Marble's $11M Bet: Compute Credit Becomes the AI Bottleneck

Two operators order the same rack of NVIDIA accelerators. One closes a syndicated loan against most of the invoice. The other has signed offtake and still cannot get past credit committee. The hardware is identical; the contracts are not.

Leon Marshall, co-CEO of Marble, describes that gap as a pricing problem. A lender wants to know who is buying the compute, at what price and for how long. Most mid-market GPU operators cannot answer with a document a credit committee will accept.

Marble launched on November 10, 2026 with $11 million in Series Seed funding led by Polychain Capital. Maven 11, Ripple, Susquehanna Crypto, Lattice and Blockwall joined, alongside AI-focused angels. The round funds two planned businesses: Marble Compute for private credit, Marble Markets for trading.

Neither business is live. Marble Group's disclosures say both lines are in formation, hold no SEC, CFTC, FinCEN or New York DFS registration, and offer no services today. First syndications are targeted for early Q1 2027.

A credit gap, not a chip gap

The largest neoclouds finance growth through public markets and large banks. Operators below them often cannot, even when customers have committed to buy capacity. Marble is built for that middle layer.

Its model avoids balance-sheet lending. Marble sources deals, runs diligence, structures the paper and syndicates it to institutional and digital-asset lenders. The company earns fees for underwriting and structuring, not spread on its own capital.

That distinction sets the size of the bet. A lender with $11 million can fund a handful of clusters. A platform that syndicates can, in theory, fund many without holding the duration.

How Marble Compute is built

Three stages, each aimed at a different lender. The first is direct loan syndication between operators and institutions. The second adds single-opportunity SPVs (special-purpose vehicles), off chain and on chain, each tied to one GPU operator. The third pools those credits into multi-opportunity funds once deal flow and track record exist.

The sequencing matters more than the round size. A seed cheque does not buy GPU fleets. It buys underwriting capacity, legal structuring and the trading rails that Marble Markets needs before any lender commits size.

Underwriting leans on three inputs: the hardware, the creditworthiness of contracted customers, and the operator's ability to service debt. Marble says it works with operators to diversify offtake so lenders are backed by contracted cash flows rather than projections.

The contract is the collateral

A GPU depreciates. Its resale market can move against a lender within one product cycle. A take-or-pay contract is a claim on a customer, and that claim is what Marble wants lenders to underwrite.

Two GPU operators can buy identical hardware and raise very different amounts of debt against it. A lender wants to know who's buying the compute, at what price and for how long.— Leon Marshall, co-founder and co-CEO, Marble

Compute forwards are the hinge. An operator locks a rental price today for capacity delivered later. The lender sees a fixed revenue line instead of a spot-market guess. The operator sees how much equity the next cluster requires.

That structure is familiar in project finance. It is new for GPUs, where contracts are shorter, hardware generations arrive faster and utilisation is the variable nobody can underwrite with confidence.

What the market already prices

Marble is not early to the idea. It is early to the mid-market.

Ornn, which routes GPU capacity from neoclouds to short-term buyers, is raising a three-tranche round at a reported valuation near $1 billion, with Thrive Capital and General Catalyst in talks. Silicon Data raised a $30.5 million Series A in August for an hourly GPU rental index built from about 100 sources. Liquid Compute raised $15 million in September and has applied to run a US exchange for compute futures and options.

The debt side has its own plumbing. USD.AI closed a $106 million senior secured facility for NexGen Cloud in November, financing 1,728 NVIDIA GPUs at a Finnish data centre under a take-or-pay agreement. IREN refinanced $363 million of deployed GPU and networking equipment in October. CCIR's public tracker now lists dozens of GPU-collateralised deals.

Sources: Marble, USD.AI, Silicon Data, CME Group, November 2026.

NVIDIA has spent the year signing financing partnerships that treat AI factories as an investable asset class. The direction is consistent: compute is being turned into a yield instrument, one contract at a time.

Two questions remain open across the category. Regulators have not approved compute futures, and no public default has tested how fast GPU collateral is marked down.

The bear case is hardware

The simplest objection is that GPU credit is a bet on a depreciation curve. A lender secured on H100s in 2024 holds a different asset in 2026. Blackwell capacity commands a premium today. Nobody knows the clearing price after the next generation ships.

Marble's own disclosure names the same risks. Hardware can lose value faster than expected. Offtake contracts can be terminated or unpaid. Re-leasing capacity is not guaranteed.

The second objection is regulatory. Marble Markets needs registrations before it can trade, and digital-asset trading is planned for January. Compute futures face the same bottleneck: CME and Silicon Data targeted October 5 for NYMEX contracts, the CFTC extended its review through November 9, and the effective date is now to be announced.

Software can ship in a quarter; credit infrastructure cannot. The $11 million buys time to prove the underwriting, not permission to skip the licences.

As we wrote in October, Hyperbolic raised $30 million on the premise that GPU capacity becomes a traded market. Marble's seed is the credit leg of the same thesis: first a price, then a hedge, then a loan against the hedge.

The operators in the middle

Marble's target is narrower than "AI infrastructure". It is the operator with signed customers and no investment-grade balance sheet. Those firms sit between hyperscaler credit and venture equity, and they pay for the difference.

For lenders, the question is not whether AI demand exists. It is whether one operator can deliver contracted capacity through a hardware cycle without breaching a covenant. Marble sells diligence on that question, plus a forward curve to hedge the answer.

Whether institutions pay for that diligence at scale is the open test. The first syndications in Q1 will show whether the paper prices at a spread that clears.

Where the capital comes from

Marble's supply side is not only traditional credit. The pitch to digital-asset holders is a yield stream tied to real hardware and contracted compute, not to token emissions. Funds, treasuries and long-term BTC, ETH and SOL holders sit on balance sheets that want durable, real-economy yield, and Marble Compute is designed to route that capital into operator loans.

The structure widens the lender base. It also adds compliance surface. On-chain vaults and single-opportunity SPVs each need their own documentation, custody arrangement and investor eligibility checks, and the company says the products will be offered only to institutional counterparties in permitted jurisdictions.

That is a slower build than a pure software launch. It is also the part that decides whether Marble becomes a broker of one or two credits, or a standard venue where mid-market GPU paper gets priced.

Does compute credit survive its first default?

Three paths from here