> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # What Does Runpod Do? The Per-Second GPU Cloud - URL: https://www.sotkn.com/stack/neoclouds/what-does-runpod-do/ - Published: 2026-11-02T04:48:04.000Z - Updated: 2026-11-02T04:48:04.000Z - Description: Runpod sells GPU hours by the second to AI developers, on machines that sit in partner data centers it does not own. - Author: Sean - Tags: #layer-neoclouds STACK · Neoclouds · Status: Tightening · Last reviewed: 2 Nov 2026 Runpod sells GPU hours by the second to AI developers, on machines that sit in partner data centers it does not own. **In short.** Runpod reported more than $325 million of annual recurring revenue on 13 October 2026, up from $120 million in January. Its chief executive calls the company a software firm and the 28 partner data centers the substrate the software sits on. Our read: the hall that holds the GPU is paid when that GPU is scarce, and Runpod is paid for the seconds a developer actually runs. The stuck step is the flagship card. On Runpod’s own data the B200 list price rose 31% between January and August, and the plan to multiply supply by 7 December is still a forecast. Why this sits on the [Neoclouds](https://www.sotkn.com/stack/neoclouds/) layer: [CoreWeave](https://www.sotkn.com/stack/neoclouds/what-does-coreweave-do/) gets paid on capacity it locks in and borrows against. Runpod gets paid when someone presses start. The same GPU can clear a multi-year contract or a per-second meter. Who holds the card decides which clock runs. ## Numbers that matter | Number | What it is | Why it matters | | ------------------ | ------------------------------------------------------------ | --------------------------------------------------------- | | **\>$325 million** | ARR on 13 Oct 2026, from $120 million in January | 2.71× the January figure, on our arithmetic | | **$1.0 billion** | Valuation at the $100 million round, 24 Jun 2026 | About 4.2× the ARR the CEO had stated that month | | **+31%** | B200 list price, January to August 2026, Runpod’s data | The hour got more expensive before the October ARR print | | **24%** | Revenue from resources agents created, report of 29 Sep 2026 | Was 12% one month earlier | | **7 Dec 2026** | Horizon of the August fleet plan | H200, B200 and H100 supply planned at 3.4×, 2.5× and 2.1× | *Company figures, dated in the sources. The 2.71× is $325 million divided by $120 million. Both are the round numbers Runpod printed, not an audited total. The December multiples are a forecast of verified supply, not GPUs already online.* ![Bar chart of Runpod ARR: $120 million in January 2026, $240 million in June, $325 million in October](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/runpod-arr-path.png) Three statements, three dates. On 20 January Runpod said it had passed $120 million of ARR, with revenue up 90% from a year earlier. On 2 June Zhen Lu wrote that ARR was $240 million, doubled in five months. The page was updated on 13 September and still carried that sentence. The 24 June funding release did not state ARR. On 13 October the company said ARR had passed $325 million, up from the January figure, with more than 1.2 million developers on the platform, from about 495,000 a year earlier. It called the revenue move nearly three times. Divide 325 by 120 and you get 2.71\. The gap is $205 million of stated ARR across 266 days, from 20 January to 13 October. A $1.0 billion valuation divided by the CEO’s $240 million is about 4.2\. The same price divided by the later $325 million is about 3.1, which is our cross-date arithmetic, not the multiple the round was struck at. Announced equity is the $20 million seed of 8 May 2024, co-led by Intel Capital and Dell Technologies Capital, plus the $100 million in June. Our sum is $120 million. Runpod has not published debt, so ARR versus that sum is not a leverage ratio. ## What does it actually do? It meters GPU time. A developer starts a Pod, a serverless endpoint, or a cluster, and pays by the second. The 24 June release says the self-serve product has no commitment minimum, that a first workload is typically running within an hour of sign-up, and that the serverless platform had processed more than 20 billion inference requests. The machines are not described as Runpod’s. In the post published 2 June