In today's deep dive, we examine the free money powering India's AI startups. AWS, Google, Microsoft and the government are offering substantial computing subsidies. But what happens when the credits run out? Are these startups genuinely profitable, or is somebody else paying their bills? We look at the hidden costs, the beneficiaries and what investors might be missing.

Something rather unusual happened after Y Combinator held its first Startup School in Bengaluru on April 18 this year. Around 2,000 selected founders, engineers and young developers attended the event, where they were offered AI tools and cloud credits worth roughly $25,000. The idea was to help them build products without worrying too much about the cost of technology. But within weeks, some attendees were trying to turn those credits into cash. In May, The Economic Times reported that messages were circulating in WhatsApp groups offering $10,000 worth of AWS credits for just $2,000 to $3,000. Another attendee wanted to sell Microsoft Azure credits. A 16-year-old developer told the newspaper that eight to ten people had approached him and his friends with offers at discounts of 20% to 40%, although there weren't many buyers.

There is something amusing about founders attending an event meant to encourage them to build businesses and then trying to make money by selling the tools they received for free. But there is a more interesting story underneath. These weren't cash grants. They were credits that could be used to pay for cloud services, AI models and software tools. The packages included $10,000 each from AWS and Azure, $2,500 from OpenAI, $500 from Anthropic and $720 from Cursor, among other benefits. Some participants had no immediate need for them, while others reportedly struggled to activate certain offers. Yash Kumar, a founding engineer at TruCommerce, told ET that his $500 Supabase credits worked, but his ChatGPT and Claude codes did not. With a Supabase bill of $35 a month, even that relatively small credit was enough to cover more than a year's usage.

The resale offers also tell us something about the actual value of these giveaways. A credit advertised as worth $10,000 is not the same as receiving $10,000 in the bank. It can be used only for eligible services, may expire, and often cannot be transferred. AWS explicitly prohibits selling its promotional credits and can invalidate them if they are sold. The Bengaluru reporting established that people were offering credits for sale, not that a large number of transactions actually happened. Still, the episode raises a bigger question about India's growing AI startup industry. If companies can get tens or hundreds of thousands of dollars in free computing, how much of their apparent financial health depends on somebody else paying the bills?

To understand why this matters, consider how much free infrastructure is now available to a young startup. AWS Activate advertises credits of up to $200,000 for qualifying startups backed by participating investors and accelerators, with further support available to selected AI companies. Google Cloud offers eligible AI startups as much as $350,000 over two years. Microsoft's startup programme can provide up to $150,000, with eligible investor-backed companies potentially receiving $200,000. These are maximum programme benefits, not amounts every startup automatically receives. Even so, the figures are substantial. An Indian founder who has raised ₹2 crore can potentially receive cloud support whose advertised value is comparable to, or greater than, the cash investment.

There is another detail in these offers that is easily missed. Google's $350,000 package is not simply a cheque for that amount. Under its AI startup programme, the first year can cover eligible spending up to $250,000. In the second year, the company provides a 20% credit against spending, subject to an additional $100,000 ceiling. To receive that entire second-year benefit, a startup would need to incur $500,000 of eligible usage and pay the remaining $400,000 itself. Google's credits also cannot be used to pay for third-party models such as Anthropic's. AWS, meanwhile, allows eligible Activate credits to cover third-party models accessed through Amazon Bedrock. The details matter because a founder may have an impressive collection of credit offers that cannot actually be used to pay the company's most expensive bills.

Why are technology giants so generous? Because they have spent years competing to become the default infrastructure provider for the next generation of businesses. AWS launched Activate back in 2013, long before the present AI boom. According to Amazon, the programme has provided more than $8 billion in promotional credits to over 350,000 startups worldwide. That figure represents the advertised promotional value, not necessarily the amount of money Amazon spent providing the services. Cloud companies can offer capacity at a cost below its retail price, depending on the services and available infrastructure. More importantly, the companies they support today could become valuable paying customers tomorrow.

Imagine a small team in Bengaluru building an AI application on Google Cloud. It starts with free credits, uses Google's databases and computing infrastructure, and gradually integrates more of the platform into its product. Six months later, it has thousands of users. A year later, it has paying customers, employees and a complicated product running on Google's systems. Moving everything to AWS might now require engineering work, testing, downtime risks and weeks of distraction. By giving the company free services when it had very little money, Google has improved its chances of keeping that customer for years. AWS and Microsoft follow much the same commercial logic. The credits are effectively a marketing expense that helps them win future business.

