billiem

Home /

Free compute worth claiming in July 2026

A July 2026 field guide to free VPS, serverless, databases, product plumbing, developer tools, AI APIs, and GPU compute.

  • OCI remains the strongest persistent free VM offer, but its current official pages disagree on whether Ampere includes 2 OCPUs and 12 GB RAM or twice that.
  • Agent sandboxes and browser automation now have useful recurring free tiers from Upstash, Vercel, Freestyle, and Cloudflare.
  • Free AI compute includes several different products: metered GPU credit, shared notebooks, public demo capacity, and hosted model APIs.
  • Recurring hard caps, billing-linked allowances, dynamic capacity, trials, and application-only credits should not be treated as equivalent.

This is the 10 July 2026 snapshot of free infrastructure worth claiming for side projects. It may become a monthly thing. The simple version is: here is some free cool stuff online.

There are repeats from the June free-compute guide. That is deliberate. A useful free VPS does not stop being useful because it appeared last month. This pass stays just as broad, from free VPSes and Lambda-style workers through databases, storage, email, observability, CI, and especially free AI compute.

Free still needs a qualifier. Some plans stop at a hard limit. Some sit on billing accounts and charge for overages. Some are shared capacity with limits that move. Others are trials or application-only credits. Those are separated below rather than presented as equivalent.

Every numeric claim was checked against an official provider page on 10 July. This is a documentation-backed guide, not a record of opening and testing every account. Quotas, regions, card checks, idle policies, and data-use terms can all move, so follow the linked provider page before building around an exact number.

What changed or turned up in July

The short version

  • Claim OCI Always Free if you want a persistent Linux VM and can get capacity in your home region.

  • Put Cloudflare in front of almost everything. Workers, Pages, R2, D1, KV, Queues, Durable Objects, Workflows, Turnstile, Browser Run, and Workers AI now cover an absurd amount of a small product.

  • Use Neon, Aiven, or Turso for data. Add Upstash when the project needs Redis, queues, workflows, vectors, search, or a small remote execution box.

  • Use Modal when $30 of recurring serverless CPU/GPU credit is more useful than a notebook. Use Lightning AI when you want an interactive GPU studio.

  • Use Cloudflare R2 for S3-compatible objects when egress matters, but know that enabling R2 requires a billing profile even when usage stays within the free allocation.

  • Wire in Grafana Cloud, Better Stack, Sentry, or Axiom before the prototype becomes mysterious.

  • Keep trials and application programmes in reserve. A recurring hard-capped allowance is infrastructure. A six-month account credit is runway.

VPS and hyperscaler compute

  • Oracle Cloud Infrastructure: Always Free. Up to two AMD E2.1.Micro VMs, plus 1,500 Ampere A1 OCPU-hours and 9,000 GB-hours per month, equivalent to 2 OCPUs and 12 GB RAM total. The account also includes 200 GB combined boot/block storage and 10 TB monthly outbound transfer. Compute must live in the home region, capacity shortages are real, and idle instances can be reclaimed. Most sign-ups need a phone and card.

  • Google Compute Engine: Recurring, card-backed. One non-preemptible e2-micro instance-month, 30 GB-month standard persistent disk, and 1 GB outbound per month in us-west1, us-central1, or us-east1. It is a tiny general-purpose VM. GPUs and TPUs are not included.

  • AWS EC2 T4g: Promotion through 31 December 2026. All customers can use 750 aggregate t4g.small instance-hours per month. The VM has 2 vCPU and 2 GiB RAM, but EBS, public IPv4, surplus CPU credits, and non-covered images can still cost money. It becomes paid after the promotion on a paid account.

  • AWS Free plan: New-account credit. $100 at sign-up and up to another $100 earned through activities. The account closes or needs upgrading after six months or when the credit is exhausted. Lightsail separately gives selected VPS, container, and database bundles three free months.

  • Azure free account: New-account trial. $200 for 30 days, then 750 monthly hours each of selected Arm and AMD burstable VMs for 12 months. The useful long-lived Azure pieces are in its serverless and database allowances rather than permanent VPS hosting.

