Every startup program from the big three clouds leads with its largest number: up to $200,000 from AWS Activate, up to $350,000 from Google Cloud for AI startups and, until Microsoft rebuilt its program in July 2025, up to $150,000 in Azure credits. Those figures behave like the headline numbers in deal announcements, and anyone who has read a breakdown of how "valued at" and "up to" figures drift away from the money that actually changes hands will recognize the shape: the first number is a ceiling, and what you collect depends on conditions printed further down. For developers the gap matters more than it does for readers of business news, because credits do more than shrink a bill. They decide which services you adopt, how large you size things and what your stack costs on the morning they run out. Fortunately, each of those conditions reduces to arithmetic.

The Arithmetic Behind the Headline

Google publishes the most detailed formula, so start there. Its Scale tier for venture-backed startups covers 100% of eligible usage up to $100,000 in the first year, then 20% of usage, up to another $100,000, in the second. AI-first startups get a first-year cap of $250,000 and the same second year, which is where the $350,000 comes from.

Read the second-year line again. A 20% match capped at $100,000 pays out in full only when second-year usage reaches $500,000, which means the last $100,000 of credit arrives after you have paid $400,000 of your own. Collecting the entire $350,000 takes $750,000 of usage in two years, a curve most seed-stage teams never draw. A few lines of Python make the gap concrete:

def realized(y1_usage, y2_usage, y1_cap=100_000, y2_pct=20, y2_cap=100_000): """Credits a two-year offer actually pays, and the cash you still spend.""" credits = min(y1_usage, y1_cap) + min(y2_usage * y2_pct // 100, y2_cap) return credits, y1_usage + y2_usage - credits print(realized(60_000, 120_000)) # (84000, 96000) print(realized(100_000, 500_000)) # (200000, 400000) print(realized(250_000, 500_000, y1_cap=250_000)) # (350000, 400000)

A team that spends $5,000 a month in year one and $10,000 a month in year two collects $84,000 of a "$200,000" offer, or 42%. That is real money and worth having. It is also the number to plan around, and it falls further if usage ramps late, because the first-year allowance is tied to first-year usage.

The Big Number Belongs to Funded Companies

The second condition is who you are, and on all three clouds the answer comes from your cap table. AWS starts bootstrapped and self-funded founders at $1,000 in Activate credits, with up to $5,000 over time, while the Portfolio package of up to $200,000 requires an Organization ID from an Activate Provider such as an accelerator, angel investor or venture firm. Google keeps its Scale tiers for startups with institutional equity: SAFEs qualify, but angel checks, friends-and-family money, grants and crowdfunding do not, and a Series A has to be less than a year old.

Microsoft made the sharpest turn. On July 2, 2025, it retired the Founders Hub name along with the open, tiered path that once let a founder with little more than an idea work toward $150,000 in Azure credits. Without an investor referral, the offer is now up to $5,000: $1,000 valid for 90 days, plus $4,000 valid for 180 days once the business is verified. With a referral from an affiliated investor, it starts at $100,000.

Put together, a credit offer prices your investors more than your product. That is a reasonable filter from the provider's side, since venture backing predicts which startups grow into large accounts, but a bootstrapped team should read every "up to" on these pages as written for someone else.

A Credit Dollar Is Not a Cash Dollar

Even under the ceiling, a dollar of credit is worth less than it looks, for three reasons.

It expires on someone else's schedule. AWS says its credits usually expire within one to two years, depending on the package, and apply only to new eligible usage, never to past bills. Nothing pauses when the balance hits zero; usage simply starts landing on your card. AWS at least sends warnings 60 and 30 days before expiry and again when 75% and 100% of the balance is gone.

It covers less of the cloud than the name suggests. AWS excludes services such as Mechanical Turk, Professional Services and Training and Certification, but lets Activate credits pay for third-party foundation models in Amazon Bedrock. Google's standard program credits cover Gemini and Gemma, while third-party models are billed directly, apart from a separate allowance of up to $10,000 that AI-tier startups may qualify for. If your product calls a third-party model through Vertex AI, the largest line on your bill can sit entirely outside the $350,000.

