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

Research8 September 2026

Why an engineering team buys a room

Coding assistants went usage-based and the per-engineer bill became visible. Why a client MSA, not a statute, sends an engineering team to buy a room.

6 min read
  • coding
  • engineering
  • use cases

An engineering team does not buy a room because it wants one. It buys a room because two things happened in the same year: the price of its coding tools stopped being a seat and became a meter, and the contract with its largest client turned out to forbid the meter's other end. This note takes the two in order, says what a pod actually serves, is plain about the model it serves it with, and answers the payback question in words rather than in numbers we cannot yet cite.

The bill became visible

For three years a coding assistant cost what a seat cost. On 27 April 2026 GitHub announced that from 1 June every Copilot plan would bill on usage: credits consumed by tokens, input, output and cached, at the published API rate of the model chosen. A Business seat stayed at $19 a month with $19 of credits; beyond that an administrator sets a budget or the work stops, and the fallback to a cheaper model when the allowance ran out was withdrawn. The Information reported that Cursor's gross margin was negative 23% in the quarter ended January 2026, at a company approaching $2 billion of annualised revenue: the frontier models underneath cost more than the subscriptions above them.

What usage pricing makes visible is how much a working engineer consumes. 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, with heavy users running $500 to $2,000 a month each. Microsoft's Experiences and Devices division, at roughly the same scale and per-engineer range, told its teams by internal memo that licences would end on 30 June; under flat seat licensing the token cost had been invisible, and under usage billing it became visible at once. McKinsey's August 2026 survey of 1,719 organisations found one in five already naming operating cost as a constraint on further AI use.

The self-hosted answer

The market's first response was the one it always gives to a volatile bill: bring the machine home. Tabnine's July 2026 price list is instructive less for the numbers, $39 a user a month for the assistant and $59 for the agentic platform on annual terms, than for what sits beside them: 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. The software vendor sells the seat; the customer brings the accelerators. That line, customer-provided compute, is what we sell: a pod in the building the team already occupies, on the board that already feeds the floor, with the weights and the index inside it.

Why the MSA, not the statute, drives the Indian buyer

In India the reason is contractual, and a buyer who expects a statute to fence code will look for one and not find it. The Digital Personal Data Protection Act defines personal data as data about an identifiable individual; source code is not that, and section 16 restricts transfers only to countries the Centre notifies, of which as of mid-2026 there are none. Nothing in the statute stops a developer in India sending a client's repository to an assistant hosted abroad.

The client's master services agreement does. A bank's or a hospital group's MSA with an engineering-services firm typically forbids processing client code by any third party not on the approved sub-processor list, requires SOC 2 evidence and, for anything touching card systems, PCI DSS, and reserves the client's right to audit. Where the client is a bank or an NBFC, the Reserve Bank's Directions on Outsourcing of Information Technology Services of 2023 flow down through that MSA: their scope names application development, maintenance and testing; they require storage of data in India as applicable, a right to audit the provider and its sub-contractors, and unrestricted access to the data and the relevant business premises. A vendor whose engineers pipe the bank's code through an assistant in another country has created an outsourcing the bank never approved. So the security team blocks the tool and the engineers lose it. The room gives it back with an address the MSA can carry: this model, this index, on this machine, in this building.

What a pod serves

A pod for an engineering organisation has the same two shifts as any other room.

What an engineering pod serves
DayNight
WorkCompletions, chat, codebase questions, review, test generationBatch review of the day's merges, test generation, migrations, re-indexing
Bound byLatency per keystroke and per questionThroughput of long agent jobs
Model classA small fast model for completions; a larger open-weight coder for review and answersThe largest open-weight coder the pod holds, at large batch
What an engineering pod serves

By day the pod serves completions, which are latency-bound and want a small fast model; chat and codebase question-answering, which are retrieval-bound and want an index of the whole repository, not the sanitised part that cleared a legal review, updated as commits land; review; and test generation. By night it takes the batch work a usage meter makes expensive: a review pass over everything merged that day, test generation across untouched modules, migrations and refactors run as long agent jobs, and the re-indexing that keeps the day's answers fresh. A codebase index and the models that read it fit comfortably in 1,536 GB, and a job whose parts talk to each other mid-step must never cross a building boundary; the pod is the unit of contiguity. A room that serves only the day is a room half used, and the modelled utilisation our unit economics rest on assumes the night is real.