  • Google Cloud trial: New-account credit. $300 for 90 days. It cannot be used to attach GPUs or call partner model APIs, and trial resources stop when the time or credit runs out unless the account upgrades.

  • DigitalOcean, Akamai Cloud, OVHcloud UK, and Vultr: Trials, not free VPSes. The current official offers are $200 for 60 days, $100 for 60 days, £175 for one month, and a time-limited Vultr promotion respectively. Active resources become billable after the credit. These are useful for testing a provider or running a short burst, not for permanent zero-cost hosting.

Functions, containers, edge, and PaaS

  • AWS Lambda: Recurring, card-backed. 1 million requests and 400,000 GB-seconds per month. The allowance survives beyond the six-month new-account plan when the account upgrades to paid, so set budgets and watch adjacent services such as API Gateway, logs, storage, and data transfer.

  • Google Cloud Run: Recurring, card-backed. Request-based services get 2 million requests, 180,000 vCPU-seconds, and 360,000 GiB-seconds per month. Instance-based billing gets 240,000 vCPU-seconds and 450,000 GiB-seconds. The allowance is aggregated across projects on the billing account.

  • Google Cloud Run functions: Recurring, card-backed. First-generation functions include 2 million invocations, 400,000 GB-seconds, 200,000 GHz-seconds, and 5 GB outbound per month. Deployments can still create Artifact Registry, build, or storage charges.

  • Azure Functions: Recurring, card-backed. Classic Consumption includes 1 million executions and 400,000 GB-seconds per month. Flex Consumption has a smaller 250,000-execution and 100,000-GB-second grant. Storage and networking are separate.

  • Azure Container Apps: Recurring, card-backed. 2 million requests, 180,000 vCPU-seconds, and 360,000 GiB-seconds monthly. Scale to zero matters. Minimum replicas create idle charges, and Azure's serverless GPU option has no free GPU allocation.

  • IBM Cloud Code Engine: Recurring, card-backed. 100,000 vCPU-seconds, 200,000 GB-seconds of memory, and 100,000 HTTP requests monthly, shared by apps, jobs, functions, and builds. This is the current IBM serverless product; the June Cloud Functions entry is obsolete.

  • Cloudflare Workers and Pages: Recurring hard caps. Workers gets 100,000 dynamic requests per day, 10 ms CPU per invocation, and 128 MB memory. Static asset requests are free and unlimited. Pages gets 500 builds per month, while Pages Functions share the Workers allowance.

  • Cloudflare Workflows: Recurring hard cap. 3,000 steps per day and 1 GB state storage on Workers Free. This is useful for durable multi-step jobs without introducing another provider.

  • Scaleway Functions: Recurring, card-backed. 1 million requests and 400,000 GB-seconds per month. Scaleway Containers and Jobs each add 200,000 vCPU-seconds and 400,000 GB-seconds. Warm or provisioned instances are charged continuously, so keep scale-to-zero workloads genuinely idle.

  • Vercel Hobby: Recurring hard caps for personal, non-commercial work. Functions include 4 active CPU-hours, 360 GB-hours memory, and 1 million invocations per month. Hobby projects pause at the caps instead of buying overage. It remains a very good Next.js demo platform and a poor fit for a commercial product pretending to be a hobby.

  • Netlify Free: Recurring shared credit pool. The plan has 300 credits per month. Spending the whole pool on one meter would buy 30 GB-hours of compute, 15 GB bandwidth, 1.5 million web requests, or 20 production deploys. In practice, every meter draws from the same pool and the projects pause when it is gone.

  • Deno Deploy: Recurring, card-verified. 1 million requests, 20 GB egress, 15 CPU-hours, 350 GB-hours memory, 1 GiB volume storage, and 20 apps each month. The current platform has US and EU regions. Anyone still using Deploy Classic has until 20 July 2026 to migrate.

  • Koyeb: Recurring with awkward verification. One 512 MB, 0.1 vCPU, 2 GB SSD web service in Frankfurt or Washington. It sleeps after one idle hour, cannot be a worker or use a persistent volume, and Koyeb documents a temporary $29 card hold.

  • Render: Recurring non-production compute. Free 512 MB, 0.1 CPU web services share 750 running hours per workspace each month. They spin down after 15 idle minutes, cold-start slowly, and have ephemeral disks. Without a payment method, over-limit services suspend instead of billing.

