Author: David Christopher, Bankless
Translated by: DeepChao TechFlow
DeepChao introduction: MegaETH, the Ethereum L2 project once known for pushing performance to the limit, has officially rebranded as MEGA and announced it will allocate $30 million from its treasury to develop its own products. Its first application, focused on the decentralized AI inference market, is expected to launch in December.
As model verification technologies on the market become increasingly homogenized, proving that nodes “haven’t tampered with anything” is no longer enough to build a moat. This article takes an in-depth look at the two most compelling differentiating ideas in the MEGA white paper: privacy-preserving de-identification without a centralized point of view, and compute futures contracts with a compelling vision for the future.
This Tuesday, the MegaETH Foundation officially rebranded as MEGA and announced a $30 million treasury allocation to build products in-house. Leading the lineup is an AI inference market, set to launch in December.
Inference markets are an emerging sector I’ve been following closely. In these open markets, anyone can buy, sell, or resell access to AI model calls, typically at prices below the official rates charged by major labs. Demand is strong, and usage across most inference markets is steadily rising.
Given that this niche is still in its early stages, and the financialization of AI services is even more nascent, I think this is a smart move—even if it’s a sharp turn—for the MegaETH team, which built an Ethereum L2 known for its “extreme speed.”
The pivot also reflects a certain real-world urgency. The chain’s TVL currently stands at less than $10 million, down sharply from its peak of $245 million in May. Co-founder Shuyao Kong also candidly admitted on X that, “after building the fastest blockchain but failing to find product-market fit (PMF),” the team had to make a change.
The specifics of the new inference market’s mechanics, compute providers, and launch model lineup remain under wraps, but a three-page white paper has already outlined the team’s core thinking.
Let’s break down the central claims behind this business narrative, examine where it’s weakest, and ask what MEGA can still use to break through.
In a post, the official MEGA account (@mega_holdings) announced: “Today, we’re committing $30 million to build new products on MegaETH, starting with AI inference. Learn more at mega.holdings.”
The Core Narrative
MEGA’s entry point is what it calls the “supply paradox”: many third-party compute providers charge far less to run the same open-source model than the original lab that developed it.
Take OpenRouter, for example. Z.ai, the developer of GLM 4.6, charges $2.20 per million output tokens, while the third-party provider Venice charges just $1.75. Despite the obvious price difference, MEGA argues that most buyers still prefer to pay more for the familiar official channel.
MEGA’s proposed solution is to introduce a suite of verification mechanisms called “Purified Intelligence.” By checking providers’ model authenticity, privacy compliance, and security, it aims to prove that unfamiliar third parties can be trusted too.
The ultimate goal is to build a marketplace where inference becomes fully standardized. As the white paper puts it, in an ideal world, “who provides the inference service will matter less and less.”
In other words, the entire narrative hinges on “building trust”—and, in particular, on “verifying that the model being claimed is actually the one being run.” That’s exactly where my doubts begin.
Verification Technology Is Advancing Rapidly—Even Without MEGA
First, the industry premise MEGA takes for granted doesn’t entirely hold up.
Low-cost compute providers aren’t being shunned by the market. Popular inference routing platform OpenRouter already prioritizes low-cost routes by default. The platform attributes most differences in output quality between providers to variations in their tool-calling software configurations, rather than deliberate model substitution.
The phenomenon of “model swapping,” which the white paper criticizes as a foundational problem, certainly exists. But the strongest evidence mostly comes from shady, unauthorized resale channels. A study published in March tested three such black-box resellers and found that identity checks failed in as many as 46% of cases.
More importantly, tools for catching this kind of model substitution are emerging rapidly from all directions:
• Fingerprinting: Last week, inference marketplace Antseed launched a new fingerprint verification layer. The system tests compute sellers with obscure questions that only specific models can answer. If it detects that a provider isn’t running the model it claims to be, it immediately routes the buyer’s request elsewhere.
• Trusted Execution Environments (TEEs): A TEE is like a secure black-box enclave inside a chip that even the machine’s owner can’t inspect. It can publicly attest to exactly what code is running inside the chip. Antseed buyers can now directly ask providers to enable this hardware verification.
• Cryptographic proofs: Earlier this month, Anthropic open-sourced a prototype called FlashProve that uses mathematical methods to verify that a response was generated using specific model weights, although each inference takes 70 to 95 times more compute than a standard request.
The Pearl team is also developing its own proof system using Attestable.
