Overview
Vectordash began in 2018 as a marketplace for cheap machine-learning compute, then entered Y Combinator's Winter 2019 batch as a cloud-gaming service. The company matched gamers with nearby GPUs, streamed a full Windows PC to ordinary laptops, and charged about $20-$28 a month. Its best demo made Apex Legends look native on a 13-inch Mac.[1][2]
The company did not fail because nobody wanted remote GPUs. It found 120,000 of them. It failed because the GPUs were attached to the wrong customers and the wrong networks. Researchers could not pay, enterprises would not trust residential hosts, and mining rigs with dozens of cards often had enough upload bandwidth for one gaming stream. Supply that looked abundant in a database was scarce at the moment and place a paying customer needed it.[3]
Founding Story
The first Vectordash transaction happened before there was a marketplace. Sharif Shameem, then doing computer-vision work at MITRE and studying computer science at the University of Maryland, wanted to train a character-level neural network on his iMessage history. Cloud GPUs were expensive. A friend named Alex was mining Ethereum on an Nvidia 1080 Ti. Shameem offered him $10 a day for the card, more than it earned from mining, and got faster compute for less than AWS charged.[3]
The model disappointed him. The trade did not. In a 2023 interview, Shameem said: "The model was really bad, but the whole trade worked really great." That arbitrage became Vectordash: miners would switch idle Nvidia cards from cryptocurrency to paid AI jobs, and researchers would rent them below cloud-provider prices.
Shameem introduced the idea publicly in March 2018 as "sort of like Airbnb but for GPUs." Hosts installed a desktop client, declared availability, and automatically switched between mining and customer work. By September, Vectordash offered GPU instances from $0.25 an hour, a prebuilt fast.ai image, a vectordash jupyter command, and multi-GPU machines.[4][5]
Abhishek Bhargava and Arbaz Khatib joined the founding team. Bhargava studied computer science, computational finance, and robotics at Carnegie Mellon; Khatib authored the public Vectordash CLI package. The public record does not establish how the three met. It does show a technically ambitious team moving unusually fast: the company launched within months, accumulated more than 120,000 listed GPUs, and became one of the cheapest GPU clouds for roughly a year.[6]
Timeline
- December 2017-2018: The team begins work on a peer-to-peer GPU marketplace; founder material places the first public launch in early 2018.[4]
- September 2018: Vectordash markets $0.25-per-hour AI instances and launches peer-to-peer cloud gaming on Hacker News.[5][7]
- Winter 2019: Vectordash joins Y Combinator with a three-person team.[1]
- March 2019: The company publicly launches its $28-per-month gaming service in the Bay Area.[2]
- June 2019: A sponsored MTGGoldfish tutorial promotes a $19.95 plan and seven-day trial for Mac users playing Magic Arena.[8]
- 2020: Users report working full cloud PCs in the United States, Europe, and South America, alongside storage and availability problems.[9]
- March-April 2021: Users report no available machines and describe the service as dead. No formal closure date was found.[10]
- By 2026: YC lists Vectordash as inactive.[1]
What They Built
Vectordash built two products on the same marketplace. The first gave machine-learning researchers remote Linux machines assembled from consumer Nvidia GPUs. A command-line client listed rentals, opened SSH sessions, transferred files, and launched Jupyter. Hosts ran a desktop client that shifted a GPU between mining and paid jobs. This was genuine infrastructure, not a landing-page marketplace.[4][6]
The gaming product turned a nearby GPU into a rented Windows PC. Customers chose a machine, connected through Parsec, signed into stores such as Steam or Epic, and installed games they already owned. Unlike catalog services, Vectordash exposed a general desktop, which meant mods, launchers, and unsupported games could work. The tradeoff was persistence: users reported reinstalling games and moving saves through cloud storage between sessions.[9]
The technical target was severe. A host needed to sit within roughly 300 miles of the player to keep interaction near 20-30 milliseconds. With a strong connection, Vectordash claimed 4K at 60 frames per second. A founder demo showed Apex Legends near 100 frames per second on a Retina MacBook with less than 20 milliseconds of latency.[2]
Market Position
Target Customers
The AI product attracted students and independent researchers. That audience valued cheap compute but had little money. Startups had promotional credits from major clouds, while larger firms would not place sensitive datasets on unknown residential machines. Vectordash had supply-market fit without a durable demand market.[3]
Gaming offered a clearer buyer: someone with a Mac, Chromebook, or weak PC who already owned games and wanted a full Windows machine. The Magic Arena campaign narrowed that further to Mac owners excluded from a Windows-only title. That was smart positioning because the alternative was buying hardware, not merely choosing another subscription.[8]
Market Size
Vectordash never published a defensible market-size estimate, subscriber count, or revenue figure. YC said about 95% of computers could not run the latest games, but that describes technical eligibility, not willingness to pay.[11] The more useful evidence is category survival: Nvidia, Microsoft, and Shadow still sell cloud gaming or cloud PCs in 2026. The demand was real, but it did not guarantee room for a small undifferentiated provider.