and updated 13 September, Lu called Runpod a software company and wrote that “the 28 partner datacenters we orchestrate across are the substrate.” On 24 September the enterprise release put the same platform at 32 regions and more than 30 GPU configurations, with reserved capacity and region-level uptime. Nine days, two footprints. We keep both dates rather than merge them. Enterprise is a contract on that platform: reserved capacity, single sign-on, cost-center billing. Runpod said it had added ISO/IEC 27001 beside an existing SOC 2 Type II report, with HIPAA and GDPR support. It also said it was processing more than 10 million serverless inference requests a day, a later count than the 20 billion cumulative figure in June. It has not said what share of the $325 million is the contract and what share is still the open meter. Names it has put on the record include Sony, Toyota, Intel, OpenAI, Perplexity, Replit, and Zillow. A name is not a dollar figure or a term. What it did quantify is supply it hopes to have. The Fall 2026 report, from an August forecast, planned verified supply by 7 December at 3.4 times for H200 SXM, 2.5 times for B200, and 2.1 times for H100 SXM. The page says that tracks a median demand forecast. A plan is not a delivery receipt. ## How does it make money? By charging for the seconds. Runpod publishes no gross margin, no utilization, and no accounts, so we do not estimate profit. The January release reported net dollar retention of 120%, meaning existing customers as a group were spending more than a year earlier, and sign-ups up 155% year over year. October did not update either number. They stay on the January date. Who gets paid, on what the company has actually given: - **The partner that owns the GPU.** Lu’s description is a software layer over 28 halls. When a flagship card is scarce, the owner of the card is the one who can raise the hour. Runpod has not published the share it keeps after those halls are paid. - **Runpod, on the meter.** The $325 million is the company’s own ARR, not a gross marketplace volume it has separated out. If some of that figure is owed straight on to a partner, the release does not say so. - **The script that now starts the job.** The 29 September report says resources created by agents were 24% of total revenue, up from 12% a month earlier, and that agent workloads were 3.7 times more likely than the rest of the platform to run longer than a month. That is the stickiest usage they have printed. It is still usage. It is not a backlog. - **Equity holders, last and unpublished.** Summit Partners led the June round, and managing director Michael Medici joined the board. Intel Capital’s Mark Rostick joined the board at the May 2024 seed. They own whatever remains after the halls and the power are paid. That remainder is not in any release. Hub documentation, last modified 20 July 2026, pays template maintainers up to 7% of compute hours as Runpod credits, which are spent back on the platform. That is not the hall’s fee. The company does not publish the hall’s fee, and it has not split the $325 million between the open meter and reserved enterprise capacity. ## What limits it? The card. On 29 September Runpod said its own data showed flagship list prices up in 2026, including 31% on the B200 between January and August. In the same sentence it said high-bandwidth memory is effectively sold out into 2027, and that some data-center capacity is waiting on a grid connection. The memory claim is theirs; the sellers of that memory are on our [SK hynix](https://www.sotkn.com/stack/semiconductors/hbm/what-does-sk-hynix-do/) page. A per-second price with no minimum has two moves when the card costs more: charge the developer more, or absorb it. Runpod has not said which one the 31% became. Our page on [what sets the price of a GPU hour](https://www.sotkn.com/explained/gpu-rental-price-explained/) is the mechanism underneath that choice. The second limit is that the halls are someone else’s. Multiplying verified H200, B200, and H100 supply by early December only happens if the partners receive the cards and the power. Runpod’s August page calls those multiples a forecast of currently verified supply. Miss the cards, and the October ARR has nothing new to meter. The 20 January release described on-demand clusters