Investing in Generative Art

On June 15, 2023, Sotheby’s sold the estate of a dead hedge fund. Three Arrows Capital had gone into liquidation the previous summer owing billions, and its liquidators put the fund’s digital art through two sales titled “Grails.” The star lot was a picture of pegs and a string: Ringers #879, one of a thousand outputs of an algorithm written by Dmitri Cherniak in early 2021, in which a line wraps around a set of points to produce a shape that a collector had long ago nicknamed “The Goose.” It hammered at $5.4M and made $6.2M including fees, roughly double its high estimate (Sotheby’s; CoinDesk; Decrypt).

Twenty months earlier, at the top of the cycle, the same fund had bought a different output of the same algorithm. Ringers #109 changed hands privately on October 2, 2021 for 2,100 ETH — about $6.9M at that day’s ether price (Artnet; The Crypto Times). Those two sales are the whole of this segment’s record at the top, and they are close: the private price paid at the peak sits about 11% above the public one realised at the bottom, in a forced sale, at a 280-year-old auction house. Nothing in the NFT market has come near either figure since. A market whose two highest prices are outputs of one program owned by one bankrupt fund is a market with a canon and almost no tape.

That is the argument for this corner of the market and, read the other way, the warning about it. The argument is that a handful of algorithmic series behave like art rather than like tokens: they are catalogued, curated, exhibited and bought by people who intend to keep them, and the institutions turned up. The Centre Pompidou accepted eighteen donated NFTs in February 2023 including Autoglyph #25, given by Larva Labs; LACMA accepted twenty-two works donated by the collector known as Cozomo de’ Medici the same month (Decrypt; CoinDesk, February 13, 2023); and in December 2025 MoMA took eight CryptoPunks and a complete set of eight Chromie Squiggles into its permanent collection (ARTnews; The Block, December 2025).

The warning is the arithmetic underneath. Art-related NFT sales on platforms outside the art trade ran to $2.9B in 2021, fell to just under $1.5B in 2022 (Art Basel and UBS, The Art Market 2023), were $197M across the whole of 2024, and by the first quarter of 2025 DappRadar counted $23.8M in the art category across the quarter — the last art-category count we can source before the firm announced its shutdown on November 17, 2025 (CoinDesk). A market that trades a few million dollars a week does not have a floor in any meaningful sense; it has a last price.

This guide is about what you would actually be buying: how the works are made and issued, what the canon is and why it is short, how price is discovered in a market with a few dozen serious buyers, what happens to the picture if a server goes dark, what the round trip costs through each venue, what the IRS does to the gain, and a $20,000 purchase followed through to the exit in dollars. The hub’s flagship guide, Investing in NFTs, covers the wider market — the bubble, the crash, the dead collections, the PFP blue chips; this one stays inside the part of it with a curator.

The object you buy is not a picture. It is a token that carries a number, and a program that turns that number into a picture the same way every time.

The mechanism is worth getting exactly right, because every valuation argument in the segment rests on it. An artist writes a program — usually in JavaScript, using a drawing library such as p5.js or three.js — that takes a single random-looking input and draws an image from it. The artist does not choose the input. When you send the mint transaction, the blockchain produces a hash from the transaction itself, and that hash is written permanently into the token as its seed. Run the artist’s script with your token’s hash and you get your work; run it with someone else’s and you get theirs. The output is deterministic, reproducible by anyone, and unknown to everybody — artist included — until the moment of purchase. This is why the trade calls the act of buying a mint rather than a purchase: you are commissioning an output, not selecting one.

Three consequences follow, and they are the whole investment case.

The first is that the series is a set, not an edition. A print edition of 250 is 250 near-identical objects, and the market prices them as one asset with a condition discount. A generative series of 1,000 is 1,000 genuinely different objects from a single authored system, so the market prices them individually, the way it prices drawings. Dispersion inside a series is enormous and permanent: the cheapest output and the most admired output of the same algorithm can differ by two orders of magnitude, and no amount of market recovery closes that gap, because the difference is in the object.

The second is that the artwork is the algorithm. The collectible convention that emerged, and that the auction houses adopted, treats the script as the work and each output as an authorized realisation of it. That is a genuinely new object in art history and it is also, commercially, the reason a curator can take the segment seriously: there is an author, an intention, a body of work and a critical vocabulary, which the PFP collections built from layered trait art largely lack.

The third is that provenance is native. Every output’s seed, mint date, mint price and full ownership chain is public and permanent. In the physical art market, provenance research is a profession; here it is a database query. That removes the single largest source of fraud in fine art and replaces it with a different one, covered in the section on permanence.

Fewer than a dozen series carry the segment, and the reason is a fact about supply rather than a fact about taste: almost every canonical work was issued in a nine-month window between late 2020 and mid-2021, and the window closed.

Autoglyphs came first and remain the technical outlier. Larva Labs, the two-person studio behind CryptoPunks, deployed the contract on April 5, 2019 and all 512 were created within about four hours, generated entirely by the smart contract itself: no server, no image file, no external library. The contract emits a grid of ASCII characters and the “artwork” is that string. Nothing about an Autoglyph can stop working while Ethereum runs, which is a claim no other series in this guide can make. Larva Labs took no fee: minting required a donation of 0.2 ETH — about $35 at the time — and Larva Labs directed the proceeds to the climate charity 350.org (larvalabs.com; contemporaneous coverage).

Art Blocks, founded by Erick Calderon — who signs his work Snowfro — launched in November 2020 and industrialized the form. Its innovation was not artistic but infrastructural: a standard contract that stores an artist’s script, mints a hash on purchase, and renders the output on demand, so an artist could issue a capped algorithmic series without writing any blockchain code. Calderon’s own Chromie Squiggle was project zero, launched on November 27, 2020, and became the platform’s signature object: a coloured sine wave whose algorithm could produce effectively unlimited outputs and which Calderon capped at 10,000.

