Revenue Per GPU-Hour: What Each AI Workload Actually Earns

Two operators each own a thousand H100s. One runs a bare-metal rental business. The other sells generated video, cloned voices and product images. Same silicon, same electricity bill, same depreciation schedule. The second one earns roughly six times more per GPU-hour.

That gap is not a secret, but it is rarely quantified in one place, because every AI workload is priced in a different unit. Language models sell tokens. Image models sell images. Video models sell seconds. Speech models sell minutes or characters. Renting raw compute sells the GPU-hour itself. Comparing a $0.30 per million tokens price to a $0.13 per second price tells you nothing until both are converted into the one unit that pays the mortgage: gross revenue per GPU-hour.

This article does that conversion for ten workload classes, using published 2026 pricing and measured single-GPU throughput. The result is a ranking that looks intuitive at the top and deeply counter-intuitive in the middle, where the most expensive-looking models turn out to be the least profitable things you can put on a card.

Why revenue per GPU-hour is the only comparable number

A rack does not care what it is running. It consumes the same power, occupies the same floor space, depreciates on the same schedule and draws the same interest expense whether it is doing batch embedding or rendering a four-second product video. The denominator is identical across every workload. Only the numerator moves.

Operators who compare on unit price instead of unit revenue make a predictable mistake. They look at a $0.57 per second video model and a $0.02 per million tokens embedding model, conclude that video is 28,000 times more valuable, and fill the hall with video capacity. Then they discover that a second of video consumes roughly a hundred seconds of GPU time, while a hundred million embedding tokens consume about an hour. The ratio collapses.

The conversion is mechanical. Three numbers per workload:

Most public comparisons stop at the first two terms. The third one decides the winner. A workload that produces $24 per GPU-hour at peak but only sells 25 percent of its capacity earns $6. A boring rental contract that produces $3.29 at peak and sells 95 percent of its capacity earns $3.13. The exciting business is barely ahead of the boring one, and it carries far more demand risk.

The cost floor every workload has to clear

Before comparing workloads to each other, establish what an hour of GPU actually costs. The decomposition below uses a fully burdened H100 SXM server at 70 percent fleet utilization, four-year straight-line amortization and a 30 percent residual value, which lands the hardware component at roughly $1.12 per GPU-hour. Power and facility add about $0.33 at Northern Virginia rates with a PUE near 1.12, because an H100 draws so little relative to the revenue it can generate.

Derived from a $28,000 street price per H100 with 1.4x server overhead, against a $3.29 median on-demand rate for September 2026. Figures are indicative, not audited.

Two floors matter. The direct floor is about $1.45 per GPU-hour - hardware plus power plus facility. Below that, you are destroying capital. The fully loaded floor is about $2.40 once network, storage, staff and customer acquisition are included. Any workload that cannot clear $2.40 with realistic utilization is a loss-making product line, whatever its headline price suggests.

The comparison, all ten workloads at once

The table below converts each workload into gross revenue per GPU-hour at published market prices and measured throughput, then expresses it as a multiple of a bare H100 rental. Throughput figures assume a single H100 80GB unless the model does not fit, in which case the per-card figure is shown after tensor parallelism.

Blended token prices assume a 4:1 input-to-output ratio. Throughput from vLLM and SGLang measurements on H100 SXM at high concurrency, 2026. Prices are public list rates, not negotiated contracts.

Read the table top to bottom and three things stand out. Speech and image are dramatically more productive per card than language models. Video, despite the highest per-unit price in the industry, sits near the bottom of the productive tier. And the largest, most expensive, most capable language model in the comparison earns less per GPU-hour than simply renting the card out to someone else.

Workload by workload: where the numbers come from

Speech synthesis: the quietest high-margin business

A modern open TTS model on one H100 produces roughly 7.2 million characters per hour at production quality, a real-time factor of 20 to 60 times depending on voice complexity and streaming settings. At the $15 per million characters list price charged by first-party APIs, that is $108 per GPU-hour. Even discounting heavily to an open-model market rate of $3 per million, the card still earns $22.

The reason is structural. TTS is not autoregressive over a long sequence of independent decisions; it is a compact acoustic model plus a vocoder, both of which batch extremely well. The compute per output character is small and predictable, so throughput scales almost linearly with batch size. Meanwhile the price has stayed high because voice quality is hard to match and buyers compare on naturalness, not on cost per character.

The constraint is demand, not capacity. Voice workloads are bursty, tied to publishing schedules and product launches. An operator with 200 TTS cards and no distribution channel will run them at 15 percent utilization and earn less than a rental contract.

