Lightning AI pricing

About

Lightning AI is a cloud compute platform for building, training, and deploying AI models, built by the team behind PyTorch Lightning — the widely used open-source training framework. Its core product is the “Studio”: a persistent, ready-to-code cloud workspace that can switch between CPU and GPU machines in seconds, connect to a local IDE or SSH, and run background jobs, all without the user managing infrastructure. The pricing page cites 350,000+ builders across research labs (MIT, EarthDaily Analytics) and startups (OctAI, Aeira).

The company sells the same multi-cloud compute fabric to three audiences at once: individual researchers and hobbyists on the Free and Pro tiers, small research teams on Teams, and enterprises that want to run a secure, customized AI platform inside their own VPC with their own AWS/GCP commitments. Compute is accessed through a GPU “marketplace” that spans AWS, GCP, Lightning’s own cloud, Lambda, Nebius, NScale, and Voltage, so customers pick from the world’s best clouds in one place.

Lightning AI (formerly Grid.ai, rebranded June 2022) is a venture-backed private company that has raised across a $40M Series B (2022, led by Coatue and Index Ventures) and a $50M 2024 round (Cisco Investments, J.P. Morgan, K5 Global, NVIDIA). In January 2026 it completed a merger with GPU infrastructure provider Voltage Park, forming a combined company reported at over $2.5B valuation and $500M+ ARR with a fleet of 36,000+ owned H100/B200/GB300 GPUs — a vertical-integration move from software-on-other-people’s-clouds toward owning the metal. It competes with managed ML compute and notebook platforms such as Modal, RunPod, Replicate, Paperspace, and the hyperscalers’ first-party notebook/training services — differentiating on the persistent Studio experience, second-level billing, and a credit pool that spans interruptible and on-demand machines.

Pricing summary : How Lightning AI’s credit-pool compute model works

Lightning AI uses a hybrid freemium-plus-usage model built around three dimensions:

1. Seat / plan tier (recurring): Free ($0), Pro ($50/mo or $20/mo billed annually), Teams ($140/user/mo or $119/user/mo billed annually), and Enterprise (custom). The plan fee mostly buys a monthly or annual credit allotment and lifts caps on concurrency, storage, and CPU/GPU size — not the compute itself.
2. Usage credits (consumption): All CPU and GPU compute is billed by the second and drawn down from the credit pool. Free credits refresh monthly (15/mo ≈ 80 GPU hours on spot) and expire if unused; purchased credits expire after 12 months. When credits run out, plans pay-as-you-go for more.
3. Per-GPU-hour marketplace rates: On-demand single-GPU rates span $0.19 (T4) to $6.53 (H200), with interruptible (spot) machines discounted up to 80%. Lightning’s model APIs are separately metered in requests/min and tokens/min.

What makes this different: the credit pool is the unit that bridges a hybrid pricing model and pure pay-as-you-go — your plan sets the allotment and the caps, but every second of compute is metered against credit-based billing regardless of tier.

Pricing by product

Studio compute plans (Individual plans)

Studio compute plans (Business plans)

Sales motions across products: PLG / self-serve for Free, Pro, and Teams (online signup and checkout); sales-led for Enterprise (custom quote, VPC, and cloud-credit deals).

Per-GPU-hour marketplace rates (single GPU, on-demand vs interruptible)

Compute is billed by the second; the rates below are per GPU per hour. “Free hrs/mo” is the approximate hours the 15 free monthly credits buy on that machine.

CPU Studio rates (CPU-only Studios for coding and data prep)

8-GPU Studio rates (per-GPU, on-demand)

For multi-GPU Studios, per-GPU rates differ from single-GPU; the 8-GPU tab adds B200 at the top end: L4 $1.26, L40S $4.74, RTXP 6000 $2.78, A100 40 GB $3.46, A100 80 GB $1.79, H100 $1.99, H200 $5.90, B200 $9.86.

Hidden costs : What credit pools and spot rates really cost in practice

The $20/mo Pro headline understates real spend, because the plan fee only buys a fixed credit allotment — every additional GPU hour is metered on top. Two representative examples:

A solo Pro researcher training on an A100 80 GB

Once the bundled credits are exhausted, the per-GPU-hour rate — not the $20 seat — dominates the bill, so a single sustained A100 training run can be 5× the plan fee.

A 4-seat Teams lab running mixed H100 + spot T4 jobs

The lesson: Lightning’s seat fees are modest, but at moderate training volumes the metered GPU hours can match or exceed the platform seats — so the credit allotment and spot vs on-demand mix drive the bill far more than the tier label. This is the classic AI cost unpredictability and bill-shock trap that per-second metering both causes and, with spend caps, helps contain.