The two works that made the segment collectible in the art world’s sense both arrived in the first half of 2021. Ringers, by Dmitri Cherniak, is 1,000 outputs of a system that wraps a string around pegs; it minted in January 2021 for under $150 an output and sold out in eighteen minutes, and it produced “The Goose.” Fidenza, by Tyler Hobbs, is 999 outputs of a flow-field system that lays coloured rectangles along invisible currents; it minted in June 2021 at a fixed 0.17 ETH, about $400, and sold out in twenty-eight minutes. Two months later #313 sold for 1,000 ETH, just over $3.3M, to the collector Punk6529 — the same output had changed hands for 0.58 ETH, about $1,400, on June 11 (Decrypt; Hypebeast, August 2021).

The same output went from 0.58 ETH to 1,000 ETH in about ten weeks, a factor of more than 1,700, and it is the clearest single illustration of what a first-mover market with no supply does to a price. Around those two series sits a second tier — Kjetil Golid’s Archetype, Matt DesLauriers’s Meridian, and work by Zancan, Ganbrood and a few dozen others — with gallery representation and thin but genuine auction histories.

The label a work carries is part of its catalogue entry and part of its price, and the labels have moved. Art Blocks organizes releases into collections, of which Curated is the juried one that matters for resale. The two open tiers of the boom, Factory (February 2021 to October 2022) and Playground (March 2021 to October 2022), were retired and folded into Presents in November 2022; Collaborations opened in June 2022 with Pace Gallery and Explorations in November 2022. A 2021 work is often still described by the tier it was issued under; the collection it sits in now is the one on Art Blocks’ own project page, and that is the one a cataloguer and a saleroom will use.

Why so short a list? Because the canon was set by a first-mover market with almost no supply, and then supply became infinite. Anyone can deploy a generative series today for the cost of gas. What cannot be manufactured is a 2021 mint date, a completed critical reception, a sale at a major house and a place in a museum’s accession record. The segment’s scarcity is not in the objects — it is in the history, and history is the one input that does not scale.

No investable index of generative-art prices exists, nobody publishes a repeat-sale series for it — we looked again in September 2026 and found none — and the figures that circulate are floor prices scraped from marketplaces: the lowest ask, not a transaction. Anyone who tells you what this segment has returned is quoting the wrong number, and the reason is structural rather than temporary.

Start with what a floor price is. It is the cheapest listing on a venue at a moment in time: one seller’s ask, on one marketplace, for the least desirable output in the series. It is not a clearing price, it is not an average, and in a series where dispersion is the point, it describes the asset almost nobody wants. Multiply a floor by the series size and you get a “market cap” that could not be realised by anyone, because selling ten Fidenzas into the same book would move the floor before the fifth trade cleared. The flagship guide makes the same point about PFP collections and it is more acute here, since the objects genuinely are not fungible.

Then look at the denominator. Prices in this market are quoted, bid and settled in ether. A collector who bought at 50 ETH and sells at 50 ETH has broken even in the unit the market thinks in and may have lost half their money in the unit their tax return is filed in. On our own tape, ether closed at $2,481.66 on September 8, 2026 against $4,305.71 on September 8, 2025 — −42.4% over the year (Invest Alternative daily series `crypto.eth_usd`, source CoinGecko). Over that same window the series ranged from a high of $4,707.47 on September 13, 2025 to a low of $1,566.01 on June 26, 2026. A flat ETH floor across those twelve months was a 42% dollar loss, and a floor that rose 70% in ETH terms was roughly flat in dollars. Every performance claim in this segment has to state its currency before it means anything.

The third problem is survivorship. The works discussed in this guide are the ones that lasted. The population they were drawn from is very large and mostly dead: a 2023 study by dappGambl examined 73,257 NFT collections and found 69,795 of them — 95% — with a market capitalization of zero (Forbes Australia; The Chainsaw, September 2023). Generative series were not exempt from that; the platforms that issued them simply have longer memories than the traders who left.

What can be said honestly is this. Volume across the whole NFT market was $5.63B in 2025, down 37% from $8.9B, at an average sale price of $96 against $124 the year before (CryptoSlam, via CoinMarketCap and Cointelegraph, January 2026). Art was a small and shrinking slice of that: DappRadar put the art category at $197M for 2024 and $23.8M in Q1 2025, a 93% collapse from the 2021 peak, with active art-NFT traders down 96% from their 2022 high to 19,575 in the quarter.

Against those numbers the S&P 500 returned about +313% with dividends over the ten years to August 31, 2026, and fine art’s long-run real return, on the academic repeat-sale literature summarized in our Investing in Fine Art guide, is about 2.4% a year before costs. Generative art has one full cycle of history, two seven-figure prices at the top of its documented record — both of them outputs of the same algorithm, both out of the same estate — and no published return series. Size the position on that basis.

Every canonical series was issued once, under one of two mechanics, and the mechanic set the spread of cost bases the secondary market has been trading against ever since. It is the part of the market an outsider never sees, because it happened once and cannot happen again.

There were two issuance mechanics. A fixed-price mint set a single price and opened the contract; the series sold out in minutes or hours, and because everyone paid the same amount, every early holder’s cost basis was identical — Fidenza, at 0.17 ETH and twenty-eight minutes, is the canonical case. A Dutch auction started high and stepped the price down on a timer until the series cleared, which raised more for the artist and produced a wide spread of cost bases among the first owners — the collectors who bought at the top of the curve and the ones who waited are in the same series at very different entry prices. Neither mechanic gave anyone a choice of output.