Fine-tuning: highest revenue, lowest predictability

Fine-tuning a LoRA adapter on an 8B model with a one-million-token dataset for three epochs occupies one H100 for 40 to 90 minutes. At $25 per million training tokens - the rate major APIs charge - a single job bills between $75 and $225, implying $60 to $180 per GPU-hour.

This is the highest revenue per card in the entire comparison, and nobody should build a business on it alone. Training demand is project-shaped: a customer appears, consumes a burst of capacity, and disappears for a quarter. Utilization is measured in the low tens of percent across a fleet sized for peaks. It works as a premium tier sold alongside inference, not as a base load.

Image generation: the widest spread of any category

A SDXL-class model at 1024 by 1024 with 30 sampling steps renders in about 1.5 seconds on an H100, or roughly 2,400 images per hour. A distilled model like FLUX Schnell cuts that to under a second and pushes past 4,000 per hour.

The price range is brutal: $0.002 at the commodity floor, $0.04 at the brand-name tier, and over $0.10 for premium multi-megapixel output. Multiply 2,400 images by that range and the card earns anywhere from $4.80 to $96 per hour. The spread is entirely a positioning decision, not a technical one: the same weights serve both ends.

What makes image attractive is that it clears the cost floor even at commodity pricing, and it batches almost perfectly. What makes it dangerous is that image demand is the most elastic and most easily substituted in the stack. Buyers routinely regenerate twenty times to get one usable frame, and they migrate to whichever provider is cheapest that week.

Speech recognition: steady, boring, profitable

Whisper-class transcription runs at 50 to 100 times real time on an H100, so one card processes 3,600 minutes of audio per hour, or sixty hours of content. At $0.006 per minute that is $21.60 per GPU-hour, a 6.6x multiple over rental.

Recognition has the best demand profile in the entire list. Call centers, media archives, compliance recording and meeting transcription generate continuous, predictable volume with long contract cycles and low price sensitivity, because the alternative is human labor at a hundred times the cost. The workload is also unusually tolerant of batching and queuing, so operators can fill idle capacity with backlog instead of letting the card sit.

Video generation: the price is real, the volume is not

Video is where intuition fails hardest. A premium model charges $0.40 to $0.57 per generated second, more than fifty times the price of an image. But generation is slow: a five-second 480p clip from an open model takes four to eight minutes on an H100, and 1080p with audio takes longer. One card produces roughly 45 to 55 seconds of finished video per hour.

At $0.09 to $0.13 per second for open-model pricing, the card earns $4 to $7 per GPU-hour. Better than rental, far worse than speech or images. The premium end of the market belongs to closed models running on infrastructure nobody else can access, with prices set by scarcity rather than by compute cost.

Embeddings: high throughput, collapsing price

An embedding model is pure prefill with no autoregressive decode, so an H100 processes roughly 108 million tokens per hour. The problem is what those tokens are worth. Large embedding models list at $0.13 per million, small ones at $0.02, and open alternatives have pushed the effective market rate toward the bottom of that range.

At the top of the range the card earns $14 per hour; at the bottom, $2.20, which is below the fully loaded cost floor. Embedding is the clearest case in this article of a workload where technical efficiency does not translate into revenue, because the output is a commodity with almost no differentiation between providers.

Practical consequence: do not buy H100s for embeddings. An L40S or even a CPU cluster serves the same volume at a fraction of the capital cost, which is exactly why the per-GPU-hour metric has to be paired with capital efficiency before it means anything.

Language models: the scale paradox

Here is the part that surprises operators coming from other compute businesses. Bigger models charge more per token but earn less per card, and the effect is enormous.

An 8B model in FP8 or INT4 at high concurrency produces around 12,500 tokens per second on one H100, or 45 million tokens per hour. At a blended $0.07 per million for a 4:1 input-output mix, that is $3.15 per GPU-hour - roughly break-even against a rental contract, and only profitable if you own the hardware rather than renting it.

A 70B model in FP8 needs two cards and produces about 2,400 tokens per second aggregate, or 1,200 per card. That is 4.3 million tokens per card per hour. At a blended $0.50 per million, the card earns $2.16 per hour - below the fully loaded floor. You are paying for two cards to serve a model whose per-token premium does not cover the second card.

The 671B-class mixture-of-experts model is worse still. Eight H100s hold the weights in 4-bit, and aggregate throughput lands near 1,500 tokens per second, or about 190 per card. Blended at $0.52 per million, each card earns roughly $0.35 per hour. Serving a frontier open model at public API prices is, on these numbers, a way to convert GPU capital into operating losses.