Want to estimate your own Lightning AI bill? Use the Lightning AI pricing calculator to model your monthly cost based on plan tier, GPU type, on-demand vs interruptible hours, and concurrency.

Pricing evolution : From Grid AI to a credit-pool GPU marketplace

Lightning AI’s pricing page is a client-rendered single-page app, so its historical plan dollar figures were not preserved in public archives — the milestones below are reconstructed from dated company and press announcements, and only the current live capture carries a verified rate card. Where a period’s plan prices are not documented, they are left unstated rather than estimated.

Cadence

Tracked range: 2022 Q2–2026 Q2. Per-quarter price-change counts are marked unknown where archived plan prices were not recoverable; the 2026 Q2 column reflects the first capture with verified dollar figures.

Notable changes

• 2022-06-16 — Grid.ai rebrands as Lightning AI and ships an open-source AI dev platform with a $40M Series B (Coatue, Index Ventures, Bain Capital); PyTorch Lightning cited at 20M downloads.
• 2024-11-21 — $50M raise (Cisco, J.P. Morgan, K5 Global, NVIDIA) lifts total funding to ~$103M; the free tier is publicly described as 22 GPU hours/month — the earliest documented free-tier capacity, since restructured into the current 15-monthly-credit (~80 spot GPU-hour) allotment.
• 2025-08-19 — Multi-cloud GPU Marketplace launches (Lambda, Nebius, Voltage Park, Nscale), pitching up to 70% cost savings via capacity routing.
• 2026-01-21 — Voltage Park merger completes, turning a software-and-other-clouds platform into an owned-GPU AI cloud (>$2.5B valuation, >$500M ARR, 36,000+ GPUs).
• 2026-06-02 — First verified rate-card capture: Free ($0), Pro ($20/mo annual), Teams ($119/user/mo annual), Enterprise (custom), with by-the-second credit billing and on-demand GPU rates from $0.19 (T4) to $6.53 (H200).

The Voltage Park merger in detail

The January 2026 merger is the most consequential structural change in Lightning AI’s pricing trajectory because it changes what the company actually sells. Before the merger, Lightning was primarily a software experience (Studios + credit pool) brokering compute across third-party clouds — its August 2025 GPU Marketplace explicitly positioned it as a router across Lambda, Nebius, Voltage Park, and Nscale. By absorbing Voltage Park’s 36,000+ owned H100/B200/GB300 GPUs, Lightning becomes a vertically integrated AI cloud that both sets the rate card and owns much of the underlying supply. That gives it far more control over the per-GPU-hour economics it publishes — the spot/on-demand spread and the up-to-80% interruptible discounts are now backed substantially by inventory it controls rather than purely arbitraged capacity. For buyers, the practical signal is durability: the published marketplace rates are less likely to be squeezed by a single supplier’s pricing, but the company now carries depreciation risk on a large owned fleet, which is the kind of cost pressure that historically pushes infrastructure vendors toward commitment-based discounting over time.

What’s unique : Credit pool, second-level billing, and a multi-cloud GPU marketplace

1. The plan fee buys credits, not compute. Unlike pure seat pricing, a Lightning plan is really a monthly/annual credit allotment plus higher caps. The same per-second GPU rate applies on Free and Teams; the tier changes concurrency, storage, and machine ceilings — a clean separation of pure-usage metering from the subscription envelope.

2. By-the-second billing across CPU and GPU. Every Studio — CPU or GPU, on-demand or spot — is billed by the second and drawn from one credit pool, so switching a Studio from CPU to GPU mid-session is a billing non-event rather than a new instance.

3. Spot economics are first-class. Interruptible machines are discounted up to 80%, and the rate table lists on-demand and interruptible side by side, encouraging users to default to spot for tolerant workloads (a T4 drops from $0.19 on-demand math to $0.52/GPU/hr interruptible band; data-prep CPUs run as low as $0.28/hr spot).

4. A multi-cloud GPU marketplace. Lightning brokers capacity across AWS, GCP, its own cloud, Lambda, Nebius, NScale, and Voltage — so the same Studio abstraction runs on whichever cloud has the GPU, and Enterprise can route it onto its own AWS/GCP commitments.

5. Free expires, paid lasts a year. Free credits reset and expire monthly, while purchased credits last 12 months — a deliberate nudge that keeps the free tier non-stockpilable while giving paying customers runway. It mirrors the broader shift from entitlements to credits in LLM billing, where the credit itself becomes the durable unit of value.

