On Tuesday, the MegaETH Foundation was renamed MEGA and committed $30 million from its cash flow to the creation of its own products, headlined by an inference market launched in December.
Inference marketplaces are an emerging category that I’ve been covering more and more recently: open marketplaces where anyone can buy, sell, or resell access to AI models, usually for less than what the big labs charge. Demand has been strong, with usage up and right in most markets.
With the category so early and the financialization of AI services even earlier, I think it’s a smart path for the team behind MegaETH, the Ethereum L2 built for speed, even if it’s a sharp turn. It’s also a shift made out of necessity. The chain TVL sits at just under $10 milliondown from May’s peak of over $245 million, and co-founder Shuyao Kong written the that the move was to take place “after having built the fastest chain and without having experienced pmf”.
Mechanisms, vendors and market models remain secret until launch, but a three-page white paper lays out the team’s thinking.
Let’s break down the pitch, where I think it’s weakest, and where MEGA could still stand out.
— (@)
The land
MEGA starts with what it calls a supply-side paradox, namely that many third-party suppliers use the same open models at a lower price than the labs that built them.
On OpenRouter, for example, GLM 4.6 costs $2.20 per million exit tokens from its developer, Z.ai, and $1.75 from Venice. Despite this difference, MEGA claims, buyers continue to pay more for the familiar option.
Its response is to introduce “purified intelligence,” a set of checks on a supplier’s model, privacy and security intended to prove that these unknown suppliers can purchase safely. The goal is a market where inference becomes interchangeable and, in the words of the journal, “who is the inference provider matters less and less.”
In other words, pitching leads with confidence, and especially with model checking. This is where my skepticism begins.
Verification progresses without MEGA
On the one hand, this hypothesis is only partially verified.
Cheap providers obviously don’t lose: OpenRouter, the popular inference router, already reward lower prices by default, and it pins most quality gaps between vendors on how they configure their tool calling software. Model swapping, the issue around which MEGA’s paper is built, is happening, although the clearest evidence comes from unofficial dealers, where a study published in March, on three of these services, found that 46% of identity checks failed.
More importantly, the tools to catch it are coming from everywhere:
-
Fingerprinting: Last week, inference marketplace Antseed activated a verification layer (fingerprinting) that asks sellers about obscure facts, each model responds in its own way, then steers buyers away from those who don’t use the model they’re advertising.
Do you like this article?
Subscribe to Bankless or log in
-
Sealed material: Trusted execution environments (TEEs), closed areas of a chip that even the machine’s owner can’t see into, can certify what software is running inside, a check for Antseed buyers. can already demand.
-
Cryptographic proofs: Earlier this month, Anthropic made available a prototype that mathematically proves which model produced a response, although it requires 70 to 95 times the computing power of a normal query. Pearl builds its own version with Attestable.
The MEGA newspaper cites all three and admits that each has limits, without specifying which ones it will use. He could adopt any one, or any combination, which would be great.
But many teams are tackling the same problem, and most of their tools are open to anyone. The reasons why buyers are wary of third-party suppliers are fading, and shipping verification won’t do much to distinguish MEGA.
Where MEGA could stand out
This shifts the weight of what could allow MEGA to stand out on two other ideas in the document.
The first is privacy. MEGA describes eliminating sensitive details from a request before a provider sees them and dividing tasks between trusted and untrusted providers. This is important because true confidentiality, where sealed hardware prevents the provider from reading your prompt, is today largely limited to open models. Scrubbing is the workaround for anyone who wants to privately access border intelligence.
It is also difficult to succeed. Freysa released a version last year in its Enchanted app, since renamed Silo, swapping private details for placeholders so users could query closed models like GPT and Claude. But Freysa said in November 2025 it was removing phone cleaning after performance varied by device and iOS app release notes listing the feature as removed in July.
The version of MEGA should also run out of sight of MEGA, on a user’s computer or inside sealed hardware. Otherwise, users simply trust MEGA with their sensitive information.
— (@)
The second idea concerns future capacity. The MEGA article describes a feature in which buyers reserve a specific model, volume, and minimum level of performance, such as speed and reliability, for a defined window at a fixed price and, optionally, redeem those reservations. Think of a company booking a hotel block months in advance of a conference. The buyer locks in capacity for a launch and can resell it if plans change. The supplier could take this committed request to a lender and borrow to purchase more GPUs.
Negotiate enough of these contracts and you get a public price for future inference. Future GPUs trading like ICE plan price of an hour on a chip, which says nothing about the model that works on it or its quality. MEGA’s contracts would set the price of the production itself.
None of the inference markets I’ve covered offer this, and this is where a blockchain is most useful. But there is a tension. Inference keeps getting cheaper and models are replaced within months, so why would a buyer set a price or model today? Maybe they’re paying for a guaranteed seat, or protecting themselves against the end of today’s subsidized prices. I don’t have a strong opinion yet.
What is still missing
For now, both ideas are just sketches. None of the codes are public, and the document does not specify who qualifies as a vendor, who performs the controls, or where the privacy tools would be run.
He also never mentions a token. Asked about the role of the MEGA token in the market, Kong told The Block“wherever there are ways to use the MEGA token, we will.” And the $30 million, more than half of MEGA’s outstanding market capitalization, comes with no published breakdown.
December should answer some of these questions. I will monitor named providers with clear entry rules, cleanup that escapes MEGA’s perspective and a first forward contract with real counterparties behind it. The check brings MEGA to the starting line. Privacy and future capability are what would set it apart.
About the Author
