Lium reported approximately $1.94 million in marketplace billings over 30 days in its October 26, 2026 snapshot. Yet customer rentals accounted for only part of the income available to its hardware suppliers: 25.2% of estimated, monthly-normalized H100 provider earnings came from Bittensor incentives, using the reported rounded figures.
The billing measures activity on the marketplace. The earnings mix shows how that activity supports the machines behind it. A customer can pay for useful compute while rewards also contribute to the provider’s return.
For Lium, the commercial question is whether rental earnings can sustain competitive GPU availability as those rewards change. Answering it requires separating customer spending from protocol incentives, then examining the costs and operating conditions each side faces.
An operating marketplace, with measurable activity
Lium connects customers seeking GPU compute with independent hardware providers. Customers rent machines for workloads such as model training, inference and experimentation, while Lium handles discovery, provisioning, metering and payments.
Its network statistics, checked on October 26, 2026, reported approximately:
- $1.94 million billed to renters over 30 days.
- 1.02 million GPU-hours served.
- 21,636 rentals started.
- 83% utilization of offered GPU-hours.
The billing totals cover the last 30 complete UTC days, ending October 25; utilization is a separate rolling measure. Lium defines gross marketplace billing as charges for GPU pods and storage volumes. That is not the same as the company's revenue, and it should not be described as profit.
A rental is a transaction, not necessarily a unique customer. The published aggregate feed does not identify how many independent customers account for the activity or how many return repeatedly.
Utilization measures the proportion of offered GPU-hours that were rented. It does not establish whether every rental completed useful work successfully. The reporting population also depends on the platform's definition of currently listed and eligible GPU models.
These supplier-reported figures come from Lium’s systems, not an independent financial audit.
What an H100 actually earns
The H100 earnings breakdown exposes the difference between the price a customer pays and the income a provider evaluates when deciding where to place a machine.
Providers receive 95% of customers’ GPU rental payments. Eligible machines can also earn subnet emissions while rented or available for rental.
The following estimates use Lium’s provider data observed on October 26, 2026. They are normalized to a 730-hour month from a 30-day observation window, rather than amounts necessarily received during those exact 30 days. Ownership and operating costs still have to be deducted; these estimates are not audited monthly profits.
| Monthly income estimate per GPU | NVIDIA H100 80GB HBM3 |
|---|---|
| Provider rental-fee share | $933 |
| Idle rewards | $198 |
| Additional rented rewards | $116 |
| Total reported earnings | $1,247 |
| Share from subnet incentives | 25.2% |
Source: Lium provider earnings and its public provider-economics feed. Model-level estimates retain the rounded October 26 snapshot; percentages use those displayed values. Idle and rented rewards are USD equivalents of Subnet 51 alpha stake, holdings of the subnet’s token, not guaranteed cash payments. Realized value depends on token prices, liquidity and conversion costs.
The customer assesses performance, reliability and the rental bill. The provider must cover hardware ownership and operation from a broader income base. An attractive customer price can therefore coexist with supplier economics that remain exposed to token rewards.
The H100 sensitivity test
The reported H100 estimate combines $933 in rental earnings with $314 in incentives: $198 of idle rewards and $116 of rented rewards. The scenarios below reduce only the incentive component.
| Scenario | Rental earnings | Incentives | Total |
|---|---|---|---|
| Reported baseline | $933 | $314 | $1,247 |
| Incentives reduced 50% | $933 | $157 | $1,090 |
| No incentives | $933 | $0 | $933 |
Scenario assumptions: Customer demand, rented GPU-hours, operating costs and rental earnings are held constant. Incentives remain USD equivalents of alpha stake. This isolates income sensitivity; it does not model how providers or customers would actually respond.
Halving incentives removes $157 from the monthly estimate. Removing them entirely reduces income to $933, approximately 25% below the baseline. With costs held constant, that is less income available to cover those costs; the table does not establish whether either outcome would leave a provider profitable.
Alternatively, to replace the lost $314 entirely through customer rentals, rental revenue would need to rise by about 33.7%, assuming the same volume of rented GPU-hours, the same provider fee share and no other changes.
At the reported average rental price of $1.90 per GPU-hour, the equivalent customer price would be approximately $2.54. The 95% provider share applies at both prices, so it cancels in this proportional calculation.
A provider cannot simply raise prices by 33.7% and assume customers will continue renting. Higher prices may reduce demand, while competing suppliers may be unable or unwilling to increase their own rates.
