BitMind Investment Memo

Sep 16

Written By Sami Kassab

Subnet 34 on Bittensor

September 2025

Snapshot

Description: BitMind’s mission is to develop state-of-the-art consumer and enterprise products for AI-content detection. Their deepfake detection models consistently outperform competitors across a variety of industry-standard benchmarks.
Subnet Output: Miners are split into two sets. One set outputs AI-generated media, creating a high-quality deepfake dataset. The other set outputs open-source multiclass classifier (deepfake detection) models to the BitMind platform.
Team: The BitMind team has a deep bench of engineering talent, along with a strong leadership team with prior experience in decentralized AI, computer vision, and financial services:

Ken Miyachi: Co-founder and CEO of BitMind. Prior to starting BitMind, Ken was Senior Tech Lead at the NEAR Foundation. He also previously founded LedgerSafe and worked at Amazon on recommendation systems.
Dylan Uys: Co-founder and Head of AI at BitMind. Prior experience includes developing ML solutions across computer vision, NLP, fraud detection, recommender systems, and search at companies like ViaSat and Poshmark.
Canh Trinh : Head of Engineering at BitMind. He previously led interoperability integrations at Axelar and held engineering and leadership roles at JP Morgan Chase and Deutsche Bank.

Market: Financial services and media platforms face the greatest risk from deepfake-driven fraud. AI-based fraud losses are projected to hit $40B by 2027, yet most firms remain underprepared and underinvested. BitMind is focused on this gap, aiming to be the technology partner for companies lacking the ability to develop in-house solutions.

Investment Rationale

BitMind is a prototypical example of how a research subnet focused on frontier AI challenges can commercialize their findings into enterprise-grade products.

On the research front, BitMind has already outperformed well-capitalized incumbents like Reality Defender in head-to-head evaluations. In a brief benchmark of 115 AI-generated images and 90 real images, BitMind consistently delivered stronger results. That track record speaks to both their technical execution and their ability to master subnet design as a discipline.

On the commercialization side, the team has launched consumer applications that now reach 50K monthly active users, alongside enterprise offerings where API usage exceeds 2M+ requests per week.

Operationally, Ken and his team have shown the ability to adapt under pressure, giving us conviction in their capacity to navigate any challenges that come next. They weren’t afraid to rip out and rebuild their incentive mechanism when it became clear the old model wasn’t sustainable. On the miner side, they’ve weathered bad actors and exploits without losing stability or momentum.

Deepfake detection is on track to become a $40B+ market, while attacks are rising rapidly, and no clear leader (whether Big Tech incumbent or new startup) has emerged. That leaves the field wide open. Deepfake detection is a perpetual cat-and-mouse game, meaning a subnet is the optimal structure to stay adaptive. That’s why we’ve backed BitMind.

We’ve developed a valuation framework for subnet tokens that accounts for their yield and Bitcoin-like emission schedule. Since subnet token inflation begins high and tapers off over time, we anchor valuations on key circulating supply milestones. For our analysis, we anchor the valuation to the 2-year post-TGE mark, when ~53% of the supply is circulating, and apply that discount to the FDV. We then factor in staking yield projections, which reduces our cost basis by 34%. Together, these adjustments give an Adjusted FDV of $37.8M, a 72% discount to the current FDV.

At that level, given the massive market opportunity in deepfake detection, the setup offers the kind of asymmetric, venture-scale upside we look for.

Overview

As AI becomes more sophisticated and accessible, so do the attackers. Deepfake incidents have surged over the past year, and even in crypto we’ve seen highly capable people fall victim. One growing vector is live “Zoom-style” attacks where impostors join calls disguised as trusted counterparts, then drop a malicious link that, once clicked, drains their wallets.

These tactics are moving quickly upmarket. With most financial transactions now initiated online and verification often reduced to voice or video, the same playbook can be used to trigger wire transfers, harvest credentials, or access sensitive deal data. What starts in crypto is a preview of systemic risk across the broader financial sector.

In response, players like Reality Defender, Clarity, and Hive AI have emerged, building proprietary detection engines for enterprise clients. The risk they face—and by extension, their customers—is failing to allocate enough to R&D. Deepfake detection is a cat-and-mouse game; generative and discriminative techniques evolve in lockstep. The optimal setup balances commercializing a product with staying at the frontier. Because without that, detection will always lag behind the latest attack methods.

BitMind, subnet 34 on Bittensor, operates with a different model than these traditional companies. Instead of relying on a small internal team, it has built a globally competitive, incentive-driven R&D ecosystem for detecting AI-generated content. Researchers around the world compete to break and defend its systems in real time, creating an adversarial environment where detection methods are constantly stress-tested against the latest generative advances.

Subnet Architecture

Over the past month, the team has rolled out a new subnet architecture termed GAS (Generative Adversarial Subnet). GAS mirrors the dynamics of generative adversarial networks (GANs) but has a custom implementation within the BitMind subnet.

The miner set is split between two categories now: Generative miners create AI content with the goal of fooling the Discriminative miners. Discriminative miners train detection models and open-source them so that validators can verify the performance of the models. Rewards are distributed based solely on model accuracy. The full Generative miner design is still in development, but they’ll earn rewards based on how successful they are at beating Discriminative miners.

