It is no longer "what can we run for free?" It is "which free dollars leave us free to leave?"
For a decade, "free tier" meant roughly the same thing everywhere: a year of small virtual machines, a few gigabytes of object storage, and a permanent allowance on serverless functions. That model is gone. In July 2025, AWS replaced its 12-month free tier for new accounts with a credit plan capped at $200 and six months. In June 2026, Oracle cut its Always Free Arm compute allowance in half without a blog post or customer notice. Google Cloud's $300 trial credit can no longer pay for Gemini API usage in AI Studio or for partner models.
The money has not disappeared. It has moved from automatic allowances into application-based startup programs worth $150,000 to $350,000, and into storage providers that compete on egress rather than on free gigabytes. For infrastructure teams, that shift changes the question. It is no longer "what can we run for free?" It is "which free dollars leave us free to leave?" This guide covers what each provider actually offers new accounts in 2026 across compute, storage, and GPU, and where the fine print quietly turns a credit into a lock-in.
Accounts created on or after July 15, 2025 get $100 in credits at signup and can earn up to $100 more by completing activities in services such as Amazon EC2 and Amazon Bedrock, according to AWS's announcement. At signup you choose a Free plan or a Paid plan. The Free plan ends after six months or when the credits run out, whichever comes first, and it cannot reach every AWS service. AWS gives you 90 days after the free period to reopen the account by upgrading. More than 30 services keep an always-free allowance on either plan. Accounts opened before the cutover stay on the legacy 12-month structure.
One detail matters for startups: AWS Activate, the startup credit program, lists an account on the Paid plan as an eligibility requirement. The Free plan and startup credits do not mix.
Oracle Cloud Infrastructure's Always Free tier was the outlier for years: four Arm-based Ampere A1 cores and 24 GB of memory with no expiry. Oracle's documentation now lists 1,500 OCPU hours and 9,000 GB hours per month for the A1 shape, which Oracle states is equivalent to 2 OCPUs and 12 GB of memory. InfoQ reported that the change took effect on June 15, 2026 with no public announcement, and that support agents gave conflicting answers about whether pay-as-you-go tenancies were affected. The rest of the Always Free set is intact: two AMD micro VMs, 200 GB of block storage, and 10 TB of outbound data transfer per month.
Google Cloud still offers $300 over 90 days, but its documentation now excludes the Gemini API in AI Studio and partner models sold as managed APIs from what the credit can pay for. While an account is still a Free Trial billing account, you cannot add GPUs to VMs, use Google Cloud Marketplace, or request a quota increase. If you neither upgrade nor stop, the trial account closes, resources stop, and after a 30-day grace period they are deleted. Azure offers $200 to spend in the first 30 days, plus 12 months of free amounts on popular services and more than 55 always-free services, but only if you move to pay-as-you-go before the credit window closes.
| Provider | New-account credit | Always-free compute | When the credit ends |
|---|---|---|---|
| AWS | $100 at signup, up to $200 total | 30+ always-free services (for example Lambda); EC2 runs on credits | Free plan closes at 6 months; 90 days to reopen by upgrading |
| Google Cloud | $300 for 90 days | 1 e2-micro VM (us-west1, us-central1, us-east1), 30 GB standard disk | Resources stop; deleted after 30-day grace period |
| Azure | $200 for 30 days | 12 months of popular services, then 55+ always-free services | Subscription disabled unless moved to pay-as-you-go |
| Oracle Cloud | Trial credit plus Always Free | 2 Arm OCPUs + 12 GB; 2 AMD micro VMs; 200 GB block | Always Free continues; idle instances can be reclaimed |
Put side by side, the programs split into two types. AWS, Google Cloud, and Azure lead with a time-boxed credit and keep a thin always-free layer underneath it. Oracle leads with a permanent allowance large enough to run real services, which is precisely why it was cut. The always-free compute on the hyperscalers is sized for a status page or a cron job, not an application: Google Cloud's permanent allowance is one e2-micro instance in one of three US regions with 30 GB of standard persistent disk.