  • Railway: Recurring $1 credit. The Free plan caps each service at 0.5 GB RAM, 1 vCPU, and a 0.5 GB volume. New accounts first get a $5, 30-day trial. Free deployments can be blocked during regional peaks, so this is a tiny-bot allowance rather than dependable free hosting.

  • Northflank Developer Sandbox: Recurring but vaguely sized. Two services, two jobs or cron jobs, and one addon/database, advertised as always-on without sleeping. A payment method is required and the official pages describe the compute only as limited, so there is no honest CPU/RAM figure to quote here.

  • Fly.io: Very short trial. Two total VM-hours or seven days, whichever comes first. There is no durable free plan for new users.

Agent sandboxes and browser compute

  • Upstash Box: Recurring Developer Preview. Ten concurrent boxes and 5 active CPU-hours per month. Each box has 2 vCPU, 4 GB RAM, and 5 GB disk, freezes after one idle hour, and currently runs only in us-east-1. The free plan also includes $1 of built-in LLM use. APIs and pricing may change.

  • Vercel Sandbox: Recurring, personal/non-commercial. 5 active CPU-hours, 420 GB-hours memory, 5,000 creations, 20 GB network, 15 GB snapshot storage, and ten concurrent Firecracker microVMs per month. Hobby sessions can run for up to 45 minutes. This is one of the better free places to execute untrusted or agent-generated code.

  • Freestyle VMs: Recurring hard cap. 20 vCPU-hours, 40 GiB-memory-hours, and 16,800 GiB-storage-hours per day, with ten concurrent and 50 saved VMs. Free VMs and snapshots are not persistent, which makes this disposable agent compute rather than a VPS.

  • E2B Hobby: One-time credit. $100 of usage, no card, one-hour sessions, up to 20 concurrent sandboxes, and up to 8 vCPU/8 GB RAM. The account blocks when the credit is gone.

  • Daytona: One-time or unclear credit. $200 of advertised compute without a card, usable against CPU and GPU sandboxes. The public pricing page does not state an expiry, so it does not belong in the recurring-free group.

  • Cloudflare Browser Run: Recurring hard cap. Ten browser minutes per day and three concurrent sessions, plus up to five crawl jobs per day. It is enough for small screenshot, PDF, crawling, and browser-agent jobs. The API returns 429 when the daily browser allowance is gone.

Relational and SQLite databases

  • Neon: Recurring, no card. Up to 100 projects, each with 100 CU-hours, 0.5 GB storage, and compute up to 2 CU/8 GB RAM. Each project has its own allowance and scales to zero after five idle minutes. An always-on 0.25 CU database would need about 187.5 CU-hours per month, so Free is designed for intermittent work.

  • Supabase: Recurring hard cap. Two active projects, each with 500 MB Postgres, 50,000 MAU, 1 GB file storage, 500,000 Edge Function calls, and realtime allowances. Projects pause after a week of inactivity. Use it when Postgres, auth, storage, realtime, and functions belong in one dashboard.

  • Aiven PostgreSQL and MySQL: Recurring, no card. One free service of each type per organization, each with 1 CPU, 1 GB RAM, 1 GB storage, and backups. The provider and region are fixed, idle services power off, and there is no SLA.

  • CockroachDB Basic: Recurring, card-backed. 50 million request units and 10 GiB storage per organization, expressed as a $15 monthly usage benefit. It scales to zero, but the account can bill beyond the benefit unless resource limits are set.

  • Azure SQL Database free offer: Recurring, card-backed. Up to ten databases per subscription, each with 100,000 vCore-seconds and 32 GB of data and backup storage per month. Choose the hard auto-pause option if zero cost matters more than continuity.

  • Turso: Recurring, no card. 100 databases, 5 GB storage, 500 million rows read, 10 million rows written, and 3 GB embedded sync per month. Quota breaches block requests rather than rolling into an invoice. It remains the most interesting free distributed-SQLite offer.

NoSQL, queues, vectors, and search

  • MongoDB Atlas M0: Recurring, no card. One free cluster per project with 512 MB and 100 operations per second. It never expires, but automatically pauses after 30 idle days.