Two secondary features shaped supply. Allowlists (a pre-approved set of addresses permitted to mint before the public) rewarded existing holders and created a class of buyers with a lower basis than the market. And artists and platforms typically retained a reserve — outputs held back at issuance — which sits over the market as an overhang until it is sold or gifted. When you research a series, the reserve is the first thing to find: a thousand-output series where the artist kept 100 has a real float of 900, and the 100 are a supply event waiting for a strong tape.

The economics of the primary sale set the tone for everything downstream. On the Art Blocks model the platform takes a share of the mint and the artist takes the rest, and the artist also receives a continuing royalty on secondary sales, split with the platform where the trade happens elsewhere. The published arrangement is a 10% platform commission on primary sales; on secondary trades routed through outside marketplaces Art Blocks takes about 2.5% and the artist royalty is typically 5%, while trades on Art Blocks’ own marketplace carry no platform fee and honour whatever royalty the artist set. The numbers have changed more than once and are set per project; treat the platform’s own current fee page as the authority and read it before you buy. (The royalty rate used in this guide’s worked example is stated as an assumption where it appears.)

The thing to carry away is that your mint price is public forever. In equities, nobody knows what you paid. Here, every prospective buyer can see your basis, the date, and whether you have been trying to sell. That transparency is why bids in this market cluster: a counterparty who can see that you minted a Fidenza for 0.17 ETH in 2021 is negotiating with information a physical-art dealer would pay for.

Price in this market is set by perhaps a few hundred people, and it is set on two entirely separate mechanisms that beginners routinely conflate: the floor, which is a liquidity price, and the grail, which is an art price.

The floor is set by whoever needs money. In a series of a thousand, at any moment a handful of holders want out, and the cheapest of them sets the number that every chart and every aggregator reports. Floor tokens are the commodity end: they trade quickly, they track ether closely, and they are bought largely by people who want exposure to the series name rather than to a particular picture. If your thesis is “generative art will recover,” a floor token expresses it, and it will behave like a leveraged bet on ether with a royalty deducted.

The grail price is set the way a painting’s is: by two or three collectors who know the artist’s catalogue, want a specific output, and will wait years for it. The vocabulary the market uses to sort outputs is the artist’s own — in Fidenza, the scarce colour palettes and the rare “spiralling” or block-heavy compositions; in Ringers, the peg counts and wrap behaviours that produce a legible figure, of which “The Goose” is the famous case. Trait-rarity tables published by marketplaces are a crude proxy for this and are frequently wrong about which outputs the connoisseurs actually want, because rarity is computable and desirability is not. The gap between the two is where an informed buyer earns their return, and it is the only part of this market where research beats beta.

The two Ringers are what a grail price looks like over a cycle. #109 was bought privately at about $6.9M in October 2021, when everything in the market was rising; #879 was sold publicly at $6.2M in June 2023, out of a liquidation, when almost nothing was. Twenty months, the deepest drawdown the asset class has had, and about 11% between the two prices. Grail buyers in this segment are not marking to the floor, which is the strongest single argument that the top of it behaves like art. It is also a sample of two, which is why nobody can build an index out of it.

Between the two sits the practical problem: there is almost no book. A serious generative series may see a few sales a month. There is no market maker, no continuous two-way price, and no obligation on anyone to bid. Collection-wide bids do exist on the trading venues — a standing offer to buy any token in the series at a stated price — and they are the true liquidation price of a floor token, usually well below the visible floor. Before you buy anything in this segment, look at the best collection bid rather than the floor, because the bid is what you can actually sell into today.

Invest Alternative runs a daily data store, and being straight about what is in it matters more here than in most markets, because the segment we are writing about is not in it.

Our NFT series track four PFP collections and an aggregate volume reading, not generative art. As of the September 8, 2026 read, CryptoPunks floored at $74,352, Bored Apes at $16,935, Pudgy Penguins at $9,189 and Azuki at $1,784; in ether at that day’s price those are 29.96, 6.82, 3.70 and 0.72 ETH respectively (our series `nft.floor_usd_*`, source CoinGecko, arithmetic ours). Our Investing in CryptoPunks guide takes the first of those collections apart trait by trait. What we publish nowhere is a Fidenza, Ringers, Squiggle or Autoglyph series, and we will not quote one from a third-party aggregator as though it were ours.

What our tape does give this guide is the two variables that drive a generative-art holder’s dollar outcome regardless of which series they own: the ether price, covered above, and the depth of the blue-chip NFT market as a whole. Across the seven daily readings we recorded between September 1 and September 8, 2026, 24-hour volume in the blue-chip basket we track ranged from $359,756 to $1,109,461 and averaged $685,896 (our series `nft.bluechip_vol24h_usd`, CoinGecko; the September 6 reading is missing). That is the most liquid corner of the NFT market. A segment whose most heavily traded neighbours turn over well under a million dollars on a normal day is not a market you can exit in size, and the generative names trade a good deal thinner than that.

Our composite index of alternative assets stood at 100.271 on September 8, 2026, provisionally, with a 30-day change of +5.74% and a one-year change of +0.29%; NFTs sit in its cultural sleeve at a weight of 1.7 and the sleeve is marked “awaiting” rather than live. That weight is a statement of our view: this is a satellite holding in a diversified alternatives book, not a sleeve.

What separates this segment from the rest of the NFT market is that the traditional art trade adopted it, partially and with visible ambivalence, and the record of that adoption is public and dated.