Hardware depreciation, electricity, networking, maintenance, financing, downtime and taxes are not included in Lium's headline earnings figures.
A provider with inexpensive, already-owned hardware could operate under very different constraints from a company financing a new GPU server.
Support for supply does not establish artificial customer demand. Rewards may make it worthwhile to list a machine, maintain spare capacity or accept a lower rental rate. Customers can still have independent reasons to buy its compute. The distinction is between what attracts a renter and what keeps the supplier available; this earnings breakdown measures the latter more directly than the former.
How incentives can influence prices
Lium's emissions documentation describes separate allocations for rented GPUs, eligible idle GPUs and emissions directed to a burn mechanism.
The rented reward pool receives 13% of the subnet's provider emissions. The remaining 87% is divided between eligible idle capacity and the designated burn allocation, according to the applicable rules.
Providers must maintain eligible hardware, pass validation checks and keep rental prices at or below model-specific soft limits. Idle rewards also depend on capacity caps and competition from other providers seeking rewards within the same GPU category.
A supplier considering a higher rental price must account not only for the possibility of losing customers, but also for the potential loss of idle rewards if the new price exceeds the relevant soft limit.
Consequently, subnet rewards may influence how machines are priced, even when a particular machine is not currently rented.
The mechanism gives rewards a potential role in pricing decisions. It does not quantify their effect on customer prices or show that a particular discount is funded by emissions.
There is another qualification: idle refers to capacity not rented by an external customer. Lium may run its own default workloads on such machines. The documentation describes these workloads separately from customer rentals.
Their economic contribution and relationship to externally paying demand need to be assessed independently.
Available capacity can help a marketplace win work that would otherwise go elsewhere. But aggregate utilization alone cannot show whether that capacity would remain available with smaller rewards: rental income and costs differ across suppliers. A high fleet-wide figure can conceal hardware operators with very different reasons to stay or leave.
Can Lium compete with Runpod and Vast.ai?
Lium competes with established services, including Runpod and Vast.ai, which already provide GPU infrastructure through different combinations of marketplaces, managed services and deployment tiers.
Lium's October pricing showed that competitive H100 offers were available. Its own comparison with Vast.ai, based on an October 17 snapshot, placed a single H100 SXM at $1.39 per hour against a $2.40 median Vast.ai listing.
Lium compared its lowest available Secure-node price with Vast.ai's market median. The company disclosed this difference, but a minimum price and a median price do not answer the same purchasing question.
A buyer could find another Vast.ai host below that median. Conversely, the lowest Lium offer may not match a competing machine's memory allocation, CPU resources, location, storage or operational characteristics.
Runpod further separates infrastructure options, including Community Cloud and Secure Cloud, alongside products designed for serverless inference and multi-node workloads.
Buyers should compare the actual configurations available when the job begins, using the cost of a successfully completed workload.
- Provisioning time and availability.
- Compute throughput and usable memory.
- Storage and data-transfer charges.
- Failed or interrupted jobs.
- Recovery and checkpointing requirements.
- Security controls and contractual commitments.
Until equivalent workloads are executed under controlled conditions, there is no sound basis for declaring one provider universally cheaper.
The operational details that can change the decision
Security depends on the machine selected.
Lium's security documentation explicitly states that providers operating standard, non-confidential nodes can technically access files, environment variables and process memory within customer containers.
Such access is prohibited by platform policy, but contractual restrictions do not remove the underlying technical capability.
Lium also supports confidential-computing configurations using Intel TDX, with attestation mechanisms intended to protect workloads from host access. Its documentation says attestation covers the node, not the customer's container image; the presence of attestation alone does not verify that image or every component of a workload.
The distinction matters for proprietary models, customer information, credentials and other sensitive workloads.
Runpod documents vetted infrastructure partners and enterprise security requirements for its Secure Cloud. Those host controls are distinct from cryptographic protection against host access. These arrangements should be compared against the exact protections offered by an eligible Lium confidential-computing node, rather than treating all marketplace machines as equivalent.
Deployment controls affect the final bill.
Lium bills GPU use by the second, without a minimum rental duration. Its pods do not have a conventional Stop function: users stop compute billing by deleting the pod.
Persistent volumes and backups are charged separately. Data left only on a deleted pod's local storage is lost.
Lium's published terms also do not provide a contractual uptime warranty.
