Last week a foundation I had never heard of closed a $100 million Series B to build "the data availability layer for the agentic economy." The deck was gorgeous. The thesis fit in one sentence: autonomous AI agents will generate transaction and data volume that no existing chain can absorb, and whoever owns that bandwidth owns the next decade. I read the technical appendix twice. On the second read I noticed what the round's participants apparently had not — the appendix contained no throughput model. Not a conservative one, not an aggressive one. None. There was a chart of "projected agent GDP," a chart of "projected DA market size," and a straight line drawn between them as though causality were a design choice rather than a claim requiring evidence.
Every hack is a lesson in trustless verification. So is every fundraise. The difference is that a hack announces its failure loudly and immediately, while a bad thesis can survive a full liquidity cycle before the market discovers there was never a payload underneath the pitch.
I have watched this specific movie three times. In 2017 it was "scalability chains" built to absorb the coming wave of ICO traffic. In 2021 it was "metaverse-native L1s" built to absorb the coming wave of virtual worlds. In 2024 it was "modular DA" built to absorb the coming wave of rollups. Each time the wave arrived smaller than advertised, and the infrastructure designed to catch it sat at single-digit utilization while the token unlock schedule did its quiet work. I am not predicting the agent economy will fail. I am predicting that data availability will turn out to be the wrong place to stand when it succeeds.
That distinction matters, because the money currently flowing into "AI-first" chains and DA layers is not being allocated on the basis of a bottleneck. It is being allocated on the basis of a metaphor — the metaphor that machines will behave like retail traders, only faster and in greater numbers. Machines do not behave like retail traders. They behave like machines. And machines, it turns out, are extraordinarily efficient at not producing the kind of data that DA layers are built to sell.
The Manufacturing of Infrastructure Demand
Crypto does not discover infrastructure needs. It manufactures them, and it manufactures them on a remarkably reliable schedule.
The pattern is consistent enough to be traded. First, a genuine technological shift appears at the edges — ICOs in 2017, AMMs in 2020, generative models in 2023. Second, analysts extrapolate adoption curves using retail-human behavior as the unit of measure. Third, a bottleneck is declared: bandwidth, throughput, storage, compute, data availability. Fourth, capital floods into whatever layer claims to own that bottleneck, because the bottleneck is the most legible thing to underwrite. Fifth, the wave arrives at perhaps five percent of the projected volume, the infrastructure idles, and the narrative quietly relocates to the next bottleneck.
I lived through the 2017 version personally. At twenty-seven I spent six weeks auditing the 0x protocol's whitepaper and early smart contract interactions, and what I concluded — in a long piece I called "The Invisible Exchange" — was that the value was in the open-source atomic swap standard, not the token issuance. Infrastructure narratives outperform issuance narratives, I argued, and the developer community responded because the argument was testable. Everyone was building exchanges to capture ICO flow. Almost nobody was modeling how much flow would actually exist once the issuance machine stopped. When it stopped, the exchanges that had been built to catch it caught dust.
The 2020 DeFi summer repeated the lesson with better branding. I spent that season interviewing liquidity providers, collecting a few hundred qualitative data points, and the thing I kept hearing was not "I need more throughput." It was "I need to understand my impermanent loss." The bottleneck was informational and psychological, not infrastructural. The people who built for the declared bottleneck — faster settlement — captured less value than the people who built for the real one: risk comprehension.
By 2024 the declared bottleneck had become data availability. The rollout of blobs on Ethereum was genuinely elegant engineering, and the modular thesis that grew around it was genuinely coherent. Rollups would post their transaction data to a dedicated DA layer, cheaply, and that layer would become the settlement substrate for everything. Billions in token value attached to that forecast. I reviewed a dozen of those models. Every single one assumed rollup demand would grow faster than DA capacity prices would fall. One of those assumptions was wrong, and it is not the one the industry bet on.
Now it is 2026, and the declared bottleneck has been upgraded again. This time it is agents. AI agents, we are told, will transact with each other autonomously, at machine speed, in volumes that dwarf human activity, and they will require a purpose-built data layer to do it. I want to examine that claim the way I would examine any other: mechanically, from the bottom up, starting with what data availability actually costs and what agents actually produce.
