Reasoning capital markets: How crypto builds AI computing power financial infrastructure

Author: Lucas Tcheyan; Source: Galaxy Digital; Translated by: Shaw, Golden Finance

Introduction

On-chain inference capital markets** refers to a range of networks, protocols, supporting infrastructure, and applications used to coordinate AI model inference businesses that sit outside frontier labs and the centralized API ecosystems of hyperscale cloud providers, as well as the financial layer that is being built on top of this business and starting to take shape. Users don’t need to route every API request through model providers such as OpenAI and Anthropic—or through the cloud vendors behind them. Instead, they can submit prompts to a network of GPU operators scheduled by crypto token incentives and on-chain settlement mechanisms. Under some architectures, users can also obtain cryptographic assurances or economic guarantees for the accuracy of outputs and data privacy.

In 2026, this sector has attracted growing attention. The global GPU demand share for inference—i.e., the computation process of generating results from already trained AI models using new data—has already surpassed model training and has become the largest compute-supply demand scenario. At the same time, autonomous agents have become a new class of inference demand: software that can complete payments automatically, with no human intervention throughout the process.

Over the past few years, decentralized GPU trading markets, inference protocols, payment channels, asset tokenization tools, capital-raising vehicles, and on-chain liquidity segments have each experienced their own development windows. The new change now is: various foundational components are converging into a unified, comprehensive system—an inference capital market. As AI inference is widely deployed across business scenarios, the market demand for this system is expected to keep rising. Current on-chain practices revolve around business activities that have real productive capacity and tangible economic value, and related demand no longer comes solely from within the crypto industry.

Multiple forces are driving this ecosystem convergence. GPU compute demand has clearly shifted from model training to inference tasks; and in business scenarios that only need to meet a “good enough” standard, open-weight models are narrowing the performance gap versus frontier closed-weight models. This enables enterprises to shift once-costly tasks to the most cost-effective compute supply side (regardless of whether they rely on crypto-native underlying channels or traditional channels), making it economically feasible.

Inference compute demand is steadily growing, which also forces users to procure compute in more innovative ways. A research report released recently by Citadel states that the AI Token consumption scale tracked by the Silicon Data LLM Index has declined, reflecting the market’s shift toward lower-cost model options. (Note: AI Tokens are billing units used by AI service providers for pricing their services. Do not confuse them with crypto tokens issued on blockchains.)

A number of companies, including Coinbase, Microsoft, and Airbnb, have also recently begun shifting to open-source models, with many being China-based open-source models. OpenRouter also completed a new round of financing recently, further confirming the market’s increasing demand for diversified model access channels—channels that can effectively reduce inference costs. Part of the reason for this trend is constrained compute supply, which has led to continuously rising marginal deployment costs for inference services.

The second force driving ecosystem convergence is financialization. As AI adoption continues to rise, intelligence capabilities have almost become a production input for all businesses, creating demand for compute to be turned into a commodity and then financialized. More and more teams are exploring solutions to turn AI compute into tradable assets and incorporate it into a more complete financial system. Early frameworks for inference capital markets are taking shape: the market is beginning to apply financialized pricing to AI hardware and compute capacity, aiming to build a complete integrated trading market.

GPU Indices and Futures Markets

Before discussing the on-chain deployment forms of inference capital markets, we should first focus on the much larger off-chain markets, the most representative of which is GPU futures. Predictions about the scale of AI infrastructure expansion vary widely among different parties. Morgan Stanley forecasts that, by 2028, global data center capital expenditures will be about $2.9 trillion (excluding power investment), with about $2.5 trillion directly used for AI-related workload. McKinsey estimates that by 2030 global data center capital expenditures will reach $6.7 trillion; of this, $5.2 trillion will go to AI compute infrastructure and $1.5 trillion to traditional IT. Its AI scenario ranges into two cases: a weaker-demand scenario of $3.7 trillion, and a demand-accelerated release scenario of up to $7.9 trillion. Goldman Sachs estimates that between 2026 and 2031, global capital expenditures on AI infrastructure across compute, data centers, and power will be about $7.6 trillion. No matter what the final precise numbers are, multiple institutions’ forecasts converge on the same conclusion: compute and hardware are the largest-spending segments. Morgan Stanley, McKinsey, and Goldman Sachs’ modeled ranges all show that this segment accounts for roughly 55%~67%.

These projections are difficult to translate into precise reality because there are many unknown variables on both the supply and demand sides. The first is demand elasticity: if the money saved from a drop in compute costs isn’t converted into retained enterprise profits, but instead continues to be reinvested to train larger-scale models and expand deployment scenarios, then efficiency gains only expand compute usage, rather than shrinking total bills. The second variable is chip useful life. The industry currently has no unified conclusion; depreciation estimates commonly fall in a range of 3 to 7 years. Although newer, stronger chips are released every year and in theory would accelerate the obsolescence of older chips, in reality older hardware still retains value. Because supply remains tight, enterprises have to rely on existing inventory hardware to meet compute demand, while older chips can also handle inference tasks with lower barriers. The final result is: a massive and continuous inflow of capital into assets with highly volatile prices; and that is precisely the typical environment that gives rise to pricing, hedging, and financing-support markets.

Today’s compute procurement ecosystem is often compared to a black-market trading venue—if you want to get a source of supply, you can only contact a certain “middleman” directly.

