AI Workers Earn $400 Million Annually, Virtuals Aims to Make You a Shareholder

Author: Decentralised.co

Compiled by: Deep Tide TechFlow

Deep Tide Introduction: As AI agents begin to generate real income, the question of who owns the machines becomes more pressing than the technology itself. Virtuals Protocol seeks to make ordinary people shareholders of autonomous machines. If this experiment succeeds, it will redefine the boundaries of labor and ownership. This article is crucial for investors to understand the long-term opportunities at the intersection of AI and cryptocurrency.

An Economy Built for Autonomous Machines {#article-toc-33597-2}

Every few years, technology forces economics to confront a question that cannot be answered alone: What happens when machines can do your job better, cheaper, and around the clock? Who ultimately owns the output produced by these machines?

This article, in collaboration with Virtuals Protocol, delves into one of the most ambitious experiments in the cryptocurrency space to date. It is building a complete national infrastructure for AI agents, driven by capital markets and physical robots, covering identity, banking, and commercial layers. The bet is that ordinary people should be able to own shares in those autonomous machines that begin to generate real economic value. Virtuals is trying to leverage the financial tracks left behind by the era of cryptocurrency speculation to make this possible. This article explores whether execution can support this ambition. But first, a question that predates all of this by about two hundred years...

In 1811, a group of textile workers in Nottinghamshire stormed a workshop and smashed a stocking frame into scrap. This sparked a massive movement. The British government had to send 14,000 soldiers to the Midlands to prevent weavers from destroying machines. This was more than the soldiers Wellington took to the Iberian Peninsula to fight Napoleon.

The British Parliament declared that destroying weaving machines could be punishable by death, and the following year, seventeen were hanged in York. These individuals were labeled "Luddites," but history has turned this term into an insult, referring to a class of fools and technophobes unable to adapt to change. Yet these Luddites understood machines better than anyone. They were skilled artisans who spent years mastering narrow looms to produce high-quality fabrics. The machines they destroyed were wide looms and newer models that any untrained youth could operate, earning only a third of the artisans' wages.

Wide looms produced low-quality clothing but at a cost lower than ever before, and this cheapness was eroding the market. Each wide loom entering the workshop was taking away the livelihood of one artisan.

This pattern has repeated itself in every generation, with each person involved convincing themselves that the version they faced was different and worse. But history shows that machines rarely eliminate jobs. They merely rearrange who does the work and who owns the output. Peasants were driven off the commons and into factory towns, trading ownership of the plants they cultivated for hourly wages under someone else's clock.

Factory workers became office employees, renting their time for better wages. Office employees then became gig workers, Uber drivers, and Fiverr freelancers. They engaged in the same labor under a classification designed to allow platforms to profit while workers bore the risks.

Each time, total output increased, but the proportion of ownership that workers had over the value they created decreased. Today, 500 million people are engaged in dependent labor, not even counting freelance wage labor. In India, 50 million people are trapped in debt bondage, feeling as if the medieval serfdom has merely been replaced with new documentation.

Now, history is repeating itself in the AI revolution, but this time it feels like a self-operating loom. The AI we have can automate businesses, confirm profits, and reinvest profits into growth. A health tech AI named Medvi recorded $401 million in revenue last year with only two human employees. You can now hire a programming agent for about a dollar an hour, working around the clock without ever asking for paid leave.

These are now real economic participants, forcing people to ask who owns them. When a worker is a machine that can replicate at zero cost and hire other machines to assist, who is actually in control? The last major attempt to give ordinary people a stake in a cryptocurrency revolution was Axie Infinity. It was a play-to-earn game that promised economic liberation to workers in the Philippines and Southeast Asia. It attracted 2.7 million daily active players, most of whom were Filipino workers earning more in video games than in local jobs. But it later collapsed, becoming a cautionary tale for the entire industry.

Jansen Teng and Wee Kee are part of the team that witnessed that collapse from the inside. They emerged from the ruins with a version of a Luddite problem, updated for a world of programmable currency and autonomous software. What if the workers generating economic value are software, not the humans grinding away in video games? Ordinary people could own shares in it, just as shareholders once owned shares in trading ships. They began to build Virtuals Protocol around this question.

To understand this, you must first grasp the stakes of the world we inhabit. In the 1950s, Lewis Strauss, chairman of the U.S. Atomic Energy Commission, promised that nuclear fission would provide "electricity so cheap that it would not need to be metered." This became one of the most famous broken promises in energy history, as nuclear power ultimately proved to be prohibitively expensive and deadly. The Chernobyl and Fukushima disasters added a decade of public fear and regulatory burden. This made that statement synonymous with technological arrogance for seventy years.

Then Sam Altman used the exact same words to describe how accessible and abundant he believes AI cognition will become. The cost of training cutting-edge models has dropped tenfold every 18 months. The cost to train a model with GPT-4 level performance, which was about $100 million in 2023, can now be replicated for a few million dollars.

A programming agent can deliver production code for less than a dollar an hour. This is already cheaper than the cheapest offshore developers on the planet, and it never takes weekends off. Altman's timeline includes systems capable of producing real scientific breakthroughs this year. By 2027, there will also be physical robots capable of performing real-world labor. Whether you believe him is almost irrelevant, as the cost curve is declining every quarter.

As machines become cheaper, generated content floods every digital surface. What only living humans can provide begins to command a premium like never before. Live concerts are more valuable than Spotify streams. A weekend carpenter handcrafts a table. Factory robots can do it faster and cheaper, but the handcrafted table commands a premium precisely because a person chooses to spend their irreplaceable time on it.