There is nothing particularly wrong with this arrangement. In fact, it has helped make technology entrepreneurship much more accessible. Fifteen years ago, an Indian founder trying to build a technology company often needed money for servers, hosting, technical staff and other infrastructure before the product had even found customers. Cloud computing removed much of that upfront burden. A team could rent what it needed and pay as the business grew. Startup credits made the process even cheaper. The difference today is that AI is changing the nature of the bill itself, and that creates a financial problem which wasn't nearly as serious for many traditional software companies.

Take a normal software product used by 1,000 customers. Adding another thousand customers might increase its server expenses, but those costs don't necessarily rise in proportion to usage. An AI business can be different because every question, generated document, voice conversation or automated task may require fresh computing work. Somebody has to pay for that work. The more customers use the product, the more it can cost the company to serve them. Some tasks are cheap, while others require expensive models, repeated attempts and considerable processing time. If a company charges everyone a fixed monthly subscription, its most enthusiastic customers can sometimes be its least profitable ones.

Consider a hypothetical Indian startup selling an AI research assistant. It earns ₹10 lakh a month from subscriptions. Its cloud services, model APIs and other direct costs would normally amount to ₹8 lakh. However, it has credits covering ₹6 lakh of those expenses, leaving just ₹2 lakh to be paid in cash. Looking only at the cash expenditure, the business appears to retain ₹8 lakh out of every ₹10 lakh of revenue before salaries, marketing and other expenses. That's an 80% cash-based gross margin. Remove the credits, and the same business retains just ₹2 lakh, or 20%. The customers haven't changed, the product hasn't changed, and revenue is exactly the same. Only the party paying the computing bill has changed.

Of course, proper financial statements may account for credits differently depending on the arrangement and applicable rules. Promotional credits can reduce recorded expenses, while certain other forms of support may receive different treatment. There is no reason to assume that Indian AI companies are automatically hiding these benefits or preparing misleading accounts. The problem arises when founders and investors discuss monthly spending, profitability or margins without separating temporary support from the normal cost of running the product. A startup could be spending very little cash on technology while still operating a business that would be expensive to run without promotional support.

This becomes especially important when companies start raising money at valuations based on their future revenue and profits. Investors generally like software businesses because, once the product has been built, additional customers can be served relatively cheaply. A business that earns ₹100 crore in revenue and keeps ₹80 crore after direct costs is naturally different from one that retains ₹30 crore. Even if both are growing at the same rate, the second company has less money available for engineers, salespeople, marketing and future development. Its value should reflect that difference. Yet AI startups are frequently discussed alongside conventional software companies, even though their costs can look quite different.

Investors have already started finding evidence of this problem. In its 2025 State of AI report, Bessemer Venture Partners examined two groups of fast-growing AI startups. The first, which it called Supernovas, included ten exceptionally fast-growing companies that averaged around $40 million in annual recurring revenue during their first commercial year and $125 million in their second. Those are extraordinary numbers, but their average gross margin was only about 25%, with some companies recording negative margins. A second group, called Shooting Stars, grew more steadily and averaged gross margins of roughly 60%. Both groups were successful by ordinary startup standards, but their financial characteristics were very different.

A separate 2026 study by ICONIQ, based on responses from more than 300 executives building AI products, found average AI-product gross margins of 45% for 2025. Respondents expected the figure to rise to 53% in 2026 and 59% in 2027. Those later numbers were forecasts, not achieved results, and the study covered a different set of businesses from Bessemer's research. Nevertheless, the two reports point in the same direction. AI products can grow quickly while keeping a smaller share of revenue after direct expenses than investors have traditionally expected from software businesses.

India is already seeing this debate among companies building AI coding tools. In March 2026, Deepak Dhanak, co-founder and chief operating officer of Indian startup Rocket, told Financial Express that many platforms in the category were operating with gross margins of 20% to 30%, while most were below 50%. He said Rocket itself maintained margins above 60% and treated margins as an important part of product development. These were his estimates and company claims, not independently audited industry statistics. But the comments are revealing because they come from someone competing in the same market. Dhanak also pointed to weak customer retention, suggesting that some users were attracted by the novelty of creating software with AI rather than a lasting need for the product.

Now consider Emergent, founded by brothers Mukund and Madhav Jha. In February 2026, the company said it had crossed $100 million in annualised revenue within eight months of launching. By July, Mukund Jha told TechCrunch that the annual run rate had reached $120 million, with more than 200,000 paying customers. Emergent also raised $130 million in a Series C round, becoming a unicorn. These numbers illustrate how quickly a new Indian AI startup can build a global business. But annualised revenue is based on the current pace of sales rather than a full year of completed revenue, and neither that growth nor the funding round tells outsiders exactly how much it costs Emergent to serve its customers. There is no public evidence here that Emergent's margins depend on free credits. The point is that revenue, paying customers and the cost of delivering the service must be examined separately.

