Skip to content
AI for businessFor building ownersHow it worksPricingResearchCompany
Let’s talk
AI for businessFor building ownersHow it worksPricingResearchCompanyLet’s talk
On this pageSources

Research11 June 2026

Distributed inference and the cost per token: reading the public numbers carefully

250 trillion tokens a month and what it does and does not say, the 52.8% and 76.1% NVIDIA figures, why placement beats parallelism, and what our own cost per token depends on.

7 min read
  • inference
  • economics
  • physics

Three public figures do a lot of work in the way we describe this market, and we would rather readers understand exactly what each measures than take any of them on trust. One is a count of tokens crossing a marketplace. One is a benchmark of cost per token on distributed hardware. One is an older count of tokens processed inside a single company. None of them is our number. All three are worth reading slowly.

The demand figure: 250 trillion a month

The largest independent routing marketplace publishes its own throughput, and that series is the cleanest public read on inference demand anyone has. It now runs at about 250 trillion tokens a month, which is three quadrillion a year, from about 0.4 trillion a month two years earlier. The growth rate implied is roughly five-fold every six months, or 625 times over two years.

What it says is narrow and strong. This is not a forecast and not a survey: it is one marketplace's own published throughput at its own published rate, and it counts tokens that were actually served by somebody, at a price somebody paid.

What it does not say: that the rate holds. Nothing compounds five-fold every six months forever, and a marketplace's share of a growing market moves for reasons that have nothing to do with the market. Token counts also inflate as models reason at length before answering and as context windows grow, so part of any multiple is tokens per answer rising rather than answers rising. We use the figure for one purpose: to establish that the demand clock runs in months while the supply clock, grid connections, runs in years.

The NVIDIA figure: 52.8% and 76.1%

On 17 March 2026 the NVIDIA Technical Blog published results from a benchmark run with Comcast on what they called an AI grid: the same small voice model served on four RTX PRO 6000 cards at distributed points of presence, compared against serving it from a centralised cluster. The reported reduction in cost per token was 52.8% at baseline load and 76.1% at burst.

What the figure says is narrow and useful. For a small model, on modest hardware, serving inference close to where requests arrive cost roughly half as much per token as serving it centrally under steady load, and roughly a quarter as much under peak load. The gap widens at burst because a central cluster has to be provisioned for the sum of everyone's peaks and sits partly idle the rest of the time; distributed nodes absorb their own local peaks, so less provisioned capacity goes unused. The arithmetic behind that is utilisation, not magic.

What the figure does not say matters just as much. It says nothing about large models. It says nothing about Indian power prices or about the hosting arrangements we use. It was run on a workstation-class card, not on the node we cost our mills on. And it describes one operator's network under one traffic pattern; another operator with flatter traffic would see a smaller burst gap.

We cite it because it is the clearest published statement, from the vendor with the least incentive to make it, that distributing inference lowers the cost per token rather than raising it. We do not cite it as our result. Our promise to buyers is sovereignty first, then a comparison: quality and total cost, compared on their own workload before they decide. We publish no number for any saving, and we do not quote this benchmark as one.

The older figure: 9.7 trillion to 480 trillion

At the Google I/O keynote on 20 May 2025, Sundar Pichai said that Google was processing over 480 trillion tokens a month across its products and APIs, up from 9.7 trillion a month a year earlier: a factor of roughly fifty in twelve months.

We keep the citation and no longer lead with it. It is one company's internal count, most of it not sold to anyone, and a year-on-year multiple from a single vendor is a weaker instrument than a marketplace's continuous published series. It remains useful as corroboration: two independent counts, on different bases, both say the same thing about the direction and the order of magnitude.

Why placement beats parallelism

The reason distribution works for inference and fails for training is bandwidth, and the numbers are not close.

Bandwidth by link, in the order a job would have to cross them
LinkBandwidthWhere
NVLink or Infinity Fabric600 to 1,800 GB/sInside one mill
InfiniBand NDRabout 50 GB/sInside one hall
City fibre at 1 Gbpsabout 0.125 GB/sBetween mills
Bandwidth by link, in the order a job would have to cross them

Inside a mill, the accelerators share a fabric that moves hundreds of gigabytes a second. Between two buildings on city fibre, the link moves an eighth of a gigabyte a second: four orders of magnitude less. Any job that needs its parts to talk to each other mid-step has to stay inside one box.

What crosses a building boundary, and what does not
WorkloadAcross buildings?
Batch inference, generation, rendering, embeddingsFine. Only inputs and outputs cross the link; a 1 Gbps link carries 7 to 9 TB a day
LoRA fine-tune, about 40 MB per stepMarginal. Viable with gradient accumulation
Full fine-tune of a 7B model, about 14 GB synced per stepNot viable. 112 seconds per step on the link alone
What crosses a building boundary, and what does not
The boundary, checked
14 GB per step ÷ 0.125 GB/s = 112 s of synchronisation per step
0.125 GB/s × 86,400 s ≈ 10.8 TB per day raw; 7 to 9 TB at realistic utilisation

So the product is scoped to what physics allows. Many independent jobs can be placed across the fleet. One tightly coupled job stays inside one machine, because no scheduler can fix the case where it would have to cross two buildings. Placement, not parallelism. The mill is the unit of contiguity: real fabric inside, city fibre between, and a boundary any technical buyer can verify.

That is also why the memory in one mill matters more than the count of machines. A mill holds 1,536 GB, enough for a 671-billion-parameter model, so the class of job that has to stay in one box mostly fits in one box.

What cost per token depends on in our model

Given that the job fits in the box, the cost per token from a mill is set by five things, in roughly this order.

