Scale unrestricted private inference
The opportunity is to own the execution layer for sensitive, unrestricted AI workloads: private prompts, useful model behavior and verifiable cache billing, delivered through a capital-efficient compute portfolio. Harnesses bring recurring sessions; OpenRouter aggregates public demand; ZAMA funds switching and integrations; structured compute contracts turn that demand into lower cost and less capital tied up.
The demand is already in APIs
Our private notes put 81% of Venice demand in APIs, spanning Fleek, OpenRouter, IDE extensions and NSFW platforms. They also record a 16% OpenRouter share. The API-heavy mix is the important signal: a large private/unrestricted inference business can be built inside other products, without first winning a standalone consumer chat habit. [S45]
Venice's founder-reported revenue milestones move from $100M annualized in August to above $150M annual revenue in October, as recorded in the supplied diligence. The notes describe consumer subscriptions as roughly 50% margin and API margins as thin. Distribution exists; keeping the spread is the harder problem. Revenue, usage share and margin have different denominators—use the API mix to locate demand, rather than multiplying it into an unsupported segment P&L. [S42][S43][S45]
The notes identify creative writing, roleplay, security work and users blocked by refusals. Agent workflows add long-lived proprietary context. These customers need the model to engage with their task and the execution path to protect it. A permissive model on a provider-visible route solves only half the problem.
Who owns each layer
| Player | Position | Leverage for Zama |
|---|---|---|
| Hermes / Nous, pi, IDE and agent harnesses | Own context, tools and repeated sessions | Native private mode and persistent session placement |
| OpenRouter | Aggregates public demand and allocates provider traffic | Win allocation through realized session cost, speed and reliability |
| Venice | Established private/unrestricted app and API; token incentives and credit bundles | Win on operational reliability, stable models, simple billing and explicit verification |
| Abliteration | Unrestricted models for security, research and training data | Compete on selected task quality; license or host appropriate models |
| NEAR, Phala, Akash and specialized clouds | Confidential execution or GPU supply | Assemble compatible supply and a second-source recovery path |
| Liquid Compute, Ornn and hedge counterparties | Contract pricing, financing and compute-market infrastructure | Structure duration, recovery rights and residual price exposure |
| Zama / Rand | Cryptography reputation, relationships and ZAMA incentive capital | Finance distribution and verification while building modern inference execution |
The private feedback from Tarun credits Venice, Prime Intellect and Nous with using crypto-market intuition to grow. It also describes Zama's AI connections as stronger before the current boom. Zama's advantage is the combination of capital, relationships and cryptography; inference operations and serving expertise still need to be built. [S45]
Cache cheating exposes a product gap
The original notes describe the trade: quote a low input price to win router traffic, then bill repeated context as uncached. They locate the economic crossover at roughly five prompts and estimate 2–10% revenue extraction. These are the notes' estimates, useful for choosing repeated-session workloads and a billing benchmark. [S45]
Robot's research finds strategic pricing and cache-related extraction incentives. Its measurements do not identify intent for every provider. OpenRouter itself describes the old pricing games and changed default routing on October 28 to consider realized price, speed and failures, while retaining warm-cache affinity. The opening now is to deliver and demonstrate honest execution under the new allocation rules. [S02][S01]
Verifiable cache usage makes price integrity part of the product. The proposed serving stack records fresh input, cache reads/writes, eviction events, output tokens, model/version and applied rates in a signed usage receipt. Bind metering code to the attested serving path and let the customer reconcile each receipt to the invoice. A signed counter alone does not prove a cache hit: verification must cover the code that observes cache events and computes the bill.
Keep receipts customer-scoped and content-free. Publish cache lifetime and eviction rules, preserve session affinity and return a visible cache miss when reuse fails. Compare total cost across repeated tasks, including retries. TEE protects the execution boundary; cache management and metering produce the billing advantage.
Compute markets create capital efficiency
The Google AI discussion's useful mechanism is the mismatch between long supplier commitments and short customer demand. An unused GPU hour expires; location, interconnect, hardware and credit terms change its value. Financial engineering connects those mismatches to a usable inference service. [S22][S31]
Buy duration against committed demand, negotiate prepayment discounts and finance the commitment. Secure supplier sellback or secondary-transfer rights before purchase. Use public router demand to fill variable capacity, retain service headroom and sell surplus hours that have a better external use. Hedge the remaining price exposure with matched futures or OTC contracts. [S31][S33]
Liquid Compute's financing structures and Ornn's term curves add ways to price and fund that book. A supplier whose downside is protected may accept a smaller deposit or lower rent; capture the benefit in our contract. Match chip, TEE configuration, region, tenor and collateral needs rather than treating index prices as interchangeable GPU capacity. [S32][S34]
The economic edge is lower cost per useful GPU-hour and less equity trapped in idle capacity. That buys predictable customer prices, redundancy and burst headroom. Venice's move toward owned clusters makes utilization and replacement capital part of the competitive comparison. Our portfolio must outperform that ownership model on delivered service and capital efficiency. [S44]
Why the product can beat Venice
The notes identify reliability failures, billing friction, frequent model/API changes, confusing privacy levels and weak roleplay depth. Those are concrete product openings. Venice already offers fiat access, TEE/E2EE, agent integrations and subsidized builder programs; matching those features will not win the account. [S45][S08][S39]
| Differentiation | Customer benefit |
|---|---|
| Verified private route across primary and auxiliary inference | Sensitive context stays within the stated execution boundary |
| Strong unrestricted models for selected workloads | Fewer unnecessary refusals and better task completion |
| Attested metering and reconciled cache receipts | Repeated context receives the cache treatment the customer paid for |
| Structured supply, recovery rights and service headroom | More predictable prices and fewer capacity failures |
| Stable versions, plain invoices and explicit privacy controls | Less migration work and fewer operational surprises |
Venice's $68 plan includes $75 API credit face value; its $200 plan includes $225. Compare effective bills and successful tasks after allowances, cache savings and incentives. [S38]
ZAMA turns capital into a stronger operating position
The planning envelope is approximately $50M of ZAMA over four years. Its role is to absorb integration effort, switching cost and initial adoption risk. The durable asset is the retained workload and its purchasing power. The original notes explicitly ask what remains after incentives: recurring customer demand must fund the service. [S46][S45]
More retained sessions improve capacity underwriting; better terms fund service quality and lower session cost; the better product retains more demand. Keep customer receipts, award cost, capacity-sale proceeds and hedge P&L separate so each part of that mechanism remains visible.