MegaETH Rebrands as MEGA, Allocates $30M to the AI Inference Market

Author: David Christopher, Bankless

Compiled by: Deep潮 TechFlow

DeepChain Summary: MegaETH, formerly known for its extreme performance, has officially rebranded as MEGA and announced a $30 million treasury allocation to transition into self-developed products, with its first application focused on the decentralized AI inference market expected to launch in December.

Faced with increasingly homogenized model verification technologies on the market, merely proving that nodes have not substituted data is no longer sufficient to build a moat. This article provides an in-depth analysis of the two most compelling differentiated concepts in the MEGA whitepaper: privacy obfuscation脱离中心化视角的隐私脱敏, and the highly persuasive compute power forward contracts面向未来说服力十足的算力远期合约.

On Tuesday, the MegaETH Foundation officially rebranded as MEGA and announced a $30 million allocation from its treasury to develop proprietary products, headlined by an AI Inference Market set to launch in December.

The inference market is a new sector I’ve been closely tracking recently: in this open market, anyone can buy, sell, or resell access to AI model inferences, typically at prices lower than official rates from large labs. Current market demand is very strong, with usage across most inference markets steadily rising.

Given that this niche is still in its early stages and the financialization of AI services is just beginning, I believe that for the MegaETH team, which previously built the “ultra-fast” Ethereum L2, this is a sharp turn—but still a smart one.

This shift carries a similar sense of real-world urgency. The blockchain’s TVL has now dropped to less than $10 million, a significant decline from its peak of over $245 million in May. Co-founder Shuyao Kong also admitted on X platform: “After building the fastest blockchain but failing to achieve product-market fit (PMF),” the team must make changes.

The specific mechanisms, compute providers, and list of launched models for the new reasoning market remain confidential prior to official release, but a three-page whitepaper has outlined the team’s core vision.

Next, let’s break down the core claims of this business narrative to identify its weakest link and determine what MEGA can still leverage to break through.

The official MEGA account (@mega_holdings) announced in a tweet: “Today, we commit $30 million to building a new product on MegaETH, with AI inference as our first focus. Learn more at mega.holdings.”

MEGA's entry point stems from its proposed so-called "supply paradox": the cost of running the same open-source model through third-party compute providers is far lower than the original lab's pricing.

Using the OpenRouter platform as an example, GLM 4.6 from the developer Z.ai is priced at $2.20 per million output tokens, while the third-party provider Venice offers it for just $1.75. Despite the clear price difference, MEGA believes most buyers still prefer to pay a premium for the well-known official channel.

To address this, MEGA has introduced a comprehensive verification system called "Purified Intelligence," which evaluates the authenticity, privacy compliance, and security of service providers' models, demonstrating that even unknown third parties can be trusted.

Its ultimate goal is to create a trading market where reasoning capabilities are fully standardized. As stated in the whitepaper, ideally, “it will become increasingly irrelevant who provides this reasoning service.”

In other words, the entire narrative is built on "establishing trust," particularly on "verifying genuine model calls." But it is precisely here that I begin to have doubts.

First, the industry assumptions preset by MEGA are not entirely valid.

Low-cost compute providers are not currently being overlooked by the market: the popular inference routing platform OpenRouter itself prioritizes low-cost channels in its default routing, and attributes most differences in output quality between providers to variations in their tool-calling software configurations, rather than deliberate model substitution.

Although the "model swapping" phenomenon criticized in the whitepaper as a fundamental issue does objectively exist, the most conclusive evidence has largely occurred through informal resale gray markets. A study published in March found that, after testing three such black-box resellers, up to 46% of identity verifications failed.

More importantly, technological tools designed to detect these substitution activities are rapidly emerging from all directions:

• Fingerprinting: Last week, the inference marketplace Antseed officially launched a new fingerprint verification layer. This system tests sellers of computing power with obscure questions that only specific models can answer; if the seller is not running the model they claim, the system immediately redirects the buyer’s request elsewhere.

• Hardware Isolation (TEE): A Trusted Execution Environment (TEE) is a secure, black-box area within the chip that even the machine's owner cannot access, enabling public verification of exactly what code is running inside the chip. Buyers of Antseed can currently directly request service providers to enable this hardware verification.

• Cryptographic Proofs: At the beginning of this month, Anthropic open-sourced a prototype solution called FlashProve, which uses mathematical methods to verify that responses are genuinely generated by specific model weights, although each inference requires 70 to 95 times the computational power of a standard request.

In addition, the Pearl team is also developing their own proof version using Attestable developing their own proof version.