What Data Availability Actually Is
Strip the marketing and data availability is a narrow, well-defined guarantee. When a rollup executes transactions off-chain, it must publish the underlying data somewhere so that any honest party can reconstruct the chain's state and prove fraud if it occurs. If that data is withheld, the rollup becomes a black box — users can be trapped, and the sequencer can steal. DA is not storage. It is a short-lived guarantee of retrievability, typically with a defined expiry window measured in weeks.
Ethereum's implementation is blobs. Each blob carries roughly 128 kilobytes. At the Dencun upgrade the network targeted three blobs per twelve-second slot, with a maximum of six, and later capacity increases pushed that ceiling higher. Run the arithmetic on the original targets and you get something on the order of two to three gigabytes per day of guaranteed DA capacity, expandable to roughly double that under congestion. That is not a lot of bytes. It is, however, more bytes than the market has been willing to buy.
This is the part of the story the pitch decks omit. In the quarters following Dencun, blob utilization on Ethereum oscillated between a low single-digit percentage and brief spikes during airdrops and inscription frenzies. Those spikes were real, and they were also transient — hours, not months. Independent DA layers tell a similar story from the other direction. Celestia's blocks carry megabytes at six-second intervals; EigenDA advertises a bandwidth ceiling measured in megabytes per second; Avail makes comparable claims. All of them are capacity-rich. None of them are capacity-constrained. The binding constraint on DA adoption has never been supply.
I want to be precise here, because "DA is overhyped" is a lazy formulation. The accurate formulation is that the DA market was underwritten against a demand curve that has not materialized, and the industry's response to that disappointment has been to search for a new demand curve rather than to re-examine the pricing model. Agents are that new demand curve. The question is whether they fit.
What Agents Actually Produce
Here is where the metaphor breaks down, and where I have actual experimental data rather than a vibes-based forecast.
For the past several months I have been running a simulation of autonomous agents operating inside a decentralized economic environment. The setup was deliberately unglamorous. I instantiated a population of agents that competed for a scarce compute resource using an escrowed token, with each agent required to post an attestation that it had actually performed the inference it claimed. I varied population size, verification latency, and the cost of lying. I ran it more times than I care to admit.
The first finding is arithmetic and boring: agents are tiny data producers. A payment between two agents is a few hundred bytes. A signed attestation that a model produced a specific output is a hash and a signature — on the order of one hundred bytes. Even a rich agent-to-agent interaction that includes an intent, a bid, a settlement, and a receipt rarely exceeds a couple of kilobytes. Assume a hundred million agent transactions per day — a number that no one has demonstrated and that would already represent a thousandfold increase over current on-chain activity — and you are looking at somewhere between ten and a few hundred gigabytes per day, depending entirely on how much state each agent chooses to carry. A meaningful fraction of that would be batched, compressed, or settled net rather than gross.
Compare that to the declared market. The Series B I opened with is not being raised to serve a few hundred gigabytes per day. It is being raised to serve the imagined output of an agent economy modeled as if every agent were a human trader posting its entire thought process to a blockchain. That model is wrong on the physics. Inference happens off-chain. What touches the chain is the proof, not the thinking.
The second finding is more interesting and, I think, more important. In the simulations, throughput did not converge on the data availability ceiling. It converged on the verification ceiling. As I increased agent population, the binding constraint was never how fast attestations could be published — blobs and DA layers had headroom to spare. The binding constraint was how long it took to verify that an inference had genuinely been performed by the model an agent claimed to be running. When verification latency rose, agents stopped transacting, not because they ran out of bandwidth but because they could not trust each other fast enough. The system throttled on trust, not on throughput.
Every hack is a lesson in trustless verification, and this is the lesson the agent economy is about to learn in public. An agent that acts on a forged attestation is a hack in slow motion. An agent that pays for inference that never happened is a hack in slow motion. The scarce resource in machine-to-machine commerce is not data availability. It is verifiable computation and cheap, reliable attestation of what actually ran. Those are compute-bound and cryptography-bound problems. They are not bandwidth problems, and no amount of DA capacity solves them.
The Deployment Order Problem
There is a subtler structural issue, and it is the one that makes me skeptical of the entire "agent DA" category rather than merely unimpressed by its sizing.
