A ₹149 subscription and a very expensive power user
In January 2024, Ola founder Bhavish Aggarwal's AI venture Krutrim raised funding at a $1 billion valuation, becoming India's first AI-labeled unicorn. About seven weeks earlier, a Bengaluru team led by Vivek Raghavan and Pratyush Kumar closed a $41 million seed round for Sarvam AI, one of the largest seed rounds in Indian history, to build large language models tuned for Indian languages. Both raised capital on a simple pitch: India needs its own AI companies, not just resellers of someone else's API. But headline valuations hide a harder question that every founder in this space has to answer with a spreadsheet, not a slide deck: when a customer pays you a flat monthly fee and your product's cost of serving them scales with how much they actually use it, do you make money on that customer or lose money on them? This chapter builds an AI startup, VaidyaAI, a hypothetical WhatsApp-based vernacular health-information assistant, from first principles, to answer that question with real numbers, then traces what happens to the founders' ownership as they raise the capital to keep the company alive long enough to fix it.
What actually makes an AI company different from a software company
A traditional SaaS company sells software whose marginal cost of serving one more user is close to zero. Once the code is written, a thousand users and a million users cost roughly the same to serve per user — this is why mature SaaS companies report gross margins of 80-90%. An AI company that wraps a large language model does not get this for free. Every query a user sends triggers an API call (or a self-hosted inference pass) that consumes GPU time, and GPU time costs real money per token, every single time. Cost of goods sold (COGS) in an AI product is not fixed — it is a variable cost that scales linearly (at minimum) with usage, the same way a restaurant's ingredient cost scales with covers served, not the way a software license's marginal cost stays flat. This single structural fact is the reason a16z's widely cited 2023 analysis of the generative AI application layer flagged compressed gross margins as the defining risk of the category, and it is the reason cap-table math and unit-economics math cannot be separated when you build an AI company: the money you need to raise, and the price you can defensibly charge, both depend on how expensive your product is to actually run.
The three layers of the Indian AI stack
Before building VaidyaAI's numbers, place it in the stack. Indian AI startups cluster into three layers, and the layer you choose determines your capital intensity, your moat, and your realistic exit.
Infrastructure and foundation-model layer. Companies here train or heavily post-train their own base models — Sarvam AI (Indic-language LLMs, selected under the government's IndiaAI Mission to build a sovereign foundation model) and Krutrim (a full-stack play spanning its own LLM, cloud, and now announced AI chip ambitions) sit here. This layer is brutally capital-intensive: pretraining a competitive model from scratch requires thousands of GPU-hours, and India's biggest structural handicap is that it has almost no domestic large-scale GPU manufacturing or hyperscale clusters comparable to the US or China, so compute is imported, taxed, and expensive. Very few founders should attempt this layer; it needs $10M+ before product-market fit is even testable.
Applied / vertical AI layer. Companies here take an existing foundation model (via API or a fine-tuned open-weight model) and wrap it around a specific workflow, dataset, and distribution channel — Haptik (conversational AI, acquired by Reliance Jio), Yellow.ai, and Observe.AI (contact-center conversation intelligence, Indian-founder-led though US-headquartered) are the well-known examples. This is where the vast majority of viable AI startups, including VaidyaAI, actually live: capital-efficient to start, but exposed to exactly the margin-compression problem this chapter works through, because you are renting your core capability from someone else's API.
Tooling and infrastructure-for-AI layer. Companies that don't build models or applications but sell the picks and shovels — GPU cloud brokers, vector databases, evaluation and observability tooling, the private data-centre operators (Yotta, E2E Networks, and others) empanelled under the government's IndiaAI Compute Portal to offer subsidized GPU-hours to startups. This layer's economics resemble classical infrastructure businesses more than AI businesses.
VaidyaAI is an applied-layer company: it does not train a model, it calls one, and its entire unit-economics problem flows from that choice.
Worked example: does VaidyaAI make money on its own customers?
VaidyaAI charges ₹149 per user per month for unlimited vernacular health-information queries over WhatsApp, answered by a general-purpose LLM API. To know whether this price is sustainable, compute the actual cost of answering one query, using representative capable-model API pricing (illustrative — check a provider's live pricing page for current rates, since these change frequently): ₹2.50 per million input tokens is roughly $2.50, and ₹10 per million output tokens is roughly $10, at an assumed exchange rate of ₹85 per US dollar (used consistently throughout this example). A typical exchange with VaidyaAI — the user's question plus recent conversation history as context, then a generated answer — runs about 500 input tokens and 300 output tokens.