In some sense, forward trading markets already exist, but they haven’t yet taken a standardized form. Large buyers have long locked in forward compute supply through private agreements. Forms include hourly on-demand rentals, multi-year compute pre-purchase contracts (equivalent to GPU take-or-pay agreements), and bilateral deals between service providers and top customers. Prices are often set through relationship-based negotiations, with low transparency. OpenAI and other frontier model companies bulk-sell API call quotas; major hyperscale cloud providers reserve compute capacity among themselves. AI-dedicated cloud providers (New Cloud / Neoclouds) pre-lock compute with public cloud providers and intermediaries, rooted in the overall supply-demand imbalance. Baseten, one of the world’s top inference service providers, bluntly stated that purchasing compute today is like black-market trading and requires specialized outreach to middlemen to obtain sources of supply. Intermediaries who profit from information asymmetry and relationships, as well as large compute holders, have little incentive to move into open and transparent order-book trading, even if a standardized market could slightly improve trading efficiency. Historically, similar frictions have also blocked the development of commodity markets. Energy giants such as Vitol, BP, and Shell resisted, causing attempts to launch liquefied natural gas exchanges to take a decade without successful deployment. GPU futures have gradually emerged atop such fragmented spot markets; their role is a standardized tool to transfer price risk and, in the short term, will not replace existing compute procurement models.

For futures markets to function effectively, they need accurate price indices that can serve as contract anchors. Compared with standardized commodities, building compute indices is more difficult. A simple “GPU hour” lacks a unified standard: the chip model, VRAM, network specifications, region, and whether the compute is on-demand spot capacity or pre-purchased locked-in resources must be defined. Before power, bandwidth, and liquefied natural gas become highly liquid markets, they also faced the difficulty of non-uniform underlying contract specifications. The industry ultimately adopted the same path: split into grades and set benchmark prices, rather than forcing every unit of a commodity to be perfectly homogeneous. Oil pricing relies on benchmark grades such as WTI and Brent from the New York Mercantile Exchange. Natural gas uses Henry Hub prices as the pricing anchor.

A GPU market is gradually forming a similar pricing framework. Galaxy portfolio company Ornn launched a compute price index, compiled based on real transaction data. Silicon Data publishes H100, A100, and B200 rental indices daily on the Bloomberg terminal. It standardizes quote benchmarks across different hardware configurations, service providers, and regions, integrating them into a single benchmark. Compute Desk is also moving in the same direction. According to Ornn’s positioning, these indices are closer to Secured Overnight Financing Rate (SOFR) than to the much-criticized old London Interbank Offered Rate (LIBOR). The index data comes from massive real market transactions rather than subjective estimates by experts. It does not track a single-model GPU; instead, it aggregates market-wide prices for a set of standardized compute resources. Only with real trading data can the heterogeneous hardware specifications be handled in a controllable way. The index does not require every GPU hour to be identical; as long as there are enough real transaction samples, it can compute a representative price.

But compute indices still face challenges not present in oil markets. For decades, West Texas Intermediate (WTI) crude oil has reliably served as a pricing anchor because standardized crude oil barrel specifications have remained unchanged. However, GPU benchmark underlying assets will keep iterating. Compute clusters will transition from H100 to H200, B200, GB200, and Rubin chips. Each hardware generation update requires revising the pricing benchmark. Hardware diversification further complicates the issue: AMD chips, Google TPU, Amazon Trainium, large cloud providers’ in-house ASICs, and demand for country-level chips are all incompatible with one another. Building a long-term stable and effective pricing benchmark becomes harder, not easier.

The second core controversy lies in delivery rules. For AI labs seeking to hedge inference compute budget volatility, or trading teams expressing price views, they only need pure price risk exposure and do not need physical hardware. For these entities, contracts with difference settlement based on indices are sufficient. But for vertical cloud providers serving end customers, they need real compute resources. The GPU futures being rolled out currently use a cash settlement model because price-hedging demand is easiest to standardize. Even for many commodity futures that support physical delivery, cash settlement is typically the norm behind the same logic. Physical delivery is feasible but has higher implementation difficulty, requiring further standardization and detailed contract terms. Given the pace of cash-settled futures adoption and new market demand, it is not surprising to see physical-delivery GPU contracts in the next year. Some also argue the current development order is backwards: when only a few suppliers hold the supply source, cash settlement based on thin sample indices may easily breed market manipulation. Commodity markets typically need physical delivery mechanisms or mature exchange-for-physical (EFP) systems to make futures prices track spot fundamentals.

A mature market also requires real trading intent from both supply and demand sides; it cannot be only a game between speculators. Natural longs: enterprises whose costs are deeply tied to compute depth and who want to lock procurement costs—including AI labs, application developers, downstream vertical cloud providers that have sold off compute and must secure supply. Natural shorts: institutions holding GPU inventory but unclear about future use—such as large cloud providers, large GPU holders, and compute intermediaries. Lending institutions providing credit capital for GPU procurement also need a unified benchmark price. Loans backed by depreciating hardware must have asset valuations aligned with fair market prices. Speculative capital and proprietary trading teams act as common participants across all markets, providing liquidity, but they are not industry trading counterparties. Today’s market’s key structural contradiction is that sellers mostly prefer signing long-term contracts to lock in revenue, while buyers prefer short-term contracts to retain flexibility for adjustments.

Despite the challenges, the beginnings of a mature GPU trading market are already visible. Prediction market platforms like Kalshi have launched multiple GPU price trading products. The parent company of the New York Stock Exchange, Intercontinental Exchange (ICE, working with Ornn), and the Chicago Mercantile Exchange (CME, working with Silicon Data) have both announced plans to launch GPU futures next year. Compute as a commodity may become real soon.

