Sources

Every number on this site, and where it comes from.

Each figure is a primary source, a derivation you can redo, a founder-audited constant, a modelled value, or a commitment. This page says which, and links to the text.

59figures
40primary sources
37linked to the text
9 September 2026last checked

Five kinds of number. Each labelled.

Primary sources

Published by someone other than Azita. Quoted as printed, with the date and a link where one exists.

Derived figures

Azita arithmetic on primary figures. The working is shown so you can redo it.

Audited constants

Engineering and cost constants, audited by the founder.

Modelled figures

Outputs of Azita's model. Not yet measured, and labelled so wherever they appear.

Stated commitments

Promises Azita makes in signed agreements. They are terms, not measurements.

What “modelled, not yet measured” means

A modelled figure is what Azita's model produces from audited inputs and stated assumptions. It has not been read off a meter at a running installation. Wherever such a figure appears on the site it carries that label, and the page that uses it says what would change it. Measured values replace modelled ones as they are recorded, and this page records the swap with a date.

What availability does not establish

RackQuilt, Token Mills and AzExchange are available today. That statement is not a deployment count, a benchmark or a certification, and no figure here should be read as one. Operating figures are modelled until they are measured, priced figures are commitments, and each carries its label. The site names no customer and shows no logo.

Index

3 entries are marked retired. The site has stopped arguing from them, and each one says what replaced it. They stay here so an old link still lands on an honest page rather than on nothing.

Audited constants

azita-design-engine
capex-289400
thermal-2625-cfm
hardware-lead-14-18
change-control-major
rackquilt-90-days(retired)

Modelled figures

azita-census
azita-unit-model
azita-plan-case(retired)
kicker-cap-15-18
utilisation-60

Stated commitments

floor-487
power-at-actuals
pilot-exit
availability-published

Primary sources

Published by someone other than Azita. Quoted as printed, with the date and a link where one exists.

01Primary

250 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.

Published throughput on the largest independent model routing marketplace

OpenRouter, read from its public rankings; Azita register Usage data through 2 September 2026

Open the source Figure on file

What we use it for: The demand clock. It is the reason a megawatt that can be reached this quarter is worth more than a megawatt that arrives in five years.
On the site: “Demand for inference is compounding five-fold every six months.”
/
“250 trillion tokens a month now cross the largest independent routing marketplace.”
/platform
Listed in the sources of “Distributed inference and the cost per token: reading the public numbers carefully”.
/research/cost-per-token-distributed-inference
The working: Five-fold every six months is four half-years in two years, so 5 x 5 x 5 x 5, which is 625 times over two years. That is not a forecast. It is one marketplace's own throughput extended at its own published rate, and the site always says so in the same sentence.
Read with: The rankings page publishes per-model usage and a running series rather than one headline monthly total, so the 250 trillion roll-up is held in the register as a dated extract with the series it was read from. It is one marketplace, not the whole market, and the site never presents it as the whole market. Independent trade reporting through mid-2026 put the same marketplace above 200 trillion tokens a month.

02Primary

The median project built in 2025 took 61 months from interconnection request to commercial operation. Over five years, and rising: the same median was 36 months for projects built in 2015.

Queued Up: 2026 Edition, Characteristics of Power Plants Seeking Transmission Interconnection

Lawrence Berkeley National Laboratory, Electricity Markets and Policy June 2026

Open the source

What we use it for: The far end of the supply calendar. It is the number a new grid connection is measured against, and the reason the site talks in weeks rather than in price.
On the site: “A new grid connection: 265 weeks.”
/platform
Listed in the sources of “The nine-month problem: where the time goes, and what a scan replaces”.
/research/nine-month-problem
Listed in the sources of “The power already exists: what sanctioned load means and how much of it is idle”.
/research/sanctioned-load-idle
The working: 61 months is about 265 weeks, which is the figure the supply-gap chart prints alongside a mill's 14 to 18.
Read with: LBNL's median is for generation seeking transmission interconnection in the United States. It is quoted as the clock the whole industry runs on, not as a claim about any one project or any one market.

03Primary

Medium-voltage switchgear 44 weeks. Large power transformer 128 weeks. A new substation more than 160 weeks. Generator step-up units 144 weeks.

Lead times for grid and data centre electrical equipment

Wood Mackenzie's quarterly supply chain survey, as reported by Power Magazine, with the Uptime Institute's supply chain commentary 2025 and 2026 survey rounds, read September 2026

Open the source Figure on file

What we use it for: The middle of the supply calendar: the equipment a megawatt has to queue behind before anyone can draw it. Money buys none of it.
On the site: “Medium-voltage switchgear, 44 weeks. A large power transformer, 128 weeks. A new substation, more than 160.”
/platform
Listed in the sources of “The nine-month problem: where the time goes, and what a scan replaces”.
/research/nine-month-problem
Read with: The registry key is historic. The 44 and 128 week figures are Wood Mackenzie's survey averages as reported in the trade press; the substation figure is the upper end of the published range and is held in the register with its round. The Uptime Institute's own supply chain notes sit behind its member wall, which is why the linked document is Wood Mackenzie's.

04Primary

A full rack draws up to 142 kilowatts, distributed internally over eight power shelves and a low-voltage busbar.

NVL72 AI factory enterprise reference architecture

NVIDIA Current edition, read September 2026

Open the source

What we use it for: The mismatch the whole company exists inside: the newest racks draw 142 kilowatts into rooms the world actually built for eleven.
On the site: “The newest racks draw 142 kilowatts into rooms the world actually built for eleven.”
/platform
Read with: 142 kilowatts is NVIDIA's figure for a full rack at load; the published band is 120 to 142. The eleven kilowatts on the other side of the sentence is the conventional per-rack design load of ordinary enterprise and commercial rooms, and it is Azita's register figure, not NVIDIA's.

05Primary

Computing capacity has been growing 3.3 times a year since 2022, with a 90% confidence interval of 2.7 to 4.1 times. A doubling time of seven months.

