Last year, a foundation model lab tried to license data from a robotics company.
The data was exactly what they needed, two years of real-world manipulation footage, a physical environment no competitor had access to. The kind of dataset you can't manufacture in simulation.
The deal stalled. Not on price.
The lab asked one question: can you prove this data is what you say it is? Not a hostile question, a standard one. They needed a chain of custody independent of the seller. Something a third party could verify, not just take on trust.
There was no answer. The data lived in an S3 bucket with timestamps in a database the same team administered. Nothing independently verifiable. No deal.
This is the situation most physical AI companies are in right now. Sitting on datasets that AI labs, research institutions, and model developers will pay serious money for, stored in infrastructure that makes those datasets harder to sell than they should be.
Physical AI Data Is Becoming an Asset Class
Physical AI is scaling faster than most infrastructure discussions acknowledge. Global robotics investment hit record levels in 2025, industry trackers put the figure in the $35–38 billion range. The Unitree G1 humanoid ships today at $13,500; EngineAI launched the T800 at $25,000 at CES 2026; 1X is taking NEO pre-orders at $20,000. Price targets that were widely cited as "by 2030" goals have arrived, roughly four years early, and the next floor, sub-$15K units, is now a 2030 question. NVIDIA announced its Physical AI Data Factory Blueprint at GTC in March 2026, a reference architecture for unifying how physical AI training data gets generated, augmented, and evaluated at scale.
The hardware story is moving fast.
The more interesting story is underneath it: the data these systems generate is starting to look less like an operational byproduct and more like a balance-sheet asset. Not just input to your own models, but tradeable, licensable, and potentially yield-generating. Sensor feeds from 40,000 physical locations. Functional imaging data from live neural tissue experiments. A humanoid robot's first 10,000 hours of real-world task execution. None of this can be reconstructed after the fact.
When data is unique, irreplaceable, and increasingly in demand from well-funded labs, it becomes an asset. The question is whether you're storing it in infrastructure designed for an asset, or infrastructure designed for server logs.
Why SkyMapper Stores Every Frame on Akave Cloud?
SkyMapper is building the world's first decentralised, blockchain-verified global telescope network, 62 SkyBridge units deployed across six continents, with the SETI Institute as a partner and a target of 1,000 connected telescopes by the end of 2026.
Every frame SkyMapper captures is irreplaceable. A satellite pass. A transient astronomical event. A comet trail at 3:47 AM. The sky doesn't offer a second take.
SkyMapper stores every one of those frames on Akave Cloud with cryptographic provenance, an on-chain record written at ingestion, verifiable by any reader without trusting SkyMapper or Akave. When a frame contributes to a published orbit determination or a transient discovery, the underlying data is defensible: this is the exact, unaltered image the telescope recorded at that moment.
That's the shift from "we have the data" to "we can prove what the data is."
For physical AI companies outside of astronomy, robotics fleets, IoT networks, neurotech labs, the same problem applies. The data exists. The question is whether it's stored in a way that makes it provable, sellable, and independently verifiable.
What Standard Cloud Storage for Ai Gets Right, and Where It Misses?
AWS S3, Azure Blob, and GCS are well-engineered products built for a specific class of problem: large volumes of data, high availability, regenerable sources. For internal training pipelines, non-commercial datasets, and reproducible data environments, they work well. Plenty of physical AI companies will continue using them for exactly those use cases.
The mismatch appears when the data becomes something you want to monetize or verify independently.
1. On provenance:
AWS offers Object Lock for tamper resistance, CloudTrail for audit logs, and external notarization options. You can approximate independent provenance on AWS with enough architectural investment. The difference is that Akave writes a cryptographic record at the storage layer by default, on an independent ledger, without requiring additional tooling or trusting the storage provider. For a physical AI company whose primary sales argument to a data buyer is "this data is what we say it is," building that verification infrastructure on top of a hyperscaler is a project. On Akave it's the baseline.
2. On egress:
AWS prices data leaving its infrastructure on a tiered schedule, $0.09/GB for the first 10TB, stepping down through $0.085, $0.07, and $0.05/GB at higher volumes. This applies to internet egress; data transferred between AWS services within the same region is free. For companies whose training compute also runs on AWS EC2 or SageMaker, intra-AWS transfers won't incur these costs. But for companies moving data to external buyers, partner labs, or non-AWS training infrastructure, the numbers add up quickly.
A 200TB pull to an external destination costs approximately $13,800 on AWS at current tiered rates. That resets monthly. Every external marketplace transaction is another egress event.
3. On custody:
Centralised storage isn't the wrong architecture universally, financial services and healthcare run petabytes of mission-critical data on centralised cloud infrastructure and pass rigorous audits. The specific challenge for physical AI data is that it can't be recreated from source. "Recoverable from backup" and "permanently lost" are not equivalent outcomes when the data was collected once, in conditions that won't recur.
