By Vishal | October 2026
The question a marketer keeps asking
I am not saying there is no AI market. The technology has proven itself, and the models are good. My question is different, and as a marketer it is my first, second, third, fourth and fifth question: show me the market. Who pays, from which budget, how much, and what do they stop paying for in return?
This piece works through that question with numbers. The figures are rough, the assumptions are stated, and where I could not verify something I say so.
1. The debt behind the boom
The scale is in trillions, not hundreds of trillions, and that is still enormous.
• About $1.35 trillion of debt sits on the balance sheets of five hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle), plus roughly $1.65 trillion of off-balance-sheet AI obligations such as long-term data centre, compute and power contracts (Nikkei Asia analysis, as reported by YourStory). The second figure is lease-like commitments, not all borrowings.
• AI-related debt issuance is estimated at $450-570 billion in 2026, against roughly $100-121 billion in 2025 (Morgan Stanley and Goldman Sachs estimates, as reported).
• Goldman Sachs reportedly estimates about a third of 2026 AI capex is debt-financed. Citrini Research expects hyperscalers to issue far more in 2027-28 than the market expects.
• AI-linked bonds have reportedly become the largest segment of the investment-grade index, so index-tracking funds, including target-date funds, hold them automatically.
This matters because the lenders are not only banks. Pension, infrastructure, hedge and sovereign funds are all exposed.
2. What revenue would it take?
I reversed the calculation: given the debt, how much revenue is needed to service it?
Assumptions: 6.5% interest, 10-year repayment (13.9% of principal per year), 50% gross margin.
|
Debt scenario |
Annual debt service |
Revenue needed |
Interest only |
Revenue needed, interest only |
|
$1.35T |
$188B |
$376B |
$88B |
$176B |
|
$3.0T |
$417B |
$834B |
$195B |
$390B |
|
$4.0T |
$556B |
$1,113B |
$260B |
$520B |
This is a floor. It excludes equity-funded capex, R&D and staff costs, and any return to shareholders.
Per customer
Take the $3T case ($834B), assume half comes from consumers and half from businesses, and spread it across eight markets (USA, EU, UK, India, Singapore, Malaysia, Indonesia, Philippines: about 2,775M people and 33M firms).
|
Paying adoption |
Per individual per year |
Per month |
Per adopting firm per year |
|
5% |
$3,000 |
$250 |
$253K |
|
10% |
$1,500 |
$125 |
$126K |
|
15% |
$1,000 |
$84 |
$84K |
A consumer plan costs about $20 a month. Most firms are small and would spend closer to $1,000 a year than $126K.
3. The market as it looks today
Using the same eight markets, $20 a month ($240 a year) and paying adoption of 5, 10 and 15% of the whole population:
|
Adoption |
Consumer revenue (headline) |
Consumer revenue (realistic regional pricing) |
Business revenue |
Combined (headline) |
|
5% |
$33.3B |
$17.3B |
$8.3B |
$41.6B |
|
10% |
$66.6B |
$34.5B |
$16.5B |
$83.1B |
|
15% |
$99.9B |
$51.8B |
$24.8B |
$124.6B |
The business figure assumes about 33M firms, an average spend of about $5K per adopting firm, and seat-based pricing only.
Benchmarks, from memory and not verified: consumer subscriptions such as Spotify and Netflix reached roughly 5-6% of their addressable users after 10-18 years, and typical freemium conversion is 2-5%. A 5% paid share of total population across a mixed-income set of markets is already at the top of precedent. 15% has none for a standalone subscription.
A gap of 7-20x between $834B and $41-125B is large. It is not zero, and it excludes the part most likely to close it: AI replacing labour and services spending, rather than selling seats.
4. India: where the money actually is
India is the most useful test case, because I can anchor it to reported profits.