Global AI computing capacity is doubling every seven months

Epoch AI Data insight, read September 2026

Open the source

What we use it for: The first of the four steps on the constraint: silicon is not the binding limit, because silicon is the part that compounds.
On the site: “Ten times more silicon by 2028.”
/platform
The working: 3.3 times a year compounded over two years is 3.3 x 3.3, which is about 10.9. That is where the site's 'ten times more silicon by 2028' comes from, and it is arithmetic on Epoch's published rate, not a separate forecast. At the bottom of Epoch's interval, 2.7 x 2.7, it is about 7 times; at the top, about 17.
Read with: Epoch measures quarterly AI chip sales converted to a common unit across the major designers. The rate is theirs; the two-year extension is ours and is shown above.

06Primary

233 GW of large load asked for a connection. 9 GW approved. 3.9 GW actually drawn.

Large load interconnection queue

ERCOT, in its large load reporting and its update to the Texas Senate Committee on Business and Commerce Queue position at the start of 2026; update of April 2026

Citation on file

What we use it for: The gap between what has been asked for and what is actually flowing, in the one market that publishes both. Demand for power is not the same thing as delivered power.
On the site: “233 GW asked. 9 GW approved. 3.9 GW actually drawn.”
/platform
Listed in the sources of “The power already exists: what sanctioned load means and how much of it is idle”.
/research/sanctioned-load-idle
Read with: ERCOT's own domain refuses requests from our build environment, so the filings are held in the register as dated extracts rather than linked. The 233 GW queue figure is corroborated by trade reporting, which put the queue above 233 GW at the start of 2026, up 269% on the year. The approved and drawn figures are read from ERCOT's large load reporting and carry their extract date in the register.

07Primary

75% of registered voters oppose a data centre near their home, 61% of them strongly. The same survey recorded 42% opposition in September 2025.

Americans now overwhelmingly oppose new data centres near their homes

Heatmap Pro, polling conducted by Embold Research Fielded 8 to 13 August 2026; the earlier wave, September 2025

Open the source

What we use it for: The fourth step on the constraint: permission. It is the one input that is getting harder rather than easier, and no amount of capital moves it.
On the site: “42% to 75% opposed, in twelve months.”
/platform
Read with: United States polling of registered voters. It is quoted as evidence that the permission to build is tightening, not as a statement about any other market. A mill needs no new permission because it enters a room that is already permitted and already energised.

08Primary

At least 75 projects worth approximately $130 billion blocked or delayed in a single quarter, the largest single-quarter concentration on record. Active opposition groups rose from 396 to 833 in three months.

Q1 2026 report

Data Center Watch Published 2026, covering January to March 2026

Open the source

What we use it for: What the permission problem costs in a single quarter, counted in projects and in dollars.
On the site: “$130bn of projects blocked in a single quarter.”
/platform
Read with: Blocked or delayed, which is the report's own pairing; the site keeps both words. One quarter of one market, and it matched almost all of the previous year in three months.

09Primary

10,500 local exchanges in the estate. A programme of just over 100 of them converted to edge data centres. 0.95% of its own estate.

FiberCop transforms exchanges to create a national network of over 100 edge data centres

FiberCop, in its own release, with the estate counts from its network disclosures August 2026

Open the source

What we use it for: The proof that estate owners will not do this themselves. One operator owns 10,500 buildings and has converted a hundred, and it already had the power, the fibre and the balance sheet.
On the site: “One operator owns 10,500 buildings. It has converted a hundred.”
/estates
“0.95% of its own estate, and it already had the power, the fibre and the balance sheet.”
/estates
The working: 100 divided by 10,500 is 0.95%.
Read with: The hundred is the programme FiberCop has announced and started: the first site is live in Rome, with Turin, Genoa, Bologna, Naples and Palermo following. The site says 'converted a hundred' because that is the programme's own count, and it states here that the programme is under way rather than finished. FiberCop is named as an example of the shape, and it is a counterparty rather than a rival.

10Primary

52.8% lower cost per token at baseline and 76.1% lower at burst, distributed grid versus centralised cluster

Building the AI Grid with NVIDIA: Orchestrating Intelligence Everywhere

NVIDIA Technical Blog 17 March 2026

Open the source

What we use it for: Published evidence that inference placed close to the user is cheaper per token, not only faster.
On the site: “The reported reduction in cost per token was 52.8% at baseline load and 76.1% at burst.”
/research/cost-per-token-distributed-inference
Read with: NVIDIA's benchmark with Comcast: the same small voice model on RTX PRO 6000 GPUs, a four-node grid against one centralised cluster. It is their result on their workload. We cite it; we do not claim it as ours.

11Primary

9.7 trillion tokens a month in May 2024 to over 480 trillion in May 2025: 50 times more

Google I/O 2025 keynote

Google, delivered by Sundar Pichai 20 May 2025

Open the source

What we use it for: A second, independent reading of the demand clock, from one vendor's own products rather than a marketplace.
On the site: “Google was processing over 480 trillion tokens a month across its products and APIs, up from 9.7 trillion a month a year earlier.”
/research/cost-per-token-distributed-inference
Read with: Demoted in September 2026. The site's demand figure is now the routing marketplace's published throughput (openrouter-2026); this stays as corroboration and as the longer series. The figure is Google's own products and APIs, not the whole market.

12Primary

One year: the standard of performance for a new connection where a new sub-station must be commissioned. Three months where distribution mains must be extended or augmented.

MERC (Electricity Supply Code and Standards of Performance of Distribution Licensees including Power Quality) Regulations, 2021

Maharashtra Electricity Regulatory Commission Notified 2021

Open the source

What we use it for: A local reading of the supply clock: how long a new connection in Mumbai takes when the network has to be built out to reach it.
On the site: “A sanctioned connection that already exists has no queue.”
/owners
“The state regulator's own standard of performance allows a year before a connection that needs a new sub-station is made.”
/research/sanctioned-load-idle
Read with: A 2024 amendment consultation proposed shortening these timelines to 90 days. The notified 2021 timelines are what we cite. The point holds either way: load that is already sanctioned waits for nothing. Mumbai is a worked example here, not the company's base.

13Primary

The average wait for a grid connection in primary data centre markets exceeds four years; about ten years in the most constrained markets

2026 Global Data Center Market Outlook

JLL, with Statista (2026) for the upper bound of the band 5 January 2026

Open the source

What we use it for: The supply clock globally: why new capacity arrives in multi-year steps.
On the site: “Grid connections wait 4 to 10 years in major markets.”
/research/sanctioned-load-idle
Read with: The four-year average is JLL's own sentence and was confirmed on the page. The ten-year upper bound comes from Statista's 2026 data centre series and is held on file; the site quotes the band as 4 to 10 years.