The Business Models That Are Opening Up
The data marketplace for physical AI training data is an early-stage bet, not a proven liquid market. Pricing models are still forming. Demand isn't fully established. Standards haven't been set.
But the directional signals are clear, and the companies that own verifiable, sovereign datasets will be better positioned when the market matures, because provenance isn't a gate that blocks monetization today, it's a factor that increases trust, pricing power, and the ability to transact at scale with multiple buyers.
Three paths are taking shape.
1. Data marketplace participation.
AI labs are actively sourcing real-world training data they can't collect themselves. 375AI, whose edge nodes cover across the U.S, each collecting multimodal data through high-resolution cameras, environmental sensors, GPS, and an NVIDIA GPU, already operates a token model where external buyers purchase data credits burned against verified observations. The edge AI sector, which 375AI themselves put at ~$66 billion with 21% annual growth through 2030 (their own market sizing, worth reading critically), is built on the premise that physical-world data collection at scale creates durable value. Provenance doesn't make that sale happen, but data without a verifiable chain of custody trades at a discount and doesn't scale cleanly across multiple buyers.
2. Direct licensing to foundation model labs.
Some datasets can't be replicated at any price: a neurotech lab accumulating biological neural network data at mouse-brain scale, a humanoid robotics company with thousands of hours of real deployment data across force-feedback and proprioception logging. Companies like Netholabs, building at the respective frontiers of neural emulation and humanoid deployment, are accumulating datasets no competitor can duplicate. Whether they're Akave customers or not, the strategic question facing any company in this position is the same: is your data stored in a way that a third-party buyer can independently validate?
3. Ongoing data access revenue through DePIN.
The decentralised physical infrastructure network model turns sensor deployment into a continuous data business. Buyers access verified, distributed data on an ongoing basis rather than purchasing a one-time dataset. The prerequisite in every version of this model is verifiability, data that demonstrably is what it claims to be, from the claimed location, at the claimed time.
The Math: What 500 Edge Nodes Actually Costs
500 edge nodes. 50GB per node per day. 750TB per month.
- AWS egress costs apply to data moving outside AWS infrastructure. Transfers between AWS services in the same region are free. If your training compute runs on AWS EC2 or SageMaker, intra-AWS egress costs are significantly lower.
On storage alone, Akave runs ~35% cheaper at this scale. Add one external training run and the gap widens to roughly $19,800/month. Add marketplace access, partner lab data shares, or a second training job and it compounds further.
Provenance Is Strongest at Ingestion
The strategic argument isn't "you can't monetize without blockchain provenance." Contracts, reputation, and validation pipelines have closed data deals without on-chain verification, and they'll continue to.
The argument is narrower and more defensible: provenance is strongest when it's captured at the moment of ingestion. Retrofitting it later is possible, you can hash historical datasets, build partial lineage, add attestations after the fact. But retrofitted provenance introduces gaps that are hard to defend in front of a serious buyer. You can approximate trust, but you can't fully reconstruct origin integrity once the chain of custody has a gap.
For physical AI data specifically, where the value proposition to a buyer is "this was captured in the real world, in these conditions, at this time", origin integrity is the core claim. Capturing it at ingestion is structurally stronger than constructing it retrospectively.
The datasets being collected today are the inventory available in 12–24 months. The window to capture provenance at the source is now, while the data is being generated.
A Note on Migration Complexity
Our S3-compatible interface means standard object-storage workflows connect by updating your endpoint and credentials. That's accurate for most use cases.
It's not the full picture.
Physical AI infrastructure often involves streaming ingestion pipelines, real-time edge-to-cloud sync, GPU-coupled data workflows, and IAM-heavy architectures. If your stack uses Lambda triggers on S3 events, KMS-managed encryption keys, or tightly coupled AWS-native tooling, migration is a project, not a credential swap. We're direct about that.
For new deployments, greenfield robotics platforms, new IoT networks, neurotech labs setting up infrastructure, Akave is designed to be the default from day one. For existing deployments with complex AWS-native dependencies, the honest answer is: evaluate modularly, start with workloads that are easy to move, and plan the rest carefully.
How We Built Akave for This?
We built Akave Cloud for AI and data workloads where provenance and sovereignty need to be requirements built into the storage layer, not added on top.Akave runs a dedicated immutable storage ledger. NVIDIA's Physical AI Data Factory Blueprint, announced at GTC in March 2026 as a reference architecture for generating and evaluating physical AI training data at scale, points toward exactly the kind of pipeline that needs a verifiable storage layer underneath it. Akave is designed to be that layer: S3-compatible on the interface, cryptographically provable on the record.