On-Chain Inference Capital Markets

Inference service providers are, in essence, like token-processing factories: they input raw material GPU compute power, process it, and generate AI Token outputs. GPU hour resources increasingly become standardized through various compute indices, but the development level of the top-layer AI Token trading market is much lower. One major difficulty is that Tokens produced by different large models do not have a unified basis for comparison. Even so, the ecosystem in this sector is gradually taking shape. China’s three major telecom operators have retail-inferred compute as a metered public utility, launching standardized monthly Token bundles—similar to mobile data plans. Reports say Amazon will adjust its settlement with Anthropic: instead of locking compute time, it will charge based on the actual amount of Tokens consumed. Other reports suggest the Shanghai Futures Exchange is in the early phase of designing a plan to launch AI Token-linked futures products, corresponding to the GPU futures upstream positioning by CME and ICE in the compute supply chain.

The crypto industry is building its own analogous ecosystem. An on-chain inference capital market relies on existing crypto + AI underlying components (GPU compute service providers, decentralized model-development networks), while also incorporating emerging tracks, including agent payment standards and tokenized inference trading markets. Related ecosystems have already deployed multiple blockchains and execution environments, but development resources are highly concentrated on Base and Solana, supported by the well-developed developer and user communities on these two chains.

The core of the entire system is the inference service providers and inference networks—i.e., all kinds of projects that convert prompts into outputs. Around the core layer, multiple levels of infrastructure are built so that inference services are usable, practical, and have financializable characteristics: model developers, GPU and compute suppliers, routing and trading platforms, agents and top-layer applications, payment channels, and capital-raising infrastructure. Those surrounding layers are crucial because they either create inference demand and produce compute supply inputs, or they transform compute consumption behavior into assets that are payable, financeable, schedulable, and claimable.

Many of the services above are not unique to the crypto industry; mature counterpart products exist in the off-chain world. In the technology stack layer, agent scheduling frameworks like Hermes and Ironclaw can call frontier lab models indiscriminately, or access compute provided by on-chain inference service providers such as Venice. Decentralized training developer models like Nous Research can connect to OpenRouter’s one-stop model aggregation platform. GPU compute suppliers correspond to unpermissioned open-source versions of hyperscale cloud providers and data centers, but generally with smaller scale. Agent payment protocols like x402 and MPP can pay OpenAI and Anthropic subscription bills, and also conveniently settle fees on the Venice platform. Programmable automated settlements are quickly becoming an industry-wide common capability and no longer just an edge unique to crypto. OpenAI and Visa have also recently launched payment infrastructure for agents.

The truly differentiated innovation is concentrated at the financialization layer: crypto technology is reshaping how inference compute is claimed (ownership), priced, and financed. In our previous reports, we discussed in depth that crypto underlying channels can accelerate capital formation, and many underlying components can also be applied directly to the inference sector.

Inference service financialization has attracted many on-chain projects. Using blockchain payment channels and tokenization tools, inference businesses are transformed into tradable assets. There are mainly three forms:

• Inference service providers (Venice, Morpheus, etc.): tokenize inference access rights, turning the right to claim future inference services into assets that can be held, priced, and resold.

• Proof of Useful Work projects (Pearl, Ambient, etc.): tokenize inference production behavior; service providers receive token rewards as long as they complete inference tasks.

• Credit service providers (USD.AI): the model differs. They do not directly tokenize inference services; instead, they provide financing for the hardware that carries inference workloads, using stablecoin deposits to fund underlying GPUs and data centers.

All of the above components integrated together jointly form the on-chain inference capital market.

Inference Service Provider Layer

Inference service providers are the core of the entire stack. The decentralized inference track is highly similar to the traditional AI API market here: users or developers choose a model, submit prompts, pay according to Tokens / request counts / subscription modes, and then receive output results. In the simplest form, the user experience is close to OpenRouter, Together AI, Fireworks, or official APIs from frontier AI labs. The difference is that native crypto inference providers can procure compute from decentralized GPU networks, accept settlement in stablecoins or native tokens, provide access to open-source / uncensored models, add privacy protections, or bind tokenized access rights with compute usage behavior.

OpenRouter is an on-chain inference track with great potential as a real-world deployment scenario. In the platform, demand is priced in AI Tokens. Users can freely switch inference service providers with each individual request. In such a market environment, service providers with better cost performance and faster responsiveness should capture more share. Over the past three months, the total Token volume carried by on-chain inference service providers was in the range of 0.5%~1% of OpenRouter’s daily processing volume, while during the same period OpenRouter’s overall Token handling scale has kept growing explosively. This indicates that on-chain service providers have already gained initial market recognition beyond native crypto communities, but their overall share remains low—reflecting that they are still constrained by factors such as channel coverage, relative costs, and others, and at this stage they cannot compete on the same stage as mature centralized services.

However, OpenRouter represents only part of the Token flow. Take Venice as an example: the platform disclosed that on June 23 it handled 100 billion Tokens across all channels cumulatively—ten times its volume processed in the OpenRouter channel. If you only count OpenRouter’s data, you cannot fully reflect the true development momentum of each independent project. Major on-chain inference service providers are continuously accumulating stable user groups through various operational methods, with some tactics relying on differentiated product features: Venice heavily emphasizes privacy protection as a core strength. When users call inference services, they don’t need to overly worry about the service provider retaining, censoring, or leaking data, enforcing content moderation, or being forced to hand over sensitive information. Chutes and AkashML let anyone connect GPUs to the network, monetizing idle compute and lowering service costs. Even though these distinctive features can help service providers win a limited share, most centralized platforms can replicate similar capabilities. Relying on product features alone is difficult to secure a meaningful market share.

On-chain inference services can build a truly defensible moat through inference financialization mechanisms: converting inference access rights into assets that buyers can hold, hoard, and resell, instead of only subscription services that can be consumed once.

Venice: Tokenizing Inference Ownership

Venice was founded by Erik Voorhees, a veteran in the crypto industry and a serial entrepreneur, and it pushes the furthest along the path of converting inference access rights into holdable assets. The project uses a two-token system—VVV and DIEM—wrapping future inference service claim rights into assets that users can mint, hold, and resell.