The coding category is particularly interesting because the product itself can become expensive to provide. Someone asking an AI coding assistant to fix a minor error might consume very little computing. Another customer building an entire application could generate hundreds of requests, test multiple approaches and ask the system to keep correcting its mistakes. Both may pay the same subscription price. The company, however, incurs very different costs. A founder can respond by limiting usage, charging for additional requests or moving simpler tasks to cheaper models. But those changes can upset customers who originally signed up believing their subscription would cover almost everything.

Cursor, one of the biggest names in AI coding, faced precisely this difficulty in June 2025. The company changed its $20 monthly Pro plan from a request-based allowance to one built around a pool of model usage. Customers complained about unclear limits and unexpected charges, prompting CEO Michael Truell to apologise publicly in July and offer refunds for qualifying unexpected charges. Cursor explained that complicated requests could cost many times more than simple ones, making the old pricing difficult to maintain. The episode matters because it shows that even a successful, well-funded AI company cannot ignore the cost of its most active customers. Free credits and favourable arrangements with model providers can help, but they don't eliminate the need to charge customers enough to cover their usage.

Indian voice AI offers another example. Companies are building automated calling systems for banks, hospitals, recruitment agencies and customer-service teams. These products can be more useful than general chatbots because businesses may save real money on calls and support staff. But a ten-minute conversation requires ten minutes of speech processing, model usage and other services. Indian startup Bolna's publicly listed standard usage pricing is around ₹5.52 per minute, with different rates available for higher volumes. That is a customer price, not a disclosure of Bolna's underlying costs or profits. It nevertheless illustrates why usage-based charging makes sense. As customers consume more minutes, the supplier earns more revenue to help cover the costs of providing those minutes.

There is also a strong argument against being overly pessimistic about these businesses. AI is becoming cheaper to run. Companies can use smaller models for simple tasks, reserve expensive ones for difficult work, store previously generated answers and improve the software that manages requests. ICONIQ's 2026 research found that roughly two-thirds of surveyed companies reported improving costs per query. A startup that spends ₹10 to complete a task today may eventually be able to do it for ₹3 or ₹2, particularly if competitors keep pushing model prices down. That could make businesses profitable even if their current margins look poor.

However, lower prices do not automatically mean lower bills. As AI systems become more capable, people ask them to perform larger and more complicated jobs. Instead of summarising one document, a customer may want the system to examine fifty documents, compare the information, identify mistakes and produce a detailed report. The price of an individual model request could fall while the number of requests needed to complete the job increases sharply. In August 2026, Gartner forecast that computing costs per complex AI-agent workflow could rise more than fivefold through 2028 as these systems take on more steps. That is a forecast rather than a certainty, but it captures the problem. Cheaper computing may encourage considerably more computing.

This is where the Indian government's role becomes especially interesting. In March 2024, the Union Cabinet approved the IndiaAI Mission with an outlay of ₹10,371.92 crore. One of its major objectives was to give Indian startups, researchers and institutions access to powerful computers without forcing them to buy expensive hardware. The government organised computing capacity through private providers and offered financial support to lower the cost. By June 2026, the shared capacity had grown to more than 45,000 GPUs, the specialised processors widely used to train and run AI models. According to government figures, 237 projects had accessed subsidised computing by August, accounting for 93.18 lakh GPU-hours. Under the standard support arrangement, the government can cover 40% of eligible computing costs.

The basic idea is understandable. India cannot build a serious domestic AI industry if only companies backed by large foreign investors can afford the infrastructure. A startup working on Indian-language technology, agricultural research or healthcare applications may need powerful computers long before it can earn meaningful revenue. Government support could help such companies develop products that would otherwise never be built. It could also help Indian researchers create technology that other startups can use. Unlike a private investor, the government has reasons to support research even when the company producing it may never become highly profitable.

But an October 20 investigation by ThePrint uncovered a rather uncomfortable detail about where the support was going. Journalist Vrinda Tulsian examined 304 compute approvals listed on the IndiaAI portal as of October 8. Of approximately ₹1,275 crore in committed compute subsidies, nine organisations building large foundation models accounted for ₹1,058 crore, or roughly 83%. Researchers, students and fellows together accounted for just ₹10.14 crore, around 0.8%. The figures were based on commitments recorded on the portal, not a calculation of all cash already paid. Still, the concentration was striking, particularly for a programme intended to make computing more accessible.