Aggregate throughput at concurrency above 64, divided across cards in the tensor-parallel group. Assumes steady-state serving, no cold starts and no speculative decoding.

Two forces produce this. First, tensor parallelism is not free: eight cards cooperating on one forward pass spend a meaningful share of their time waiting on interconnect instead of computing, so aggregate throughput per card falls well below the single-card case. Second, the market price for frontier open models is set by the cheapest efficient server, not by the average one. If your implementation is 30 percent less efficient than the price setter, your margin is gone before you start.

The correction that reorders everything: billable ratio

Everything above is peak capacity revenue. Real fleets do not sell peak capacity. Apply realistic utilization and the ranking changes shape.

Utilization ranges reflect 2026 market conditions for operators without captive distribution. Contracted or vertically integrated demand runs higher.

After the correction, only speech and premium image generation clearly beat a rental contract by a wide margin. Video lands at or below rental. Small-model inference sits just above break-even. Large-model inference is under water in every column.

This is not an argument against selling inference. It is an argument that distribution, not silicon, is the scarce input. The operators earning 30x rental on speech are not running better kernels; they have a sales channel that keeps 200 cards busy with call-center transcription. The operators losing money on 671B models have excellent infrastructure and no pricing power.

A worked 1,000-GPU portfolio

Put the numbers together. A hypothetical operator with 1,000 H100s and four product lines, using conservative midpoints from the realized table:

Monthly figures assume 730 hours. Revenue is gross; power, facility, staff, network and customer acquisition are not deducted.

Against a fully loaded cost of $2.40 per GPU-hour, the portfolio returns about $1.15 million per month before financing, a 40 percent gross margin. Respectable, and roughly in line with what specialized GPU clouds report on a before-depreciation basis.

Now run the same 1,000 cards as a frontier-model inference service at $0.35 per GPU-hour realized. Monthly revenue is $256,000 against $1.75 million of cost. The fleet loses $1.5 million a month, and no amount of engineering improves the outcome, because the constraint is the market price per token, not the implementation.

Revenue quality, not just revenue quantity

A dollar of video revenue and a dollar of rental revenue are not the same dollar. Before committing capital, score each workload on four axes that the per-hour metric deliberately hides.

Contract duration and notice

Committed capacity, whether bare metal or reserved inference, is financeable. GPU-backed debt is priced against contracted revenue, and lenders discount spot revenue to near zero. A fleet with 70 percent committed capacity borrows at a materially lower rate than one with 30 percent, which changes the cost structure more than any optimization.

Demand variance

Speech recognition and embedding demand are smooth. Image demand follows marketing calendars. Video demand follows product launches and social platform trends. A workload with 3x seasonal variance requires 3x the capacity of its average load, which divides its realized revenue per card by three. Smooth demand is worth a lower price.

Hardware generation risk

Every workload is locked to a GPU generation for a different length of time. Embedding and speech models saturate memory bandwidth and run efficiently on older cards, so an H100 fleet stays competitive for years. Frontier language models are bound by HBM capacity and interconnect, and each new generation resets the efficiency baseline, which shortens the economic life of the asset serving them. A workload that pays back in fourteen months is safe; one that needs thirty-six is exposed to a generation change you cannot control.

Customer concentration

A single customer representing 40 percent of an inference book is not a revenue stream; it is a single point of failure with an invoice attached. Speech and embedding workloads tend to fragment across many buyers, which is a hidden strength the revenue table does not show.

What this means for the facility

The workload mix determines the building, not the other way round. Four consequences matter at the design stage.

Power density follows the model, not the card count. A hall serving 671B mixture-of-experts models runs eight-way NVLink domains, which means 40 to 80 kW per rack and liquid cooling at the rack or row level. A hall serving embeddings and speech can run 15 to 25 kW per rack on air, at a much lower cost per megawatt. Mixing both in one room forces the high-density design on the low-density load.

Storage and egress are first-order costs for media workloads. An image or video service writes terabytes per day and pushes the output over the network. A video product can spend more on egress than on the GPU it renders on. Transcription and embedding are the opposite: small inputs, small outputs, negligible bandwidth.

Batch-tolerant workloads decouple from grid peaks. Transcription backlog, embedding reindexing and fine-tuning jobs can be scheduled into low-price hours or curtailed during demand response events. Interactive video and voice cannot. If the facility has a constrained power envelope, the batch-tolerant share of the portfolio is what makes the envelope usable.