of up to 64 H100s. Later releases talk about reserved multi-node capacity and do not restate a larger on-demand cap. Frontier training jobs that need thousands of GPUs in one contract still sit with the clouds that pre-buy the rack. The third limit is the books. There is no public margin, utilization, debt balance, or contracted backlog. A usage book can leave, and January’s 120% net dollar retention has not been updated. On 13 October Lu said the growth “demands a different level of financial and operational discipline,” and hired Brian Erickson, whose earlier work includes DigitalOcean’s public listing and finance roles at Amazon and Microsoft. The next useful document is a margin, not another sign-up count. ## Who else does this? [CoreWeave](https://www.sotkn.com/stack/neoclouds/what-does-coreweave-do/) sells reserved GPU capacity and finances the cards against the contract. [Lambda](https://www.sotkn.com/stack/neoclouds/what-does-lambda-do/) draws a GPU loan as the hardware and the customers show up. [Nebius](https://www.sotkn.com/stack/neoclouds/what-does-nebius-do/) is the public version of a prepay cloud, and our [Wire note on Nebius lifting GPU prices](https://www.sotkn.com/wire/nebius-raises-gpu-prices-october-2026/) is the same hour-price clock on a different balance sheet. [TensorWave](https://www.sotkn.com/stack/neoclouds/what-does-tensorwave-do/) is the all-AMD case, where the chipmaker’s credit stands behind a lease the tenant cannot carry alone. [Together AI](https://www.sotkn.com/stack/neoclouds/what-does-together-ai-do/) sells the inference platform on top of the hour. Runpod’s difference is the open meter: no minimum on the self-serve door, and a software layer that Lu says can, later, sit on GPUs a customer already owns. That portability is a sentence in the June post. It is not a product with a revenue line yet. Lu wrote that in one of every three of Runpod’s first-quarter sales conversations, AWS was what the team was leaving. That is his account of those calls, not a market share. The map of this folder is [who the neoclouds are](https://www.sotkn.com/explained/who-are-the-neoclouds/), and the split with the big clouds is [neocloud versus hyperscaler](https://www.sotkn.com/explained/neocloud-vs-hyperscaler/). ## What would change this view? - **Easing:** a published gross margin or utilization, a contracted backlog with a dollar figure, or the 7 December supply plan showing up as cards online rather than a forecast. An ARR print that keeps rising after the B200 list price rose would say the increase was passed through. - **Tightening:** the next ARR figure stalls or falls, net dollar retention reprinted under 100%, or the December multiples missed because the partners did not get the cards or the power. - **A split we do not have:** any number for enterprise reserved capacity versus the per-second meter, or the share of the bill that partner halls keep. Either one would move our read of who gets paid. ## Practical takeaways Read the $325 million as the developer bill Runpod reports, not as a fleet it owns. The dates that test the read are 7 December 2026, the horizon of the August supply plan, and whatever ARR sentence comes next, because that sentence is the first one a new chief financial officer has to stand behind. Until then the binding object is the flagship GPU, and the memory inside it is the upstream check on the HBM pages. ## FAQ **What does Runpod do?** Runpod is a GPU cloud for AI developers. It sells Pods, serverless endpoints, and clusters by the second, with no commitment minimum on the self-serve product. The company describes itself as a software layer over partner data centers, and its chief executive has put that footprint at 28 halls. **How does Runpod make money?** Developers pay for GPU time by the second. Runpod said on 13 October 2026 that annual recurring revenue had passed $325 million, up from $120 million in January. It does not publish gross margin, utilization, or how much of that bill is passed to the halls that own the GPUs. **Who owns Runpod?** It is private. Announced funding is a $20 million seed on 8 May 2024, co-led by Intel Capital and Dell Technologies Capital, and a $100 million round on 24 June 2026 led by Summit Partners at a $1.0 billion valuation. Our sum of those two rounds is $120 million. **Why did the B200 price matter?