The credits themselves also behave differently at expiry. Google Cloud stops trial resources and deletes them after the grace period. AWS closes Free plan accounts at the end of six months. Azure disables the subscription when the 30-day credit ends unless you upgrade. None of them bill you by surprise on the trial, but all of them put your data on a deadline.
Free storage allowances are small everywhere. Google Cloud gives 5 GB-months of regional storage in three US regions. Backblaze B2 makes the first 10 GB free permanently. Cloudflare R2 gives 10 GB-months, 1 million Class A operations, and 10 million Class B operations every month, on the Standard storage class only. Oracle's Always Free tier includes 20 GB of object storage once the trial ends. Wasabi has no permanent free tier, only a 30-day trial.
None of those numbers will carry a production dataset. What carries forward is the egress policy, because it decides what happens when the credits expire and you want to move. R2 charges nothing for egress on any storage class. Backblaze B2 includes free egress up to three times your average monthly stored data, then $0.01 per GB, with unlimited free egress to partners including Cloudflare, Fastly, CoreWeave, and Vultr. AWS includes 100 GB of free data transfer out per month, aggregated across all services and regions, then starts at $0.09 per GB. Oracle's 10 TB of monthly outbound transfer is the largest free egress allowance of any hyperscaler by two orders of magnitude.
| Provider | Free storage | Egress policy | The catch |
|---|---|---|---|
| Cloudflare R2 | 10 GB-month, 1M Class A, 10M Class B ops | $0 egress on all storage classes | Free tier covers Standard class only |
| Backblaze B2 | First 10 GB | Free up to 3x average stored data, then $0.01/GB; unlimited to CDN and compute partners | Partner list decides your free routes |
| Wasabi | 30-day trial only | Free while egress stays within stored volume | 1 TB minimum and 90-day minimum storage duration |
| Amazon S3 | No permanent allowance for new accounts (credits apply) | 100 GB/month free, then from $0.09/GB | Exit waiver only when moving all data off AWS |
| Google Cloud Storage | 5 GB-month (3 US regions) | 100 GB/month from North America, then standard rates | Free Tier limits can change with 30 days notice |
| Oracle Object Storage | 20 GB (Always Free only state) | 10 TB/month outbound | Objects above 20 GB deleted if trial ends without upgrade |
We covered the economics behind this in our podcast with Backblaze CEO Gleb Budman: egress fees are the toll booth that turns cheap storage into expensive storage the day you try to leave.
No hyperscaler free tier includes GPUs. Google Cloud's documentation says so directly: GPUs and TPUs are not included in the Free Tier, and Free Trial accounts cannot attach them at all. Oracle's Always Free set has no GPU shapes. The AWS Free plan does not include every service, and new AWS accounts typically need a service quota increase before they can launch GPU instances at all. For GPU work, "free" means an application-based credit program.
The main programs in 2026:
Headline credit amounts are denominated in the issuing cloud's list prices. AWS's p5.48xlarge, with eight NVIDIA H100 GPUs, lists at $55.04 per hour on demand in us-east-1, or $6.88 per GPU-hour. A $100,000 Activate grant spent there buys about 14,500 H100 hours. Neoclouds publish H100 rates at less than half that. Thunder Compute, a GPU provider that publishes comparisons against AWS, lists its H100 at $3.20 per hour, which would stretch the same $100,000 to 31,250 hours. Treat any vendor's comparison with the scepticism it deserves, and check current rates on our GPU Price Explorer before you plan a training run around a credit balance. The point stands either way: a $150,000 neocloud credit can be worth more compute than a $200,000 hyperscaler credit.
Credits are designed to be spent where the data lands. Once a training corpus, a feature store, or a vector index is sitting in one provider's object storage, every later decision is priced against the cost of moving it. That is the real product being sold.