Strengths & weaknesses

Billing UX : Credit meters, spend caps, and Monthly/Annual toggles

• Credit pool meter — every plan shows a monthly/annual credit allotment (15/mo Free, 240/yr Pro, 600/yr Teams) that all CPU/GPU compute draws down by the second.
• Monthly vs Annual toggle — per-plan switch on the pricing cards surfaces the annual discount inline (Pro $50→$20, Teams $140→$119).
• Real-time cost controls + spending limits — Teams and above expose organization-level cost management and hard spend caps to prevent runaway GPU bills.
• Interruptible (spot) selection — the rate table lets users pick on-demand vs interruptible per machine, with the discount shown side by side.
• Credits Auto-Reload — automatic credit top-ups (listed in the feature matrix) keep Studios running once the allotment is exhausted.
• AWS Marketplace billing — Teams can route Lightning spend through their existing AWS Marketplace contract; Enterprise can apply AWS/GCP commits.

Strategic wins : Why the credit-pool-plus-marketplace structure works

1. Decoupling the seat from the compute lets one price serve five segments

By making the plan an allotment-and-caps envelope and metering compute identically across tiers, Lightning serves hobbyists, developers, labs, and enterprises with the same rate card. This is a textbook application of separating the access fee from the usage meter, and it keeps the free tier genuinely useful without cannibalizing paid compute revenue.

2. Publishing the full GPU rate card builds trust in a murky market

GPU cloud pricing is notoriously opaque; Lightning publishes T4-through-H200 on-demand and interruptible rates openly, which lowers buyer friction and makes the platform comparable. Transparent metered rates are exactly the signal usage-based buyers look for, as we cover in usage invoicing and billing cycles.

3. Spot-first defaults align cost incentives with Lightning’s supply

Showing interruptible rates beside on-demand — and discounting them up to 80% — nudges price-sensitive users onto spare capacity, improving Lightning’s marketplace fill rate while genuinely lowering customer cost. It’s a rare win-win in usage-based pricing for SaaS and AI where the cheap option is also the operationally efficient one.

4. Vertical integration via the Voltage Park merger backs the rate card with owned supply

The January 2026 merger that absorbed Voltage Park’s 36,000+ owned GPUs turned a capacity-router into a vertically integrated AI cloud, so the published per-GPU-hour rates and up-to-80% spot discounts now sit on inventory Lightning controls rather than purely arbitraged third-party capacity. This is the same supply-side leverage that lets cost-led infrastructure players sustain aggressive metered rates, and it strengthens the credibility of the marketplace’s “70% cheaper” positioning — a durability signal usage-based buyers care about when they price in forecasting and budgeting AI usage costs.

5. Open-source flywheel (PyTorch Lightning) lowers the cost of acquiring paying compute users

Lightning’s biggest community moments are technical, not promotional: a how-to on finetuning GPT-like models drew a 498-point, 122-comment Hacker News thread (2023-05-25), “LoRA from scratch” hit 339 points (2024-01-22), and “Takeaways from hundreds of LoRA experiments” reached 258 points (2023-10-13) — all published as runnable Lightning Studios. That content-as-distribution motion seeds the free tier with already-warm developers, so the freemium credit allotment functions as a conversion surface for an audience the open-source framework already acquired, a pattern we explore in usage-based pricing for SaaS and AI.

Areas to improve : Predictability gaps and gated cloud-credit access

1. Give buyers a credit-to-dollar and credit-to-GPU-hour estimator on the page

The plan cards quote credits (15/mo, 240/yr, 600/yr) but the dollar and GPU-hour value of a credit is only implied (“15 credits ≈ 80 GPU hrs on spot”). A built-in estimator mapping credits → on-demand vs spot hours per machine would cut the biggest source of bill-shock anxiety for new users.

2. Offer a mid-tier path to bring-your-own-cloud before Enterprise

AWS/GCP credit usage and VPC deployment are Enterprise-only, which forces any cost-conscious team with existing cloud commits straight into a sales motion. A self-serve “Teams + BYOC” add-on would capture the large middle of startups that already hold cloud credits but aren’t ready for a custom contract.

3. Let Free and Pro users carry a small credit balance forward

Monthly expiry of free credits is a sound anti-stockpiling rule, but a modest rollover cap (e.g. one month) would reduce the “use-it-or-lose-it” friction that pushes bursty researchers toward competitors with pure pay-as-you-go and no expiry. The free tier has also already shifted once — from the 22 GPU hours/month documented at the November 2024 raise to today’s 15-credit (~80 spot GPU-hour) framing — and re-expressing the free allowance in credits rather than hours makes period-over-period comparison harder for the same researchers the open-source community brings in. (No pricing-backlash thread above Hacker News / Reddit trust-event thresholds was found in this review; Lightning’s high-engagement community activity is technical tutorial content, not pricing complaint.)