Utilisation. A mill at 60% of sellable capacity, our modelled steady state, spreads its fixed costs over three-fifths of the hours; the same mill at 30% costs twice as much per token. Nothing else in the model moves the number as far. This is why the after-hours cooling question matters so much: a mill that cannot run overnight cannot reach 60%.

The design choice. The engine can choose the silicon for a room as well as the layout. What that choice does to the delivered cost per token is part of our cost model, it is not yet measured, and we print no figure for it. It only matters at fleet scale: choosing silicon per room is only rational across many rooms.

Power. We model a blended ₹9.0 per kWh with a band of ₹8 to ₹11 by state, reimbursed to the host at actuals at the host's own tariff. It is a real cost and it is never one flat number fleet-wide.

Silicon, measured as dollars per terabyte per second of memory bandwidth, because for inference the bandwidth is what serves tokens. A ten-kilowatt mill lands at $289,400, and the design is card-agnostic inside a frozen envelope, so the number tracks the market rather than a part.

The host floor, which is fixed per mill per year and is the reason a building says yes.

What the cost per token does not depend on is the GPU-hour list price. The IndiaAI compute portal lists H100 SXM capacity at ₹153 per GPU-hour on demand and ₹117 reserved, and any buyer will quote those numbers at us. They are the right price for shared, resident, allocation-gated capacity. They are not the price of a machine a buyer can name in a contract, in a building the buyer chose, that nobody else touches. What we offer is a comparison on delivered output: quality and total cost on the buyer's own workload, set against what the buyer runs today, not an hourly rate.

What we will publish

The published figures say that distributing inference lowers cost per token when the work fits in one box. They do not say by how much it will for our fleet, on our silicon, at our power prices. That number can only be measured on real work, installation by installation, and it is what a comparison of quality and total cost on a buyer's own workload rests on. Until it is, the honest position is the one above: a demand series that explains the urgency, a benchmark that points the right way, and a physics table that says where the line is.

Sources

  1. 01Published throughput on the largest independent model routing marketplaceOpenRouter, read from its public rankings; Azita register, Usage data through 2 September 2026250 trillion tokens a month now cross the largest independent routing marketplace. Three quadrillion a year. The rate is a five-fold increase every six months.
  2. 02Building the AI Grid with NVIDIA: Orchestrating Intelligence EverywhereNVIDIA Technical Blog, 17 March 202652.8% lower cost per token at baseline and 76.1% lower at burst, distributed grid versus centralised cluster
  3. 03Google I/O 2025 keynoteGoogle, delivered by Sundar Pichai, 20 May 20259.7 trillion tokens a month in May 2024 to over 480 trillion in May 2025: 50 times more
  4. 04Why one machine is the unit of contiguityAzita, derived from published link bandwidths, September 2026Inside a mill 600 to 1,800 GB/s (NVLink, Infinity Fabric); inside a hall about 50 GB/s (InfiniBand NDR); between buildings about 0.125 GB/s (1 Gbps city fibre)
  5. 05IndiaAI Compute Portal rate cardIndiaAI Mission, Ministry of Electronics and Information Technology, Portal listing as audited for the deck, September 2026H100 SXM at ₹153 per GPU-hour on demand and ₹117 reserved, with up to 40% subsidy subject to IndiaAI approval
  6. 06Utilisation, modelledAzita proof-of-concept model, September 202660% of sellable capacity contracted is the modelled steady state; operating band 45 to 65%; ceiling 80%
  7. 07Power reimbursed at actualsAzita, in the master hosting agreement, September 2026Every unit the machine draws is metered on a revenue-grade meter and reimbursed at the host's own tariff, as a separate line, never netted, never resold
  8. 08Landed capex per millAzita proof-of-concept cost model, founder-audited, September 2026$289,400 (₹2.52 Cr) per ten-kilowatt mill, fully loaded
  9. 09The design engine, as it runs todayAzita design engine (RackQuilt), founder-audited, September 2026340,200 rack configurations swept for every single room, in about fifteen seconds. Running today, assisted, and instrumented at every installation; never licensed.
  10. 10The mill: Pod-10/K, the volume SKUAzita unit model, September 2026A ten-kilowatt factory. 1,536 GB of memory in one mill, enough to hold a 671-billion-parameter model. Electrical intake 415 V three-phase, twenty amps a phase, off the board that already feeds the floor; no new connection. 14.0 to 14.4 kW at the meter, not ten: the room has to reject that heat. Water: zero, air-cooled, no structural work and no plant room.

Every number on this site is listed with its citation on the sources page. Figures marked modelled are outputs of our model, not measurements.

Earlier · 14 May 2026Certification-ready is not certified: the regimes that actually govern an in-building millLater · 9 July 2026The nine-month problem: where the time goes, and what a scan replaces

Let’s find the right fit.

AI for your business

Tell us what you want to run and where your data needs to stay.

Discuss your requirements

Put your building to work

Share a few details about your property. We’ll help you understand its potential and what comes next.

Assess your property

AI factories in buildings that already have power.

Platform

  • How it works
  • RackQuilt
  • Token Mills
  • AzExchange

For buyers

  • AI for business
  • Voice agents for BFSI floors
  • Private coding pods
  • Hospitals and diagnostics
  • VFX assist for post houses
  • Promo and product clips
  • Creators
  • Agencies
  • Engineering
  • Finance
  • Healthcare & BFSI

For buildings

  • Estates
  • For building owners

Tools

  • Pricing
  • Room illustrations

Proof

  • Trust & compliance
  • Sources
  • Research

Company

  • Company
  • Team
  • Careers
  • Press
  • Investors
  • Contact
AI enquirieshello@azitalabs.comProperty enquirieshello@azitalabs.com

© 2026 Azita Labs Private Limited

PrivacyTermsHosting terms

New Delhi, India