The result of the system is a virtuous cycle of progress. Generative miners are rewarded for producing higher quality datasets, creating better test data for Discriminative miners to develop models against. These new models will begin to outperform against existing datasets produced by Generative miners, which lowers their reward capture, pushing them to create higher quality data.

Market

The market for deepfake solutions is still early, and funding levels reflect that. A handful of companies have raised modest but notable rounds that point to growing demand:

Pindrop: Raised a $100M debt financing round in 2024. The company focuses primarily on audio deepfake detection;
Hive AI: Raised a $50M series D in 2021, valuing the company at $2B;
Reality Defender: Raised a $33M series A last year, with reports estimating that the firm had revenues just north of $8M in 2024;
GetReal Labs: Raised a $17.5M series A earlier this year;
Clarity: Raised a $16M seed round last year, with the firm primarily focused on deepfake video detection.

Meanwhile, the underlying technology driving deepfakes is advancing quickly and becoming cheaper to use. As we’ve seen with LLM inference, costs are trending down even as quality improves. Google’s state-of-the-art image models, for example, now generate outputs for as little as $0.039 per image with striking realism. With the attack surface expanding rapidly and AI being weaponized at scale, we expect capital flows into detection to accelerate meaningfully.

The most compelling path to scale for BitMind is through enterprise customers within the finance, media, and cybersecurity sectors, along with companies that sell privacy and identify verification services into a broader range of companies.

Just in the Financial Services sector, fraud losses stemming from AI-generated attacks are expected to reach $40 billion in the US by 2027. Beyond the immediate financial losses from fraud, banks and financial institutions also risk losing customer trust and facing widespread attrition after even a single, isolated incident. Given the asymmetric upside in achieving a single successful fraud attempt, institutions (e.g., banks, insurance companies, etc) will be forced to dramatically ramp up security budgets to protect their customers.

Risks around attrition and consumer flight extend out to media companies and platforms as well. The largest of these (e.g., YouTube, TikTok, Meta) are either developing their own proprietary solutions for combating deepfakes or working with large established firms, but other firms such as small-medium sized news publications and more niche social platforms will need to lean on companies and partners that specialize in developing these solutions.

Regulatory pressure for media and social platforms is already underway as well; the Take It Down act was put into law earlier this year, which mandates platforms remove intimate or sensitive deepfakes within 48 hours of notification, and the broader No Fakes Act is still making its way through the Senate and the House.

We see BitMind’s strongest in-roads coming in financial services, where fraud prevention and transaction security are existential priorities. Beyond finance, its detection capabilities also extend naturally into e-commerce, operational security, and communication software.

Product

The BitMind team currently has multiple products split across both developers and individual consumers.

Most companies simply don’t have the resources to build out their own deepfake detection tools or host open source solutions. BitMind’s API enables developers to integrate deepfake detection functionality within their own applications, with the API capable of processing image and video files. Results are returned with minimal latency (typically under 1 second) and include the prediction, confidence scores, and a similarity metric. The API also includes a pre-processing layer to better handle lower-quality images/videos and reduce false positives/negatives. Requests served by the API have nearly doubled since July, now reaching nearly 2.5M+ requests/week.

Consumers are still effectively fighting their own battles on social media platforms when it comes to determining what’s real. BitMind has shipped two products to help social media users out on the front lines: a browser extension and a new mobile app. While scrolling on Twitter, for example, the browser extension simply lets a user hover over an image and it will give a detection score indicating the authenticity of an image.

With the mobile app, users can share images or videos from social media apps with the BitMind app, where it will detect whether the file is a deepfake. The mobile app was launched less than a month ago, and has already reached over 3,500 downloads.

Across all their product lines, BitMind now has 50K+ monthly active users, and are pacing to see continued growth given the upward trends in their API data along with the initial positive reception of their mobile app.

The BitMind Token

Within crypto, buyback models are often introduced to create a vague sense of tokenholder alignment. Lately, sentiment has turned against them as the prevailing view is that early-stage projects should reinvest in growth rather than distribute capital—directly or indirectly—to investors.

While we generally agree with that view, subnets operate differently than most crypto networks. Because of their proof-of-work component, subnet tokens emit heavily in the first two years of their life. That makes daily sell pressure far greater than a typical crypto project, where supply unlocks are limited to team and investor schedules. The result is a structural urgency for subnet owners to find product-market fit quickly and generate revenue that can be directed back into the subnet to offset miner emissions.

BitMind faces this same dynamic, but its architecture is distinct. Miners output open-sourced models, meaning the subnet token is not required to access the service. That reality pushes BitMind to adopt more of a commercial open source software (COSS) playbook, similar to Red Hat, which built a business by packaging, supporting, and scaling open-source software for enterprise customers.

BitMind is following a similar path. The subnet produces cutting-edge detection models, while the team layers on the enterprise-facing capabilities that make those models usable at scale. That includes building turnkey products, delivering custom integrations, and offering the kinds of onboarding, advisory, and managed services that large customers need.