One of the cases involved Zenteiq Aitech Innovations, founded by IISc professor Sashi Kumaar Ganesan. The company was listed under a research and academia category and had an approval for 2,080 Google TPUs, another type of AI processor, worth approximately ₹162.86 crore. ThePrint reported that this single approval accounted for most of the value recorded under that category. It also found that smaller academic projects often faced substantial delays, while the large model-building projects received approvals much closer to their planned start dates. The figures raise a reasonable question about whether the system is better suited to well-established organisations than to individual researchers trying to get started.

The investigation included an especially revealing example from Dhrumil Barot, founder of bootstrapped startup Vynara Technologies. Barot had developed Vynaris, a service intended to help developers access different AI models, and said it reached $1,000 in monthly revenue within days of launch. He tried obtaining computing support through IndiaAI but struggled with the payment requirements. Even after a subsidy, the remaining expense was too high for him to fund upfront. He eventually abandoned the project because he could not afford the computing required. Another founder seeking less than ₹10,000 worth of computing told ThePrint that he had been asked to provide evidence of substantial annual revenue or funding. He moved to AWS instead.

There is an obvious contradiction here. A programme designed to help founders who cannot afford computing may require those same founders to demonstrate that they already have sufficient financial resources. Some verification is necessary because government-funded capacity should not be handed out without checking applicants. But the requirements should reflect the actual stage of the businesses applying. A professor working on a promising new model or a founder testing an unproven idea cannot reasonably be expected to have the financial strength of an established technology company. Otherwise, public subsidies risk becoming another advantage enjoyed mainly by companies that already have access to capital.

That doesn't mean the government's decision to support larger model builders is necessarily wrong. Training an Indian foundation model can require far more computing than a small startup's experiments, and developing domestic models is an explicit national objective. The scale is evident in the government's February 2026 disclosure of support for selected organisations. BharatGen, the IIT Bombay-led initiative, was listed for ₹990.92 crore in compute support, Sarvam AI for ₹246.72 crore, Gnani AI for ₹177.27 crore and Soket AI for ₹162.47 crore. These are announced support values under the foundation-model programme and shouldn't be added directly to the October portal totals, which measure a different set of recorded commitments.

There are already visible results from this spending. On November 3, the government announced the release of EKA-1, a model developed by Soket Labs with 26.3 billion parameters and support for all 22 scheduled Indian languages, as well as Zenteiq's Dhīr 2B and Dhīr 5B open-source multilingual models. These products are designed for uses ranging from language processing to work in government and industry. Their release doesn't by itself prove that the programme has delivered good value for taxpayers. That requires evidence of adoption, quality and real-world usefulness. But it does show why public computing support shouldn't be judged only through the profitability of the recipients. The country may benefit from useful technology even when the organisation building it needs substantial assistance.

Another question concerns who ultimately receives the money. Companies such as Yotta Data Services, E2E Networks and other infrastructure suppliers provide the computing capacity used by startups and researchers. ThePrint's analysis found that Yotta and E2E together accounted for approximately 67% of the subsidy routed through compute providers. Their role is different from that of startups receiving discounted computing, because they are supplying the infrastructure and receiving payments for its use. Nevertheless, the arrangement means that a government programme supporting AI development also creates business for the companies renting out powerful machines. Infrastructure suppliers can earn revenue whether the startups using their services eventually succeed or fail, provided their bills are paid.

There is a similar pattern at the international level, although the amounts are vastly larger. Microsoft has invested heavily in OpenAI, while OpenAI has committed to buying enormous amounts of computing from Microsoft. Amazon has invested billions in Anthropic, which also uses AWS infrastructure. According to Microsoft's 2026 annual report, it had funded $11.9 billion of its $13 billion investment commitments related to OpenAI by June 30. The same report disclosed $24.1 billion in revenue from commercial arrangements with OpenAI, including revenue-sharing payments, during Microsoft's financial year. These figures cover different financial relationships, but they show how closely the largest AI companies and their infrastructure suppliers have become connected.

Such arrangements can make commercial sense. Developing powerful AI models requires extraordinary amounts of computing, and the companies building them need reliable access to infrastructure. Suppliers have an incentive to finance promising customers who may eventually purchase billions of dollars of services. But there is also a risk when investors, suppliers and customers depend heavily on one another. If future demand falls short of expectations, a model developer could struggle with computing commitments while its investor and infrastructure provider simultaneously lose a valuable customer. Bloomberg examined these relationships in May 2026, describing how investments and purchasing agreements had tied the fortunes of major AI companies together. This is not evidence that the transactions are improper, but it shows how financial support and commercial demand can become difficult to separate.

