The most aggressive customer-acquisition weapon in artificial intelligence right now is not a product feature. It is free money, in the form of compute credits, and the amounts have quietly become enormous.
According to Wall Street Journal reporting now circulating widely among founders and investors, OpenAI is offering every Y Combinator startup $500,000 in free token credits with no equity attached, with an option of a further $1.5 million in exchange for equity. Anthropic has reportedly raised its own YC offer to $500,000 in API credits, up from $30,000. Google is offering up to $500,000 in cloud credits plus Gemini early access and DeepMind engineering support. Some startups are said to be stacking more than $3 million in combined credits, roughly a US median seed round, before raising a dollar.
🛑 US AI giants enter price war for startup customers
— jck✨ (@Alea_) July 7, 2026
🛑 The Wall Street Journal reports AI leaders are offering large free compute and API/token credits to capture startup share; some startups say they have received more than $3m in combined cloud and token credits, roughly…
The land-grab is a lock-in strategy
None of this is charity. A startup that builds its product on subsidized tokens embeds one provider’s models, data formats and workflows into its architecture. By the time the credits run out, switching means re-engineering. It is the same compounding dynamic we described in enterprise deployments, where the real cost of agentic AI is vendor lock-in, playing out one layer down the stack, at the startup level. The hyperscalers know this playbook well: free AWS credits onboarded a generation of startups that never left.
The founder playbook
Taking the money is rational. Taking it naively is not. Three moves preserve leverage: stack credits from more than one provider and keep the architecture model-agnostic from day one; treat the credit expiry date as a migration deadline and price the switch before accepting; and negotiate data portability while the provider is still courting you, because leverage never gets better than the moment before you sign.
What the subsidy math says about the sellers
Read from the providers’ side, the escalation from $30,000 to $500,000 offers is a confession about switching costs. If moving between model APIs were genuinely hard, no one would need to pay startups half a million dollars to choose a side; the technical truth is that swapping one model’s API for another is often a weekend’s work, and the real lock-in has to be manufactured through accumulated context, fine-tuning investments, and workflow integrations that deepen over time. The credits buy the time for that accumulation to happen. It is customer acquisition priced like enterprise sales, aimed at companies that do not yet have revenue.
The accounting deserves its own scrutiny. Credits cost the provider marginal inference capacity but book as customer traction, and a cohort of startups whose usage is subsidized is a cohort whose demand signal is partly synthetic, worth remembering when usage growth statistics appear in IPO narratives. History says the strategy works anyway: AWS credits onboarded a generation that never left, because by expiry the migration cost exceeded the annoyance of staying. Whether that repeats depends on the one variable the providers cannot subsidize, how good open-weight alternatives are when the credits lapse, and that frontier keeps moving.
What to watch
Watch whether credit offers keep escalating into 2027, which would signal the model providers still cannot win on switching costs alone, and watch for the first prominent startup to publicly migrate providers when its credits lapse. That story will tell the market what these subsidies actually bought. The pricing side of the same strategy is already visible in Anthropic’s introductory pricing on Claude Sonnet 5. The capital side of it shows up in funding rounds too: Cognition’s $26 billion valuation is, in part, a bet that developer habits formed on subsidized usage are hard to unwind later.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