VVV is positioned as a “capital-type asset.” Holding VVV is not the same as owning Venice equity (the platform equity is financed independently and completed a $65 million Series A round in June), but token holders can theoretically share in the project’s development upside. The most direct mechanism is: Venice uses part of its revenue to buy back and burn VVV. Buyback-and-burn is split into two channels: one is discretionary burn using regular revenue; the other is a programmed burn rule—in every new subscription revenue, a fixed portion of funds will be used for buyback-and-burn. To date, 42% of VVV has already been burned.

VVV also has utility. Users can stake any amount of VVV to earn annual token inflation rewards. Staking 100 VVV unlocks a Pro-tier subscription right. But its most core value lies in the linkage mechanism with DIEM (Venice’s “compute asset”). Holders can stake locked VVV to mint DIEM, where each 1 DIEM permanently corresponds to $1 of Venice inference quota. Holding 100 DIEM means having $100 of API quota, usable for all models on the platform and valid indefinitely (as long as Venice continues operating normally).

The amount of VVV required to mint a single DIEM follows a curve set by the platform. As the circulating supply of DIEM approaches the project’s preset target cap, the required staked VVV increases exponentially. The logic is that each DIEM creates a perpetual $1 settlement liability on the balance sheet. Currently, DIEM supply is nearing the target threshold. The minting cost has risen from roughly 90 VVV/DIEM at launch to now several hundred VVV to mint one DIEM. This mechanism slows the speed of new DIEM minting, meaning early minters acquire DIEM at a lower VVV cost, while late entrants face significantly higher costs.

When VVV staking is locked to back DIEM, stakers can only receive 80% of regular staking yield; the remaining 20% goes to the Venice platform. Moreover, the staked VVV locked in this way can only be unlocked when the corresponding DIEM is burned. If a minter has already sold DIEM and wants to retrieve the staked VVV, they must buy back DIEM on the secondary market. If DIEM’s market price rises, redeeming will result in a loss.

The two-token design forms a mutually reinforcing feedback loop. DIEM can only be minted via staking-locked VVV, so when market demand for DIEM rises it continuously withdraws circulating VVV, giving VVV real-use scenarios beyond pure speculation. Conversely, DIEM also benefits from Venice’s platform growth: the stronger the platform’s practicality and the more widespread it becomes, the higher the value of this transferable daily inference service claim right. Holding DIEM is not just holding a claim quota to resell—it is also akin to betting on Venice’s long-term development.

Even if end users never touch crypto assets, the platform’s main business can still continuously empower the token economy. Venice’s team says most users are not crypto-native; many have no interest in the token itself. But as long as users open subscriptions, buy inference quotas, and use platform services, those actions will still drive VVV buybacks and burns and generate demand for Venice inference services. The token economy is built downstream of product business, not replacing the product itself. Venice is not a crypto token hunting for AI use cases everywhere; it is an AI product, and part of the platform’s usage rights and access privileges are routed into a tokenized inference trading market.

The most distinctive aspect of Venice’s DIEM lies in its ownership attribute: users can own the inference resources they consume, rather than simply renting them.

DIEM is an experiment around tokenizing inference access rights and the delivery model. Its core differentiating highlight is ownership. Users can hold the inference resources they consume, not merely rent them. Under a pay-per-use model, once inference resources are consumed they no longer exist. But holding tokenized access credentials is like owning an asset that can be held long-term, transferred, or sold. This gives rise to multiple use cases:

• Because the claim right is tradable, holders facing fluctuating demand can retain baseline usage entitlements, sell or rent out quota during idle periods, and recover the cost that would otherwise be fully lost under a strict pay-per-use model. AI agents can directly hold DIEM, obtaining a no-permission, self-owned balance of inference resources for calling. Relevant trading venues include spot trading platforms such as Aerodrome, or longer-duration rental markets like Surplus, UsePod, AntSeed, and CarpeDiem.

• The Venice team often mentions another scenario: users buy DIEM, use inference services for that day, and then sell it the next day. If the price stays stable, it’s equivalent to free use of inference compute; if the price rises, users can also make additional profit. Of course, risks exist as well: if the price drops, losses incurred by holders could far exceed the cost of directly buying inference services for a single use. For some users, this means that while consuming compute, they can also speculate on the price of inference resources.

• DIEM can also lock in costs. Enterprises or agents with stable and predictable compute demand can use DIEM to lock costs along the compute dimension. The logic is similar to multi-year cloud compute pre-purchase contracts. Enterprises cannot predict how much inference quota $1 will be worth two years later, but they can lock in rights today. If they transfer close to the cost price after use, the actual compute usage cost can be very low. As of July 7, 2026, DIEM is priced at $1,270. One DIEM is roughly equivalent to a daily $1 quota over four years—meaning the buyer effectively pre-pays about three and a half years of perpetual compute usage rights. But the downside is obvious: to buy this kind of cost certainty, users must hold a highly volatile, perpetual, dollar-denominated asset, which undermines the stability users originally sought. DIEM’s pricing is built on a perpetual redemption commitment, implying that the market’s discounted rate for Venice’s continued operational capability reaches double digits; and this claim right only has value if Venice continues providing services normally.

This mechanism is still in an early stage and has tangible weaknesses:

• Inference tokenization is best suited for issuers that need to create upfront demand and raise funds. AI labs with top-tier models and real pricing power lack incentive to push tokenization, because the model would lose the ability to implement differentiated pricing for different customers, could not capture upside from expired but unredeemed quotas, and would weaken flexibility in adjusting prices later.