# cost of one VaidyaAI query, in INR
input_tokens, output_tokens = 500, 300
usd_per_million_input, usd_per_million_output = 2.50, 10.00
inr_per_usd = 85
cost_usd = (input_tokens * usd_per_million_input
+ output_tokens * usd_per_million_output) / 1_000_000
cost_inr = cost_usd * inr_per_usd
print(f"{cost_inr:.3f}") # -> 0.361
Trace it: input cost is 500 × 2.50 / 1,000,000 = $0.00125; output cost is 300 × 10.00 / 1,000,000 = $0.003; total $0.00425 per query; at ₹85/$1 that is ₹0.36125, which prints as ₹0.361 per query at three decimal places. Add non-LLM infrastructure — WhatsApp Business API messaging fees, hosting, a small vector database for retrieval — at a flat ₹8 per user per month regardless of usage.
Now compute gross margin for two different users on the same ₹149 flat plan. An average user sends 40 queries a month:
COGS = 40 × ₹0.361 + ₹8 = ₹14.44 + ₹8 = ₹22.44. Gross profit = ₹149 − ₹22.44 = ₹126.56. Gross margin = 126.56 / 149 = 84.9% — a healthy, SaaS-grade margin.
But flat pricing means a power user — say, someone using VaidyaAI as their primary source of health information, sending 400 queries a month, ten times the average — is served at the same ₹149 price:
COGS = 400 × ₹0.361 + ₹8 = ₹144.40 + ₹8 = ₹152.40. Gross profit = ₹149 − ₹152.40 = −₹3.40. Gross margin = −3.40 / 149 = −2.3%. VaidyaAI loses money on this exact customer, every single month, for as long as they stay subscribed.
The break-even query volume — the usage level at which a ₹149 flat-plan customer stops being profitable — solves from ₹0.361·q + ₹8 = ₹149: q = (149 − 8) / 0.361 = 141 / 0.361 ≈ 391 queries per month. Any user above roughly 13 queries a day costs VaidyaAI more than they pay.
Build vs. buy: when self-hosting beats the API
The obvious fix — route power users to a smaller, self-hosted open-weight model — has its own crossover math. Assume a dedicated GPU instance capable of serving fine-tuned inference at meaningful capacity costs a flat $500/month (≈₹42,500), independent of query volume within its capacity. The API path costs ₹0.361 per query with no fixed cost. Self-hosting wins once the fixed cost is cheaper than the variable API bill it replaces:
Break-even total query volume = ₹42,500 / ₹0.361 ≈ 117,730 queries per month. At VaidyaAI's average of 40 queries per user, that is 117,730 / 40 ≈ 2,943 users. Below roughly 2,900 active users, paying per query via API is cheaper than running your own GPU; above it, self-hosting (and the upfront engineering cost of fine-tuning a smaller open model on your own conversation data — itself a data-moat investment, not just a cost cut) becomes the rational move. This crossover, not "which model is smartest," is usually the real build-vs-buy decision an Indian AI founder makes, because talent to fine-tune and operate a self-hosted model is comparatively cheap in India (an experienced ML engineer in Bengaluru typically costs a fraction of the equivalent Bay Area salary), which shifts this crossover earlier than it would sit for a US-only team.
Funding the runway, and what it costs in ownership
None of the fixes above — usage-based tiers, model routing, fine-tuning a cheaper model — ship for free. They need runway: cash in the bank divided by monthly burn. If VaidyaAI raises a ₹1.5 crore pre-seed round and burns roughly ₹10 lakh a month on a lean five-person team (salaries, the API bill, WhatsApp fees), runway = ₹1,50,00,000 / ₹10,00,000 = 15 months to find the fix and reach the next round. Raising that runway costs equity, and the cost compounds every round — this is the part founders consistently underestimate.