Today, BitMind is focused on user growth across its product lines. Over time, the team intends to implement monetization and direct a majority of revenues into a token buyback program. Token buybacks for BitMind function as reinvestment. Miners, while separate entities within the network, are functionally part of the team. They test and develop the models that ultimately power BitMind’s products; miners serve as the R&D department. Token emissions to miners are effectively BitMind’s R&D budget. To expand that budget, the value of BitMind’s token must increase, and buybacks provide a direct way to achieve this.

In other words, BitMind’s buybacks are a structural mechanism to push value back into the token, strengthen miner incentives, and improve the quality of the detection models that the products ultimately rely on.

Key Reasons for Investing

SOTA detection models: BitMind is outcompeting some of its largest and most well-funded competitors (e.g., Reality Defender). On their internal benchmarking, BitMind outperforms Reality Defender on measurements around accuracy and precision, with latency that’s an order of magnitude lower. BitMind’s recent move to restructure their subnet and introduce a winner-take-all competition (a daily prize value of ~$20,000) adds another catalyst to the miner competition, ensuring they keep pushing the pareto frontier of what’s possible with deepfake detection.
Clear product usage growth: BitMind has both an enterprise service (API) and consumer products (browser extension and mobile app). Across the board, the company has already reached 50K monthly active users. Their API usage is seeing steady growth, and is now serving over 2.5M+ requests/week. Consumer product development is notoriously difficult, yet BitMind’s browser extension already has over 19K downloads, and the recent mobile app launch beat internal expectations and has over 3,500 downloads.
Emerging, massive market: The financial services sector alone is estimated to reach $40B in fraud losses within the next 2 years. Yet, nearly half of businesses within the financial (and adjacent) industries view themselves as unable to reliably detect/defend against deepfake attacks. Increased frequency of attacks, regulatory/compliance pressure, and broad public scrutiny after reported incidents will push these firms out of their historical pattern of severely underspending on fraud-prevention solutions. Unlike the big tech platforms (e.g., YouTube, Meta, etc), these firms are generally unequipped to develop their own in-house solutions, making them the ideal customers for BitMind.

Key Risks

Financial sector conversion challenges: For BitMind to reach massive scale and land multi-million dollar contracts they’ll need to successfully break into the financial sector. Selling products into banks or financial institutions typically requires mature, established relationships and tenured industry professionals to help close deals, which BitMind is currently lacking. However, these institutions will face continued compliance and public pressure to create safeguards against fraud attacks. As they look to move more quickly, product performance will begin to outweigh traditional relationship-based sales processes. BitMind has internal benchmarking demonstrating their outperformance relative to their competitors, which is what these customers will ultimately be attracted toward.
Established media companies building in-house: Social media companies like YouTube and TikTok will continue to be some of the most impacted players by deepfakes and AI-created content in general. Their scale, engineering talent, and massive relevant datasets mean they will likely develop their own proprietary products deepfake detection along with adjacent creator tools. While these companies will likely be out of reach for BitMind, smaller media platforms or news publications lack the requirements to develop their own tech stack, and will choose a buy over build approach. Unlike the financial sector, these companies will likely rely less on established/trusted tech providers and will seek out the best technical solutions, which makes BitMind the primary contender for their business.
Execution risk around rollout of new subnet design: BitMind is beginning to roll out their new GAS subnet design. The architecture is relatively novel for Bittensor, as it now splits the miner set into two categories. Given there are both large technical and economic changes associated with this new design, there’s a high probability of bugs and exploits that will negatively impact the BitMind subnet, and by extension, slow down their product development. However, Ken and team are veterans within Bittensor and now have extensive experience designing incentive mechanisms. We’ve seen firsthand how they battle through issues with miners and incentive mechanisms. We have full confidence in their ability to maintain their resilience and resolve issues as they arise.

This content is provided for informational purposes only and does not constitute investment advice or a recommendation to buy or sell any security. Unsupervised Capital holds a position in TAO and may hold positions in the subnet tokens or other digital assets discussed herein and may buy, sell, or change positions at any time. Past performance is not indicative of future results. Digital assets involve substantial risk, including potential total loss of capital. Consult your own advisers regarding any investment decisions.

Investment in digital assets and blockchain technology involves substantial risk and may result in partial or total loss of investment. Digital assets are subject to extreme price volatility, regulatory uncertainty, and market manipulation. Past performance is not indicative of future results. This memorandum is for informational purposes only and does not constitute investment advice, an offer to buy or sell securities, or a solicitation of any offer to buy or sell securities. Recipients should consult their own advisors before making investment decisions. This memorandum contains projections and estimates based on current expectations. Actual results may differ materially from those projected. All forward-looking statements are subject to risks and uncertainties. Unsupervised Capital Management, LLC makes no representations regarding the accuracy or completeness of information herein. Third-party information has not been independently verified. To the maximum extent permitted by law, Unsupervised Capital Management, LLC shall not be liable for any damages arising from this memorandum. By reviewing this memorandum, you acknowledge these disclosures. This information is subject to change without notice.

Sami Kassab