The houses went in hard in 2021. Christie’s sold Beeple’s Everydays: The First 5000 Days for $69,346,250 on March 11, 2021, and transacted roughly $150M of NFTs across that year — a figure its own 2022 NFT total of $5.9M, reported as a 96% decline, implies almost exactly. Sotheby’s did roughly $100M, including CryptoPunk #7523 at $11.75M on June 10, 2021 in Natively Digital: A Curated NFT Sale, which made $17.1M across twenty-seven artists (CoinDesk, December 15, 2021; CNBC). Sotheby’s launched its dedicated digital venue, Sotheby’s Metaverse, in October 2021; Christie’s followed with the on-chain Christie’s 3.0 in September 2022.

Then they diverged. Sotheby’s ran the June 2023 “Grails” sales of the Three Arrows Capital estate that produced the Ringers auction record and firm prices for Fidenzas and Autoglyphs in the worst month of the cycle. It has kept going since: its Gen Art Program, run on Art Blocks Engine, commits to two or three generative artists a year in fully on-chain sales and opened on July 26, 2023 with Vera Molnár’s Themes and Variations — the first Dutch auction in the house’s history — and it continued to hold Digital Art Day sales through 2025.

Christie’s went the other way: The Art Newspaper reported on September 9, 2025 that the house had closed the digital-art department it founded in 2022, though the Christie’s 3.0 platform remained live and the house said it would keep selling digital art inside the larger 20th- and 21st-century category. Read the two decisions together and you get the fair summary of the segment’s institutional position — one house has found a repeatable business in it, one has folded the business into a bigger department, and the category’s future at the top end depends on a small number of specialists staying employed.

The museum record is thinner than the marketing suggests and more real than the sceptics allow, and it is worth reading in date order. In December 2022 the Buffalo AKG Art Museum bought tokenised editions of all sixteen works in the Peer to Peer series, the first NFT collection acquired by a major American museum, and did not disclose what it paid (buffaloakg.org). In February 2023 the Centre Pompidou accepted eighteen NFTs by thirteen artists, including CryptoPunk #110 donated by Yuga Labs and Autoglyph #25 donated by Larva Labs, and LACMA accepted twenty-two works by thirteen artists from the collector known as Cozomo de’ Medici in the same month, among them Cherniak’s Ringers #962 and work by Matt DesLauriers (Decrypt; CoinDesk and The Art Newspaper, February 13, 2023).

Then the record changed shape. On March 27, 2025 the Pompidou went further and bought a digital work into the national collection of France — Robert Alice’s 382181_Garden City, a lightbox-and-NFT piece of 2023 — which was reported as news precisely because purchases in this field are rare (nftnow; Right Click Save, 2025). In December 2025 MoMA took eight CryptoPunks and a complete set of eight Chromie Squiggles into its permanent collection, in the Media and Performance department, as gifts from SquiggleDAO and a group of collectors coordinated by the Swiss collection 1OF1 (ARTnews; The Block, December 2025).

Read together, those five accessions say one thing clearly and decline to say another. Four institutions on two continents have now put on-chain work into permanent collections, and a museum of MoMA’s weight taking a complete set of Chromie Squiggles is about as strong an art-historical endorsement as this category is going to receive; the sceptics’ line that no serious institution has touched it stopped being true in 2022.

What the record still does not contain is a price. Three of the five accessions were gifts, and a gift costs a museum nothing but storage and a curator’s judgement, so it validates significance rather than value. The two that were purchases — Buffalo AKG in 2022, the Pompidou in 2025 — disclosed no figure, and neither work was from the Art Blocks canon this guide is about. So the accession record establishes that institutions collect this material and establishes nothing whatever about what a canonical generative work is worth. That is why section 16 keeps a priced acquisition on the watch list rather than treating the museum question as answered.

The practical consequence for a seller is the fee structure, and it is severe. A buyer at Sotheby’s or Christie’s pays 28% on the first $2M of hammer, 22% from $2M to $8M and 15% above, on schedules effective February 13, 2026 and September 1, 2026 respectively; Phillips charges 29% to $2M, 22% to $8M and 15% above on a schedule effective April 12, 2026 (Antiques Trade Gazette; ARTnews; the houses’ rate cards). Sellers pay a commission in the region of 10%, negotiable and often waived on major consignments — in practice at roughly the $1M level and above, not on the small lots most of this segment produces. Our Investing in Fine Art and How Art Auctions Work guides take those mechanics apart in detail; the number that matters here is the total, and it is set out in the venue section below.

Every generative work sits somewhere on a spectrum from “cannot break” to “breaks when a company stops paying a bill,” and where it sits is the single most important technical fact about it. Almost nobody checks before buying.

The token itself is a row in a smart contract: an owner address, an identifier, and a token URI — a pointer to a metadata file that in turn points to an image. Three arrangements exist. Fully on-chain works generate or store the image in the contract itself; Autoglyphs are the canonical example, and no external party can affect them. Content-addressed off-chain works point at an IPFS or Arweave hash, which fingerprints the file so any copy can be verified as the right one, but which still requires that somebody, somewhere, keeps hosting it. Plain-URL works point at an ordinary web address on a server the issuer pays for; when the bill stops, the token points at nothing.

One published sample of roughly 498,000 tokens, assembled by Guy Harrison for the alwaysNFT project, found about one in five whose metadata or media link no longer resolved. Nobody can count the dead links across the whole market, and that is the honest answer to anyone who calls an NFT permanent. Our Investing in Bitcoin Ordinals guide covers the opposite design, in which the image itself is written into the block and there is no pointer to break.

Generative works have an extra dependency that static ones do not, and it is the one that trips up buyers who think they have bought something on-chain. On the Art Blocks model the artist’s script is stored on-chain, which is a genuine and unusual commitment. But the script has to run to produce a picture, and running it requires a JavaScript engine and a drawing library — p5.js, three.js — plus a renderer that assembles the two. Art Blocks has gone further than most here: it maintains an on-chain dependency registry that can hold library releases on-chain and name preferred storage networks, and its newer contracts support fully on-chain storage as well as IPFS and Arweave for external assets.