Standard practice carves out an ESOP (employee stock option) pool before the first outside money comes in, so investors aren't diluted by future hires. VaidyaAI's two founders start at 100%; a 10% ESOP pool is carved out first, leaving Founders 90%, ESOP 10%. Each subsequent round dilutes every existing holder by the new investor's percentage of the post-money company, then adds the new investor's stake:
def dilute(table, new_pct):
factor = 1 - new_pct
return {k: v * factor for k, v in table.items()}
cap_table = {"Founders": 90.0, "ESOP": 10.0}
cap_table = dilute(cap_table, 0.15) # pre-seed: Rs 1.5cr for 15%
cap_table["Pre-seed"] = 15.0
cap_table = dilute(cap_table, 0.20) # seed: $1.2M for 20% (~Rs 51cr post)
cap_table["Seed"] = 20.0
cap_table = dilute(cap_table, 0.20) # Series A: $6M for 20% (~Rs 255cr post)
cap_table["Series A"] = 20.0
for holder, pct in cap_table.items():
print(f"{holder}: {pct:.2f}%")
Trace it by hand to confirm the output before trusting it. After the pre-seed round (factor 0.85): Founders 90 × 0.85 = 76.5, ESOP 10 × 0.85 = 8.5, Pre-seed investors take 15 — sum 100. After seed (factor 0.80): Founders 76.5 × 0.8 = 61.2, ESOP 8.5 × 0.8 = 6.8, Pre-seed 15 × 0.8 = 12, Seed investors take 20 — sum 100. After Series A (factor 0.80 again): Founders 61.2 × 0.8 = 48.96, ESOP 6.8 × 0.8 = 5.44, Pre-seed 12 × 0.8 = 9.6, Seed 20 × 0.8 = 16, Series A investors take 20 — sum 100. The code's printed output is exactly:
Founders: 48.96%
ESOP: 5.44%
Pre-seed: 9.60%
Seed: 16.00%
Series A: 20.00%
Three funding rounds, none of them individually alarming (15%, 20%, 20% dilution — all standard), and the founders have fallen from 100% to below half the company, while still legally controlling it only if their board seats and voting agreements say so — ownership percentage and control are not the same thing, and a founder who only tracks the first stops noticing the second.
The regulatory floor: DPDP Act and the IndiaAI Mission
VaidyaAI processes health queries, which puts it squarely inside India's Digital Personal Data Protection Act, 2023 (DPDP Act), the country's first comprehensive data-protection law. Unlike the EU's GDPR, the DPDP Act does not carve out a separate "special category" tier for health data with automatically stricter rules; instead it gives the central government power to notify certain data fiduciaries as Significant Data Fiduciaries based on factors including the volume and sensitivity of personal data they process, the risk to individual rights, and potential impact on India's sovereignty and electoral integrity. A health-focused consumer AI product handling large volumes of medical queries is a strong candidate for this classification, which brings extra obligations: appointing a Data Protection Officer based in India, conducting periodic Data Protection Impact Assessments, and independent data audits. Every fiduciary, significant or not, must obtain clear, specific, and revocable consent before processing personal data, report breaches to the Data Protection Board, and — notably — the Act takes a "blacklist" approach to cross-border data transfer rather than the EU's "whitelist" (adequacy) approach: data can flow out of India to any country except ones the government specifically restricts, which is a materially lighter compliance burden than GDPR imposes on companies serving Indian users. Separately, the government's IndiaAI Mission — a ₹10,372 crore outlay approved by the Union Cabinet in March 2024 — runs an IndiaAI Compute Portal that empanels private data-centre operators to offer subsidized GPU-hours to Indian AI startups and researchers, directly targeting the compute-cost handicap described earlier in this chapter. A founder building VaidyaAI should treat both of these — the compliance floor and the compute subsidy — as inputs to the same spreadsheet used to plan runway and pricing, not as separate legal and technical concerns.
Common misconception
Students consistently assume that because an AI startup is "just calling an API," it inherits the same 80-90% gross margins as a normal SaaS company, since there's no factory, no inventory, no shipping. The VaidyaAI numbers above correct this directly: a flat-price AI product's gross margin is not one number, it is a distribution across your user base, because COGS scales with usage while price does not. A company can report a healthy 84.9% blended margin while quietly losing money on every one of its heaviest, most engaged users — the exact users a normal SaaS company would be happiest to have. The fix is never "just use a cheaper model" alone (though that helps, as the Active Recall below shows); it is architectural — usage-based pricing tiers, routing simple queries to cheap models and only escalating complex ones to expensive ones, and building toward self-hosting once volume crosses the fixed-cost break-even. Treating inference cost as a rounding error, the way software cost is traditionally treated, is the single most common reason technically strong Indian AI products fail to reach profitable scale even after finding genuine product-market fit.
Active recall
Attempt each question before reading its answer.
- VaidyaAI switches its output-generation calls to a cheaper model at $5 per million output tokens (input pricing unchanged at $2.50/million, still 500 input / 300 output tokens per query, ₹85/$1). Recompute the per-query cost, the 400-query power user's gross margin, and the self-host break-even user count.
- Just before the Series A, the lead investor insists on refreshing the ESOP pool back up to 10% of the company on a pre-money basis — i.e., before Series A cash arrives, taken pro-rata from all existing holders (Founders 61.2%, ESOP 6.8%, Pre-seed 12%, Seed 20%, at that point). Compute the resulting cap table after this refresh and after the Series A round closes.