The profit pool (FY26)
|
Segment |
Profit (₹ lakh crore) |
Basis |
|
Enterprise (Nifty 500) |
18.2 |
Reported. The Nifty 50 alone is about 9.2 |
|
Medium |
~5.0 |
Estimate |
|
Small |
~3.0 |
Estimate |
|
Unorganised |
~12 |
Estimate, owner earnings |
|
Total |
~38 |
About $400B |
From profit to an AI tools market
Tech spend as a share of profit is not a published benchmark, so I built it from the revenue-based norms (2-4% of revenue) and stepped it down for smaller firms: Enterprise 20%, Medium 12%, Small 6%, Unorganised 2%. That gives about $50B of corporate tech spend. Gartner's figure of $176B for all India IT spending in 2026 (software $24.7B, IT services about $35B, data centre systems $9.4B) is a reasonable cross-check once government and multinational subsidiaries are added.
AI is only a slice. One survey says AI is below 20% of the IT budget for 95% of Indian enterprises. Applying 5%, 10% and 20%:
|
AI share of tech budget |
Domestic AI tools market |
|
5% (today-ish) |
~$2.5B |
|
10% (about 2028-30) |
~$5B |
|
20% (ceiling) |
~$10B |
IDC projects India's AI spending at $9.2B by 2028, which fits this range. Enterprise accounts for about 78% of it. Separately, Nasscom puts AI-related revenue in the Indian tech industry at $10-12B in FY26, but that is mostly AI-led services sold abroad, not domestic tools.
If AI displaced just 2% of the Nifty 500 wage bill (about ₹16.5 lakh crore, or $175B), it would add about $3.5B on top.
5. The telecom lesson
India has already lived through a layered, capital-heavy market collapse. In July 2026 TRAI data, Jio held about 39% of wireless subscribers, Airtel about 38%, Vodafone Idea about 15.5% and BSNL about 7%. It is a duopoly with a weak third player. A free-data price war from a cash-rich conglomerate drove out a dozen operators.
Where the analogy holds:
• Capital-intensive foundation layers end in price wars and consolidation. Open-weight models and falling token prices are the AI version.
• Survivors had fallback revenue. In AI, the largest model owners are conglomerates. The exposed ones are the pure-play labs, and their risk reaches lenders through compute contracts.
• Value moves up the stack. Cheap data produced UPI apps, streaming and delivery. Cheap tokens should do the same.
Where it breaks:
• Telecom is licensed country by country. Models are borderless, though regulation will fragment them (EU, India, China).
• A failed telco's spectrum and towers pass to survivors. GPUs lose value over 5-6 years, so recovery is worse.
• Consumer subscriptions may not be the main monetisation. Assistants are being bundled into phones, search and messaging.
6. Application and middleware: a map for India
This is where I expect the money to go. The totals are anchored to the table in section 4. The split across areas is my assumption.
|
Area |
Layer |
Low (5%) |
Mid (10%) |
High (20%) |
Why India |
|
Customer support and voice agents |
App |
$0.5B |
$1.0B |
$2.0B |
Indic voice, BPO labour pool |
|
Software engineering and IT ops |
App |
$0.45B |
$0.9B |
$1.8B |
Large developer base and GCCs |
|
BFSI back-office (KYC, underwriting, claims) |
App/middleware |
$0.38B |
$0.75B |
$1.5B |
UPI and Aadhaar rails |
|
Marketing, content and AI-search visibility |
App |
$0.25B |
$0.5B |
$1.0B |
Agency and SME base |
|
Finance, accounting and GST compliance |
App |
$0.2B |
$0.4B |
$0.8B |
MSME compliance load |
|
Healthcare (documentation, triage, imaging) |
App |
$0.18B |
$0.35B |
$0.7B |
Doctor shortage, vernacular |
|
Manufacturing and supply chain |
App |
$0.18B |
$0.35B |
$0.7B |
Export manufacturers |
|
Legal, HR and internal knowledge |
App/middleware |
$0.18B |
$0.35B |
$0.7B |
Large enterprises first |
|
Education and upskilling |
App |
$0.1B |
$0.2B |
$0.4B |
Vernacular tutoring |
|
Agriculture and public services |
App |
$0.1B |
$0.2B |
$0.4B |
Government stacks |
|
Total |
|
$2.5B |
$5.0B |
~$10B |
|
By layer in the mid case, I assume foundation and API fees are about 35% ($1.75B), middleware about 15% ($0.75B) and applications about 50% ($2.5B).