  • Cloudflare D1: Recurring hard cap. Ten databases, 500 MB each, 5 GB total, 5 million rows read and 100,000 rows written per day, plus seven-day Time Travel. It is the obvious SQLite choice when the rest of the app already lives on Workers.

  • Cloudflare KV, Durable Objects, and Queues: Recurring hard caps. KV has 1 GB and 100,000 reads/day but only 1,000 daily writes, deletes, and lists. SQL-backed Durable Objects bring free stateful coordination. Queues gets 10,000 operations/day, which is roughly 3,300 normally delivered messages once send, receive, and acknowledgement are counted.

  • Aiven Kafka, Valkey, and OpenSearch: Recurring, no card. The free Kafka tier has modest throughput and five two-partition topics. Valkey gets 1 CPU/1 GB RAM. OpenSearch gets 4 GB RAM and 20 GB disk. Each is limited to one per organization, uses fixed infrastructure, and can power off when idle.

  • Upstash Redis, QStash, Workflow, Vector, and Search: Recurring. The bundle covers a 256 MB Redis database, 1,000 QStash messages/day, 1,000 workflow steps/day, 1 GB vector data with daily query limits, and a preview search database with 200,000 documents. Free Redis databases can be archived after extended inactivity, and Search is explicitly preview-grade with no SLA.

  • Qdrant Cloud: Recurring, no card. One 1 GB RAM, 0.5 vCPU, 4 GB disk node, estimated for roughly 1 million 768-dimensional vectors. It suspends after one idle week and can be deleted after four unless reactivated.

  • Pinecone Starter: Recurring. Five indexes, 2 GB database storage, monthly read/write units, and 1 GB egress, plus free embedding, reranking, and Assistant allowances. The June promotion that doubled Assistant input tokens expired on 30 June, so the normal 500,000-input-token allowance applies again.

  • Convex Free: Recurring. 1 million function calls, 20 GB-hours of actions, a 0.5 GB database, files, text search, vector search, and realtime in one backend. Prolonged quota excess can return errors.

  • Firestore, DynamoDB, and Azure Cosmos DB: Recurring, card-backed cloud primitives. Firestore includes 1 GiB plus daily read/write/delete quotas. DynamoDB includes 25 GB and provisioned read/write capacity. Cosmos DB includes 1,000 RU/s and 25 GB on one free-tier account per subscription.

Storage, CDN, DNS, and media

  • Cloudflare R2: Recurring, billing profile required. 10 GB-month Standard storage, 1 million Class A operations, 10 million Class B operations, and free egress. The free allocation does not apply to Infrequent Access storage.

  • Tigris: Recurring. 5 GB Standard storage, 10,000 Class A and 100,000 Class B operations per month, with free egress. It is a smaller but genuinely global S3-compatible alternative.

  • Backblaze B2: Recurring. The first 10 GB storage is free, most transaction allowances are generous, and monthly egress is free up to three times average stored data. It remains useful for backups and CDN-backed assets.

  • Google Cloud Storage: Recurring, card-backed. 5 GB-month plus request allowances in selected US regions, and 100 GB outbound from North America per month outside excluded destinations.

  • Cloudflare DNS and CDN: Recurring. Authoritative DNS queries are uncapped and unbilled on the normal plans, with CDN, universal TLS, and unmetered DDoS protection. Free zones created since September 2024 are limited to 200 DNS records, which is still plenty for most side projects.

  • ImageKit: Recurring hard cap. 20 GB delivery, 3 GB DAM storage, and image/video transformation allowances per month. Delivery and uploads stop at the cap.

  • Cloudinary: Recurring, no card. 25 monthly credits shared across transformations, storage, and bandwidth. One credit can mean 1,000 transformations, 1 GB storage, or 1 GB bandwidth, not all three at once.

Email, auth, realtime, and product plumbing

  • Resend, Brevo, and Mailgun: Recurring. Resend includes 3,000 sent and received emails per month with a 100/day limit. Brevo includes 300 sends/day but adds branding. Mailgun includes 100 messages/day and one custom domain. Amazon SES separately gives 3,000 monthly message charges for only the first 12 months after first use.