That narrows the gap without closing it. The chain guarantees the recipe, the seed and, increasingly, the library; it does not guarantee the kitchen. If the platform’s rendering service disappears and the referenced library version becomes hard to obtain, the work is reconstructible by a competent developer but is not displayable by a normal person, which for an art object is a distinction with real consequences.

The remedy is cheap and almost nobody does it. On the day you buy, archive four things and keep them where you keep your other records: the full script as stored on-chain, the token hash, the exact library and version the script depends on, and a rendered master file at the highest resolution the work supports, together with a note of the software used to render it. That package makes the work reconstructible without any third party. It costs an hour.

The “right-click, save” objection — that anyone can copy the image — is correct and beside the point, in the same way that anyone can print a poster of a Rothko. What the token conveys is a publicly verifiable, transferable claim to be the canonical owner of this entry in this contract, with an immutable record of every prior owner. In the physical art market that record is the expensive part; here it is free and the display copy is the cheap part. Whether that inversion is worth paying for is a social question, not a technical one.

The part of this market that most resembles the traditional art trade is the part that produces objects, and it is also where the tax and logistics of ordinary art collecting come back in through the side door.

Several of the canonical artists work in both media. Tyler Hobbs opened two solo shows in the spring of 2023 — Mechanical Hand at Unit London and QQL: Analogs at Pace Gallery in New York, on view from March 30 to April 22 — at which paintings and physical plotter drawings, images drawn on paper by a pen-holding machine following the same kind of algorithm, were shown and sold alongside the tokens.

The pattern that emerged across the segment is a set of variations on the same idea: a token holder is offered the right to have their output produced as a physical work; a gallery exhibits the plotter drawings and treats the tokens as the artist’s primary practice; or a series is minted live at an in-person event, so that attendance becomes part of the provenance. The most complete version of the last idea was Bright Moments, which ran ten in-person minting events between Venice, California in June 2021 and a finale in Venice, Italy in April 2024, issuing 10,000 CryptoCitizens whose city is part of their identity; Christie’s offered the project’s complete works, 216 pieces, on June 27, 2024.

For a buyer, the physical seam changes three things and each one costs money. First, a plotter drawing is tangible personal property, so a sale of it is unambiguously a collectible for federal tax and, unlike the token, it can attract state and local sales or use tax on purchase — in New York City, 8.875% on the invoice including any premium. Second, it needs the things art needs: framing, ultraviolet-filtering glazing, climate, insurance at roughly 1% of value a year, and shipping. A token costs nothing to hold; a drawing costs about 1% a year forever. Third, it introduces authentication risk that the token does not have. The whole point of the token is that its provenance cannot be forged; a signed and numbered print on paper can be, and the market for artist-authorized physical outputs of digital works has no catalogue raisonné, no Pest Control, and in most cases no third party who will confirm anything.

The right way to think about the pairing is that the token is the asset and the physical object is the display, and that the two are separate assets rather than one wrapped in the other. A token here conveys no claim on the drawing and the drawing conveys none on the token, which is precisely what distinguishes this from the redemption structures in our Investing in Tokenized Real-World Assets guide. So buy the drawing because you want it on a wall, price it as a print — our Investing in Prints and Editions guide covers what edition size, signature and condition do to that price — and keep the investment case on the token, where the provenance is free and the carry is zero.

Moving one of these works from one owner to another costs about 6% of the price on a marketplace and 38% of hammer through an auction house — a factor of six, decided entirely by where you consign, and almost nobody calculates it before signing.

On the general marketplaces, the seller pays. OpenSea’s fee moved from 2.5% for years to 0.5% with the launch of its OS2 platform on February 13, 2025 and to 1.0% from September 15, 2025 (OpenSea’s own fee changelog). Blur, the trading-oriented venue that took the professional flow, charges 0% and enforces a 0.5% minimum royalty on collections that enforce one.

Buyers pay gas — the fee the Ethereum network charges to include a transaction in a block — and gas has collapsed. Since the Dencun upgrade in March 2024 the average mainnet transaction fee has run well under a dollar, about $0.30 in late 2025, so a marketplace purchase, which is heavier than a plain transfer, costs cents in normal conditions and a few dollars when the network is busy. At the 2021 peak the same purchase cost $50 to $500; the network friction that once made a small trade pointless has largely gone, and none of the costs that matter in this section is a network cost any more.

On top of the venue fee sits the creator royalty, and its history is the most important commercial fact in the NFT market. Royalties were never part of the token standard; they were a marketplace convention, honoured because the venues chose to honour them. When a venue chose not to, they collapsed, and by the mid-2020s enforcement ranged from full to nil depending on where a trade was routed — industry surveys in 2026 put the share of secondary trades that actually pay a royalty at roughly 60% to 70%, with curated art venues enforcing and the volume marketplaces mostly not. For generative art this cuts both ways: an artist’s continuing royalty is what makes the segment resemble a creative economy rather than a trading pit, and a buyer planning an exit cannot assume the royalty they paid on the way in will be paid by the person who buys from them. Assume the royalty when you sell and do not assume it when you value.

The artist-platform primary market — a new release on Art Blocks or a comparable platform — is where the mint economics in section 4 apply, and it is a different game: you are buying an unknown output at a fixed or descending price with no secondary history to check.