• DIEM has no maturity principal-protection mechanism. There is also no underlying collateral or reserve asset support—unlike the GPU collateralized lending products discussed below. Holding DIEM is equivalent to an indefinite bet that Venice will keep operating normally years later. If the platform stops providing services, holders have no contractual safeguards or recourse paths.

• The claim right represented by DIEM is defined unilaterally by Venice as $1 of inference rights, not a fixed quantity of compute quota. Venice sets the token consumption standard for each model unit, and that standard changes with supply-demand conditions. The holder’s risk comes not only from secondary-market price volatility, but also from discretion between the platform and holders. In theory, when model costs decline, $1 should buy more compute—but holders only benefit from that upside if Venice chooses to share the reduction.

More deeply, the question is whether an equity-like form such as DIEM—dollar-denominated and perpetual—represents the risk exposure that buyers of inference compute actually want, or whether market participants prefer other debt instruments with fixed terms and pricing based on compute or AI Tokens (or both).

At present, DIEM is mostly held as a speculative asset and is not actually used to call inference services. The actual inference quotas consumed weekly are less than 50% of its total capacity. Venice’s official materials define DIEM as a range-volatile perpetual claim certificate and categorize buyers into three groups: end API users, holders who do not sell tokens and continue capturing value (VVV holders), and speculators who arbitrage spreads. The latter two groups make up the vast majority.

The closest centralized counterpart is OpenAI’s scaled offerings (Scale Tier). Users commit in advance to model throughput, priced per minute in AI Tokens and locked for a period. But scaled offerings do not provide compute ownership—quotas are bound to accounts and can only be used within OpenAI’s platform, not transferable. DIEM’s advantage is exactly the opposite: it can be held long-term, resold for a second time, and combined with other components in the crypto inference ecosystem. A more ideal financial instrument might blend scaled offerings’ fixed-term compute-priced quotas with DIEM’s ownership and transferability attributes.

For Venice, each DIEM circulating outstanding represents $1 of compute that the platform must keep redeeming continuously—and cannot be resold again to other customers. Therefore, it is a liability. Thus, the platform uses revenue to buy back and destroy tokens, which is not simply a reward to token holders.

Ultimately, VVV and DIEM were not designed as equity-like tools for Venice. They were originally used as cold-start mechanisms to build platform users. Now their value foundation comes from the tokenized right to claim compute. VVV holders can mint DIEM to obtain a perpetual inference service claim right for Venice. As the platform grows and compute value rises, the value of this right will also increase. From Venice’s perspective, each DIEM that has not been redeemed is a compute liability awaiting settlement and cannot be sold repeatedly. This is the core reason the platform uses revenue to buy back tokens—not to treat holders preferentially. One side holds the claim right and expects upside. The other side bears the settlement obligation and wants to reduce liability scale. The benefit linkage created around Venice’s compute (not any equity relationship) forms a coordinated bond. This is also why VVV is an exploration with special meaning: it builds the inference business on top of an application-based token mechanism.

Tokenization at the Inference Output Side

Venice tokenizes inference access rights; the Proof of Useful Work network focuses on tokenizing inference production processes, subsidizing inference service supply costs through token inflation. Traditional proof-of-work networks bootstrap by paying token rewards to miners who solve random puzzles, and Bitcoin uses that mechanism to secure the network—but much of the compute power is then used solely for meaningless hash calculations without real output. Proof of Useful Work replaces the random puzzles with real inference tasks, securing blockchain security through the same batch of compute and producing AI services that customers are willing to pay for. Pearl and Ambient are the two on-chain deployment routes currently live, and their underlying design philosophies are completely different.

Pearl

Pearl Network is a one-layer public chain built by forking Bitcoin code. It preserves Bitcoin’s UTXO ledger model and difficulty adjustment mechanisms, but it replaces the SHA-256 hash algorithm with matrix multiplication—the core computation underlying AI inference and model training. Pearl’s core proposition: the matrix operations used to fulfill user inference requests can simultaneously function as proof-of-work.

In AI model processing, the underlying operation that handles prompts is essentially the multiplication of two large numeric matrices—matrix multiplication. In Pearl’s mechanism, miners add a layer of random noise onto the original matrix to perturb it, and then perform multiplication on the perturbed matrix. Matrix multiplication forms a computation-heavy task that participates in a mining race across the network. During execution, the system continuously checks whether intermediate results satisfy the difficulty threshold. Miners who reach it first obtain block rewards under rules identical to Bitcoin. The difference is that the work being verified is the real model inference computation, not meaningless hash computation in traditional mining. After the matrix operations complete, the system quickly removes the injected random noise to restore the standard inference result the customer needs. One compute operation yields two benefits: it both generates AI output content and competes for block rewards.

Two key design choices make this “two-in-one” mode feasible for real deployment. First, Pearl adapts to vLLM (a widely used model runtime software in AI enterprises) in a plugin form, enabling service providers to integrate quickly without restructuring their existing architecture. Second, the winning computation results need to be publicly verifiable by the entire network, so Pearl packages data using zero-knowledge proofs to protect users’ prompts and the service provider’s proprietary model weights from being leaked. The added overhead from these auxiliary mechanisms is relatively low. Pearl discloses that running models under this scheme increases computational load by 0.5%~10%. In live testing based on the popular open-weight model Llama-3.3-70B, Pearl’s optimized version matches or even exceeds the original version’s runtime speed, because the team’s refactored core operation module is more efficient than the standard implementation in some deployment environments.

As one of the first networks combining proof-of-work with AI inference, Pearl initially attracted strong miner attention, and network-wide compute power rose quickly. However, the protocol cannot distinguish between valid computations (the computation tasks that serve real inference requests) and invalid computations. Whether or not a customer needs the corresponding computation output, the execution itself is deemed valid. Pearl’s whitepaper anticipated this issue, and its model assumptions already included a miner group that simply executes meaningless computations to collect block rewards. The market performance after launch also confirms this. Early mining hype drove compute power sharply upward, but there were almost no signs that this compute power was actually being applied to real inference business.