  • Clerk, Auth0, Kinde, and WorkOS AuthKit: Recurring auth. Clerk includes 50,000 monthly retained users per app, Auth0 25,000 MAU, Kinde 10,500 MAU, and AuthKit 1 million MAU. WorkOS requires billing information for production even when AuthKit remains within the free allowance, and its paid SSO/directory products are separate.

  • Firebase Authentication: Recurring, no payment method on Spark. 50,000 MAU for non-phone auth and 50 SAML/OIDC MAU. Phone authentication is not part of that free allocation.

  • Ably and Pusher Channels: Recurring realtime. Ably includes 6 million messages/month, 500 messages/second, and 200 concurrent connections. Pusher's sandbox includes 200,000 messages/day and 100 concurrent connections.

  • Cloudflare Turnstile: Recurring. Twenty widgets, ten hostnames per widget, unlimited challenges, and seven-day analytics. It works without putting the site behind Cloudflare.

  • Svix and Hookdeck: Recurring webhook plumbing. Svix includes 50,000 attempted messages/month with 30-day payload retention. Hookdeck includes 10,000 events/month with three-day retention. Use Svix for sending webhooks and Hookdeck for receiving, buffering, inspecting, and retrying them.

  • Novu: Recurring, no card. 10,000 notification workflow runs per month across email, SMS, chat, push, and in-app channels. Delivery-provider charges still exist outside Novu.

  • Flagsmith and PostHog: Recurring. Flagsmith includes 50,000 flag API requests/month. PostHog's free cloud tier includes 1 million analytics events, 1 million flag requests, replay, errors, logs, AI observability, surveys, and warehouse rows in one unusually broad product-plumbing bundle.

Observability and uptime

  • Grafana Cloud: Recurring, no card. 10,000 metrics series, 50 GB each of logs, traces, and profiles, synthetic tests, k6 capacity, three active users, and 14-day retention. Official pages currently disagree on the separate frontend-observability session allowance, so that number is deliberately omitted.

  • Better Stack: Recurring. Ten uptime monitors, ten heartbeats, a status page, 100,000 exceptions, replay, and small logs, traces, metrics, and web-event allowances. It is substantially broader than the uptime-only description in June.

  • Sentry Developer: Recurring hard cap. 5,000 errors, 5 GB logs, 5 million spans, 50 replays, and small uptime, cron, and metric-monitor allowances. It is one user and does not buy overage.

  • Axiom Personal: Recurring, no card. Up to 500 GB monthly ingest, 25 GB stored, 30-day maximum retention, three datasets, and three monitors. The free deployment is US-only, but the current allowance is much more generous than the vague June entry suggested.

  • New Relic and Honeycomb: Recurring. New Relic includes 100 GB ingest and stops access at the free limit rather than charging. Honeycomb includes 20 million events and 100 million metrics points monthly, making it a good fit for high-cardinality traces and agent telemetry.

  • UptimeRobot, Checkly, and Healthchecks.io: Recurring. UptimeRobot includes 50 five-minute monitors and explicitly permits commercial projects. Checkly adds ten two-minute uptime monitors plus API and Playwright runs. Healthchecks.io covers 20 cron and background-job checks.

CI and cloud development environments

  • GitHub Actions: Recurring. Standard hosted runners are free for public repositories. Private repositories on GitHub Free get 2,000 minutes and 500 MB artifact storage per month. Larger runners are paid even for public projects.

  • GitHub Codespaces: Recurring. Personal GitHub Free accounts get 120 core-hours and 15 GB-month storage. A two-core machine turns that into 60 wall-clock hours. Stopped codespaces still consume storage.

  • GitLab.com CI: Recurring. 400 compute minutes per month for Free namespaces. Runner multipliers mean one wall-clock minute is not always one billed minute.

  • Azure Pipelines: Recurring. Private projects get one hosted job and 1,800 minutes/month. Public projects can get up to ten hosted parallel jobs with no monthly minute cap, although new organizations may need to request the grant.

  • CircleCI: Recurring, no card. 30,000 credits/month, worth up to 6,000 small-Docker minutes or 3,000 medium-Linux minutes. Eligible open-source organizations can receive much more.