The auction houses are the expensive route and occasionally the right one. Add the buyer’s premium (28% to $2M at the majors on their 2026 schedules) to a seller’s commission around 10%, and the two together come to 38% of the hammer price — which is close to 30% of everything the buyer pays. Do the arithmetic and the conclusion is uncomfortable: for a seller to net the same amount, an auction bidder must pay 33.7% more than a marketplace buyer if a 5% royalty is enforced on the marketplace sale, and 41.5% more if the sale routes through a zero-fee venue at the 0.5% royalty minimum (our arithmetic on the published schedules). That premium is sometimes available, because a catalogue, a specialist and a saleroom reach buyers who will never open a crypto wallet. It is not available for a floor token, ever.

The practical rule that falls out of that arithmetic is to route anything at or below the series floor through the cheapest marketplace and accept the standing bid, because a commodity output will never draw a 34% saleroom premium; and to reserve the auction house for outputs with a name, a documented exhibition history or a trait the specialists ask for, negotiating the seller’s commission down before signing. Waivers are routine on the seven-figure consignments houses compete for; on a lot of a few tens of thousands you should expect to pay the commission, and the arithmetic above assumes you do.

The carrying cost of a generative artwork is close to zero and its transaction cost is close to a stock’s, which makes it structurally cheaper to hold than any physical collectible in this hub — and that advantage is smaller than it looks, because the risks it replaces are not priced in dollars.

Take the round trip first. Buying ether on a regulated exchange costs roughly 0.4% to 1.5% depending on tier and payment method. Withdrawing it costs a dollar or two. Buying the work costs gas — cents in normal 2026 conditions, a few dollars when the network is busy — plus, if you are crossing the spread rather than bidding, the gap between the floor and the best collection bid, which in a thin series is commonly 3% to 8% and occasionally far more. Selling costs the venue fee (0% to 1%), the royalty where enforced (0.5% to 5%), gas again, and the exchange fee to convert back. Total round trip: roughly 5% to 15% at an unchanged price, before tax. Against a physical artwork’s 25% to 40% that is a large and genuine advantage, and against a listed security it is expensive.

Now the carry, which is where the comparison becomes striking. A $20,000 painting costs about 1% a year to insure, plus framing, plus storage or wall space, plus condition risk: over five years that is $1,000 of insurance and several hundred dollars of everything else, and the object can be damaged. A $20,000 token costs nothing a year. There is no storage fee, no insurance premium — because, for a self-custodied retail holder, no insurance is available at all — and no physical deterioration. Custody is a one-off: a hardware wallet from Ledger or Trezor at $80 to $250, bought from the manufacturer and never from a reseller.

That zero carry is real, and it is paid for in a currency that does not show up on a spreadsheet. What replaces insurance and storage risk is key risk and signature risk, and both are total-loss events with no recovery. A physical painting that burns is an insurance claim. A wallet whose seed phrase is compromised, or whose owner signs a malicious transaction, is gone in one block, permanently, with a public record of where it went and no mechanism to reverse it. Phishing losses across the crypto ecosystem ran to about $494M across roughly 332,000 victims in 2024 and fell sharply to $83.85M across 106,106 victims in 2025 — a fall of 83% — an average of $790 each (ScamSniffer annual reports, via Cointelegraph). The fall is real progress. The distribution is not: the losses that matter to a collector are the six-figure ones, and they cluster on people who signed something.

There is one further cost that belongs in this section rather than the tax one, because it is a cash cost of the purchase itself. Buying with ether is a disposal of the ether, so the purchase can generate a tax bill before the artwork has done anything at all. The worked example puts a number on it.

United States, federal, general, and the treatment is worse than most buyers assume in one respect and better in another. Nothing here is advice; the rules below have hard citations and your facts will differ.

The purchase is a taxable event

Buying an artwork with ether disposes of the ether at its dollar value that day. If you acquired ether at $1,600 and spend it when it is worth $2,481.66, you realise $881.66 of gain per unit at the moment of purchase — taxable that year, at long- or short-term rates depending on how long you held the ether, before you own the artwork for a single day. Ether is not itself a collectible, so that gain runs at the ordinary long-term rates, up to 20% plus the 3.8% net investment income tax. Keep the lot record for every unit of ether you buy.

The 28% collectibles rate

IRS Notice 2023-27, issued March 21, 2023, announced an intention to issue guidance and, in the meantime, a “look-through” analysis: an NFT is a collectible under section 408(m) if the underlying right or asset it represents is one — a gem, a coin, a work of art. Pending final guidance, which had not arrived as of this writing, the Service applies that test. A digital image issued, catalogued, exhibited and sold as art is about as squarely inside the “work of art” limb as anything in the market, so a generative artwork is the most likely NFT to draw collectible treatment: long-term gains at a maximum 28% rather than 20%, plus the 3.8% NIIT, for a top federal rate of 31.8%. The same Notice confirms that a collectible cannot be held in an IRA without a deemed distribution. Plan for 28% and be pleasantly surprised.

Reporting arrived in stages

Form 1099-DA, under the 2024 digital-asset broker regulations, requires custodial brokers to report gross proceeds on sales from January 1, 2025, with the first forms issued in early 2026 and transitional penalty relief for good-faith brokers in that first year. Cost-basis reporting follows, but the trigger is the acquisition rather than the sale: it applies only to assets acquired on or after January 1, 2026 and held in the broker’s custody. Anything bought before that date, or bought elsewhere and transferred in, stays uncovered, and the basis on it is yours to prove — which for a 2021 mint means your own records or nothing.

Two carve-outs sit around that. Custodial marketplaces are brokers under the rule, but the instructions carry a de minimis exception for specified NFTs: under the optional method a broker need not report a customer whose aggregate NFT proceeds for the year do not exceed $600 — reporting relief, not a tax exemption. And the extension of the rule to non-custodial front-ends was repealed by Congress under the Congressional Review Act and signed on April 10, 2025, before it took effect, so a trade on a decentralized venue generates no form at all. The obligation to report it is unchanged. The chain is public; the Service does not need the form.