Competition
Vectordash sat between catalog streaming and a full remote PC. That openness was valuable, yet it inherited the hardest parts of both models: game compatibility, Windows storage, low-latency video, fraud, host reliability, and geographic coverage. Google Stadia and Nvidia GeForce Now could fund data centers and negotiate game rights. Shadow could charge more for a persistent cloud PC.
The current contrast is blunt. Nvidia advertises more than 2,000 install-to-play titles and RTX 50-series streaming up to 5K/120 or 360 frames per second. Shadow starts at $37.99 a month for a persistent Windows PC with 512 GB of storage.[12][13] Vectordash's 2019 price was attractive, but a startup could not match an incumbent's hardware purchasing, network, and catalog power with residential supply alone.
Business Model
The AI marketplace charged renters by the hour and paid hosts for usage. A contemporary forum discussion inferred a margin near 50% on one configuration, though Vectordash did not confirm that figure.[5] Gaming replaced variable demand with a $28 monthly subscription and promised hosts $60-$105 per month.
Those prices concealed a capacity trap. One $28 subscriber could not cover a host earning up to $105, support, payment costs, and idle capacity. The model needed several customers per host, but latency and upload limits prevented mining rigs from serving several simultaneous streams. No reliable revenue or gross-margin data was disclosed.
Traction
The strongest number was supply: more than 120,000 Nvidia GPUs listed on the compute marketplace. Gaming evidence was smaller but concrete. The company reached YC, earned TechCrunch coverage, demonstrated low-latency Apex Legends, and paid for a niche MTGGoldfish campaign. Users in 2020 described functional full desktops and quick provisioning, including access outside the original Bay Area.[3][8]
Post-Mortem
Abundant hardware was not usable capacity
Vectordash counted GPUs, but customers bought sessions. A mining rig might expose 60 or 100 cards while sharing one residential connection. Shameem later said: "you could really only utilize one of those GPUs." A 4K stream consumed 15-20 Mbps, close to the upload capacity he observed on typical US connections.[3]
The team tried geographic matching, host-uptime scoring, adaptive resolution, and coastal launches. Those measures improved individual sessions but did not change the denominator. Every market needed enough reliable hosts with spare upload, the right GPU, and competitive economics. Expansion multiplied cold starts instead of pooling capacity.
The first customers could not finance the marketplace
Cheap AI compute solved a real founder problem, but Vectordash recruited the buyers least able to support a two-sided infrastructure company. Students had low budgets. Startups had credits. Enterprises had money but distrusted residential machines with proprietary data. The gaming pivot found customers willing to subscribe, yet it added real-time latency and persistent-storage expectations to an already fragile host network.
This was not simply a poor customer choice. Trust and service quality are capital goods in cloud computing. AWS, Nvidia, and Google can spend on data centers, security certification, peering, and reserved capacity before demand arrives. Vectordash tried to replace that capital with marketplace coordination. The approach lowered hardware cost but transferred reliability work to software and support.
Openness created product debt
A full Windows desktop let people install almost any game, which distinguished Vectordash from closed catalogs. It also made every launcher, anti-cheat system, save file, and storage requirement part of the experience. Users described reinstalling games, managing cloud saves, and losing access when no machine was nearby.[9]
The team used preloaded machines and Parsec to shorten setup. Those remedies could not deliver persistence at a $20 subscription while keeping capacity fungible. By early 2021, reports of unavailable machines exposed the structural failure: without reliable local supply, the broadest game catalog was irrelevant.
Stadia was a warning, not the cause
Google announced Stadia as Vectordash launched, and contemporary coverage predicted that a giant could undercut the startup.[14] But Stadia later closed, so incumbent size alone cannot explain Vectordash. The stronger counter-narrative is that cloud gaming was difficult for everyone. Vectordash's specific disadvantage was trying to meet data-center service levels with residential economics. Nvidia and Shadow survived by controlling the machines and network more tightly.
Key Lessons
- Count deliverable units, not listed assets. Vectordash's 120,000 GPUs sounded like liquidity. Shared upload links reduced many rigs to one usable gaming stream, so the marketplace's headline supply overstated what customers could buy.
- The cheapest buyer can be the most expensive market. Students loved low-cost GPUs, but supporting them required marketplace infrastructure their spending could not finance. Enterprise budgets came with security requirements the residential network could not meet.
- A broad catalog shifts work onto operations. Letting users install any Windows game removed licensing limits but introduced storage, launcher, anti-cheat, and support burdens. Openness was a feature and an operating cost.
- Distribution should match a narrow exclusion. The Magic Arena campaign targeted Mac owners blocked from a specific game. That wedge was stronger than selling generic cloud gaming, but the company did not prove it could repeat the motion across enough communities.
Sources
- Y Combinator company profile
- TechCrunch launch coverage
- Latent Space interview with Sharif Shameem
- Founder's original GPU-mining launch post
- fast.ai launch thread
- Vectordash CLI on PyPI
- Hacker News gaming launch discussion
- MTGGoldfish sponsored walkthrough
- Cloud gaming user discussion on storage and persistence
- April 2021 user report on service availability
- YC W19 company introduction
- Nvidia GeForce NOW memberships
- Shadow PC offers
- Contemporaneous Stadia competition analysis