- Instead of $6M for 20%, suppose Series A is $9M for 25% of the company. Using the original (no ESOP refresh) post-seed cap table — Founders 61.2%, ESOP 6.8%, Pre-seed 12%, Seed 20% — compute the post-Series-A cap table and the round's post-money valuation.
- Under the DPDP Act, 2023, what obligations does VaidyaAI likely face beyond a generic AI wrapper handling non-sensitive data, and why?
- Name the three layers of the Indian AI startup stack and place one real Indian company in each, explaining what distinguishes their capital needs.
- At VaidyaAI's original ₹0.361-per-query cost and ₹8 flat infra cost, on the ₹149 plan, what is the exact monthly query volume at which gross margin crosses from positive to negative?
Answers
1. New per-query cost: input unchanged at $0.00125, output = 300 × 5.00 / 1,000,000 = $0.0015, total $0.00275 × ₹85 = ₹0.234. Power-user (400 queries) COGS = 400 × 0.234 + 8 = 93.6 + 8 = ₹101.60; gross profit = 149 − 101.60 = ₹47.40; margin = 47.40/149 = 31.8% — a swing from −2.3% to a healthy positive margin on the same customer, from one pricing lever. This also shifts the self-host break-even: 42,500 / 0.234 ≈ 181,624 queries, ÷ 40 ≈ 4,541 users — self-hosting now only pays off much later, because the API alternative got cheaper. Both downstream numbers move from the single upstream change.
2. Non-ESOP total pre-refresh = 61.2 + 12 + 20 = 93.2. ESOP needs +3.2 points (6.8 → 10), taken pro-rata: Founders lose 61.2/93.2 × 3.2 ≈ 2.10 → 59.10; Pre-seed loses 12/93.2 × 3.2 ≈ 0.41 → 11.59; Seed loses 20/93.2 × 3.2 ≈ 0.69 → 19.31; ESOP = 10.00 (sum ≈100). Series A then dilutes everyone by 0.8 and adds its own 20%: Founders 59.10 × 0.8 ≈ 47.28%, ESOP 10 × 0.8 = 8.00%, Pre-seed 11.59 × 0.8 ≈ 9.27%, Seed 19.31 × 0.8 ≈ 15.45%, Series A 20.00%. Versus no refresh (Founders 48.96%), founders end up almost 1.7 points lower — the pool refresh is paid for by existing shareholders, not by the new investor, which is exactly why founders negotiate over it.
3. Pre-money = $9M/0.25 − $9M = $27M, post-money $36M. Dilution factor = 1 − 0.25 = 0.75. Founders 61.2 × 0.75 = 45.9%, ESOP 6.8 × 0.75 = 5.1%, Pre-seed 12 × 0.75 = 9.0%, Seed 20 × 0.75 = 15.0%, Series A = 25.0% (sum 100).
4. Health-query volume and sensitivity make VaidyaAI a strong candidate for classification as a Significant Data Fiduciary, adding a mandatory India-based Data Protection Officer, periodic Data Protection Impact Assessments, and independent data audits, on top of the baseline duties every fiduciary carries: specific and revocable consent, purpose limitation, and breach reporting to the Data Protection Board. Cross-border transfer is still permitted by default under the Act's blacklist model unless the destination country is specifically restricted.
5. Infrastructure/foundation layer (Sarvam AI, training its own Indic-language LLMs — most capital-intensive); applied/vertical layer (Haptik, Observe.AI — wraps an existing model around a workflow, where VaidyaAI itself lives); tooling/infra-for-AI layer (IndiaAI Compute Portal-empanelled GPU providers like E2E Networks — sells capacity, not a model or an application).
6. 0.361q + 8 = 149 → q = 141/0.361 = 390.58, so gross margin crosses zero at 391 queries in a month (roughly 13 a day) — the last query below that count is still profitable, the query that pushes the count to 391 and beyond is not.
Think About It
Think about this: How would you explain ai startups: building an ai company in india to a friend who has never seen a computer? What real-world analogy would you use? Imagine you had to build a system using these concepts — what would be your first step? Try this: before moving on, write down three things you learned and one question you still have.
Key Takeaways — Summary and Recap
Let us recap what we covered: the core ideas behind ai startups: building an ai company in india, how they connect to real-world applications, and why they matter for your journey in computer science. Remember these key points as you move forward. For competitive exam preparation (CBSE, JEE, BITSAT), focus on understanding the WHY behind each concept, not just the WHAT.