It is spent at prices you would never pay. Facing a real bill, teams rightsize instances, buy commitments, move batch work to interruptible capacity and question every managed service. Under credits none of those habits form, so the architecture you settle on is the one that is cheapest when compute is free, which is rarely the one that is cheapest when it isn't.

Leaving Got Cheaper. Your Architecture Didn't.

The usual fear about credits is lock-in, and the price of the exit itself has fallen a long way. Google moved first, and its January 2024 announcement that customers leaving Google Cloud would stop paying data transfer fees was matched by AWS on March 5 and Microsoft on March 13. In the EU, the Data Act goes further: from January 12, 2027, providers may not charge for the switching process at all, egress included, although data flowing between clouds you run in parallel can still be billed at cost.

The same Google post contained a more revealing admission. Removing the fee, Google wrote, "does not solve the fundamental issue" that keeps many customers from moving, and Google meant software licensing. For a startup, the fundamental issue is usually closer to home. Copying terabytes takes a weekend. Rewriting code that leans on a proprietary database, a managed queue, a vendor identity system or a model endpoint with its own request format takes a quarter, and those are exactly the services that are easiest to say yes to when someone else is paying.

The Same Two Words at the Top of the Market

Startup credits are the retail version of something the AI industry now does at enormous scale: swapping investment for compute commitments and announcing the ceiling. In September 2025, Nvidia said it intended to invest up to $100 billion in OpenAI in tranches, one for each gigawatt of Nvidia systems OpenAI deployed. Four months later, The Wall Street Journal reported that the megadeal was on ice, Jensen Huang said the figure had never been a commitment, and Nvidia went on to invest $30 billion in OpenAI's next funding round as a plain equity deal. In November 2025, Anthropic committed to buy $30 billion of Azure compute while Microsoft and Nvidia committed to invest up to $5 billion and up to $10 billion in the company. In both announcements, investment flows one way between the same companies and purchase commitments flow the other.

Nobody hid the qualifiers; "intends" and "up to" were right there in the press releases. The lesson for a developer is the one deal pages keep teaching: the headline tells you the most that could move, and the structure tells you what will.

Run a Shadow Bill From Day One

The fix is a habit rather than a spreadsheet: every month, look at what you would have paid without credits next to what you actually paid. On AWS, Cost Explorer can split a month by record type, which puts credits on their own negative line:

aws ce get-cost-and-usage \ --time-period Start=2026-10-01,End=2026-11-01 \ --granularity MONTHLY \ --metrics UnblendedCost \ --group-by Type=DIMENSION,Key=RECORD_TYPE \ --query 'ResultsByTime[0].Groups[].[Keys[0], Metrics.UnblendedCost.Amount]' \ --output table

Everything except the Credit row is your shadow bill. On Google Cloud, the billing export to BigQuery stores each line's gross cost separately from a credits array, with promotional credits itemized under the type PROMOTION, so the same view is a single query away. Put the shadow bill on a dashboard you already check, next to latency and error rates, and the cliff stops being a surprise.

Before accepting an offer, or the next tranche of one, pin down six numbers:

  • Realized value: your usage forecast run through the offer's actual formula, not its headline.
  • Expiry dates: when each tranche lapses, and whether your usage will reach it before then.
  • Excluded spend: the services and third-party models the credits won't touch.
  • Cash-rate cost: what your current architecture would cost per month with commitments and rightsizing, paid in real money.
  • Exit cost: the engineering weeks needed to move your stateful services, not just your bytes.
  • Cliff month: the month the shadow bill becomes the real one, marked on the calendar with a budget alert set well ahead of it.

Take the Credits, Price the Promise

None of this argues against cloud credits. If you would have spent the money anyway, $84,000 of free compute is $84,000 of runway, and the providers have sound reasons of their own for handing it out. The mistake is letting the headline stand in for the value, then letting free compute pick an architecture you can only afford while it lasts. Discount every offer by who funded you, how fast you will grow into it, what it leaves out and when it expires, and build as though the bill were already yours. The number on the program page is the most you could ever collect. The shadow bill is the company you are actually building.