The model tier, said plainly

The models that fit on eight accelerators and serve a team at interactive latency are a tier below the frontier on the long agentic tasks that generate the largest bills, and developers will notice. We would rather say that here than have it discovered in an evaluation.

No parity claim

We do not claim the pod matches a frontier model on long agent tasks, and will not until we have measured it on a team's own repository. The narrower claim: most of a team's daily consumption is within reach of the open-weight coding models the pod holds.

The consequence is a division of labour. The long agent task a frontier model does better stays with the cloud tool the team already has, on the code the client has agreed may leave the building, and the meter for that lane stays visible. The gap between the tiers has narrowed every quarter we have watched it, and we pin every model version and index snapshot so an auditor can reproduce what the pod did on the day.

The payback question, in words

A buyer's bill today has two parts: seats, which are predictable, and usage above the seat, which is not, and which the reporting above suggests can be many times the seat. A room can carry the work behind both under one proposal, based on the team's workload and deployment, with a separate usage line for the frontier lane that stays with the cloud tool. Whether that comes out ahead depends on two things the buyer controls and we do not. The first is how many engineers share the pod: a pod is sized for one engineering organisation, and a team too small to fill it pays more per seat than it should, which is why we sell to organisations of the size that fills one. The second is how full the night is: a pod that runs the batch shift earns the modelled utilisation; a pod that sleeps at seven does not. We will not print a payback period, because the only honest inputs are the buyer's own headcount and the buyer's own night. What we offer instead is a comparison of quality and total cost on the team's own repository, against what the team pays today, before anyone decides.

What we will measure before quoting

Before a quote we will run the team's own repository through a pod and measure token consumption by task class; completion and codebase-answer latency at the team's concurrency; batch throughput on the night jobs; retrieval quality on the team's own questions; and the energy the work draws at the meter. Then the proposal is based on the team's workload and deployment, night shift included, with quality and total cost set beside what the team pays today.

That is why an engineering team buys a room. The meter made the bill visible, the contract forbade the meter's other end, and the room is the only place where the code, the index and the model can all sit in the building the MSA names.

Sources

  1. 01GitHub Copilot is moving to usage-based billingGitHub Blog, 27 April 2026, effective 1 June 2026Credits 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.
  2. 02Cursor AI pricing in 2026 (reporting The Information)CloudZero, 18 May 2026, updated 4 September 2026Gross margin of minus 23 per cent in the quarter ended January 2026 while approaching $2 billion in annualised revenue.
  3. 03Uber burns its 2026 AI budget in four months on Claude CodeForbes, Janakiram MSV, 17 May 2026About 5,000 engineers; the year's budget spent in four months; heavy users at $500 to $2,000 a month.
  4. 04Microsoft cancels Claude Code after token costs exceed budgetEnterprise DNA, citing an internal memo first reported by The Verge, 31 May 2026About 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.
  5. 05Tabnine pricingTabnine, read July 2026Code 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.
  6. 06The State of AI: Global Survey 2026McKinsey, QuantumBlack, August 2026; fielded 4 May to 8 June 2026, n=1,719 across 97 countriesOne 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'.
  7. 07Reserve Bank of India (Outsourcing of Information Technology Services) Directions, 2023Reserve Bank of India, RBI/2023-24/102, 10 April 2023, effective 1 October 2023Scope 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.
  8. 08Digital Personal Data Protection Act, 2023 (No. 22 of 2023)Ministry of Law and Justice, Government of India, 11 August 2023Personal 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).
  9. 09The 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.
  10. 10What an eight-billion-parameter model weighsAzita analysis, derived; retrieval economics, Azita analysis, September 20264.5 GB for an eight-billion-parameter model, quantised. 1,536 GB in one mill.
  11. 11Why 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)
  12. 12Utilisation, modelledAzita proof-of-concept model, September 202660% of sellable capacity contracted is the modelled steady state; operating band 45 to 65%; ceiling 80%

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 · 6 August 2026A floor, a meter and a cap: how a building gets paidLater · 9 September 2026Why the voice floor needs a named room

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