But more and more signals suggest the project is starting to land real business. The most notable progress is that in May, Pearl announced a partnership with Together.ai, a top inference and compute service provider, launching inference access nodes with pricing at least 25% lower than Together’s usual rates; the price gap is subsidized by Pearl token rewards generated from the same compute. At the end of the day, a compute-for-two-use architecture like Pearl can only produce meaningful output when paid real inference demand dominates compute deployment. If there is no end-user demand, block rewards alone only attract mercenary miners, and the network will ultimately become another kind of proof-of-work mechanism without real production—similar to Bitcoin.

Ambient

Ambient takes a completely opposite design approach from Pearl. Instead of allowing miners to run any model, Ambient requires the entire network to use a single large open-weight model, and it builds a consensus mechanism around verifying the output results of that model.

Pearl uses a brute-force contest mode where all miners compete to solve the same problem; Ambient uses an auction mechanism for competition. Users or AI agents submit inference tasks, set deadlines and bids, which is essentially “finish this inference within X minutes and I’ll pay Y,” after which miners participate in bidding to take on the work. The winning miner runs the query using the network’s unified model and posts a bond. If the miner fails to deliver results on time, the bond is forfeited—constraining miners to provide quality service and response speed. The system randomly selects a set of verifiers to check results. Verifiers’ priority is weighted based on past records of valid work, rather than the size of staked assets. Miners handle a large number of diverse tasks in parallel rather than all competing for one single block, enabling the network to avoid the performance bottlenecks traditional proof-of-work has. The project is developed via a Solana code fork, using a proof-of-useful-work mechanism to replace staked consensus, aiming to achieve near-Solana-level execution speed.

Ambient is a service provider on the OpenRouter platform for the Kimi K2.7 model, with Token input/output quotes that are the second lowest.

The auction mechanism is also the core reason Ambient can make inference prices competitive. Typical API service providers need to rely on the fees paid by users to cover the full cost of each request. Under Ambient, miners completing the same amount of work can earn double benefits: first, the fee paid by the winning user or agent submitting the query task; second, protocol rewards issued for verified useful work. Since miners bid around tasks with explicit quotes and latency requirements, their bids are anchored to the net cost after subtracting expected token rewards, not the gross cost before subtracting rewards. In effect, token inflation subsidies supply-side are transferred mostly to the demand side through the auction mechanism, reflected as lower inference service prices. The most important difference from ordinary mining subsidies is that rewards are bound to real tasks that someone has actually posted and paid for. If this mechanism runs smoothly, token issuance won’t be used solely to buy compute anymore; instead it will buy lower-priced, verifiable inference services—attracting more users and giving miners more business volume, which further reinforces the network’s demand foundation for its tokens.

This auction mechanism is also what Ambient claims it solved the problem Pearl couldn’t overcome. In Pearl, miners can earn block rewards as long as they perform matrix multiplication regardless of whether there is a customer needing the computation output; this is why a large amount of compute with no real demand gets pulled into the network. In Ambient’s system, miners can only obtain Ambient tokens (not yet issued) by winning and taking on tasks that others have published and paid for. The mechanism design, from the root, merges mining behavior and taking on real inference tasks into one.

Ambient also uses an original approach for inference output verification. If a miner claims that it used the agreed model to execute a query request, how can users confirm the miner didn’t secretly switch to a cheaper, lower-quality model to cut costs? Even in centralized services today, this kind of risk exists—many institutions have been accused of secretly lowering output quality to compress costs. Ambient’s solution leverages the core runtime characteristics of large language models: as a model generates text, each step outputs log probabilities (logits)—i.e., before choosing the next token, it scores all candidate tokens with raw values. This stream of scored data acts like a fingerprint of the model computation process and can be compressed via hashing into a short numeric value for comparison and verification.

For miners generating thousands of Tokens of output, verifier nodes do not need to fully rerun all tasks. The verifier randomly selects a node position in the generated text, requiring the miner to provide the computation fingerprint at that point. Then the verifier runs one single round of the model only at that position to generate one Token, and compares whether the fingerprints match. With only a single computation, the system can verify the full result of thousands of Tokens. This logic resembles Bitcoin: producing the work is expensive, but verifying the work is cheap. Ambient claims this verification adds overhead of around 0.1%. By comparison, other projects’ zero-knowledge proof approaches add overhead typically reaching 10x to 1000x.

How much value does Proof of Useful Work actually have?

The key difference between these projects and other decentralized compute networks is that the work used to secure the blockchain is exactly the inference business customers actually need. When the mechanism runs smoothly, one unit of energy investment yields both network security and a product that can be sold externally. For compute service providers, mining becomes an additional revenue stream for existing hardware. At the same time, inference outputs are verifiable, so when AI agents purchase inference services, they don’t need to blindly trust that the service provider won’t degrade models or arbitrarily stop services.

Once there isn’t enough real terminal demand, block rewards alone are enough to attract miners; the entire PoW network will be filled with idle compute that has no customer attached—only a shell of “useful work,” lacking real output.

Besides various technical difficulties, there are two major barriers to realizing the vision above. First is competition on the demand side. Decentralized inference networks must directly face competition from centralized service providers and pure GPU leasing businesses. The latter don’t need to bind with crypto tokens, and are often faster and cheaper. To break through, decentralized networks must capture a specific category of buyers: those seeking the lowest trust barrier, results that can be verified, resistance to censorship, and neutrality, without platform-side unilateral rug-pull risk. The current market size willing to pay a premium for this is still limited, but there is potential for rapid expansion in the future: provided relevant projects can continuously and stably provide cost-effective inference services, or if market trust in centralized AI service providers continues to decline. Pearl’s launch experience is a warning case: in the absence of enough real demand, relying only on block rewards attracts mercenary miners, leading to large amounts of compute being stacked without actual customers, making useful work merely nominal.