14Primary

Mumbai office stock 15.74 mn sq m (169.4 mn sq ft) in H1 2025, vacancy 17.4%. Bengaluru 229.5 mn sq ft; Pune 105.7 mn sq ft.

India Real Estate: Office and Residential Market, H1 2025

Knight Frank Research July 2025

Open the source

What we use it for: The office floor area behind the Mumbai worked example of stranded load.
On the site: “Mumbai's 169 million square feet of office stock is the base of the estimate.”
/research/sanctioned-load-idle
Read with: Confirmed from the report's Mumbai office market table. The three-city total (Mumbai, Bengaluru, Pune) is 504.6 mn sq ft.

15Primary

Retired

An additional 26.3 GW of load from AI data centres projected by 2031-32, to be integrated into the grid

Written reply in Parliament on power demand from AI data centres

Ministry of Power, Government of India (Minister of State Shripad Naik) July 2026, reported 29 July 2026

Open the source

What we use it for: The scale of AI load India expects to connect this decade.
On the site: No page on this site quotes this figure any more.
Retired, and what replaced it: Nothing on the site cites this any more. The supply-gap argument is now made with lbnl-queued-2026, uptime-lead-times, ercot-queue and epoch-ai-2026. The entry stays so old links land somewhere honest.
Read with: This is India-wide projected AI data-centre load, not the sanctioned commercial load of any one city or state, and the site never presented it as the latter. The link is a press report of the reply; the Parliament record is on file.

16Primary

H100 SXM at ₹153 per GPU-hour on demand and ₹117 reserved, with up to 40% subsidy subject to IndiaAI approval

IndiaAI Compute Portal rate card

IndiaAI Mission, Ministry of Electronics and Information Technology Portal listing as audited for the deck, September 2026

Open the source Figure on file

What we use it for: The subsidised price floor for accelerator hours in India. Azita never sells on that unit; it appears so a reader can see the floor we are not competing on.
On the site: “The IndiaAI compute portal lists H100 SXM capacity at ₹153 per GPU-hour on demand and ₹117 reserved.”
/research/cost-per-token-distributed-inference
Read with: The portal exists and was opened. The rate card itself is behind the portal's application flow, so the figures are held as a dated screenshot in the data room rather than linked.

17Primary

Work by a licensed electrical contractor; inspection by a Chartered Electrical Safety Engineer; a test report to the Electrical Inspector before energisation

Central Electricity Authority (Measures relating to Safety and Electric Supply) Regulations, 2023

Central Electricity Authority, Ministry of Power In force 8 June 2023

Open the source

What we use it for: The electrical route to energising a mill on a building's existing connection: who may do the work, who inspects, what is filed.
On the site: “Installed under the CEA (Safety) Regulations 2023 by a licensed electrical contractor, with a Chartered Electrical Safety Engineer engaged per site and the test report filed with the Electrical Inspector before energisation.”
/trust
“Installed only by a state-licensed electrical contractor under the CEA Safety Regulations 2023, with the test report filed with the Electrical Inspector before energisation.”
/owners
“The CEA 2023 route: a licensed contractor, a Chartered Electrical Safety Engineer per site, the test report before energisation.”
/research/certification-ready-not-certified
Listed in the sources of “The nine-month problem: where the time goes, and what a scan replaces”.
/research/nine-month-problem
Read with: Azita's compliance memo maps the route to Regulations 31, 33 and 45 of the 2023 text. The regulation numbers are the memo's reading and sit on file; the regulation itself is linked.

18Primary

Part 4 governs fire and life safety in occupied buildings. The mill enters an existing building as equipment under the building's current fire NOC.

National Building Code of India (SP 7), Part 4: Fire and Life Safety

Bureau of Indian Standards Current edition, as notified by BIS

Open the source

What we use it for: The fire regime the mill is designed into, and the reason the building's existing NOC continues to govern.
On the site: “Your existing fire NOC continues to govern. The mill enters as equipment under Part 4 of the National Building Code, on your existing NOC file.”
/trust
“The design reference behind all of this is Part 4 of the National Building Code of India, the fire and life safety part, in its current edition as notified by BIS.”
/research/certification-ready-not-certified
Read with: The registry key is historic. The 2016 edition has been withdrawn and superseded, so the site cites the current edition of SP 7 as notified by BIS and never writes a year against it.

19Primary

The wiring code the mill's internal wiring and its tie-in to the building are specified to

IS 732:2019, Code of Practice for Electrical Wiring Installations (fourth revision)

Bureau of Indian Standards 7 May 2019

Open the source

What we use it for: The wiring specification in the type file and the per-site test results.
On the site: “Wiring to IS 732, detection to IS 2189, in the type file and tested per site.”
/trust
“Wiring to IS 732.”
/research/certification-ready-not-certified

20Primary

The code the mill's smoke detection and its alarm interface to the building are selected and installed to

IS 2189:2008, Selection, Installation and Maintenance of Automatic Fire Detection and Alarm System, Code of Practice (fourth revision)

Bureau of Indian Standards 2008, reaffirmed 2018

Open the source

What we use it for: The detection layout inside the enclosure and its interface to the building's fire panel.
On the site: “Wiring to IS 732, detection to IS 2189, in the type file and tested per site.”
/trust
“Detection to IS 2189.”
/research/certification-ready-not-certified

21Primary

Legacy safety standards under CRS (IS 13252 Part 1 and IS 616) run alongside IS/IEC 62368-1:2023 until 1 November 2028, then are withdrawn

Compulsory Registration Scheme for Electronics and IT Goods

Bureau of Indian Standards, under MeitY's Electronics and Information Technology Goods (Requirement for Compulsory Registration) Order, 2021, which replaced the 2012 Order Order S.O. 1248(E) of 18 March 2021; migration guidelines of 9 March 2026

Open the source

What we use it for: Which components in the mill must carry a BIS registration mark, and the date after which a legacy registration no longer counts.
On the site: “Registration of the electronics and IT goods inside the mill. The legacy safety standards are being withdrawn; the date is on the sources page.”
/trust
“BIS CRS registrations, with the legacy standards withdrawn on 1 November 2028.”
/research/certification-ready-not-certified
Read with: The sunset date is from BIS's migration guidelines of 9 March 2026, issued after the MeitY notification of 29 October 2025; the CRS page carries the scheme, not the date. Registrations granted under the 2012 Order carried over and renew under the 2021 Order.