  • Google Cloud Build: Recurring, billing account required. 2,500 e2-standard-2 build-minutes per month in the default pool. Google Cloud Shell separately gives 50 hours/week and a 5 GB persistent home directory.

  • AWS CloudShell: Recurring. No additional charge, 200 hours/month per account per region, 1 GB persistent home per region, and ten concurrent shells. Any AWS resources created from it still cost money.

AI, GPU, and model APIs

Free AI compute is several different products wearing one label. A notebook gives temporary hardware. ZeroGPU gives a queued slice of a GPU inside somebody's public demo. Modal sells real metered GPU time using a recurring credit. Model APIs provide inference without exposing the machine at all.

GPU, CPU, and notebook compute

  • Modal Starter: Recurring $30 compute credit. Real serverless CPU and GPU jobs, notebooks, schedules, and web functions, with 100 containers and ten-GPU concurrency. Functions are preemptible by default, and explicit region placement costs more. This remains the cleanest free burst-compute offer.

  • Lightning AI Free: Recurring, no card. Fifteen credits/month plus one active CPU Studio that restarts every four hours. At current listed rates, those credits are roughly 75 T4 hours, 31 L4 hours, 10 A100 40 GB hours, or three H200 hours. Phone verification, country availability, capacity, and hardware pricing all apply. The marketing headline says 80 free GPU hours; the current rate table tops out at 75 T4 hours, so use the credit figure as the durable promise.

  • Hugging Face CPU Spaces: Recurring public demo compute. 2 vCPU, 16 GB RAM, and 50 GB non-persistent disk. Free Spaces sleep after roughly 48 idle hours and their source is public.

  • Hugging Face ZeroGPU: Shared GPU demo access. A free account gets five GPU minutes/day when using existing ZeroGPU Spaces; unauthenticated users get two. Free users cannot host their own ZeroGPU Space, which requires PRO or an organization plan. The current hardware is shared RTX Pro 6000 Blackwell capacity, but this is Gradio demo hosting, not an arbitrary GPU box.

  • Google Colab Free: Shared, dynamic notebooks. CPU, GPU, and TPU access with no stable public quota or hardware guarantee. Free sessions run for at most 12 hours. It is useful for experiments and useless as an infrastructure promise.

  • Kaggle Notebooks: Shared, dynamic notebooks. Free GPU and TPU runtimes remain available, but a dependable weekly allowance is not present in current public documentation. The June claim of 20 TPU hours/week should not be repeated as a firm July number.

  • Databricks Free Edition: Shared, non-commercial serverless compute. Limited CPU and capacity-dependent GPU notebooks, one small SQL warehouse, jobs, three apps, and one scale-to-zero Lakebase project. Exact compute quotas are unpublished, outbound internet is restricted, apps stop after 24 hours, and fair-use exhaustion can disable compute for the rest of the day or month. Databricks also reserves the right to train on Free Edition data, so keep private datasets out.

  • SageMaker Studio Lab: Free notebook access with a deadline. Up to four GPU hours or eight CPU hours per 24 hours, with 15 GB storage and no AWS account or card. AWS stops accepting new Studio Lab customers on 30 July 2026. Existing users continue, but no new features are planned.

Hosted model APIs

  • Cloudflare Workers AI: Recurring. 10,000 Neurons per day. The 8 July pricing update shows more granular model-specific token equivalents but still meters the backend in Neurons. Cloudflare says customer content is not used to train or improve models without consent.

  • GitHub Models: Recurring public preview. Low-tier models on Copilot Free commonly allow 15 requests/minute and 150/day; higher-tier models get smaller limits. Embeddings have their own quota. It is useful for learning, prompt comparison, and proof-of-concept model switching rather than production traffic.

  • Gemini Developer API: Model-dependent free tier. Selected models have free input and output tokens, but rate limits vary by model, project, account, and region and should be read from AI Studio. Free-tier content is used to improve Google products; paid-tier content is not.

