Rent, reserve or buy GPUs for voice AI in 2026: L4, L40S, H100 and B200 prices compared (US and India)

Why the same GPU has so many prices

Three kinds of seller rent data-centre GPUs, and they price differently. Hyperscalers such as AWS, Google Cloud and Azure charge the most on demand but discount heavily for commitments. GPU-focused clouds, often called neoclouds, such as Lambda, RunPod and CoreWeave, publish on-demand prices close to a hyperscaler's three-year rate, without the commitment. Indian clouds such as E2E Networks and Yotta sell capacity inside India, which matters for data residency.

The term you sign moves the price nearly as much as the seller. On-demand means pay by the hour with no commitment. A reservation, which AWS also sells as a Savings Plan and Google as a committed use discount, trades a one- or three-year promise for a lower rate: 25 to 37% off on demand for one year at AWS, Google and Azure, and 45 to 57% off for three. Spot or interruptible capacity is cheaper still, but the provider can take it back at short notice. Serverless platforms such as Modal bill per second of use.

The tables below were checked on 24 October 2026. Prices for 8-GPU instances are divided by eight to give a per-GPU figure. Rupee prices for US providers are a straight conversion at ₹95.04 per dollar, before taxes, and Indian providers' prices are shown as listed, which for E2E and JarvisLabs means before 18% GST.

Sonex Labs sells the per-use alternative to renting any of these cards, speech and LLM APIs priced per character, second and token, so it has an interest in how this comparison reads. Every price below is a third-party list price with its source.

Which card does which job in a voice stack

For voice, each card tends to do one job, and the job decides which price matters. The L4 (24 GB, 72 W) carries 48 low-latency speech-to-text streams in NVIDIA's Riva tests (NVIDIA Riva ASR performance). The L40S (48 GB) carries 32 text-to-speech streams for NVIDIA's own multilingual TTS model, with a 133 ms average first audio chunk.

Language models need more. Softcery models 40 to 75 concurrent voice sessions per H100. In NVIDIA's benchmark of an 8-billion-parameter LLM, first-token time on an L40S was 618 to 778 ms at 25 to 100 concurrent requests, against 268 to 615 ms on an H100, so the L40S is a budget LLM card at best. The B200 is the newest and most expensive card; Pipecat ran 44 agent processes on each one for its PhoneLLM model.

An hourly price only becomes a cost per call once you divide by the streams the card carries and by how busy your calls keep it, which for business-hours traffic is often under a quarter of the time. The self-hosting cost model does that division for every component, and how many concurrent calls a 24 GB GPU carries shows how to measure the stream count.

L4 prices: the cheapest card for speech to text

An L4 costs $0.44 to $0.81 an hour on demand, and a three-year AWS reservation brings it to $0.369. Because one card carries dozens of speech-to-text streams, the L4 is usually the cheapest self-hosted component in a voice stack. No sourced price for buying a single L4 server was found, so there is no owned row.

L40S prices: the usual card for text to speech

An L40S rents for $0.79 to $2.25 an hour, and owning one in a small server works out at about $0.74. Text to speech is the most expensive line in a voice API bill, so this is the card where self-hosting tends to pay off first.

The owned figure assumes a $6,199 card plus $3,800 for the rest of the server, $10,000 in all, drawing 0.6 kW at $300 per kW-month for colocation, with 10% a year for maintenance and no resale. That is $278 + $180 + $83 = $541 a month, or $0.74 per GPU-hour. It beats AWS's three-year price of $0.804 only if the card stays in service and busy for the full three years.

H100 prices in the US and India

H100 prices spread the widest: $1.29 an hour for Azure spot capacity, $2.69 to $4.29 on demand at Lambda and RunPod, and $11.06 on demand at Google. A committed hyperscaler price can beat a neocloud's on-demand one: AWS's three-year Savings Plan at $2.97 per GPU costs less than RunPod Secure at $3.99, but only if you can forecast three years of load. Even at that rate an LLM GPU has to be busy about 43% of the time to beat a cached mini-class API, as critical utilisation shows.

In India, E2E Networks lists ₹379.05 an hour on demand and says committed contracts start at ₹347.70, both excluding GST (E2E Networks H100). JarvisLabs, which added an Indian region in 2026, lists an H100 SXM at ₹330.48 an hour plus GST, billed per minute (JarvisLabs, updated 20 October 2026). Yotta's dedicated H100 at ₹1,96,113 a month works out at ₹269 an hour over 730 hours, though its page does not say which term that price is for.

B200 prices: the most throughput per card, at the highest hourly price

A B200 rents for $5.98 to $14.24 an hour, 1.4 to 2.1 times an H100 at the same provider. AWS does not publish reserved prices for its B200 instance. New 8×B200 servers cost about $400,000 to $500,000 according to third-party estimates, but no sourced power figure was found, so no owned cost is worked out here.

Rent, reserve or buy: what each choice costs

The cheapest hourly rate is not the cheapest choice unless the load is there to use it. Renting on demand costs the most per hour but nothing when idle; reserving costs less per hour but commits you to paying for hours you may not use; owning costs least per hour and commits you furthest. When to move from one to the next, as your call data firms up, is covered in the break-even-first deployment guide.

amortised $ per GPU-hour =
  [ (price - resale) / 36  +  kW x colo $ per kW-month  +  price x maintenance% / 12 ]
  / (GPUs x 730 x availability)

8x H100, no resale:   (7,917 + 3,000 + 2,375) / (8 x 730)   = $2.28
8x H100, $195k resale: (2,500 + 3,000 + 2,375) / (8 x 730)  = $1.35
1x L40S server:        (278 + 180 + 83) / (1 x 730)          = $0.74