When Lium Makes Commercial Sense
Temporary experimentation and checkpointed workloads. Lium deserves consideration where a team needs temporary GPU access and can bound the cost of interruption. Compare the complete job bill, keep reproducible environments and checkpoint or export valuable outputs. A low hourly rate is useful when the workload can tolerate the operating conditions attached to it.
Proprietary or sensitive models and datasets. Assess the exact node’s isolation and attestation before uploading data. Platform rules prohibiting host access do not remove a standard provider’s technical capability to inspect a workload. If the available confidential-computing configuration and data-handling controls do not meet the requirement, a more controlled infrastructure arrangement may be appropriate.
Production-critical workloads. Require explicit answers on uptime commitments, incident response, recovery and data preservation. Lium’s published terms do not warrant uptime. Compare its documented protections with the specific Runpod or Vast.ai service under consideration, rather than inferring guarantees from a platform name or a “Secure” label. Where an outage exceeds the team’s recovery capacity, stronger contractual guarantees or a different deployment arrangement may justify a higher price.
A commercial opening for AI agents
Lium’s API, CLI and SDK let an agent discover capacity, rent a GPU, execute a task and terminate the pod without supervising each step through a browser. Restricted permissions, server-enforced spending budgets and scheduled termination make temporary compute a resource an agent can procure within defined limits.
Automation changes the purchasing risk: a missed cleanup step can prolong billing, while automatic termination can destroy unsaved local outputs. Budget limits and persistent storage therefore belong in the workload design. These documented controls merit testing; an API by itself establishes no advantage over competing services.
Where Lium stands
A marketplace can have real customers before it has self-sustaining supply. Lium’s reported billing and H100 earnings mix are consistent with that distinction. The commercial test is whether rental income can keep competitive hardware worth supplying as rewards change, after ownership and operating costs. More GPU-hours alone cannot settle it: customer retention, the distribution of rental earnings across providers and the cost of maintaining availability matter. The present evidence supports an operating marketplace with economically material incentives; it does not yet establish supply that can sustain itself on customer fees.
Research methodology and sources
Research date: October 26, 2026.
Snapshot verification: The public statistics feed dated October 26 reports $1,938,349 in billing, 1,019,915 GPU-hours and 21,636 rentals for September 26–October 25, with 83.1% rolling utilization. The provider feed dated October 26 at 12:30:27 UTC corroborates the H100 estimates; minor same-day differences do not replace the original snapshot.
This analysis uses public data and documentation. It does not represent an independent financial audit, a paid GPU benchmark or a verified sample of individual customer accounts.
Primary sources
- Lium Network Statistics — marketplace activity, utilization, pricing and provider earnings. October 26, 2026 snapshot.
- Lium Provider Earnings — model-level earnings divided into rental fees, idle rewards and rented rewards. Monthly estimates based on trailing 30-day observations, checked October 26, 2026.
- Lium Public Statistics API Documentation — definitions of GMV, GPU-hours, rentals and utilization.
- Lium Rewards Calculator Methodology — sources, aggregation rules and limitations of provider earnings.
- Lium Rental Fees and Subnet Emissions — provider reward structure and eligibility.
- Lium Provider Rental Fees — fee allocation, price limits and penalties.
- Lium Pod Security, Billing and Terms of Service.
- Lium AI Agent Documentation.
- Lium's October 2026 Vast.ai Comparison.
- Runpod Pricing and Security Documentation.
- Vast.ai Pricing Documentation.
Calculation notes
- H100 incentive income: $198 idle rewards + $116 rented rewards = $314.
- Incentive share of H100 reported provider earnings: $314 / $1,247 = 25.2%.
- Rental revenue increase required to replace those incentives, holding rented hours and the provider's fee share constant: $314 / $933 = 33.7%.
- Equivalent reported average H100 rental price: $1.90 × ($1,247 / $933) ≈ $2.54 per GPU-hour.
These calculations measure income sensitivity, not profitability. They do not model customer price elasticity, capital expenditure, operating costs or changes in subnet reward allocation. Source figures are rounded. The 730-hour normalization is distinct from the 30-day observation window. Using the same-day unrounded H100 feed yields approximately 25.1%, 33.6% and $2.54; the displayed 25.2% and 33.7% follow from the retained rounded snapshot inputs.
Editorial disclosure: Bittensor.zone is an independent publication. This article is not sponsored by Lium, Runpod or Vast.ai and does not constitute an investment recommendation.
