Data reviewed 12 September 2026
What running one yourself would cost
Your own GPUs bill every hour, busy or idle. An API bills only for what you use.
What the model does
Model
What they are doing
Adjust the details1,000 tokens in, 500 out, INT4, rupees
Currency
Cost of sovereignty
Nothing
At 200 concurrent users, running Kimi K3 yourself costs ₹27,97,711 a month less than paying per token.
Cost reaches zero at
146 concurrent users
A month, at your 200 concurrent users
A month, at your 200 concurrent users. Your own GPUs, 28 × H100, ₹73,99,280. Paying per token, ₹1,01,96,991. Self-hosting is ₹27,97,711 a month cheaper at this point.
Self-hosting here is GPUs only, a , and neither side prices jurisdiction, withdrawn access or where the data goes.
Config setup
28 × H100 80GB SXM
One copy of the model, spread across 28 GPUs that work on each request together. One copy is bigger than a single server of eight GPUs, so it has to run across more than one.
one copy
No official run for this card.
The official runs cover Kimi K3 on the only.
The best configuration we found
Cheapest here: 24 × RTX PRO 6000, at ₹208 per million tokens.
- 01₹31,88,640 a month₹208 per million tokens
- 02₹33,63,840 a month₹220 per million tokens
- 03₹34,33,920 a month₹224 per million tokens
- 11₹73,99,280 a month₹484 per million tokens
This compares cards, not models, and it prices GPUs and nothing else. Whether to hold the weights at all is a different question, and this page does not answer it.
How this was worked out
- 01
What one user sends
1,000 tokens in + 500 tokens out = 1,500 tokens a request
Roughly 1,125 words in total.
- 02
How long that holds a session open
2.536s to the first word + 500 ÷ 37.185 tokens a second = 16.0s
Read from the official measurement of a hosted endpoint. That endpoint's hardware is not published, so this figure does not move when you change the card. No official run covers this model on this card at this load. This sets how long a session stays open, which bounds how many requests one concurrent user can get through in a day. It does not set the traffic, which you give, and it never changes how much memory the model needs.
- 03
The work you asked for
200 concurrent users × 1,700 requests a day each = 3,40,000 requests a day
86,400 seconds ÷ 16.0s = 5,406 a day is the most one concurrent user could get through
The second line is a check, not an input. It is what this cluster could clear if nobody ever idled. Both halves of the comparison describe the same workload.
- 04
Tokens a month
3,40,000 × 1,500 tokens × 30 days = 15,300 million tokens
per user: 1,700 requests a day, 76.5 million tokens a month
- 05
What that costs on the API
₹286 in and ₹1,428 out per million, blended at 1,000:500 = ₹666 per million
15,300 million × ₹666 = ₹1,01,96,991 a month
Output is dearer than input, so the moves when your token mix does. Prices are published in dollars, converted at ₹95.21 to the dollar.
- 06
What has to sit in GPU memory
the model itself: 1454 GB at INT4, as released, plus 10% working room = 1599 GB per copy
each open conversation: 27.0 KB a token × 8,192 tokens on 24 growing layers = 0.21 GB
+ 0.42 GB of fixed state on 69 layers
= 0.63 GB for each open conversation
24 latent-attention layers grow with the conversation. 69 linear-attention layers keep a fixed state. loads once per copy. is what everything left over has to hold.
- 07
The cheapest cluster that can actually be built
28 × H100 gives 1807 GB usable once the fullest of 7 stages sets the limit, less 1599 GB for the model = 208 GB left
208 GB ÷ 0.63 GB a user = 330 concurrent users a copy
28 × H100, holding 330 concurrent users
A mixture of experts is split by handing each GPU a share of the experts, so any power of two up to 16 is a legal width. A split can reach across two servers, and a copy too big even for that runs as stages one after another. The fewest GPUs that hold the model and every open conversation in one copy win; a second copy is added only when no single copy can.
- 08
What speed those users were measured to get
The official runs cover Kimi K3 on the B300 288GB only, not on the H100 80GB SXM.
Nothing measured means nothing shown.
- 09
What the cluster costs
28 GPUs × ~₹362 an hour × 730 hours = ₹73,99,280 a month
₹73,99,280 ÷ 15,300 million tokens = ₹484 per million self-hosted
Approximate market rate, September 2026. Paid around the clock.
- 10
Where the two meet
the API costs ₹50,985 a month for each concurrent user
at 146 concurrent users it has caught up with the cluster serving them
The same two calculations, run at more than one demand. Self-hosting is a staircase because GPUs come in whole units, so the two meet where a step finally sits under the line.
Seeds reviewed 12 September 2026. Model prices captured between 7 August 2026 and 12 September 2026, the exchange rate 7 August 2026, hardware rates between 6 September 2026 and 11 September 2026. Ranked by the official intelligence index. Architecture and released sizes are read from each model's own repository and dated on the line they change. Where a model publishes no speed, 30 tokens a second and 0.5 s to the first word are assumed, and the line it changes says so.
Every figure was collected by hand from its official source and dated.
Measurements republished here belong to the people who ran them, and are used under their own terms: Artificial Analysis, InferenceX (SemiAnalysis). Every figure taken from them links the run it came from.