The second challenge is the token value-capture mechanism. All projects outline a growth flywheel: real business usage drives demand for the native crypto token; token demand supports mining rewards and secures the network; network security further attracts more business usage. Yet so far, no project has truly run this closed loop. After mining outputs tokens, miners typically sell tokens to cover operating costs. On the demand side, there is no mechanism that forces users to buy tokens at scale. When users use inference services at large scale, verify credentials, and utilize core products, they mostly do not need to hold crypto tokens. Users can pay Pearl inference fees using USD; the project plans to launch a token-to-compute exchange trading market in the future, which also indirectly admits that the current flywheel has not formed. Ambient, meanwhile, delays publishing its token economics plan and has not clarified whether inference services will be priced in native tokens. The end result is that tokens mainly come from mining and are then sold off, rather than being actually consumed by business scenarios.

The most feasible path for these networks is to build the native token as the underlying payment channel for inference services. That is the most direct way to complete the value loop. Combined with the price advantage brought by token inflation subsidies, this strategy has strong attractiveness. Lower inference prices attract real traffic, and if services must settle using native tokens, then business usage could become native token demand. But for the flywheel to work positively, there are prerequisites: users’ usage habits must be retained; as subsidies gradually decline, naturally generated native token demand must eventually exceed mining sell pressure.

AI Inference Hardware Financing Track

Venice tokenizes inference access rights; Pearl and Ambient tokenize inference production; and at a deeper layer, an entirely new on-chain market is emerging, dedicated to providing financing for the GPUs that carry inference workloads. The mode described in this chapter most reflects the advantages of crypto technology. The key to having this system run smoothly is that the project does not issue new tokens and does not need to cultivate token demand for cold starts. The platform absorbs conventional funds through hardware assets, converts stablecoin deposits into procurement loans for compute operators, and then uses GPU rental cashflows to repay depositors principal and interest.

Top compute operators have long financed using mechanisms such as bank credit lines, asset securitization, and private credit. CoreWeave’s multi-billion-dollar GPU collateralized loans are a typical example. However, it’s harder for smaller and mid-sized AI vertical cloud providers to finance. They hold hardware assets and have contracts generating stable rental income—meeting the loan base conditions—but they lack complete balance sheets, treasury/capital management teams, and access to credit institutions resources, making it difficult to obtain financing quickly. USD.AI is designed to provide lending services for these kinds of entities. Deposit users provide loan capital, GPU leasing income is used to repay debts, and the interest generated is returned to depositors as yield. Compared with traditional banks, this model has three difficult-to-replicate advantages:

• The supply of funds is open to all stablecoin holders, not limited to a small number of closed lending funds.

• Every loan becomes a composable on-chain financial instrument, supporting staking, trading, or serving as collateral in other protocols.

• The beneficial rights of collateral are recorded on-chain for claimability, while also relying on mature traditional legal frameworks to ensure recourse rights.

USD.AI uses a two-token mechanism. Depositors can mint USDai—this is a synthetic dollar stablecoin backed by PYUSD issued by PayPal (and PYUSD is itself backed by U.S. short-term Treasury bills and cash reserves). USDai does not generate yield itself. Its original design intention is to maintain high liquidity and composability. If depositors want to earn yield, they can stake their USDai to convert it into sUSDai; when the staked position earns, value accumulates in parallel. Yields come from two parts: interest paid by the GPU borrowers on outstanding loans, and Treasury yields from reserve assets that remain idle before being deployed. Currently, loan deployment size is about half of the reserves, and the annualized staking yield remains around 8%; as more funds complete deployment, the protocol’s target yield range is 10%~15%.

The core difficulty of physical GPU-collateralized loans lies in enforcing creditors’ claims if borrowers default. USD.AI states that it will handle related processes through the CALIBER system. This name is an acronym for “Collateralized Asset Ledger: Insurance, Bailment, Evaluation, and Redemption.” Under this framework, the GPUs that complete financing will complete information备案 and be tokenized into ERC-721 standard NFTs. USD.AI says these NFTs have legal title certificate efficacy under Article 7 of the Uniform Commercial Code. Based on bailment arrangements, borrowers can continue to use the hardware devices, while the corresponding NFTs can be pledged as collateral. The binding relationship between tokens and physical hardware does not automatically take effect and cannot be achieved purely through technology. The whole linkage relies on written备案 documentation, on-site inspection, device installation certificates, insurance policies, ongoing collateral monitoring, lien documents, and coordination with data centers or asset custodians. If a default occurs, on-chain auctions can only transfer legal claims; recovering physical assets still requires off-chain legal and operational systems support. This framework has not yet undergone stress testing through a full default-asset disposal cycle, and long-term reliability still needs verification.

There is naturally a maturity mismatch problem when liquidity tokens are connected to three-year installment amortizing loans. Many real-asset (RWA) credit protocols rely on commitment-based instant redemptions to mask this risk, but under market stress they easily blow up. The USD0++ depeg event is a typical example. USD.AI does not promise instant redemption. Redemption requests are processed in 30-day settlement cycles, using only the principal already amortized and repaid during that period for redemption, and it follows FIFO (first-in, first-out). The protocol will not dispose of normally performing loans to meet users’ withdrawal needs. The protocol’s upper layer borrows the Flashbots MEV-Boost mechanism by introducing a competitive redemption queue: users who want priority redemption can bid for the right to jump the queue, and related fees will be allocated to the holders who choose to wait in line. Loan term design is similar to commercial mortgage-backed securities (CMBS): loan-to-value ratio is 70%~80%; borrowers must maintain reserves covering roughly three months of principal and interest payments; if payments are late, assets will be forfeited. Hardware comes with insurance, ongoing monitoring, and assets can be recovered through specialized partner institutions.

