Unit Economics

The economic stack

A single GPU-hour rental at Vast involves three parties earning money or paying costs:

1. The client pays the listed price (or wins an interruptible bid).
2. The provider receives the listed price minus Vast's take rate.
3. Vast keeps the take rate as gross revenue.

The provider also has costs (hardware, power, bandwidth, ops). The client also has alternatives (other clouds, marketplaces, or running their own GPU). Vast also has costs (platform engineering, dispute resolution, fraud, payment processing). Each layer's margin is what's left after their respective costs.

Vast's take rate

Public sources put Vast's take rate in the 15-25% range depending on tier and payment method. The company hasn't officially disclosed specific numbers; the range is consistent with marketplace industry norms.

The take rate is one of the strategic levers Vast can pull. Raising it boosts gross margin per transaction but risks pushing providers to competitors. Lowering it grows the supply corpus (more providers find Vast worth listing on) but reduces revenue per GMV. The equilibrium settles somewhere that keeps both sides reasonably happy.

Compare with related marketplaces:

• Airbnb: roughly 15% (host) + 6-12% (guest), so ~20-25% blended.
• Uber: roughly 20-30% (varies by market and time).
• eBay: ~10-13% (final value fees).

Vast's take rate is in the middle of this range. The pricing makes sense: high enough to fund platform operations and trust/safety systems, low enough that providers don't aggressively defect to competitors or set up direct relationships.

Provider unit economics

From the provider's perspective, the unit economics depend heavily on which archetype they are.

Small fleet operator with H100s

• Hourly listed price: $2.50/hour.
• Vast take (20%): $0.50.
• Provider gross: $2.00/hour.
• Power cost (700W at $0.08/kWh): $0.06.
• Bandwidth + colocation amortized: $0.10.
• Hardware amortization (~30k card, 3-year depreciation, 24/7): $1.14.
• Provider net per GPU-hour (at 100% utilization): ~$0.70.
• Provider net per GPU-hour (at 60% utilization, common): ~$0.40 actual revenue per hour available, minus same power+bandwidth cost.

That's positive but tight. The operator needs high utilization, good DLPerf to command premium prices, and disciplined ops to make the math work.

Hobbyist with a single RTX 4090

• Hourly listed price: $0.40/hour.
• Vast take (20%): $0.08.
• Provider gross: $0.32.
• Power cost (450W at residential $0.15/kWh): $0.07.
• Provider net per GPU-hour: $0.25.
• Hardware amortization: less relevant — they bought it for gaming and the rental is incremental.

For the hobbyist, the calculation is incremental: every rented hour is pure incremental income above what they'd otherwise earn. Even at low utilization, the rental pays back over a couple of years.

Why the math is structurally tight

The marketplace is competitive. Providers undercut each other to win bookings. Equilibrium prices settle at levels where the marginal provider just breaks even on costs (excluding hardware amortization). The infra-marginal provider (with lower costs or better hardware) earns the difference. This is standard commodity-market behavior.

The good news for providers is that as the GPU generation cycle turns and depreciation accelerates, the "marginal provider" rolls forward — old hardware exits the market or moves to deeper discount tiers — and prices for current-generation hardware can sustain better.

Client unit economics

From the client perspective, Vast is one option among several:

• Hyperscalers (AWS / Azure / GCP): 3-5x more expensive per GPU-hour at list. Reserved instances can narrow this.
• Enterprise neoclouds (CoreWeave / Crusoe / Lambda): 1.5-2.5x more expensive than Vast at on-demand rates. Reserved deals at scale narrow this further.
• Direct GPU purchase: If the client has steady, high-utilization need, buying GPUs becomes cheaper than renting at some break-even (typically 12-18 months of high utilization).
• Other marketplaces (RunPod / TensorDock): Within 0-30% of Vast pricing, depending on segment and time.

For clients whose workload tolerates Vast's variability, the math favors Vast strongly. For clients whose workload requires consistency, the implicit "cost of variability" — checkpoint overhead, retry logic, occasional lost runs — has to be amortized into the cost calculation, and at some point the alternatives become competitive.

Vast's own economics

What do Vast's economics look like as a business?

Revenue

Take rate × GMV. With GMV in the tens-of-millions to low-hundreds-of-millions range and take rate at 20%, gross revenue is in the eight-to-low-nine-figure range. The exact number isn't published, but the order of magnitude is consistent with the company's footprint.

Cost structure

• Engineering / platform. Probably the dominant cost. A marketplace at Vast's scale runs on a smart software stack; engineering investment to keep it humming.
• Payment processing. Credit card fees, crypto on-ramps, banking. A real expense at marketplace scale.
• Dispute and fraud handling. Some humans, mostly automation.
• Customer support. Light — the platform is mostly self-serve.
• Marketing / sales. Essentially zero by industry standards. The platform grows by word-of-mouth and organic search.
• Compliance and legal. Some baseline; not enterprise-cloud heavy.

Margin

The asset-light marketplace structure means Vast can be profitable at far lower revenue than capital-intensive competitors. Some industry observers estimate Vast is profitable; the absence of large fundraises tracks with this hypothesis. The exact margin profile depends on engineering headcount, which Vast keeps small.

Sustainability through cycles

How does Vast's business model hold up through the inevitable cycles?

GPU price cycle

When new generations launch, old generations drop in price. Vast's take rate on transaction value means total revenue compresses if average GPU-hour pricing falls. But volume often picks up at lower prices — more users can afford GPU compute — so GMV holds up better than per-unit pricing.

Supply abundance cycle

Crypto-mining transitions, GPU oversupply, and similar events flood Vast with cheap supply. Prices fall; Vast's per-transaction revenue falls; volume rises. Net: GMV can be flat or up.

Demand cycle

Open-source model releases, conference cycles, and similar demand surges raise prices and Vast's revenue. Demand droughts compress pricing.

Competitive cycle

Competitor marketplaces (RunPod, TensorDock) and competitor business models (managed inference, reserved deals at enterprise neoclouds) put pressure on Vast's take rate and value proposition. The marketplace defensibility is real but not absolute.

The composite picture is a business that's structurally less cyclical than capital-intensive competitors. Vast doesn't have $30 billion of leased GPUs that need to amortize regardless of demand — its costs scale roughly with revenue. Through hard cycles, that's a survivor's posture.

Contrast with traditional cloud economics

To anchor: a traditional cloud's GPU pricing reflects:

• Hardware cost (~10-20% of customer price).
• Datacenter cost, power, networking, software stack (~15-25%).
• SLA, support, account management, compliance (~15-20%).
• Gross margin (~40-60%).

A marketplace's GPU pricing reflects:

• Hardware cost + ops + bandwidth (~70-85%, captured by the provider).
• Marketplace operation, trust, payment processing (~15-25%, captured by Vast).
• Gross margin to the platform: 15-25%.

The marketplace structure is much thinner per transaction but with much lower fixed cost. At scale, both models can be profitable. The strategic question is which scales better given competitive and demand dynamics.

Takeaway

Vast's unit economics work because the platform doesn't have to fund datacenters or GPUs. Per-transaction margin is modest; total revenue is moderate; operating margin can be strong because cost-to-operate is low. The structure makes Vast resilient through cycles in ways capital-intensive competitors aren't, but also caps the upside.

The next chapter looks at customer segments — who actually populates the demand side and where the segment ceiling sits.