Key takeaways

1. Separate the access fee from the usage meter. Lightning’s plan tier sets the credit allotment and caps while every second of compute is metered identically — letting one rate card serve hobbyists through enterprises without per-segment repricing.
2. Bundle usage as credits, then meter against them. Quoting plans in credits (15/mo, 240/yr, 600/yr) instead of dollars lets the vendor adjust per-machine rates without re-pricing the plan, but teams must translate credits to dollars to budget.
3. Make spot a visible, defaulted choice. Showing on-demand and interruptible rates side by side, with up-to-80% discounts, steers price-sensitive demand onto spare capacity — improving marketplace economics while lowering customer cost.
4. Use credit expiry as a behavioral lever. Free credits expire monthly (anti-stockpiling) while purchased credits last 12 months (runway for buyers) — a deliberate split that protects free-tier economics.
5. Publish the rate card in a transparency-poor market. Open per-GPU-hour pricing is a competitive wedge against opaque GPU clouds and a trust signal for usage-based buyers.

UBP implications

1. Credit pools are the bridge between subscription and pure usage. Lightning shows how a credit allotment can wrap pay-as-you-go metering so customers get a predictable monthly envelope while the vendor still captures heavy-usage upside second by second.
2. Marketplace pricing makes the unit rate a supply lever, not just a price. When the same meter routes across seven clouds, the per-GPU-hour rate doubles as a yield-management tool — spot discounts move demand to spare capacity.
3. Expiry policy is part of the pricing model. Differential credit expiry (monthly for free, 12 months for paid) is a usage-based-pricing design choice that shapes stockpiling, churn, and free-tier cost as much as the headline rates do.

Sources

• Lightning AI pricing page — independently re-rendered to confirm plan prices and the H100 single-GPU on-demand rate (accessed 2026-06-02)
• Lightning AI billing FAQ docs (accessed 2026-06-02)
• Lightning AI GPU marketplace docs (accessed 2026-06-02)
• Lightning AI enterprise (accessed 2026-06-02)

Bottom line

Lightning AI prices a multi-cloud GPU/CPU Studio platform as a credit pool: a modest freemium seat tier (Free, $20 Pro, $119 Teams, custom Enterprise) buys a monthly credit allotment and higher caps, while all compute is billed by the second from those credits at publicly listed rates from $0.19 (T4) to $6.53 (H200), with spot discounts up to 80%. The structure is buyer-friendly and transparent, but the plan-fee-plus-credit-pool-plus-per-second-rate stack means the GPU hours, not the seat, decide the bill.

Want to compare Lightning AI against other GPU cloud and AI infrastructure pricing? Browse the pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Credit-pool plans with per-GPU-hour marketplace rates

Jun 2026

Lightning AI lists Free ($0, 15 monthly credits), Pro ($20/mo billed annually, 240 annual credits), Teams ($119/user/mo billed annually, 600 annual credits), and Enterprise (custom). Compute is billed by the second and drawn from credits, with on-demand GPU rates from $0.19 (T4) to $6.53 (H200) and spot discounts up to 80%.

Voltage Park merger — combined AI-native cloud

Jan 2026

Lightning AI completes a merger with GPU infrastructure provider Voltage Park, forming a combined company valued at over $2.5B with reported ARR above $500M and a fleet of 36,000+ owned H100/B200/GB300 GPUs. A Wayback snapshot from 2026-02 confirms the live 'Lightning merged with Voltage Park, adding 36K GPUs' banner; the SPA pricing body still did not hydrate, so no archived rate card was captured.

GPU Marketplace uniting NeoClouds and hyperscalers

Aug 2025

Lightning AI launches a multi-cloud GPU marketplace uniting hyperscalers and NeoClouds (Lambda, Nebius, Voltage Park, Nscale), claiming AI cost reductions up to 70% by routing the same Studio across the cheapest available capacity. The release positions per-GPU-hour rate transparency and no vendor lock-in but does not publish specific rates.

$50M raise; pay-as-you-go GPU pricing with free GPU hours

Nov 2024

Lightning AI announces a $50M raise (Cisco Investments, J.P. Morgan, K5 Global, NVIDIA), bringing total funding to ~$103M, and reports 240,000 users across 2,000 organizations. Press describes pay-as-you-go pricing with a free tier of 22 GPU hours per month — the earliest publicly documented free-tier capacity, later restructured into the current 15-monthly-credit (~80 spot GPU-hour) allotment. Archived pricing snapshots did not preserve plan dollar figures.

Grid.ai rebrands to Lightning AI + $40M Series B

Jun 2022

Grid.ai rebrands as Lightning AI and launches an open-source AI development platform alongside a $40M Series B led by Coatue and Index Ventures (with Bain Capital, Mantis VC, First Minute Capital), bringing total raised to ~$58.6M. PyTorch Lightning is cited at 20M downloads. Archived pricing snapshots from this period did not preserve plan prices (the SPA pricing page did not hydrate in the Wayback capture), so no period rates are asserted.