, and Coinbase's stablecoin revenue alone — $305M in Q1 2026 — is 5.5× Circle's entire net income. The distributor eats the issuer. Meanwhile Block's bitcoin business turned $1,796M of revenue into $68M of gross profit, a 3.8% margin: distribution owners run crypto as a retention feature, not a profit centre. Which is exactly why they will buy capability rather than build it.

And they already do. Coinbase has originated $2.17 billion of bitcoin-backed loans running on Morpho. Société Générale's FORGE became the first regulated bank to do the same. Morgan Stanley's E*TRADE outsourced custody and settlement to Zerohash. Yet the word "Morpho" appears zero times in every relevant SEC filing — Coinbase's, Robinhood's, PayPal's, Block's, Circle's, JPMorgan's. The infrastructure is contractually load-bearing and narratively invisible. Morpho's own name for the pattern is the DeFi Mullet: fintech in the front, DeFi in the back. That is simultaneously the winning position and its curse — you are essential, and nobody knows your name.

VII. The evidence against me

Every thesis should be attacked by its author before it is attacked by a reader, so I had the "there is no retail" premise stress-tested directly. It partly failed.

Nine crypto-native venues currently clear more than $100M in annualized protocol revenue with no custodial incumbent in front of them — Hyperliquid at roughly $785M, pump.fun and PumpSwap at $448M, Axiom $244M, edgeX $236M, Phantom $142M, and others. Retail is not absent. It is present, direct, self-custodial, and paying real money.

But look at what it is buying. All nine sell speculation. Across the market, speculation out-earns credit-and-yield by 3.6 to 1 — $3.92B against $1.10B — and lending across 146 protocols generates $328M, less than pump.fun alone. The revealed preference of on-chain retail is not the sober credit product that every institutional thesis (including mine) assumes is coming. It is the casino.

Two more data points sharpen the discomfort. Robinhood's event-contract revenue went from $3M to $104M year over year — now a tenth of net revenues — while its crypto revenue fell 47%. And ICE put $1.6 billion into Polymarket for roughly 23% of the company, describing its own role as distributing Polymarket data to institutions. The incumbents are not only absorbing crypto's infrastructure; they are buying into its speculation funnel, which is the one part everyone assumed they would refuse to touch.

I cannot fully rebut this. The most I can say honestly is that user counts for those nine venues are private, so nobody can demonstrate they serve a large number of people rather than a small number of very active ones — and that a 3.6:1 revenue ratio measures what exists today, not what a permission change makes possible. But a builder who ignores this is choosing a thesis over a tape.

VIII. The fork

Put it together and the strategic landscape reduces to two coherent positions. They are nearly opposite, and most projects are confusedly halfway between them.

Sell speculation, directly. Retail is there, native, and paying — the nine venues prove it. No incumbent stands in the way, distribution is organic, and revenue arrives immediately. The costs: you are in the casino business with the worst regulatory surface, your revenue is cycle-hostage, and the incumbents are now buying into this lane too, which means your window is a window and not a moat.

Sell capability to whoever owns the customer. This is the mullet, and every structural force in this essay points at it: it matches the one-and-a-half costs the technology genuinely lowers, it matches the servant role Basel assigns to banks, and it already has named existence proofs. The costs: you are invisible, you do not own the customer, and your margin is whatever a distributor consents to pay you.

The named panel splits on exactly this axis, which is how you know it is the real question. Duffie, Shin, Gorton and the BIS school hold that the value of a tokenized claim comes from what backs it — reserves — so the winner is whoever holds the reserve account. Fink and Buterin hold that the value comes from where it can go — 24/7, composable, global — so the winner is whoever owns the venue. Coinbase's Q1 2026 votes for both simultaneously: consumer transaction revenue −48% year over year, stablecoin revenue +11%. The speculation business shrinks while the dollar-rail business grows, inside the same company.

My reading is that the winner is the venue that also owns a reserve account — but that is a description of Coinbase, not a plan for anyone else. The buildable version for a small team is narrower and, I think, correct: own a capability that both the reserve-holders and the venue-owners need and neither wants to build. The market map from this research has specific empty cells — hedging and structuring for every buyer type, credit rails for brokers and neobanks, and anything at all for corporate treasuries. Those cells are empty not because they are unimportant but because they require options and credit expertise that neither a bank's crypto desk nor an exchange's product team currently has.

And who is walking into the wall: standalone retail apps whose educated user is precisely the user a 439-million-account firm will serve without education; emission-funded protocols, whose yields dilute with their own success; non-dollar strategies, which MiCA already ran as an experiment; chains selling blockspace priced at the protocol floor; and — the newest entry, from the counter-evidence above — anyone building sober retail credit products directly, against a tape that says retail wants the casino by 3.6 to 1.

IX. The scoreboard

Five falsifiable claims, so this can be scored rather than admired.

1. The plumbing stays invisible: no top-ten fintech or bank names a DeFi protocol as material in its filings before mid-2027. If one does, the mullet is graduating into a recognized supplier relationship and the negotiating leverage of infrastructure is better than I claim.
2. Basel stays unimplemented in the US: no Federal Register cryptoasset capital rule before mid-2027 — meaning bank behavior continues to be governed by anticipation, not law.
3. The speculation-to-credit revenue ratio stays above 2:1 through mid-2027. If it narrows below that, real credit demand has arrived on-chain and the sober thesis is winning earlier than the tape suggests.
4. Coinbase's monthly transacting users do not return to growth before mid-2027; crypto-native distribution continues to lose share to custodial incumbents.
5. If the 1.83× clock holds, the crypto equivalent of the smartphone era — the product that makes the infrastructure ordinary — lands around 2028, not this cycle.

X. 1996 or 2008

Here is why the year matters so much. If Fink is right and this is 1996, then everything is early, the shakeout is ahead, and the correct posture is to build anything and wait. If the fit is closer to right, we are somewhere after the crash and inside the absorption — the infrastructure is built and priced at its floor, the aggregators are assembling, the incumbents have arrived to sell custody and wrapping, and the surplus is already migrating away from the people who laid the fiber.

There is a detail from that period I keep returning to. At the equivalent moment on the dot-com clock, the product that would define the next twenty years already existed and looked like nothing: the iPhone did $123 million of revenue in FY2007 and $1.8 billion in FY2008, inside a company that now does $416 billion. Nobody in the fiber business saw it as their vindication. It was a phone.

So the useful question is not whether the crypto equivalent is coming. It is which currently unimpressive thing it is. My candidate is on the record above and it is deliberately unglamorous: dollar rails and credit capability, sold to whoever owns the customer, invisible inside products that will never mention a protocol — $2.17 billion of loans originated through infrastructure that appears zero times in anyone's filings. That is exactly what a defining product looks like at this stage, which is to say: like a rounding error, with the wrong people's names on it.

The disillusion in this reading is real, and I would rather state it than sell around it. The people who build the next financial system will very likely not be the people who own it, and the word "DeFi" will retire the way "information superhighway" did. But Cisco's shareholders and Apple's shareholders were both looking at the same internet in 2002. Only one of them was positioned for what it became — and the difference was not who understood the technology better. It was who understood where the customer would be.