No wash-sale rule, for now

Section 1091 applies to “stock or securities,” and neither crypto nor NFTs are covered as of 2026, so a holder can sell at a loss and rebuy the same position immediately. Successive budget proposals have sought to close this since 2021 and none has passed; at least two were live in Congress as of mid-2026. This is the single most valuable tax feature of the asset class and it should be assumed temporary; verify it for the tax year in question before relying on it.

Donation of a work held more than a year to a qualified charity is deductible at fair market value, with a qualified appraisal required above $5,000 under section 170(f)(11)(C) — and the IRS said in CCA 202302012, released January 13, 2023, that a value reported by an exchange is not a qualified appraisal for a digital asset, and that failing to obtain one is not excused by reasonable cause. The related-use limitation that cuts art deductions back to basis applies to tangible personal property; an NFT is intangible, which is an argument in the donor’s favour that has not been tested. Note the asymmetry with the physical seam: a plotter drawing is tangible and would be caught.

States mostly tax the gain as ordinary income at their top rate — 13.3% in California, 10.9% in New York — with no collectibles distinction. Washington’s capital-gains excise tax reaches Washington-domiciled sellers of long-term capital assets other than real estate: 7% above an annual standard deduction ($278,000 for 2025, indexed), rising to 9.9% on taxable gain above $1M from tax year 2025.

A full round trip on a single work at a realistic ticket, with every fee, every tax and the arithmetic shown, so you can substitute your own inputs.

The entry

You buy one work for $20,000. You pay in ether at $2,481.66, which is 8.059 ETH, and you acquired that ether earlier at $1,600. We allow $10 of purchase gas — generous at 2026 fee levels, and deliberately so — making your basis in the artwork $20,010. But spending appreciated ether is a disposal: you realise $7,105 of long-term gain on the ether, and at 20% plus the 3.8% NIIT that is a cash tax bill of $1,691, payable for the year of purchase. Add a hardware wallet at $150. Your total cash commitment is $21,851, of which only $20,010 is invested in the artwork.

The hold

Five years. Insurance: none available and none paid. Storage: none. Carry: zero. This is the part where the asset class genuinely wins.

The exit

You sell for $34,000, a 70% gain on the ticket. On a marketplace at a 1% fee with a 5% creator royalty enforced (royalty assumed; enforcement is venue-dependent, per section 10), fees are $2,040 plus $10 of gas, so net proceeds are $31,950. Your gain is $31,950 − $20,010 = $11,940. At the collectible rate of 28% plus the 3.8% NIIT, tax is $3,797, leaving $28,153.

The result

Against the $20,010 invested in the artwork, that is a compound return of 7.07% a year. Against the $21,851 of cash the purchase actually required, it is 5.20% a year. Of the $13,990 of gross gain, $2,050 went to fees and $3,797 to tax — 41.8% — and adding the $1,691 of ether tax paid at entry takes the total leakage to $7,538, or 53.9% of the gross gain. The work had to appreciate 70% over five years to produce a return in the mid-single digits.

The same trade at auction

Suppose instead you consign, and a bidder pays the same $34,000 all-in. At a 28% buyer’s premium the hammer is $26,563; a 10% seller’s commission takes $2,656, and you receive $23,906 — $8,044 less than the marketplace route on identical buyer money. For the auction route to leave you level, the saleroom bidder has to pay 33.7% more than the marketplace buyer.

The downside

Sell at $12,000 instead. Net proceeds after the same 6% of fees are $11,270, a capital loss of $8,740 against your $20,010 basis — usable against capital gains, but only $3,000 a year against ordinary income. Measured against your $21,851 of committed cash you are down 48%, and you paid $1,691 of tax on the way in for the privilege.

The currency trap

You paid 8.059 ETH. If ether is $5,000 at your exit, your $34,000 sale is 6.80 ETH — you sold for fewer units than you bought, and the entire dollar gain came from the currency, not the artwork. If ether is $1,200, the same $34,000 is 28.33 ETH, a 3.5× in the market’s own unit. Two identical dollar outcomes, two opposite verdicts on the investment.

The risks that end a position in this market are not the ones that end a position in equities, and they are ranked here by how completely they destroy capital rather than by how often they occur.

Key and signature loss is first, because it is total and irreversible. There is no custodian to call, no chargeback and no insurer for a self-custodied individual. The loss modes are a compromised seed phrase, a malicious transaction signed by the owner, and a “wallet drainer” contract that obtains a blanket approval. The defence is structural rather than vigilant: a hardware wallet bought from the manufacturer, a seed phrase generated on the device and written on paper or steel, a browser wallet that simulates transactions before signing, and — the part almost everyone skips — two addresses on the same device, one that holds and signs nothing but transfers, one that interacts with the world. A collector who lost more than $1M of Squiggles, Autoglyphs and Punks to a single signature is a documented case in our flagship, and the pattern repeats every cycle.

Liquidity is second. A handful of sales a month sets the price for an entire series, and the standing collection-wide bid is usually well below the visible floor. In a stressed tape, the bid falls faster than the ask and the spread widens until nothing trades. Assume you cannot sell for a year at anything like the quoted number.

Platform and rendering dependency is third, and it is the risk unique to this segment. If the entity that hosts a series’ rendering service and metadata stops operating, tokens whose scripts are on-chain remain reconstructible and tokens whose images sit behind a plain URL do not. Section 8 sets out the four-part archive that neutralizes most of this for a few hours of work.