  • GroqCloud Free: Recurring rate-limited inference. Current examples range from 14,400 requests/day and 500,000 tokens/day for Llama 3.1 8B to 1,000 requests/day and 200,000 tokens/day for GPT-OSS 120B. Limits are organization-level and can vary by account. Groq says inference data is not retained by default outside limited abuse and reliability cases.

  • FreeInference: Experimental research service. No-card OpenAI- and Anthropic-compatible access to open models for open-source, research, and education. The quota is only described as generous and can move. More importantly, all prompts and responses may be logged, and anonymized derived data may be published or open-sourced. Never send it secrets, private code, or personal data.

  • OpenRouter Free: Recurring shared routing. More than 25 free models at 20 requests/minute and 50 requests/day. Buying at least $10 credit raises the free-model daily allowance to 1,000. Free routing can cross providers with different data policies.

  • Hugging Face Inference Providers: Tiny recurring credit. Free accounts receive $0.10/month. PRO and paid organization seats receive $2. It is model API credit, not GPU access.

  • Cerebras Inference and NVIDIA NIM endpoints: Useful but ambiguously durable. Cerebras publishes model-specific limits for a plan that its pricing calls Free and its rate page calls Free Trial, without a public duration. NVIDIA advertises unlimited prototyping but applies unquantified, load-dependent rate limits. Treat both as evaluation access, not infrastructure.

Startup, academic, research, and open-source credits

These can be much larger than the normal free tiers, but only after eligibility and approval.

  • AWS Activate: up to $5,000 for founders and self-funded companies, up to $200,000 for qualifying portfolio companies, and larger invite-only AI paths.

  • Google for Startups Cloud: up to $350,000 over two years for qualifying AI-first startups, or up to $200,000 through the normal Scale path.

  • Microsoft for Startups: currently advertises up to $150,000, although the public landing page does not expose the full tier and expiry mechanics.

  • DigitalOcean Hatch for AI/ML: up to $100,000 over a 12-month programme.

  • Modal academic grants: up to $10,000 for graduate students, labs, and researchers. Modal also runs a startup programme but does not publish the award amount.

  • AMD AI Developer Program: approved developers can choose $100 of AMD Developer Cloud credit that expires after 30 days or a separate $50 Fireworks inference offer lasting 90 days. Those are alternatives, not $150 of one kind of compute.

  • Google Cloud research credits and AWS Cloud Credit for Research: up to $5,000 through the published student/research paths, with proposal and institutional eligibility.

  • DigitalOcean open-source credits: annual awards from $60 for small active projects up to $20,000 for large production projects, based on stars, activity, licence, and reapplication.

  • Hugging Face Community GPU Grants: discretionary GPU upgrades for selected public Spaces, with no published amount, duration, or hardware guarantee.

Stacks that fit together

  • A small real server: OCI for the VM, Cloudflare for DNS/CDN/TLS, R2 for objects, Aiven or Neon for managed Postgres, Upstash for queues and Redis, Resend for mail, and Grafana Cloud plus Sentry for visibility.

  • A no-server product: Cloudflare Pages, Workers, D1 or Neon, R2, Queues or Workflows, Turnstile, Clerk or AuthKit, and Workers AI. Keep every component on a hard cap where possible.

  • A Postgres-first SaaS prototype: Vercel Hobby for a personal demo or Deno/Cloud Run for a more general container, Neon or Supabase for data, Upstash for async work, Resend for mail, and Better Stack or Sentry for failures.

  • An AI demo: Hugging Face CPU Spaces or a ZeroGPU grant for the public surface, Modal or Lightning for real burst compute, Workers AI, Gemini, Groq, or GitHub Models for hosted inference, and R2 plus Neon or Turso for state.

  • An agent experiment: Vercel Sandbox, Freestyle, or Upstash Box for recurring disposable machines; Browser Run for browser work; E2B or Daytona when a one-time credit is acceptable; Axiom, Honeycomb, or Sentry for the traces that explain what the agent actually did.

The most interesting part of doing this again is not making the list longer. It is seeing what survives, what becomes credit-shaped, and what quietly disappears. July has enough genuinely new compute to justify another snapshot: agent sandboxes are now a real category, AI notebooks now include recurring GPU credit and documented serverless GPU access, and Scaleway's recurring serverless tier deserves more attention.