** Runpod’s own data, in its 29 September 2026 report, showed the B200 list price up 31% between January and August. On a per-second cloud with no minimum commit, a higher GPU price has to be passed on or absorbed. The company has not said which. **What is stuck for Runpod?** The flagship GPU. The company said high-bandwidth memory is effectively sold out into 2027, and its plan to multiply verified H200, B200, and H100 supply by 7 December 2026 is a forecast, not a delivery. It also publishes no contracted backlog. **Is this a recommendation to buy or sell anything?** No. This page explains how Runpod makes money and where it can get stuck. It gives no buy or sell calls. Not investment advice. Do your own research. ## Keep reading on the map Layer: [Neoclouds](https://www.sotkn.com/stack/neoclouds/). On the same layer: [CoreWeave](https://www.sotkn.com/stack/neoclouds/what-does-coreweave-do/), [Lambda](https://www.sotkn.com/stack/neoclouds/what-does-lambda-do/), [Nebius](https://www.sotkn.com/stack/neoclouds/what-does-nebius-do/), and [TensorWave](https://www.sotkn.com/stack/neoclouds/what-does-tensorwave-do/). Explained: [what sets the price of a GPU hour](https://www.sotkn.com/explained/gpu-rental-price-explained/) and [neocloud versus hyperscaler](https://www.sotkn.com/explained/neocloud-vs-hyperscaler/). Markets: [United States](https://www.sotkn.com/markets/us/). Wire: [Nebius raises GPU prices](https://www.sotkn.com/wire/nebius-raises-gpu-prices-october-2026/). Upstream of the 31% move: [SK hynix](https://www.sotkn.com/stack/semiconductors/hbm/what-does-sk-hynix-do/). Next layer down: [Semiconductors](https://www.sotkn.com/stack/semiconductors/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when GPU hours, loans, and handovers move. Not investment advice. Do your own research. ## How we check this Figures come from Runpod’s releases of 20 January, 24 June, 24 September, 29 September, and 13 October 2026, from Zhen Lu’s post of 2 June 2026 (updated 13 September), from the Fall 2026 report page, from Intel Capital’s 8 May 2024 seed release and Runpod’s note on that round, and from the Hub revenue-share page last modified 20 July 2026\. Read on 2 November 2026\. Runpod files no accounts. We do not estimate margin or debt. The 2.71×, the $205 million gap, the 266 days, the $120 million equity sum, the 4.2×, and the 3.1× are our arithmetic. Who gets paid, and the stuck step, are our read. The featured image is an illustration, not a photograph of a Runpod site. Not investment advice. Do your own research. **Last reviewed:** 2 Nov 2026 · **Next review:** after the 7 December 2026 fleet-plan horizon, or the next ARR figure, whichever comes first Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## Sources 1. [Runpod — $325 million ARR and CFO Brian Erickson (13 Oct 2026)](https://www.runpod.io/press/runpod-325m-arr-new-cfo?ref=sotkn.com) 2. [Runpod — $120 million ARR (20 Jan 2026)](https://www.runpod.io/press/runpod-ai-cloud-surpasses-120m-in-arr?ref=sotkn.com) 3. [Runpod — $100 million led by Summit Partners (24 Jun 2026)](https://www.runpod.io/press/runpod-raises-100m-summit-partners?ref=sotkn.com) 4. [Zhen Lu — The chips got faster. The stack didn’t (2 Jun 2026, updated 13 Sep 2026)](https://www.runpod.io/blog/the-chips-got-faster-the-stack-didnt?ref=sotkn.com) 5. [Runpod — Fall 2026 State of AI Compute (29 Sep 2026)](https://www.runpod.io/press/runpod-fall-2026-state-of-ai-compute?ref=sotkn.com) 6. [Runpod — State of AI Compute report page, August fleet plan](https://www.runpod.io/reports/fall-2026-state-of-ai-compute?ref=sotkn.com) 7. [Runpod — enterprise platform (24 Sep 2026)](https://www.runpod.io/press/runpod-expands-enterprise-platform?ref=sotkn.com) 8. [Intel Capital — $20 million seed, co-led with Dell Technologies Capital (8 May 2024)](https://www.intelcapital.com/runpod-raises-20m-in-seed-funding-co-led-by-intel-capital-and-dell-technologies-capital/?ref=sotkn.com) — company note: [Runpod on the $20 million round](https://www.runpod.io/blog/runpod-raises-20mm?ref=sotkn.com) 9. [Runpod docs — Hub revenue sharing (last modified 20 Jul 2026)](https://docs.runpod.io/hub/revenue-sharing?ref=sotkn.com) S [**Sean**](https://www.sotkn.com/author/sean/) Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)