Bankless: After On-Chain TVL Declines, Pivots to a New Sector, MEGA Bets $30 Million on AI Inference Market

Author: David Christopher, Bankless

Compiled by: TechFlow

TechFlow Editorial: Ethereum L2 project MegaETH, once focused on extreme performance, has officially rebranded to MEGA and announced a $30 million treasury allocation to pivot toward proprietary products. Its first application targeting the decentralized AI inference market is expected to launch in December.

With model verification technology becoming increasingly homogenized across the market, simply proving that nodes "did not swap models" is no longer enough to build a moat. This article provides an in-depth analysis of two standout differentiating concepts in the MEGA whitepaper: privacy sanitization detached from a centralized perspective, and highly compelling future compute capacity contracts.

This Tuesday, the MegaETH Foundation officially rebranded to MEGA and announced a $30 million treasury allocation to build proprietary products, spearheaded by an AI Inference Market slated to launch in December.

The inference market is an emerging track I have been closely following recently: in this open market, anyone can buy, sell, or resell AI model invocation rights, typically at prices lower than official quotes from major labs. Current demand is very strong, with usage volumes across most inference markets steadily climbing.

Given that this niche segment is still early-stage and the financialization of AI services is in its infancy, I believe that for the MegaETH team behind the "extreme speed" Ethereum L2, this sharp turn is nonetheless a smart choice.

This pivot also carries a sense of realistic urgency. Currently, the network's TVL remains below $10 million, a significant drop from its May peak of over $245 million. Co-founder Shuyao Kong also admitted on X: "After building the fastest public blockchain yet failing to find product-market fit (PMF)," the team must change course.

The specific mechanisms, compute providers, and initial model list for the new inference market remain confidential ahead of the official launch, but a three-page whitepaper already outlines the team's core strategy.

Next, let's break down the core claims of this business narrative, identify its weakest links, and examine what else MEGA might rely on to break through.

The official MEGA account (@mega_holdings) announced via tweet: "Today, we commit $30 million to build entirely new products on MegaETH, with our first entry point being AI inference. Details at mega.holdings."

Core Narrative Claim

MEGA's entry point stems from its proposed so-called "supply paradox": many third-party compute providers run the same open-source model at a much lower price than the native lab that developed it.

Take the OpenRouter platform as an example: GLM 4.6 costs $2.20 per million output tokens from developer Z.ai, while third-party provider Venice charges only $1.75. Despite the clear price difference, MEGA believes most buyers still prefer paying the premium for familiar official channels.

To address this, MEGA's proposed solution is to introduce a suite of verification mechanisms called "Purified Intelligence," which verifies providers' model authenticity, privacy compliance, and security, demonstrating that unfamiliar third parties are equally trustworthy.

Its ultimate goal is to create a trading market where inference capabilities are completely standardized. As stated in the whitepaper, in an ideal scenario, "it will become increasingly irrelevant who provides this inference service."

In other words, the entire narrative rests on "building trust," particularly on "verifying the true invoked model." But it's precisely here that my skepticism begins.

Without MEGA, Verification Technologies Are Also Racing Ahead

First, MEGA's assumed industry premise does not fully hold.

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

While the "model swapping" phenomenon criticized as a foundational flaw in the whitepaper does objectively exist, the most conclusive evidence largely occurs in unofficial secondary resale gray markets. A study published this March found that after testing three such black-box resellers, identity verification failed in up to 46% of cases.

More crucially, technical tools designed to expose these swapping behaviors are accelerating from all directions:

• Fingerprinting: Last week, inference market Antseed officially launched a new fingerprint verification layer. This system tests compute sellers with obscure feature questions only specific models can answer; if they fail to confirm the seller is actually running the claimed model, buyer requests are immediately redirected.

• Closed Hardware (TEE): Trusted Execution Environments (TEEs) act as black-box secure zones within chips where even the machine owner cannot窥探, capable of publicly proving exactly what code runs inside the chip. Antseed buyers can currently directly request providers to enable this hardware verification.

• Cryptographic Proofs: Earlier this month, Anthropic open-sourced a prototype scheme (FlashProve) that uses mathematical methods to verify answers were indeed generated by specific model weights, although each inference requires 70 to 95 times the compute of a standard request.

Additionally, the Pearl team is also developing a proprietary proof version based on Attestable .

