In the pursuit of powering artificial intelligence (AI) and the race to come out on top, data has become the new oil.
Companies that control the pipelines are pulling in record venture capital (VC), with centralised AI web infrastructure startups raising more than $280 million over the past 18 months.
Nimble secured $47 million in February, Browserbase closed a $40 million Series B in June 2025, TinyFish and Airtop followed with $47 million and $38.8 million, respectively, but all these companies share a flaw: they all operate on a centralised model.
Perceptron is the exception to this pattern.
While competitors burn through investor cash to build out their server farms and data center networks, Perceptron assembled a decentralized infrastructure powered by 700,000 user-run nodes across 150 countries.
The network is already processing approximately 10 terabytes (TBs) of data daily through real residential endpoints — a metric that hits harder than any funding announcement.
The centralized bottleneck
Centralized AI data infrastructure poses a structural problem from the outset, limiting geographic diversity, resulting in data center IP addresses getting blocked, and hindering scaling.
In the centralized model, scaling requires the purchase of increasingly more data servers, which means further capital expenditure and the introduction of more single points of failure.
Two elephants in the room must be addressed. First, AI companies are hungry for training data and the appetite is bottomless.
Specialised datasets can result in costs up to 40 times more than generic alternatives, and the demand is exploding, with examples like Reddit’s $200 million+ agreements with Google and OpenAI.
Second, AI agents are hungry for access. As autonomous agent workflows become central to how AI delivers value, those agents need reliable, unblocked pathways to the open internet.
When data center IP addresses get flagged and blocked, an increasingly common reality for centralized providers, it’s not just an infrastructure inconvenience.
It’s an agent that can’t complete a task, a workflow that breaks, a customer that churns.
Pair both realities with the constraints of centralized AI models and the situation becomes critical.
Perceptron discards that centralized approach for a better and more sustainable one without burning the VC cash.
Instead of building data centers, the Perceptron network incentivises users to share idle bandwidth through browser extensions and mobile applications (apps).
These residential endpoints look like real human users, making them harder to block and far more valuable for AI training purposes.
The skepticism around decentralized infrastructure is understandable when many projects promise scale and efficiency but then never deliver.
In the case of Perceptron’s metrics, it suggests otherwise, circumventing bottlenecks and instead, creating a network with over 700,000 registered nodes of which 100,000 are bandwidth-verified across 150 countries.
Perceptron is already generating revenue through live client contracts, with the company reporting 92% lower costs compared to legacy centralised providers.
This figure is vital to AI startups operating on already thin margins that can only become thinner in a centralized AI environment.
The company already has production relationships with three partners that validate its infrastructure’s real-world utility:
Everlyn, an AI video model that receives video data feeds from Perceptron’s network.
BrickRoad, which operates as a marketplace partner, reselling Perceptron’s datasets to companies — including OpenAI.
Aethir, which has integrated Perceptron into its broader ecosystem to support AI agent infrastructure and decentralized compute workloads, also providing an infrastructure development grant to support early compute development.
These are not testnet deployments or pilot programs. They are proof that Perceptron’s network delivers data at scale without doubt and in a real-world setting.
Weighing the costs
The cost differential is perhaps the most compelling argument for decentralized infrastructure on the whole.
Scale AI, a centralised competitor, employs over 100,000 people on payroll to label and curate data.
Perceptron achieves similar scale through token-based incentives, with contributors self-selecting and getting paid per task. The company itself carries zero payroll burden.
The structural advantage is undeniable and irreplaceable and as more users join the Perceptron network, the infrastructure becomes stronger without proportional cost increases.
The 700,000-node distribution is not something competitors can simply replicate overnight and instead, it represents a defensive moat built through organic growth rather than mere capital deployment.
Perceptron is preparing to launch a data-questing platform that will enable users to contribute data, or conduct structured annotation and labeling tasks through its node network.
The company has also launched a $10 million AI Data Fund to support emerging AI projects, offering selected teams five weeks of data support and up to 5 TB of real-world data free of charge.
With mobile expansion beyond the current 10,100 Android downloads, iOS access lists already forming, and a target of 5 million nodes by the end of the year, Perceptron demonstrates that the decentralized data economy surpasses any centralized VC cash-burning model.
For AI companies weighing infrastructure options today, choosing decentralized networks that work at scale over centralized alternatives can mean the difference between success and liquidation. Choose wisely.
The post While VCs burn cash, Perceptron built AI infrastructure with 700,000 users appeared first on Invezz