22Primary

Producers and bulk consumers of IT equipment register on the CPCB EPR portal and route end-of-life equipment only to registered recyclers

E-Waste (Management) Rules, 2022

Ministry of Environment, Forest and Climate Change; administered by the Central Pollution Control Board through the E-Waste EPR portal In force 1 April 2023

Open the source

What we use it for: How hardware leaves service: registration, recycler certificates, nothing to a scrap dealer.
On the site: “At end of life, hardware goes only to a CPCB-registered recycler under the E-Waste EPR rules, with the recycler's certificate on file.”
/trust
“End of life under the E-Waste EPR rules, through a registered recycler.”
/research/certification-ready-not-certified
Read with: The link is the CPCB's own FAQ on the rules, which was confirmed. The EPR portal at eprewaste.cpcb.gov.in could not be opened from the build environment and is not linked for that reason only.

23Primary

Class A3 allowable inlet temperature 5 to 40 °C

Thermal Guidelines for Data Processing Environments, equipment class A3

ASHRAE Technical Committee 9.9, Mission Critical Facilities, Data Centers, Technology Spaces and Electronic Equipment Fifth edition, 2021; guidance now maintained in the TC 9.9 Datacom Encyclopedia

Open the source Figure on file

What we use it for: The design point for the mill face: air-only cooling sized to the warmest class ordinary IT hardware is rated for.
On the site: “We design to A3 at the face.”
/research/air-only-mills
“Designed to 40 °C at the face (ASHRAE A3).”
/platform/token-mills
Read with: The committee page was opened; the A3 envelope is from the 2021 reference card, held on file. Designing to the allowable limit rather than the recommended one is a choice, and the research post says why.

24Primary

In 2026 spending on inference inside AI-optimised infrastructure as a service, $23.3bn, overtakes spending on training, $19.0bn, for the first time. 55% of that spending supports inference in 2026, rising to 59% in 2027.

Gartner forecasts worldwide AI-optimised IaaS spending to grow 96% in 2026

Gartner, press release 10 August 2026

Citation on file

What we use it for: The first year enterprises spend more running models than building them: the demand-side reason the site talks about rooms rather than campuses.
On the site: No page on this site quotes this figure.
Read with: Gartner's newsroom refuses requests from our build environment, so the release of 10 August 2026 is held in the register as a dated extract rather than linked; the figures were corroborated in trade reporting of 11 August 2026. It is a forecast of one spending category, AI-optimised IaaS, not of all AI spend, and the site says so wherever it quotes it.

25Primary

The modal rack density operators report is 11 kW in 2026, up from 9 kW in 2025. The large majority of racks still run under 30 kW; a growing minority of operators report peak densities of 30 kW or higher.

Global Data Center Survey 2026: rack density

Uptime Institute Intelligence Survey published July 2026; press release 28 July 2026

Open the source Figure on file

What we use it for: The other half of the mismatch: the rooms the world actually built take nine to eleven kilowatts a rack, and the newest AI racks cannot be placed in them at all.
On the site: No page on this site quotes this figure.
Read with: The press release was opened and is linked; it states the direction. The modal figures sit in the survey report behind Uptime's registration wall and are held as a dated extract. The 'nine to eleven' on the site is the 2025 and 2026 modal densities together.

26Primary

One in five respondents says their organisation is limiting its AI use because of operating costs, tokens included; the share is broadly consistent across organisation sizes and industries. McKinsey calls cost 'not yet a widespread constraint'.

The State of AI: Global Survey 2026

McKinsey, QuantumBlack August 2026; fielded 4 May to 8 June 2026, n=1,719 across 97 countries

Citation on file

What we use it for: The part we will not overstate: cost throttles a fifth of organisations today, not most of them, so Azita sells where inference sits in cost of goods rather than to everyone.
On the site: “McKinsey's August 2026 survey of 1,719 organisations found one in five already naming operating cost as a constraint on further AI use.”
/research/why-an-engineering-team-buys-a-room
Read with: McKinsey's page timed out from our build environment, so the survey is held in the register as a dated extract rather than linked. It is a survey of respondents, not a measurement of spend, and the site quotes it with its own caveat attached.

27Primary

About 9.4 million tokens a year per subscriber for a basic chatbot, and as much as 356 million tokens a year per user for super agents: roughly thirty-eight times the tokens from the same user, with no new sign-ups.

AI token economics for CFOs

Deloitte US 22 April 2026

Open the source

What we use it for: Why the demand clock keeps ticking without a single new customer: the work per user multiplies as agents replace chatbots.
On the site: No page on this site quotes this figure.
The working: 356 million divided by 9.4 million is about 37.9, which the site rounds to thirty-eight.
Read with: Deloitte's figures are illustrative usage tiers from its own token-economics guide, confirmed on the page on 9 September 2026; the multiple is Azita's arithmetic on them.

28Primary

You pay only when the software achieves specific, valuable outcomes; unresolved conversations in most cases carry no charge.

Outcome-based pricing for AI agents

Sierra 10 December 2024

Open the source

What we use it for: The US pattern for contact-centre AI: priced per resolution, not per seat or per token.
On the site: “In December 2024 Sierra set out outcome-based pricing for its agents: the customer pays when a conversation is resolved and, in most cases, nothing when it is not.”
/research/voice-floor-first-room
Read with: Read on the page on 9 September 2026.

29Primary

1.2 tokens per Agentic Virtual Agent interaction; 50 AI summaries per token.

Genesys Cloud tokens model

Genesys Cloud Resource Center 12 July 2026

Open the source Figure on file

What we use it for: Per-interaction pricing of contact-centre AI in the US.
On the site: “An automated call summary costs one fiftieth of a token, and an agentic interaction consumes 1.2 tokens.”
/research/voice-floor-first-room
Read with: The US list price per token is on the Genesys pricing hub, not on this page; held on file.

30Primary

Cascaded voice stacks cost $0.007 to $0.091 a minute; in the cheapest stack the language model is under 3 per cent of the minute, speech recognition and speech synthesis the rest.