Wash trading corrupts the price history you are underwriting. Chainalysis identified 262 sellers who had each sold NFTs more than 25 times to self-funded addresses; 110 of them made $8.9M and 152 lost about $417,000 (2021 data, Chainalysis Crypto Crime Report). The analyst hildobby’s filter classified 58% of 2022 Ethereum NFT volume as wash trading (CoinDesk, December 23, 2022). Curated generative series were less affected than incentive-farmed collections, because there was no token airdrop to farm — but “less affected” is not “clean,” and a sale history with repeated round trips between two addresses is a fabrication, not a comp.

Artist and reputation risk is the ordinary art-market risk, and it applies here with less protection than usual. A living artist can produce work that dilutes the earlier series, can leave the field, or can behave in a way that makes their work unsellable. In the physical market an estate, a foundation or a catalogue raisonné committee manages that over decades; this segment has none of that apparatus and no way to build it for a living artist, so the recourse when a reputation turns is the same as the recourse when a series stops trading — sell into whatever bid exists, or hold.

Regulatory risk is real and has receded. The SEC brought settled actions against NFT issuers in 2023 — Impact Theory on August 28, 2023 over roughly $30M raised, and Stoner Cats on September 13, 2023 over $8.2M — with Commissioners Peirce and Uyeda dissenting in both. The Commission served OpenSea with a Wells notice in August 2024 and closed that investigation on February 21, 2025. The posture since has been permissive. A genuinely curated, capped art series sold to collectors is a poor fit for the Howey test — the four-part test that makes an arrangement an investment contract, and so a security — in any event, but an issuer who promises a roadmap, a treasury and returns is selling something else.

If, having read the cost stack and the risk stack, you still want a position, here is the sequence. Do it in this order; the order is the point.

1. Decide the currency question first. Write down whether you are trying to own an artwork or trying to own ether. If it is ether, stop here and buy ether.
2. Set the size before you look at anything. This is a satellite position in an alternatives sleeve — ours carries NFTs at a 1.7 weight in a composite index and marks the sleeve as awaiting rather than live. Choose a number you can write off entirely without changing any other decision.
3. Build the custody first, with no assets in it. Hardware wallet from the manufacturer, seed generated on the device and recorded on paper or steel, never photographed and never typed. Create two addresses: a holding address that signs only transfers, and an interacting address that touches marketplaces. Test both with $50 before either holds anything.
4. Learn one artist properly before buying anything. Read the artist’s own writing on the algorithm, look at all thousand outputs — it takes an evening and it is the single highest-return hour in this market — and form your own view of which outputs are good, in writing, before you look at a price.
5. Check the plumbing. For the specific series: is the script on-chain, is the metadata content-addressed, who renders it, and what happens if that party stops. Section 8’s four-part archive is the deliverable.
6. Read the actual sale history of the actual token. Every prior transfer is public. Look for round trips between related addresses, and for a seller whose basis you can see.
7. Price the exit before the entry. Find the best standing collection-wide bid, subtract the venue fee and the royalty, and treat that as today’s realisable value.
8. Buy on the cheapest venue that has the token, and move it the same day to the holding address.
9. Record the tax lots on the day of purchase: the ether lots you spent, their acquisition dates and dollar cost, the gain realised on the purchase, and the artwork’s basis including gas.
10. Write the key-succession plan. The token dies with the key. A sealed letter naming the device’s location and the recovery procedure, held by an executor, is not optional and is the step collectors skip most often.

Specific, checkable readings that would change our view, each with a threshold rather than a direction, and each stated as of the date of this guide.

The ether price, first and always

As of September 8, 2026 our tape has ether at $2,481.66, down 42.4% over twelve months. A generative position is roughly one-for-one exposed to this. A move that takes ether decisively above its September 2025 level near $4,300 or below its June 2026 low of $1,566 changes the dollar arithmetic of every holding in the segment without a single artwork changing hands.

Depth in the neighbouring blue chips

Our blue-chip 24-hour volume series averaged $685,896 across the seven readings from September 1 to 8, 2026. Sustained readings above roughly $3M a day would be the first evidence in years of a market that could absorb size; readings that settle below $250,000 would mean the segment has no exit at any price.

A museum purchase at a published price

Section 7 sets out the accession record: five institutional acquisitions since December 2022, of which two were purchases and neither disclosed a figure. The threshold is narrower than “a museum bought one.” It is a canonical generative work — an Art Blocks Curated series, an Autoglyph — bought from an acquisitions budget with the price stated in the museum’s own annual report or announcement. That one datapoint would give a valuer the first institutional comparable this category has ever had, and it would do more for prices at the top of the segment than another year of gifts.

Whether a second auction house rebuilds a digital department

Christie’s closed its digital-art department in September 2025 and folded digital sales into its 20th- and 21st-century category (The Art Newspaper). Sotheby’s continues to hold dedicated sales and runs its Gen Art Program at two or three artists a year. A second major house committing specialists and catalogue space to generative art would materially widen the buyer base; a withdrawal by Sotheby’s would remove the segment’s only reliable route to non-crypto buyers.

Royalty enforcement

Watch whether the venues carrying the majority of generative-art flow enforce creator royalties; industry counts in 2026 put the share of secondary trades that pay one at roughly 60% to 70%. Enforcement above roughly 4% sustains the artist economy that makes this segment resemble art; a move to zero across the board turns it into a trading market with a curator’s label on it.

Final IRS guidance on Notice 2023-27

The look-through analysis has been provisional since March 21, 2023. Final guidance confirming collectible treatment would lock in the 28% rate and settle the IRA question; guidance going the other way would cut the top federal rate on a long-term gain from 31.8% to 23.8%, which is worth 8 percentage points of every gain in the segment.

A wash-sale rule for digital assets

Section 1091 does not reach crypto or NFTs as of 2026. Enactment would remove the loss-harvesting flexibility that partially compensates for the segment’s volatility, and it has been proposed in successive budgets.