The MEGA white paper explicitly mentions all three of the above solutions and acknowledges their respective technical limitations, but does not clearly state which one (or combination) will ultimately be adopted. The team will most likely integrate one or more of these solutions, which is certainly commendable.

However, the reality is that numerous teams are already focusing on the same pain point, and most verification tools have been open-sourced and are directly accessible to the entire industry. The psychological barrier preventing buyers from trusting third-party service providers is rapidly fading, and simply launching a “model verification” feature is no longer enough to set MEGA apart.

Given this, the true differentiators for MEGA lie in the other two concepts outlined in the whitepaper.

The first direction is privacy protection. MEGA envisions sanitizing sensitive details before the compute service provider sees the user request, and splitting complex tasks to be processed separately by trusted and untrusted nodes. This design is highly practical, as current privacy protections that can truly prevent service providers from viewing prompts (Prompt) through closed hardware are still largely limited to open-source models.

To securely invoke proprietary state-of-the-art large models while preserving privacy, "local sensitive data cleansing" is nearly the only viable compromise. However, the engineering implementation of this solution is equally highly challenging.

Last year, Freysa offered a similar feature in its Enchanted app, now renamed Silo, by replacing sensitive phrases with placeholders to allow users to securely access closed-source models like GPT and Claude.

But Freysa announced in November 2025 that, due to inconsistent real-world performance across mobile devices, the team had to move the obfuscation computations off the mobile platform; its iOS app’s July update log directly listed the feature as removed.

Similarly, the MEGA version of the de-identification process must also be executed entirely outside MEGA’s own visibility—such as being fully conducted on the user’s local computer or within closed hardware—otherwise, users are merely transferring the sensitive information they originally entrusted to the inference provider to MEGA instead.

MegaETH co-founder Yilong Li (@yilongl_megaeth) also noted when releasing the whitepaper: “Expanding more AI computing power only solves half the problem. Even if service providers have the right models, ample spare computing capacity, and lower prices, users still won’t pay unless they can be confident that the providers are truly running the models as advertised and protecting their data.”

The second key highlight is Future Capacity. The MEGA whitepaper envisions a mechanism whereby buyers can, within a specific time window, pre-book specific models, fixed invocation volumes, and minimum performance standards (including response speed and stability) at locked prices, and later support secondary market trading of these future capacity contracts.

You can think of it like a company booking hotel rooms in bulk months in advance before organizing a major industry summit. Buyers can secure sufficient computing power ahead of a major product launch and resell it if plans change; meanwhile, computing power providers can use these locked-in demands as collateral to secure financing from lenders, enabling them to purchase more GPUs for expansion.

Once this type of contract trading reaches scale, the market will give rise to an open and transparent "forward pricing curve." While traditional financial giants like Intercontinental Exchange (ICE) plan to launch GPU computing power futures that can hedge against the cost of renting chip hours, they cannot specify which model will be run or the quality of the output; in contrast, MEGA’s forward contracts are directly anchored to the model’s final “output itself.”

Among all the on-chain prediction markets I have evaluated so far, none offer a similar mechanism—and this is precisely where blockchain technology most effectively enables ownership and value transfer.

However, a deeper contradiction remains: AI inference costs are plummeting rapidly, and leading models are being completely replaced every few months—so why would buyers lock in fixed prices or specific model versions months in advance today?

Perhaps enterprise buyers are not purchasing low prices, but rather a “guaranteed seat that never goes offline during peak hours,” or a hedge against the surge in usage bills following the reduction of massive subsidies by major AI vendors. At this point, I have not yet formed a clear conclusion on this matter.

At least for now, the two key highlights mentioned above remain only at the sketch stage in the whitepaper. The team has not yet released any code, and the whitepaper does not clearly specify which computing power providers are eligible for access, who will conduct the verification and approval, or where the privacy sanitization tools will ultimately run.

The entire white paper makes no mention of the native token. When asked about the role of the MEGA token in the reasoning market, Shuyao Kong told The Block: “We will leverage any scenario that empowers the MEGA token.”

Regarding the $30 million treasury allocation, which exceeds half of MEGA’s current circulating market cap, no detailed budget plan has been disclosed yet.

Some of these answers should begin to emerge in the upcoming December. I will be closely monitoring three key milestones: the first named service providers to sign contracts and their admission criteria, the anonymization solution operating outside MEGA’s scope, and the first forward computing power contract executed by real counterparties.

Ultimately, verifiability can only get MEGA to the starting line of the reasoning race; only by solving true privacy and long-term compute delivery issues can it truly stand out.