India has seen an earlier version of the broader subsidy problem in food delivery, ride-hailing and online commerce. Companies spent heavily to attract customers, often offering discounts that made services cheaper than they could sustainably provide. The hope was that customers would stay, usage would increase and the businesses would become profitable as they grew. Some companies improved their economics considerably, while others struggled when incentives were reduced. AI credits are different because they subsidise the cost of supplying the service rather than directly lowering the customer's price. But the underlying question is familiar. Does the business make sense when the outside support stops?

The dangers of relying too heavily on cheap cloud infrastructure aren't new either. Back in 2020, Andreessen Horowitz partners Martin Casado and Matt Bornstein warned that AI companies often had lower margins than traditional software businesses because of their computing requirements and the human work needed to make products reliable. In 2021, the investment firm examined how cloud spending could become a major burden as companies grew. It highlighted Dropbox, which had moved substantial workloads onto its own infrastructure and reported approximately $75 million in cumulative savings over two years before its IPO. Dropbox's gross margin rose from 33% in 2015 to 67% in 2017, with infrastructure improvements playing a major role. The lesson wasn't that every startup should own its servers. It was that infrastructure decisions made early in a company's life can become expensive to reverse.

For Indian AI startups, there may eventually be three different outcomes. Some will use credits to build a product, find customers and improve their technology sufficiently that they can pay their full bills without difficulty. Others will discover that their initial prices were too low and introduce usage limits, higher subscriptions or charges based on completed work. A third group may find that customers simply aren't willing to pay enough to cover the underlying costs. Those companies could struggle even with growing revenue because every additional customer brings more expenses without enough additional income.

The challenge for investors is to identify which companies belong in which group before the credits expire. We would start by asking founders for their monthly infrastructure usage before promotional deductions, the amount actually paid, the credits still available and their expiry dates. We would also want to understand how expensive the most active customers are to serve. A company with a thousand subscribers may look profitable on average, but the busiest hundred could be consuming most of the revenue. It is equally important to distinguish money spent developing new models from the recurring expense of serving paying customers. A company investing heavily in research is different from one that loses money each time an existing customer asks a question.

We would also examine whether the product has any meaningful advantage beyond access to an existing AI model. A startup that simply places a convenient interface over a model supplied by another company faces a difficult problem. If its supplier increases prices, its expenses rise. If the supplier lowers prices, competitors may pass those savings to customers. And if the model provider launches a similar product itself, the startup could find its main service competing directly with its supplier. Companies that solve complicated customer problems, build useful integrations or develop specialised expertise may have a better chance of protecting their revenue and improving their margins.

None of this means India's AI startup boom is built on an illusion. Many companies are finding genuine customers, solving expensive business problems and creating products that would have been almost impossible to develop a few years ago. Free computing has helped make that experimentation possible. The government also has legitimate reasons to support Indian AI research, particularly in languages and areas that may not attract enough private investment. The concern is not that subsidies exist. It is that founders, investors and sometimes even customers can get used to economics that depend on those subsidies continuing.

We also shouldn't assume that a young startup with weak margins is automatically a bad investment. Some of the best companies in the world spent years losing money while developing technology and building customer relationships. The relevant question is whether there is a believable path to profitability. If the cost of serving customers is falling, customers are willing to pay more and the product becomes more valuable with use, temporary credits can be extremely useful. If the company needs another grant every time its free allocation runs out, the situation is quite different.

The Bengaluru resale episode is a useful reminder of how unusual the present moment has become. Founders were offered computing benefits worth thousands of dollars before they had built anything, and some tried to convert those benefits into cash. Meanwhile, much larger startups are receiving substantial support from cloud companies and model providers, while the Indian government is committing public money to expand access to computing. The result is an ecosystem in which the cost of building and running AI products is being shared among founders, investors, technology suppliers and taxpayers.

For most of the last decade, Indian startup investors have spent considerable time asking whether companies can acquire customers without losing money on discounts and incentives. AI introduces another question that may prove just as important. Once customers have been acquired, can the company afford to serve them at the price it charges?

The answer will become clearer as the first large batches of promotional credits expire and companies begin paying their normal bills. Some startups may find that their businesses are stronger than expected because model costs have fallen and customers are using their products more profitably. Others may discover that the attractive margins they discussed during fundraising were partly the result of temporary support. At present, there is no reliable public data showing how many Indian AI startups fall into either category.

That uncertainty is precisely why investors should be asking for the numbers now rather than waiting for a funding slowdown to expose the problem. Revenue growth is encouraging, but it means very little if the company cannot eventually earn a reasonable profit on the service it sells. India has already experienced one generation of startups whose growth depended heavily on someone else subsidising the customer. It would be unfortunate if the industry repeated the same mistake with AI, only this time with the subsidy hidden inside the computing bill.