This report includes USD.AI as a core reason because it connects into a business closed loop with the pricing layer. Lenders that provide financing for GPUs need fair market reference standards for collateral: how much revenue the hardware can generate, depreciation pace, safety loan ratios, and how risk exposure is hedged. Compute price indices and GPU futures provide exactly these pricing benchmarks. But the GPU collateralized lending business gives real credit risk exposure to the price market, so that the quoted prices have real financial meaning. In short, the pricing market provides valuation references for lenders’ hardware.

All types of risks are clearly visible, and a high yield rate is itself market pricing of risk. Whether the entire business architecture can continue depends on whether GPU rental cashflows are sufficient to cover loan principal and interest—this is also the key variable running through this report, manifested here as credit risk. If inference demand weakens, or if GPU supply becomes abundant and rental rates decline, borrowers’ cashflows will continue to tighten and default rates will rise. And when the protocol needs to liquidate collateral, the value of the hardware will simultaneously decline. USD.AI loans are amortized over a three-year cycle and correspond to a labeled 7-year equipment usage lifetime, reserving safety buffers; but if the hardware iteration cycle accelerates, the safety margin will narrow. Because this model relies on crypto-market fundraising that is invested into AI capital expenditure cycles, if the industry enters a downturn, collateral value, borrower business demand, and depositors’ willingness to contribute capital will all come under pressure. This high correlation is a risk point worth focusing on.

Two project case studies have reference value. Maple Finance proves that when professional credit risk control is packaged into yield tokens with liquidity and composability, on-chain credit can expand at scale. The business model validation shows that DeFi users are willing to invest via tokenized debt claims managed by institutional private credit assets. USD.AI uses a similar distribution vehicle, but the underlying targets are less liquid, steadily depreciating physical GPU assets rather than more liquid crypto-collateral or shorter-term institutional credit. OnRe applies a similar distribution vehicle in a higher-threshold real market segment: reinsurance. Users can use a composable dollar asset to obtain cashflows formed from both premiums and collateral asset returns. The common thread across these cases is distribution-channel capability. Crypto underlying channels broaden the path for ordinary investors to participate in private markets, but they do not reduce the risks embedded in the underlying assets themselves.

Conclusion

At this stage, whether on-chain or off-chain, the inference capital market remains very small compared to the overall growth scale of the artificial intelligence industry. For on-chain related services to achieve scaled growth, they must prove that their advantages are sustainable and have long-term viability.

The sector’s advantages are very clear. Tokenized access rights (Venice) convert the right to claim inference services into bearer assets. Holders can hold, resell, rent out, or have AI agents call them. Unlike subscription quotas bound to a single account where the service provider can unilaterally revoke, effective proof of useful work (Pearl, Ambient) leverages token inflation subsidies to make inference pricing below market average while also making results verifiable, so buyers don’t need to simply trust that the service provider won’t secretly replace cheaper low-quality models. The hardware financing business (USD.AI) transforms illiquid GPU credit assets into composable financial instruments. All stablecoin holders can participate in contributions and redemptions, with higher efficiency than traditional credit industries. Beneath all three solutions, the underlying features include permissionless, programmable capabilities, which precisely fit the likely future primary group of AI agents in on-chain inference capital market demand. The value of crypto technology is concentrated in scenarios where ownership protection, neutrality, composability, and access to inclusive capital are crucial.

The obstacles to deployment cannot be underestimated. Currently, no project has been able to convert real compute consumption demand into a true necessity for corresponding crypto token demand. Compute production networks keep minting tokens and then selling them off, relying on token inflation subsidies to offer inference services below market price—yet after token issuance, the tokens mostly flow into the market for immediate liquidation. Trading of tokenized access rights revolves more around speculation driven by project development expectations; DIEM holders are mostly speculators, and its price essentially reflects a bet on Venice’s development prospects rather than the value of actual compute usage. The financing track is an exception and the only segment with real industry customers: various AI vertical cloud providers have financing needs, and they also have rental cashflows that can be used to repay debts. Therefore, revenue comes from real business demand rather than from issuing more tokens to create hype. Overall, the current financial system’s ability to absorb speculative capital is significantly stronger than its ability to build a self-sustaining ecosystem driven by real usage demand.

Amid the booming AI industry expansion, the true core competitive advantage of on-chain inference capital markets is not direct competition in the areas where centralized giants are best—massive scale and low-cost inference service provision. Its opportunity lies in building capital channels that traditional finance can’t cover well, channels that respond slowly and are limited in scale, as well as serving emerging markets. This is also a repeatedly validated development pattern in crypto: crypto rarely wins directly at terminal products, trading platforms, large model layers, and application layers. But it often can build supporting financial infrastructure the fastest, covering asset pricing, fragmented splitting, financing, settlement, and more.

Compute inference is the newest and most scalable example. A massive asset category reaching tens of trillions of dollars is quickly taking shape. Yet to date, there is almost no complete market system that treats compute as a financial asset—indices, futures, credit, and tokenized compute quotas. This blank space contains enormous opportunities. The reason the financing business can already run is that it is the first link in the entire system to land real demand. The rest of the ecosystem is betting that as compute continues to be financialized, this unique advantage will extend upstream.

It may take years for the inference market to mature. But the financial layer built around compute is already taking shape right now.