MEGA's whitepaper names all three approaches and acknowledges their respective technical limitations, but does not specify which one will ultimately be adopted. The team will likely integrate one or a combination of these, which is certainly commendable in itself.

However, the reality is that numerous teams are currently tackling the same pain point, and the vast majority of verification tools have been open-sourced for direct industry use. The psychological barrier of buyer distrust toward third-party providers is quickly fading, making it hard for MEGA to stand out solely on launching a "model verification" feature.

Two Directions Where MEGA Could Truly Break Through

If so, the real leverage that could truly set MEGA apart lies in the other two major concepts outlined in the whitepaper.

The first direction is privacy protection. MEGA envisions sanitizing sensitive details before compute providers see user requests, and splitting complex tasks between trusted and untrusted nodes. This design is highly practical, as current privacy protections that effectively prevent providers from peeking at prompts (Prompts) via closed hardware mostly remain limited to open-source models.

To securely invoke closed-source frontier large models while preserving privacy, "local sensitive information cleaning" is nearly the only compromise path. However, the engineering implementation difficulty of this approach is equally extremely high.

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

But Freysa announced in November 2025 that due to inconsistent mobile device performance, the team had to move sanitization computing off-mobile; its iOS app's July changelog explicitly listed the feature as deprecated and removed.

Similarly, MEGA's sanitization process must operate completely outside MEGA's own visibility—for instance, residing entirely on the user's local machine or within closed hardware—otherwise, users are essentially just handing over sensitive info meant for inference providers directly to MEGA instead.

MegaETH co-founder Yilong Li (@yilongl_megaeth) also pointed out when releasing the whitepaper recently: "Building more AI compute only gets half the job right. Even if providers have the right model, ample spare compute, and lower prices, users won't pay unless they can be certain providers actually run the advertised models and protect their data."

The second core highlight is Future Compute Capacity. The MEGA whitepaper imagines a mechanism where buyers can book specific models, fixed invocation volumes, and minimum performance standards (including response speed and stability) at locked-in prices within specific time windows, with future support for secondary market transfers of these forward compute contracts.

You can compare it to companies booking hotel rooms in bulk months in advance before a major industry summit. Buyers lock in ample compute before major product launches and can resell if plans change, while sellers can use these locked demands as collateral to borrow funds for purchasing more GPUs for expansion.

Once contract trading scales, the market will birth a transparent "forward inference pricing curve." While traditional financial giants like ICE plan to launch GPU compute futures that can hedge chip hourly rental costs, those cannot dictate which specific model runs or the output quality; in contrast, MEGA's forward contracts directly anchor to the model's final "output itself."

Among all the on-chain inference markets I have evaluated to date, none offers a similar mechanism, which is precisely where blockchain technology excels in establishing ownership and facilitating circulation.

Yet a deeper contradiction remains hidden here: AI inference costs are plummeting rapidly, and flagship models undergo complete generational shifts every few months. Why would buyers want to lock in fixed quotes or specific model versions today for delivery several months later?

Perhaps corporate buyers aren't just purchasing low prices, but a "guaranteed seat that never drops offline during peak hours," or using it to hedge against skyrocketing invocation bills after major AI vendors' massive subsidies dry up. On this point, I haven't reached a definitive conclusion either.

The Missing Key Puzzle Pieces Still Lack Visibility

At least for now, the above two highlights remain in the sketch phase of the whitepaper. The team has not released any code, nor does the whitepaper clarify which compute providers qualify for access, who will execute verification audits, or where the privacy sanitization tools will ultimately run.

The entire whitepaper also omits any mention of the native token. When asked about the role of the MEGA token in the inference market, Shuyao Kong responded in an interview with The Block: "We will leverage any scenarios that empower the MEGA token."

As for the $30 million treasury allocation, which exceeds half of MEGA's current circulating market cap, no detailed budget plan has been announced yet.

The upcoming December should gradually reveal some of these answers. Next, I will focus on three key implementation metrics: the initial roster of named providers and access criteria, a sanitization solution truly operating outside MEGA's visibility, and the first forward compute contract undertaken by a real counterparty.

Ultimately, verifiability only puts MEGA at the starting line of the inference race; only by solving true privacy and forward compute delivery can we determine whether it truly stands out.