Voice Agent Cost Per Minute 2026: a worked cost model

Inworld AI 8 July 2026

Open the source

What we use it for: Why the language model is a small share of a voice minute, and why the pod must host speech as well as language.
On the site: “The language model is under 3% of the minute; transcription is about 24% and synthesis about 73%.”
/research/voice-floor-first-room
Read with: Figures read on the page on 9 September 2026.

31Primary

Regulated entities and their recovery agents shall not call a borrower or guarantor before 8:00 a.m. or after 7:00 p.m.

Outsourcing of Financial Services: responsibilities of regulated entities employing recovery agents

Reserve Bank of India, RBI/2022-23/108 12 August 2022

Open the source

What we use it for: The calling window that shapes a collections floor's day shift.
On the site: “Since August 2022 a lender and its recovery agents may not call a borrower before 8 a.m. or after 7 p.m.”
/research/voice-floor-first-room
Read with: Circular read on 9 September 2026.

32Primary

Recovery calls must be recorded, the borrower informed, and recordings preserved for six months or until related litigation concludes; contact only between 08:00 and 19:00 unless the borrower authorises otherwise.

Responsible Business Conduct amendment directions, 2026 (recovery calls)

Reserve Bank of India, RBI/2026-2027/223 6 August 2026, effective 1 January 2027

Open the source Figure on file

What we use it for: Why every recovery call is recorded and retained, and why the bank must be able to name where the recording sits.
On the site: “From 1 January 2027 every recovery call must be recorded, the borrower told, and the recording kept for six months or until any litigation about it ends.”
/research/voice-floor-first-room
Read with: Confirmed through two secondary summaries; the RBI page itself to be opened once by the founder.

33Primary

Scope includes application development, data centre and cloud services; data storage only in India as per extant requirements; the regulated entity's right to audit the provider and its sub-contractors, with unrestricted access to data and relevant business premises.

Reserve Bank of India (Outsourcing of Information Technology Services) Directions, 2023

Reserve Bank of India, RBI/2023-24/102 10 April 2023, effective 1 October 2023

Open the source

What we use it for: Why a bank's vendor must be able to name the room and let the bank audit it.
On the site: “An agreement under them must provide for storage of data only in India as applicable, give the regulated entity the right to audit the provider and its sub-contractors, and ensure the provider grants unrestricted access to the data and the relevant business premises.”
/research/voice-floor-first-room
“Their scope names application development, maintenance and testing.”
/research/why-an-engineering-team-buys-a-room
Read with: Directions read on 9 September 2026.

34Primary

The entire data relating to payment systems is to be stored in a system only in India.

Storage of Payment System Data

Reserve Bank of India, DPSS.CO.OD.No 2785/06.08.005/2017-18 6 April 2018

Open the source Figure on file

What we use it for: The data-localisation floor for payments in India.
On the site: “Since the April 2018 circular on storage of payment-system data, payment data has to sit in a system in India.”
/research/voice-floor-first-room
Read with: Notification id held on file; to be opened once by the founder.

35Primary

Personal data is any data about an individual identifiable by or in relation to it (s.2(t)); transfer outside India may be restricted to countries the Central Government notifies (s.16).

Digital Personal Data Protection Act, 2023 (No. 22 of 2023)

Ministry of Law and Justice, Government of India 11 August 2023

Open the source

What we use it for: India's personal-data law as it bears on where inference runs.
On the site: “A voice recording is data about an identifiable individual, so it is personal data under section 2(t).”
/research/voice-floor-first-room
“Section 16 restricts transfers only to countries the Centre notifies.”
/research/why-an-engineering-team-buys-a-room
Read with: Act text read on 9 September 2026.

36Primary

Credits are consumed on input, output and cached tokens at published rates; Business at $19 per user per month includes $19 of credits; fallback experiences are no longer available; administrators set budgets.

GitHub Copilot is moving to usage-based billing

GitHub Blog 27 April 2026, effective 1 June 2026

Open the source

What we use it for: The US shift from flat seats to usage billing for coding assistants.
On the site: “On 27 April 2026 GitHub announced that from 1 June every Copilot plan would bill on usage.”
/research/why-an-engineering-team-buys-a-room
Read with: Read on the page on 9 September 2026.

37Primary

Gross margin of minus 23 per cent in the quarter ended January 2026 while approaching $2 billion in annualised revenue.

Cursor AI pricing in 2026 (reporting The Information)

CloudZero 18 May 2026, updated 4 September 2026

Open the source

What we use it for: Why coding-assistant vendors are pushing inference cost back onto the buyer.
On the site: “The Information reported that Cursor's gross margin was negative 23% in the quarter ended January 2026.”
/research/why-an-engineering-team-buys-a-room
Read with: Secondary report of The Information's figures; read on 9 September 2026.

38Primary

About 5,000 engineers; the year's budget spent in four months; heavy users at $500 to $2,000 a month.

Uber burns its 2026 AI budget in four months on Claude Code

Forbes, Janakiram MSV 17 May 2026

Open the source Figure on file

What we use it for: What a per-engineer inference bill looks like once it is visible.
On the site: “Forbes reported in May that Uber, having put Claude Code in front of about 5,000 engineers, had spent its 2026 AI budget in four months.”
/research/why-an-engineering-team-buys-a-room
Read with: Forbes page not reachable by our fetcher; figures confirmed from four reports quoting it.

39Primary

About 5,000 engineers from December 2025 at $500 to $2,000 per engineer per month; licences end 30 June 2026; the cost was invisible under flat seats and visible under usage billing.

Microsoft cancels Claude Code after token costs exceed budget

Enterprise DNA, citing an internal memo first reported by The Verge 31 May 2026

Open the source Figure on file

What we use it for: The same lesson from a second large engineering organisation.
On the site: “Under flat seat licensing the token cost had been invisible, and under usage billing it became visible at once.”
/research/why-an-engineering-team-buys-a-room
Read with: Secondary report; the memo itself is not public.

40Primary

Code Assistant at $39 per user per month and Agentic Platform at $59, annual; offered as SaaS, in a VPC, on premises or fully air-gapped, on customer-provided or vendor-provided compute.

Tabnine pricing

Tabnine read July 2026

Open the source

What we use it for: The self-hosted answer already being sold in the US.
On the site: “SaaS, VPC, on-premises or fully air-gapped, you choose where your code lives, and the compute is customer-provided unless you pay the vendor to host it.”
/research/why-an-engineering-team-buys-a-room
Read with: Read on the page on 9 September 2026.