Take a 200 TB dataset built up on Amazon S3 during an Activate credit period. At S3 Standard list rates in us-east-1, it costs about $4,450 per month to store once the credits are gone. Moving it once to a neocloud for a training run costs about $13,800 in data transfer out at AWS's standard internet tiers, after the 100 GB free allowance. That single transfer is worth 69 times the maximum AWS Free Tier credit. A team that follows GPU availability between providers every quarter would pay roughly $55,000 a year in egress on that dataset alone.
AWS and Google Cloud both waive exit egress, but read the terms. AWS's program covers data transfer out when you move all of your data off AWS (or the entire usage of a single service), requires you to finish within 60 days, and requires you to delete what remains. Google Cloud's program requires you to terminate your agreement. Neither covers the multi-cloud pattern most AI teams actually run, where the data stays and the compute moves.
The same 200 TB costs about $3,000 per month on Cloudflare R2 with no egress, about $1,390 on Backblaze B2 with up to 600 TB of free egress per month, and about $1,598 on Wasabi at its published starting rate, where egress is free as long as it stays within your stored volume. The storage decision you make during the credit period sets the price of every compute decision after it.
| Storage provider | Monthly storage | Move 200 TB out once | Free egress headroom |
|---|---|---|---|
| Amazon S3 Standard | $4,450 | $13,795 | 100 GB per month |
| Cloudflare R2 | $3,000 | $0 | Unlimited |
| Backblaze B2 | $1,390 | $0 | 600 TB per month (3x stored) |
| Wasabi | $1,598 | $0 | Up to 200 TB per month (equal to stored) |
Spend compute credits wherever they are largest, but land datasets, checkpoints, and model artifacts on zero-egress or low-egress object storage from day one. S3-compatible APIs on R2, B2, and Wasabi mean the application code does not change.
Credits expire whether you use them or not. Map each grant's expiry date before you apply for the next one, so two large balances are not burning down in the same quarter while your workload can only consume one.
AWS emails Free plan accounts when 50, 25, and 10 percent of credits remain. That is a floor, not a plan. Set a budget at 100 percent of expected monthly spend on every account, including trial accounts, so the day credits end is not the day finance discovers the run rate.
Google Cloud reserves the right to change Free Tier limits with 30 days of notice. Oracle changed its limits with none. Oracle also reclaims Always Free instances that sit idle for seven days. Anything that must stay up belongs on paid capacity with a known price.
Not for new accounts. Accounts created on or after July 15, 2025 receive $100 in credits at signup and can earn up to $100 more. The Free plan lasts up to six months or until the credits are used. More than 30 services keep an always-free monthly allowance. Accounts created before that date keep the legacy 12-month structure.
Oracle Cloud still offers the largest permanent free compute: 2 Arm OCPUs with 12 GB of memory plus two AMD micro VMs, 200 GB of block storage, and 10 TB of monthly egress. Google Cloud's permanent offer is one e2-micro VM. AWS and Azure lead with time-limited credits of $200.
No. Hyperscaler free tiers exclude GPUs, and Google Cloud Free Trial accounts cannot attach GPUs at all. Free GPU compute comes from startup credit programs, including AWS Activate (up to $200,000), Google for Startups (up to $350,000 for AI-first startups), Microsoft for Startups (up to $150,000), and Nebius AI Lift (up to $150,000 for NVIDIA Inception members).
Cloudflare R2 includes 10 GB of storage per month with zero egress on all usage. Backblaze B2 makes the first 10 GB free and includes free egress up to three times average monthly storage. AWS includes 100 GB of free data transfer out per month across all services, and Oracle includes 10 TB.
It depends on the provider. Google Cloud stops trial resources and deletes them after a 30-day grace period if you do not upgrade. AWS closes Free plan accounts after six months and allows 90 days to reopen by upgrading. Azure disables the subscription unless you move to pay-as-you-go. Moving the data elsewhere is billed as egress unless the provider's rules say otherwise.
Free credits are priced to get your data in. Choose where that data lives as if you will have to pay to take it out, because eventually you will.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