MEGA’s white paper names all three approaches and acknowledges that each has technical limitations, but it doesn’t say which one the team will ultimately use. MEGA will most likely adopt one of them, or some combination. That would certainly be a positive step.
The reality, however, is that many teams are already tackling this same pain point, and the vast majority of verification tools are open source and available for the entire industry to use. Buyers’ reluctance to trust third-party providers is rapidly fading, and simply launching a “model verification” feature is no longer enough for MEGA to stand out.
The Two Areas Where MEGA Could Really Break Through
Given that, the real differentiators that could set MEGA apart lie in the white paper’s other two big ideas.
The first is privacy protection. MEGA envisions scrubbing sensitive details before computing providers can see user requests, then splitting complex tasks between trusted and untrusted nodes. This design addresses a real need: privacy protections that use secure hardware to prevent providers from snooping on prompts are currently mostly limited to open-source models.
For users to securely access closed-source frontier models while preserving privacy, “local scrubbing of sensitive information” is almost the only viable compromise. But implementing this approach is also extremely challenging.
Last year, Freysa introduced a similar feature in its Enchanted app, now renamed Silo. It allowed users to securely access closed-source models such as GPT and Claude by replacing private words and phrases with placeholders.
But in November 2025, Freysa announced that inconsistent performance on phones had forced the team to move the scrubbing process off mobile devices. Its iOS app’s July release notes went further, listing the feature as discontinued and removed.
Likewise, MEGA’s scrubbing process would have to run entirely outside MEGA’s view—for example, entirely on the user’s local computer or in secure hardware. Otherwise, users would essentially be handing sensitive information that was meant for an inference provider over to MEGA instead.
When publishing the white paper recently, MegaETH co-founder Yilong Li (@yilongl_megaeth) also noted: “Expanding AI compute capacity gets only half the job done. Even if providers have the right models, ample spare compute, and lower prices, users still won’t pay if they can’t be sure providers are actually running the models they claim to be and protecting their data.”
The second key highlight is Future Capacity. The MEGA white paper envisions a mechanism through which buyers can reserve a specific model, a fixed volume of calls, and minimum performance standards—including response speed and reliability—at a locked-in price for a specified future window. It would also eventually support the resale of these future-compute contracts on a secondary market.
You can think of it like a company booking a large block of hotel rooms months in advance of a major industry conference. Buyers can lock in enough compute ahead of a major product launch and resell it if plans change. Compute providers, meanwhile, can use these committed bookings as collateral to borrow money and buy more GPUs to expand capacity.
If trading in these contracts reaches scale, the market could give rise to a transparent, publicly available “forward inference pricing curve.” Traditional financial giants such as the Intercontinental Exchange (ICE) may plan to launch GPU compute futures that can hedge the cost of renting chip-hours, but those contracts can’t specify which model will run or what quality of output it will produce. MEGA’s forward contracts, by contrast, are directly anchored to the model’s final “output itself.”
Among all the on-chain inference markets I’ve reviewed so far, none offers a similar mechanism. And this is precisely where blockchain technology can deliver the most value through establishing and transferring ownership rights.
Yet a fundamental tension remains: AI inference costs are falling rapidly, and leading models are completely replaced every few months. Why would buyers lock in a fixed price—or even a specific model—for months in advance?
Perhaps enterprise buyers aren’t purchasing a lower price so much as “a guaranteed seat at the table during peak demand, with no risk of downtime.” Or perhaps they want to hedge against skyrocketing inference bills once today’s massive subsidies from major AI providers are scaled back. I don’t yet have a clear answer.
The Key Pieces Still Missing
For now, at least, these two big ideas remain no more than sketches in a white paper. The team has yet to release any code, and the paper doesn’t clarify which compute providers will qualify, who will conduct verification and audits, or where the privacy-scrubbing tools will ultimately run.
The white paper also says nothing about a native token. Asked what role the MEGA token would play in the inference market, Shuyao Kong told The Block: “If there are use cases that can give the MEGA token utility, we’ll make use of them.”
No detailed budget has been disclosed for the $30 million treasury allocation, which is worth more than half of MEGA’s current circulating market cap.
Some of these questions should start to be answered in December. I’ll be watching three key measures of execution: the named providers that sign on first and their eligibility requirements; a scrubbing solution that truly runs outside MEGA’s view; and the first forward-compute contract backed by real counterparties.
Ultimately, verifiability can get MEGA to the starting line in the inference race. Only solving the problems of genuine privacy and future compute delivery will determine whether it can truly stand out.