Derived figures

Azita arithmetic on primary figures. The working is shown so you can redo it.

41Derived

4.5 GB for an eight-billion-parameter model, quantised. 1,536 GB in one mill.

What an eight-billion-parameter model weighs

Azita analysis, derived; retrieval economics, Azita analysis September 2026

Citation on file

What we use it for: Why the room is the right unit. The models are getting smaller, and a hospital's entire knowledge stack, index and models, fits inside one room.
On the site: “An eight-billion-parameter model, quantised, is 4.5 GB against 1,536 GB of memory in one mill.”
/trust
“4.5 GB for an eight-billion-parameter model, quantised.”
/platform/azexchange
Listed in the sources of “Why the voice floor needs a named room”.
/research/voice-floor-first-room
Listed in the sources of “Why an engineering team buys a room”.
/research/why-an-engineering-team-buys-a-room
The working: Eight billion parameters at four bits each is 4 GB of weights. Add the quantisation scales, the embedding and output layers kept at higher precision, and a small runtime allowance, and the file lands at about 4.5 GB. Against 1,536 GB of memory in one mill (azita-unit-model), that is roughly 340 such models, or one model and a very large local index, in a single room.
Read with: A sizing derivation, not a benchmark. Accuracy turned out to be a retrieval problem rather than a model-size problem, and the retrieval argument that follows from it (coverage, freshness, reproducibility) is Azita's analysis and is labelled as such wherever the site makes it.

42Derived

1.7 to 4 GW

Sanctioned commercial load, Mumbai Metropolitan Region

Azita, derived September 2026

Citation on file

What we use it for: One city, worked through end to end, as a local check on the worldwide census.
On the site: “1.7 to 4 GW of sanctioned commercial load sits in the Mumbai Metropolitan Region.”
/research/sanctioned-load-idle
The working: 169.4 mn sq ft of office stock in Mumbai (Knight Frank, H1 2025) multiplied by the band of utility demand norms the model applies to air-conditioned commercial space, roughly 10 to 24 W per sq ft of sanctioned load (about 110 to 250 W per sq m). 169.4 mn sq ft x 10 W is about 1.7 GW; x 24 W is about 4 GW. The band is wide on purpose: it spans the norms utilities use rather than picking one.
Read with: Demoted in September 2026: this is a local derivation on a research post, not the site's market sizing. The worldwide figure is azita-census. A derivation, not a survey. Headroom at any real site is verified against half-hourly meter data and measured thermal margin, never against sanctioned-load paperwork.

43Derived

1,700 to 4,000 rooms

Rooms in Mumbai, if 1% of the band is idle and eligible

Azita, derived September 2026

Citation on file

What we use it for: A sense of scale for one city, stated with its assumption in the same breath.
On the site: “1,700 to 4,000 rooms, if 1% of the band is idle and eligible at 10 kW.”
/research/sanctioned-load-idle
The working: 1% of 1.7 to 4 GW is 17 to 40 MW. At 10 kW per mill that is 1,700 to 4,000 rooms. The 1% is an assumption chosen to be conservative, not a measured eligibility rate.

44Derived

Inside 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)

Why one machine is the unit of contiguity

Azita, derived from published link bandwidths September 2026

Citation on file

What we use it for: The boundary any technical buyer can verify: independent jobs can be placed across the fleet; one tightly coupled job stays inside one machine.
On the site: “Placement, not parallelism: real fabric inside the machine, city fibre between.”
/research/cost-per-token-distributed-inference
Listed in the sources of “Why an engineering team buys a room”.
/research/why-an-engineering-team-buys-a-room
The working: A full fine-tune of a 7B model syncs about 14 GB per step; 14 GB divided by 0.125 GB/s is 112 s of sync per step, so it is off by orders of magnitude across buildings. A LoRA step (about 40 MB) is marginal and viable with accumulation. For staging, 0.125 GB/s x 86,400 s is about 10.8 TB per day raw, 7 to 9 TB realistic, which is ample for batch inference, generation, rendering and embeddings.
Read with: Link figures are vendor-published nominal bandwidths; the fibre figure assumes a 1 Gbps line at full rate. The product is scoped to what physics allows and never claims distributed training.

Audited constants

Engineering and cost constants, audited by the founder.

45Audited constant

340,200 rack configurations swept for every single room, in about fifteen seconds. Running today, assisted, and instrumented at every installation; never licensed.

The design engine, as it runs today

Azita design engine (RackQuilt), founder-audited September 2026

Citation on file

What we use it for: What the engine actually does per room. A design firm draws four configurations by hand; the engine sweeps the space, refuses most rooms and names the constraint that killed each one.
On the site: “340,200 rack configurations swept for every single room, in about fifteen seconds. A design firm draws four by hand.”
/platform/rackquilt
“The engine evaluates 340,200 configurations for every single room, in seconds.”
/research/nine-month-problem
Listed in the sources of “Distributed inference and the cost per token: reading the public numbers carefully”.
/research/cost-per-token-distributed-inference
Read with: The configuration count is the size of the swept space per room, read off the engine. Azita publishes no module or test counts. What choosing the silicon as well as the layout does to the delivered cost per token is part of the cost model, is not yet measured, and is not published as a figure.

46Audited constant

$289,400 (₹2.52 Cr) per ten-kilowatt mill, fully loaded

Landed capex per mill

Azita proof-of-concept cost model, founder-audited September 2026

Citation on file

What we use it for: What one mill costs to land in a room, fully loaded, as one input to the cost per token.
On the site: “Costed on eight accelerators per node; the buying metric is dollars per TB/s, not a part number.”
/platform/token-mills
“A ten-kilowatt mill lands at $289,400, and the design is card-agnostic inside a frozen envelope.”
/research/cost-per-token-distributed-inference
Read with: Node $210k, completion $66.5k, trade $10.5k, IGST carry $2.4k. Card-agnostic: a card swap is a patch to the type file, not a redesign. The larger build costed in the earlier model is withdrawn with that envelope; the volume SKU is the ten-kilowatt mill. One cross-check: eight accelerators of 192 GB per node is the 1,536 GB the unit model publishes. This is a cost input, not a price or a return, and the site prints no payback period.