Retrofitting for density is the expensive mistake. Converting an air-cooled 20 kW rack position to a 60 kW liquid-cooled position means new CDUs, new manifolds, new floor loading and often a new electrical topology. Operators who start with a flexible high-density design pay for capability they may not use for a year; operators who start with air pay three times as much to convert later.

A checklist before committing capital

1. Convert every candidate workload to dollars per GPU-hour. Refuse to compare on unit price. A $0.57 second and a $0.006 minute are not the same unit.
2. Use measured throughput, not vendor peak. Run the model, the quantization and the serving engine you plan to deploy, at the concurrency your customers will actually produce.
3. Apply a realistic billable ratio. Peak capacity revenue is fiction. Model utilization by hour of day and day of week, not by annual average.
4. Check against both floors. $1.45 per GPU-hour is survival; $2.40 is sustainability. Anything below the second needs a strategic reason that survives a board review.
5. Score revenue quality separately. Contract duration, demand variance, generation risk and customer concentration determine the discount rate applied to that revenue.
6. Match the facility to the mix. High-density liquid capacity for frontier inference, standard density for speech and embedding, and enough egress for whatever the media share turns out to be.
7. Do not build a business on public API prices. Those prices are set by the most efficient operator in the market and by strategic subsidy. Your cost basis is your own.

The short version

The revenue ranking of AI workloads per GPU-hour, from most to least productive at realistic utilization, is roughly: contracted speech, premium image generation, fine-tuning project work, small-model inference, bare rental, video, embeddings, large-model inference, frontier mixture-of-experts inference.

That ordering is not intuitive. It puts a technology that charges six-tenths of a cent per audio minute at the top and one that charges fifty-seven cents per video second near the bottom. It says the cheapest-per-token models are the most profitable per card and the most expensive-per-token models are the least, because throughput and parallelism matter more than sticker price.

Most of all, it says that the scarce resource in an AI compute business is not GPUs. It is sold demand. A hall full of H100s with no distribution earns the rental rate minus whatever it costs to keep the lights on. A hall with a voice pipeline, a transcription contract and an image product earns several times that, on identical silicon.

References

1. Coloprice. "GPU Rental Prices: Live Tracker." Updated 25 September 2026. Median on-demand H100 $3.29/GPU-hr, B200 $6.01, GB200 NVL72 $13.25. coloprice.com/gpu
2. Silicon Analysts. "What You're Actually Paying For in a GPU-Hour: Hardware, Power, and the Scarcity Premium." 2026. Hardware amortization $1.12/hr, energy plus facility $0.33/hr at 70% utilization. siliconanalysts.com
3. Silicon Analysts. "LLM API Pricing: per Million Tokens by Model." 2026 price ledger. siliconanalysts.com/data/llm-pricing
4. Techsy. "vLLM vs SGLang 2026: We Benchmarked Both on H100s." Llama 3.3 70B FP8 at 2,400 tok/s aggregate at concurrency 100; Llama 3.1 8B at 12,500 tok/s on vLLM. techsy.io
5. Bruneaux, C. "Local vs Cloud: LLM Hardware and Inference Economics." August 2026. Llama 8B at 11,200 tok/s aggregate on one H100; 70B on four H100 at 3,400 tok/s; 671B MoE 4-bit on eight H100 at 620 tok/s. cc.bruniaux.com
6. CSuite. "What one AI image actually costs: cloud pricing decoded." August 2026. Per-second video rates for Veo, Kling, Seedance, Wan; per-image rates for FLUX, GPT Image, Seedream. csuite.so
7. CloudZero. "OpenAI API pricing in 2026: every model from Astra to Luna." `text-embedding-3-small` at $0.02/M tokens, `3-large` at $0.13/M, Whisper at $0.006/min, TTS at $15/M characters. cloudzero.com
8. Luminix. "Deep dive on the neocloud GPU-rental industry." 2026. Break-even utilization near 70%, depreciation 4-6 years, gross margins before depreciation 55-65%. useluminix.com
9. ECN Apps. "Private AI Infrastructure Costs 2026: Cloud GPU Rental vs On-Premise TCO." September 2026. 8x H100 SXM server at ~$335,000; 10.2 kW draw per 8-GPU chassis at 1.35 PUE. ecn-apps.com
10. Presenc AI. "AI GPU Supply and Pricing 2026." Q2 2026 on-demand H100 $1.80-3.50/hr, spot $1.20/hr; on-prem breakeven 6-14 months at moderate utilization. presenc.ai