TTornaCurb

Pre-IPO stock tokens, on an order book

There is no exchange for OpenAI stock.

An automated market maker quotes against a price formed somewhere else. For these companies there is nowhere else. TornaCurb is where the price gets made — a real order book for shares that never had a venue.

Never used an order book? Three minutes, no wallet needed →

Live on devnet · test tokens, not shares · real SPL escrow, no indexer · external audit pending · notice

Before a company was listed on the New York Stock Exchange, its shares traded on the curb — literally on the sidewalk outside the exchange. The Curb Market existed for one reason: there were shares people wanted to trade, and no exchange willing to list them. That gap is open again.

Why a book, not a pool

A curve has to guess. A book already knows.

An automated market maker quotes off a formula anchored to a reference price. For a private company there is no reference price to anchor to, and the float is thin, so the curve is steep exactly where it matters. A book quotes where makers are actually willing to trade.

TornaCurb order book
for 50 shares
1,021
average fill price
slippage vs best ask0.00%
depth reached50 / 50
Constant-product pool (modelled)
for 50 shares
—
average fill price
slippage vs spotunfillable
pool depth48 shares

Modelled, not measured: no such pool exists for a pre-IPO name. Half the stated TVL in each leg, priced at the book's best ask.

The gap is not a trick of the numbers — it is the instrument. A pool must hold inventory across every price at once, so thin liquidity means a steep curve. A maker on a book commits capital only at the price they chose, which is why order books work in markets that are too thin to support a pool at all.

The venue

Ten listings, two kinds

The split is the argument. For most of these private companies nothing publishes a price at all; Pyth does index OpenAI and Anthropic, but an index reports discovery happening on secondary venues rather than performing it, and it gives you no limit order and no price-time priority. The listed pair is a control group: where a price is formed on a real exchange, you can watch how closely the book tracks it — and only then is it reasonable to trust the same machinery where no exchange exists at all.

Pre-IPO · no exchange anywhere
OPENAIOpenAIANTHROAnthropicANDURLAndurilNEURANeuralinkKALSHIKalshiPOLYMKTPolymarketSPACEXSpaceXFIGUREFigure AI
Listed · a real exchange price, used to check the book
NVDANVIDIASPYS&P 500 ETF
OPENAI, live on devnetlive, real escrow
Best bid
1,020
Best ask
1,021
Spread
1
Resting orders
12
Open the trading terminal

Real limit orders

“Buy at 180 or better” is the most basic instruction in equities. A curve cannot express it; a book is made of them.

Price-time priority

The maker who quotes first and tightest gets filled first. That is what makes anyone quote tight.

Makers keep their edge

A pool cannot move its quote when the world reprices, so it gets picked off. A maker cancels and re-quotes.

None of this works if the book cannot take concurrent quotes. TornaCurb runs on Torna, our own open-source index, which is what makes that possible on Solana.

The moat

Why the writes run in parallel

The usual on-chain book keeps each side in one slab account, so every order on a side takes the same write lock and serializes, one per slot. Torna puts each B+ tree node in its own account, so orders that fall in different leaves touch different accounts and Solana commits them in the same slot. A deep book spreads its price levels across many leaves.

The usual bookone slab account per side$100$180$260one slabone account, one lockslot 1slot 2slot 33 orders, 3 slotsone shared account serializes themTornaprices that fall in different leaves$100$180$260leafaccount Aleafaccount Bleafaccount Cslot 13 orders, 1 slotdisjoint accounts run in parallel
Three makers whose prices fall in different leaves. Left, the side's one slab serializes them across three slots; right, Torna gives each leaf its own account, so they commit in one slot. Adjacent prices in a sparse book share a leaf and still serialize, and matching itself is always serial.
4.6-7.1xmore committed tx/slot when maker writes
land in different leaves vs. one shared leaf

Measured on a single-node solana-test-validator, the real Agave banking stage, via torna/bench (./run.sh reproduces it; method and data). Devnet is shared and noisy, so the controlled number is the honest one. This parallelizes book maintenance, not matching: top-of-book is price-time serial by definition, and nothing can change that.

Look under it yourself

TornaCurb on GitHub Torna on GitHubTornaCurb is built on Torna, our own open-source primitive.