47Audited constant

Air-only, no water draw. About 1,755 CFM at a 10 °C rise on a ten-kilowatt mill; one 5 TR room unit per mill; designed to 40 °C at the face (ASHRAE A3).

The mill's thermal envelope

Azita type-level design, founder-audited September 2026

Citation on file

What we use it for: How much air the mill moves, and how much heat the room has to reject on the cooling it already has.
On the site: “The mill moves air; the room rejects heat. About 1,755 CFM at a 10 °C rise, one 5 TR room unit per mill.”
/platform/token-mills
“About 1,755 CFM at a 10 °C rise, with the derivation shown in full.”
/research/air-only-mills
Read with: The airflow is the heat balance solved at the mill's intake, and /research/air-only-mills shows that working line by line. The earlier design solved the same relation at a larger intake and got 2,625 CFM; the relation is what carried into v8, not the number. The 10 °C rise, the 5 TR room unit and ASHRAE A3 at the face are unchanged, and a 5 TR unit is 17.6 kW, so a mill drawing 14.0 to 14.4 kW at the meter still sits under it. The airflow above is worked at the ten-kilowatt IT load; the heat the room actually rejects is the at-the-meter figure, and the airflow line will be restated against it when the founder restates the efficiency ratio. No efficiency ratio is quoted here: the one the v3 model published was computed against a feed and a socket budget that no longer exist, and it will not reappear until it has been restated against the mill's own intake. The room's own cooling runs on the building's HVAC circuit, outside the mill's own supply. Duty assumptions against after-hours cooling are checked room by room; the survey instrument carries an after-hours screen for that reason.

48Audited constant

14 to 18 weeks from order to first token, into a room that is already energised

Order to first token

Azita supply-chain workstream, founder-audited September 2026

Citation on file

What we use it for: The mill's own clock, and the number every other line on the supply-gap chart is measured against. It only reads that way if silicon is ordered against a graded pipeline rather than after a signature.
On the site: “14 to 18 weeks, order to first token.”
/platform/rackquilt
“A mill into a room already energised: 14 weeks.”
/platform
Listed in the sources of “The nine-month problem: where the time goes, and what a scan replaces”.
/research/nine-month-problem
Read with: The clock is silicon lead time plus build, burn-in and the electrical tie-in, and it runs into a room that already has its connection and its permission. It is not a construction schedule, because there is no construction in it.

49Audited constant

A silicon swap is a patch: no lab cost, inherits the type file. A new card class is a minor: a delta qualification. Touching the feed or the envelope is a major: $180 to 420k, 6 to 12 months, a board decision.

Change control on the type file

Azita engineering policy, founder-audited September 2026

Citation on file

What we use it for: Why the envelope is frozen and the card is free: the cost of changing each.
On the site: “Touching the feed or the envelope is a major: $180 to 420k and 6 to 12 months, a board decision.”
/research/air-only-mills
“A card swap is a patch, not a redesign.”
/platform/token-mills
Read with: The major-change cost band is the model's estimate of re-qualification, not a quotation. No third-party scheme certifies a machine of this kind; the classes describe how much of the type file a change re-opens.

50Audited constant

Retired

Today a site takes about nine months of consultants, cooling plant and re-qualification. We take a scan.

Walked room to running mill

Azita RackQuilt programme September 2026

Citation on file

What we use it for: The time story the design engine exists to shorten, stated as the industry's clock against ours.
On the site: “Today a site takes about nine months of consultants, cooling plant and re-qualification. We take a scan.”
/platform/rackquilt
Retired, and what replaced it: The 90-day target is retired. The clock the site quotes is now the mill's own, 14 to 18 weeks from order to first token (hardware-lead-14-18), and the engine's own sweep (azita-design-engine). The nine-month figure above is the only part of this entry the site still argues from.
Read with: Nine months is the industry's own conventional programme for a site of this kind, held in the register with the programmes it was read from. A licensed engineer still stamps every tie-in, on every site, and that does not change with the clock.

Modelled figures

Outputs of Azita's model. Not yet measured, and labelled so wherever they appear.

51Modelled, not yet measured

2,000,000 rooms worldwide with power, fibre and spare capacity already in them. About 15 GW of stranded capacity sitting inside them, energised and paid for. 0 new grid connections, substations or transformers required to reach it.

The bottoms-up room census

Azita bottoms-up census, modelled September 2026

Citation on file

What we use it for: The size of the supply that needs no new permission, and the reason the company counts rooms rather than gigawatts.
On the site: “The megawatts already exist. They need no new permission.”
/estates
“2,000,000 rooms worldwide with power, fibre and spare capacity already in them.”
/platform
“Our bottoms-up census counts roughly 2,000,000 rooms worldwide that already have power, fibre and spare capacity in them.”
/research/sanctioned-load-idle
Read with: Two of the three numbers are a model and the third is a definition. Rooms are graded one by one, and a grade is a desk assessment against building type, connection class and fibre presence; Azita publishes no count of rooms graded. It is not a survey and it is not a commitment: no room counts as eligible until it has been walked, and the site never presents the census as a pipeline.

52Modelled, not yet measured

A 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.

The mill: Pod-10/K, the volume SKU

Azita unit model September 2026

Citation on file

What we use it for: The whole specification of the machine: its electrical intake, its memory, what it draws at the meter and how the room carries its heat.
On the site: “A ten-kilowatt factory in a room that is already on.”
/platform/token-mills
“A ten-kilowatt mill on 415 V three-phase, twenty amps a phase, 14.0 to 14.4 kW at the meter, air-cooled, no plant room.”
/trust
“Ten kilowatts from the building's existing sanctioned connection, at 415 V three-phase and twenty amps a phase, 14.0 to 14.4 kW at the meter.”
/legal/hosting-terms
“Ten kilowatts from your existing sanctioned load, on your existing connection: 415 V three-phase, twenty amps a phase. At the meter, 14.0 to 14.4 kW.”
/owners
“At the meter it draws 14.0 to 14.4 kW, not ten: the fans, the supplies and the losses are real, and the room has to reject that heat.”
/research/air-only-mills
“An eight-billion-parameter model quantised weighs about 4.5 GB, and a mill carries 1,536 GB.”
/research/voice-floor-first-room
“A codebase index and the models that read it fit comfortably in 1,536 GB.”
/research/why-an-engineering-team-buys-a-room
Listed in the sources of “A floor, a meter and a cap: how a building gets paid”.
/research/floor-meter-cap
Listed in the sources of “The nine-month problem: where the time goes, and what a scan replaces”.
/research/nine-month-problem
Listed in the sources of “Distributed inference and the cost per token: reading the public numbers carefully”.
/research/cost-per-token-distributed-inference
Listed in the sources of “Certification-ready is not certified: the regimes that actually govern an in-building mill”.
/research/certification-ready-not-certified
Listed in the sources of “The power already exists: what sanctioned load means and how much of it is idle”.
/research/sanctioned-load-idle
Read with: The electrical, memory and thermal figures are a type-level design specification, audited. They do not establish measured performance at a particular site. The unit model's commercial outputs are not published on this site: pricing is a proposal based on the workload and deployment. This entry supersedes the earlier electrical envelopes, which are withdrawn: the intake is stated at the board, and the at-the-meter figure is the honest one the room's cooling is sized against, because a ten-kilowatt machine does not draw ten at the meter.