Every object you store in Akave gets a cryptographic proof at ingestion, an on-chain record verifiable by any reader, independent of us. Verification metadata and erasure set commitments go on-chain so integrity can be confirmed without moving the raw data. This is what SkyMapper uses to make every telescope frame defensible as scientific evidence: not a claim backed by institutional trust, but a proof anyone can check.
Durability is RS(32,16) Reed–Solomon erasure coding, 32 fragments of which16 parity fragments, distributed across a decentralised node network. The system tolerates 16 simultaneous node failures. Eleven nines. No single institution as gatekeeper.
Where This Goes?
Humanoid deployments are scaling from thousands to millions of units. IoT sensor networks are instrumenting physical environments at densities that didn't exist five years ago. The volume of irreplaceable real-world data is compounding faster than the storage conversation tracks.
The data marketplace for physical AI is forming but not mature. Standards are being set. Pricing models are being established. The companies entering it with clean, verifiable, sovereign datasets will have stronger negotiating positions, not because provenance is a prerequisite for every deal, but because it expands the pool of buyers, raises the defensibility of pricing, and makes the data usable across multiple transactions at scale.
The infrastructure decisions being made now will determine whether the data you're collecting is an asset you can fully control and monetize, or a resource tied to an architecture that wasn't built for this use case.
FAQ
What is physical AI data, and why is it different from regular cloud data?
Physical AI data is sensor, behavioral, and observational data generated by robots, IoT devices, telescopes, neurotech instruments, and other physical-world systems. Unlike software logs, you can't regenerate it, if a sensor feed from a specific location is deleted, that moment is permanently gone. That irreversibility makes data ownership a strategic decision, not an infrastructure afterthought.
Who actually owns the data that physical AI robots and sensors collect?
You do, legally. But legal ownership and commercially useful ownership aren't the same thing. Data stored under standard hyperscaler terms is subject to egress fees that constrain access, data residency rules that create compliance exposure, and service terms that shift over time. Commercially useful ownership means data stored with cryptographic provenance, zero-egress access, and no single-vendor gatekeeper, so you can access, license, and sell it on your terms.
How can robotics and IoT companies monetize their sensor data?
Three paths are forming: data marketplaces where AI labs purchase verified real-world datasets; direct licensing to foundation model developers who can't collect that data themselves; and DePIN-style ongoing data access models where buyers pay for verified access to a distributed network. Provenance doesn't make these deals possible on its own, but data with a clean, independently verifiable chain of custody commands higher trust, better pricing, and scales more cleanly across multiple buyers.
Isn't hyperscaler storage good enough for physical AI data?
For internal training pipelines, non-commercial datasets, and reproducible data environments, yes. The specific challenge for physical AI data is independent verifiability and egress economics when the data becomes something you want to monetize externally. Hyperscalers can approximate provenance with additional tooling, but it requires architectural investment. Egress costs apply to any data moving outside AWS infrastructure, for companies transacting with external buyers, that compounds quickly.
How does Akave provide verifiable provenance?
We write verification metadata and erasure set commitments to an immutable storage ledgerat ingestion, on an independent ledger verifiable by any reader, without trusting us. A data buyer, auditor, or research consortium can independently confirm a dataset is what the seller claims, without routing through us. This is what SkyMapper uses to make telescope observations defensible as scientific evidence.
What if I'm already on AWS with a complex pipeline?
For standard S3 object-storage patterns, updating your endpoint and credentials is the primary step and our S3-compatible interface handles the rest. For pipelines using Lambda triggers, KMS-managed keys, or custom IAM logic, those integrations need separate planning, migration there is a real project, not a credential swap. The cleanest path for most teams is to start with new deployments or modular workloads that don't carry complex AWS-native dependencies.
Why is provenance strongest when captured at ingestion?
You can hash historical datasets and add attestations retrospectively, but you can't fully reconstruct origin integrity once a gap exists in the chain of custody. Retrofitted provenance can approximate trust but leaves questions a serious buyer will ask. Capturing the proof at the moment of ingestion, before the data has moved through systems, before ownership has changed hands, produces the cleanest, most defensible record.
How does Akave's pricing compare to AWS S3?
At 750TB/month, Akave Cloud storage runs ~$11,250/month ($14.99/TB) versus ~$17,250/month on AWS S3, roughly 35% cheaper on storage alone. Add external egress for training runs, data sharing, or marketplace access (priced at zero on Akave versus ~$13,800 per 200TB pull on AWS) and the gap widens substantially at operating scale.
Get Started
See how we approach physical AI data storage at akave.com/product.
Start a free trial at akave.com/free-trial or review the S3-compatible integration docs at docs.akave.xyz.