Middleware worth watching: model routing and gateways, enterprise retrieval and permissions, evaluation and AI security, and the Indic language layer (speech, translation, tokenisation). Their main risk is being absorbed into cloud platforms.
What separates survivors: proprietary data or deep workflow integration; pricing per outcome, paid from labour and services budgets rather than software seats; not depending on one model vendor; and distribution or regulatory moats. Thin wrappers are exposed, and the labs are moving into apps themselves.
India's mid-case foundation revenue of about $1.75B is roughly 0.2% of the $834B needed in the $3T scenario. India cannot service global AI debt. It can support a domestic application layer.
7. Is it a bubble?
I think the financing risk is real, but the evidence is mixed.
Signals for:
• Valuations far ahead of revenue. One large AI-linked IPO this year sought about a $1.75 trillion valuation on about $18.7 billion of revenue and an operating loss, roughly 94x revenue.
• Speculative narratives in prospectuses. The same filing pitched data centres in orbit, deploying "as early as 2028". Its own risk factors admit the technology is unproven and may never be commercially viable. It was disclosed as speculative, not hidden, but investors are still being asked to pay for it.
• Debt growing fast, partly off the balance sheet, and held by index-tracking funds.
Signals against:
• Revenue is real and growing. TCS reports about $1.8B of annualised AI revenue and Infosys about $275M.
• Most hyperscaler capex is funded from operating cash flow.
• Bubbles coexist with real technology. Rail and fibre were both overbuilt, and both were later used.
Contagion: debt that cannot be repaid is restructured, and the losses fall on the weakest tranches. The fragile parts are leveraged data-centre vehicles, neoclouds and private credit, not bonds of cash-rich issuers. Signals to watch: refinancing spreads, GPU rental prices, data centre utilisation, and lab revenue growth against compute commitments.
8. The test that answers the marketer's question
The first company to show the market takes the prize. A credible answer has to pass five checks:
1. Who pays? A named buyer with a budget line, not "users".
2. From which budget? Labour, outsourcing or compliance spend is large. Experimental innovation budgets are not.
3. How much? Per-outcome pricing (per call resolved, per claim processed, per return filed) that beats the cost it replaces.
4. Replacing what? A measurable cost that disappears, such as a BPO seat, a manual process or a compliance team.
5. Can it repeat? Revenue that recurs year on year without a new pilot each time.
The market that passes these tests is not a $20-a-month chat subscription. It is a vertical, outcome-priced application layer built on cheap models, and in India that means voice, compliance, back-office and vernacular services at scale.
Caveats
• Population, firm counts, price points, the 50% margin, the 6.5% rate and the 10-year repayment are all assumptions.
• The medium, small and unorganised profit figures and the split by application area are my estimates, with no published source.
• Benchmarks marked "from memory" are not verified.
• Off-balance-sheet obligations are commitments, not guaranteed borrowings.
• This is analysis, not investment advice.
Sources
Gartner (India IT spending forecast, Nov 2025); Nasscom Strategic Review FY26; IDC and EY-CII (as cited by CXOToday); TRAI subscriber data, July 2026; Business Today (Nifty 50 and Nifty 500 FY26 profits); MOSPI ASUSE 2025 and Economic Survey 2025-26 (MSME and unorganised sector data); Morgan Stanley, Goldman Sachs and Citrini Research (AI debt estimates, as reported by YourStory, Crypto Briefing and AIWeekly); CRISIL (IT services vertical mix); company prospectus coverage (Trending Topics, Benzinga, Douglas Research Insights).
Disclaimer:
Comments
Post a Comment