53Modelled, not yet measured

Retired

Not published on this site.

Plan-case revenue per megawatt-year

Azita model, plan case September 2026

Citation on file

What we use it for: Nothing on the site argues from it.
On the site: No page on this site quotes this figure any more.
Retired, and what replaced it: Withdrawn from public copy. The plan case is an investor-model output, and the site does not reproduce modelled revenue, earnings or returns. A buyer's price is a proposal based on the workload and deployment; a host's terms are set out in the hosting agreement. The entry stays so an old link lands somewhere honest.

54Modelled, not yet measured

Capped at ₹15 to 18 L per mill per year; ₹7.2 L modelled at 60% duty

The host's metered kicker

Azita host model September 2026

Citation on file

What we use it for: What a host may earn above the floor, read off the meter, with its ceiling.
On the site: “Above the floor, a metered kicker capped at ₹15 to 18 L a year. The cap is modelled, not yet measured.”
/owners
“A metered kicker read off the same revenue-grade meter, capped at ₹15 to 18 L per mill per year.”
/legal/hosting-terms
“A metered kicker capped at ₹15 to 18 L a year, modelled at ₹7.2 L at 60% duty.”
/research/floor-meter-cap
Read with: Never a share of what the mill earns. The kicker is a function of metered duty, and the duty is the thing no one has measured yet.

55Modelled, not yet measured

60% of sellable capacity contracted is the modelled steady state; operating band 45 to 65%; ceiling 80%

Utilisation, modelled

Azita proof-of-concept model September 2026

Citation on file

What we use it for: The cell the unit economics and the host kicker are modelled on.
On the site: “A mill at 60% of sellable capacity, our modelled steady state, spreads its fixed costs over three-fifths of the hours.”
/research/cost-per-token-distributed-inference
“The modelled steady state is 60% of sellable capacity, and a mill that cannot run overnight cannot reach it.”
/research/voice-floor-first-room
Listed in the sources of “Why an engineering team buys a room”.
/research/why-an-engineering-team-buys-a-room
Read with: Modelled, not measured. It is the steady state, not a build gate. Duty against after-hours cooling is checked room by room, which is why the survey carries an after-hours screen.

Stated commitments

Promises Azita makes in signed agreements. They are terms, not measurements.

56Commitment

₹4.87 L per mill per year, paid monthly, before the meter adds a rupee

The host's contracted annual floor

Azita, fixed in the hosting terms September 2026

Citation on file

What we use it for: What a host is paid whether or not the machine is busy.
On the site: “A contracted floor of ₹4.87 L per mill per year, paid monthly, before the meter adds a rupee.”
/owners
“The floor: ₹4.87 L per mill per year, fixed in the hosting terms, paid monthly by NEFT.”
/legal/hosting-terms
“A contracted floor of ₹4.87 L per mill per year, paid monthly.”
/research/floor-meter-cap

57Commitment

Every 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

Power reimbursed at actuals

Azita, in the master hosting agreement September 2026

Citation on file

What we use it for: Why the host never funds the machine's power and never sells it.
On the site: “Every unit the machine draws is measured on a revenue-grade meter and paid by Azita at your own utility tariff, as a separate line, never netted against your fee.”
/owners
“Power at actuals on a revenue-grade meter, never netted, never resold.”
/research/floor-meter-cap
Listed in the sources of “Distributed inference and the cost per token: reading the public numbers carefully”.
/research/cost-per-token-distributed-inference
Read with: The power is Azita's own consumption, never resale. The modelled blended tariff is ₹9.0 per kWh in a band of ₹8 to 11 by state, and it is never one flat number fleet-wide.

58Commitment

No lock-in; 30 days' notice; the machine out in 48 hours; the room returned as found. Erasure of on-site media is set out in the agreement.

Exit terms for hosts

Azita, in the master hosting agreement September 2026

Citation on file

What we use it for: What leaving looks like for a building that hosts a mill.
On the site: “Thirty days' notice from either side; the pod is out within 48 hours of the exit date; the room is returned as found.”
/legal/hosting-terms
“Thirty days' notice, and the mill is out in 48 hours, the room returned as found.”
/research/floor-meter-cap
Read with: The registry key is historic. These are hosting terms. A buyer's exit and erasure terms are set out in the buyer's own agreement, and this entry is not cited for them.

59Commitment

No uptime figure is quoted anywhere on this site. Availability commitments are agreed for each site and workload.

No uptime figure

Azita operating policy September 2026

Citation on file

What we use it for: The reason no page here carries a percentage with a nine in it.
On the site: “We quote none. Availability commitments are agreed for your site and workload.”
/trust
Read with: The registry key is historic. The commitment is the one in the agreement for a site and workload, not a figure on this page.

Found a number without a source?

Tell us the page and the figure. A source is added or the figure comes down; either way this page is updated with the date.

Send a correction

Five kinds of number. Each labelled.

Primary sources

Published by someone other than Azita. Quoted as printed, with the date and a link where one exists.

Derived figures

Azita arithmetic on primary figures. The working is shown so you can redo it.

Audited constants

Engineering and cost constants, audited by the founder.

Modelled figures

Outputs of Azita's model. Not yet measured, and labelled so wherever they appear.

Stated commitments

Promises Azita makes in signed agreements. They are terms, not measurements.

What “modelled, not yet measured” means

What availability does not establish