DA layers are, by design, the most commoditized component in the stack. Erasure coding and sampling are solved problems. The engineering is public. The cost floor is dominated by hardware and by the security budget you are willing to pay for, and because the service is only as valuable as the settlement layer it serves, DA providers are structurally pushed toward a race to zero on price. That is fine if you own the settlement layer. It is a terrible place to be if you are a standalone token whose only product is the commodity.
The agent economy, if it arrives, will settle where composability is cheapest and where identity and reputation already live. Agents will not care which chain sells the cheapest bytes; they will care which environment lets them find counterparties, verify them, and enforce contracts without bridging risk. That is a network-effects question, not a bandwidth question. The chains that win agent settlement will win because they accumulate trusted agents and verifiable compute markets first. DA is an implementation detail they will buy at spot prices, the way you buy electricity.
Which brings me back to the money. Roughly two billion dollars, by my rough tally of announced rounds and token sales over the past four quarters, has been committed to infrastructure explicitly branded for the agentic economy — and the largest single category within that is data availability and "agent-native" execution layers. This capital is being deployed to the wrong layer of the stack, in the wrong order, against a demand curve that the deployers have modeled as a straight line between two unrelated charts.
That is not to say the capital is worthless. It is to say it is early in a way that is indistinguishable from wrong. In 2017 the scalability chains eventually found uses. In 2021 the metaverse L1s did not. The difference was whether the underlying demand curve was real. Agents are real. The question is what they consume.
Where the Bottleneck Actually Is
If I were allocating a hundred million dollars this cycle — and occasionally people ask me to — I would not fund a DA layer. I would fund three things, none of which are bandwidth.
I would fund verifiable inference. The ability to prove, cheaply and quickly, that a specific model produced a specific output is the load-bearing primitive of any machine economy. Today it is too expensive and too slow by an order of magnitude, and that gap is what is actually throttling agent-to-agent commerce in every simulation I have run.
I would fund agent identity and reputation. Before two machines can transact at scale they need a way to know who they are dealing with and a way to price the risk. This is the same problem DeFi faced in 2020, when the real innovation was not faster blocks but the ability to reason about counterparty risk.
And I would fund the boring settlement netting layer — the streaming and channel infrastructure that lets a million micro-interactions collapse into a handful of on-chain finalizations. This is unsexy and it is where the volume actually gets absorbed. It also happens to destroy the demand case for dedicated DA, which is precisely why nobody is raising a Series B for it.
Contrarian: The DA Thesis Is Backwards
The consensus view is that agents will arrive first at the infrastructure layer, then discover they need somewhere to put their data, and that this somewhere must be purpose-built because general chains are too expensive and too slow. I think this is exactly backwards, and the reversal is instructive.
Agents will arrive at the settlement layer last, not first. They will transact in private, off-chain, and in channels for as long as possible, because that is the rational behavior of any economic actor minimizing cost. They will touch a public chain only when they must — for netting, for dispute resolution, for the finality that makes a contract enforceable. By the time that happens, the data footprint per interaction will have been compressed and batched down to a rounding error, and the DA capacity they require will be a fraction of what the current models assume. The infrastructure built to serve the anticipated flood will be serving a trickle, and it will be serving it at commodity prices.
The blind spot in the bull case is the assumption that machine activity looks like human activity, only larger. It does not. Humans produce noise, sentiment, and redundant data. Machines produce exactly the data required to close a transaction and nothing else. That is not a smaller version of the human market. It is a structurally different one, and its most distinctive feature is that it is thin.
Takeaway
The agent economy is coming. I have spent months inside a synthetic version of it, and I believe the machine-to-machine market will eventually be larger, faster, and stranger than anything retail ever produced. But the value will not accrete to whoever sells the cheapest bytes. It will accrete to whoever solves verifiable inference, to whoever owns agent identity, and to whoever can make trust scale faster than transaction volume.
So the next time a foundation raises a hundred million dollars to build a data availability layer for autonomous machines, ask one question the appendix will not answer: how many bytes does an agent actually produce? The number is smaller than the deck assumes, and the gap between the deck and the number is the entire investment thesis. Watch where the two billion dollars goes next quarter. If it moves from bandwidth to verification, the market has learned something. If it moves to the next metaphor, it has only learned to be early again from a different direction.