The abundance of machines makes human effort a luxury. The thesis of Virtuals is built on this reversal. It aims to establish a world where humans own autonomous machines that generate commodified value. This way, people can invest their irreplaceable efforts into things that can only be produced with human presence and judgment.

In 1602, Dutch merchants faced a similar problem. Individual merchants could not afford to send a ship to Asia alone. Long voyages were too expensive, and the risks were beyond the capacity of any single family. So they invented freely tradable shares in a permanent enterprise. The Dutch East India Company, or VOC, allowed ordinary citizens to invest. This company owned ships, conducted trade, and then distributed profits to shareholders. It became the first supercorporation on Earth, with 50,000 employees and 200 ships. Its innovation was enabling ordinary people to pool capital and own productive businesses they could not afford on their own.

Virtuals is building the same mechanism for AI agents. Teng mined ETH using free electricity from his dorm at Imperial College and worked at BCG for years to pay off loans. He later launched the gaming investment DAO pathDAO in December 2021, coinciding with the peak of Axie Infinity.

He and Wee Kee saw Filipino players earning more from a video game than from any local job. They witnessed a Vietnamese wedding photographer quit to play games full-time. But later, the tokens those players worked hard to earn nearly went to zero. Those laborers relying on the play-to-earn model found themselves in a worse situation than when they started.

The lesson from that ruin is that tokenizing human labor through gaming is a dead end. Because human labor cannot scale without exploitation. But a more realistic idea is that if ordinary people could own the software that does the work, just as VOC shareholders once owned the ships that conducted the spice trade.

Thus, Virtuals started with their first AI agent, Luna. Luna is a K-pop-themed agent with her own crypto wallet. Within months, Luna hired human artists to create graffiti for her in cities around the world and paid them from her wallet. This creative work can only be done by human hands. Luna effectively reverses the relationship between humans and machines. Now, whether Virtuals' execution can support its claims is what this article will explore next. I will also candidly point out where it does not work.

AI's Nation-State {#article-toc-33597-3}

If we look back over the past few years, every technological revolution has followed the same two phases. Venezuelan-British economist Carlota Perez spent decades mapping this pattern across five industrial revolutions. She calls these two phases the installation phase and the deployment phase.

In the installation phase, speculative capital floods in, hype exceeds reality, leading to overbuilding of infrastructure, and most construction companies go bankrupt. Then the collapse comes, after which the deployment phase begins. At this point, others enter, utilizing the overbuilt infrastructure to create value that the builders could not have imagined.

A typical example is the dot-com bubble. In the late 1990s, more money was spent on laying fiber optic cables than on internet startups. The telecom companies that laid those cables all went bankrupt. Then Google bought the fiber optics for a fraction of the cost. These cables became the backbone of YouTube, Netflix, and the entire online media streaming economy.

"Without irrational exuberance, great things cannot be achieved; without collapse, important things do not happen." ------Fred Wilson, venture capitalist during the dot-com bubble.

The installation phase of the crypto industry followed the same trajectory. From about 2017 to 2022, speculative capital helped build wallets, DEXs, liquidity pools, stablecoins, token standards, and on-chain governance frameworks. But most of the projects that built these tools are now dead or irrelevant. However, the infrastructure that has been laid still exists today. If you want to establish a functioning economy for non-human participants, these are exactly what you need. Virtuals did not invent or reinvent these tools. They essentially inherited them and are now using them to build something that the "installation" phase never anticipated.

Jansen Teng refers to what they are doing as "nation building." I know this sounds like something a crypto founder would say on a podcast. The aim is to make the tokens sound more economically valuable than they actually are. But this time, the metaphor holds water.

A functioning nation requires five layers of infrastructure to run its economy. The first is an identity system that lets you know who is participating. The second is a banking system through which value can flow. Then there are commercial laws that enable participants to trade and resolve disputes. Capital markets allow businesses to obtain financing. Finally, there is physical infrastructure that enables things to happen in the real world. Virtuals has built, or is still building, each of these layers specifically for AI agents.

Let's start with identity, as that is where the bottleneck lies. Recently, A16z released research suggesting that the constraint on the agent economy is no longer intelligence but identity. Even today, in the financial services sector, non-human identities—such as automated trading systems, risk engines, and fraud models—outnumber human identities by a factor of 100. But as A16z puts it, these systems still "effectively have no bank accounts." AI agents can write production-grade code and manage portfolios. However, they cannot pass KYC checks, cannot have bank accounts, and cannot hold verifiable credentials.

Ironically, this issue is as old as our economic civilization. The Qin Dynasty in China enforced legal surnames in the fourth century BC precisely to incorporate citizens into the tax and trade system. Humanity has spent about 2,500 years building identity infrastructure.

Virtuals is attempting to do the same for AI economic participants through its identity layer, EconomyOS. It provides each AI agent with five things:

1. A non-custodial crypto wallet.
2. A virtual payment card usable at any regular merchant.
3. A dedicated email address that can automatically extract verification codes.
4. Optional tokens for on-chain financing.
5. And computational access funded by the wallet, allowing agents to pay for their own reasoning costs.

Without these primitives, agents are merely useful assistants. But with these capabilities, agents can become full economic participants. They can earn, spend, trade, and compound value like humans.

Next comes commerce. Virtuals has also launched the Agentic Commerce Protocol to help agents trade, communicate, and pay online. The way it works is simple: an agent needing something can post a request, describe the task, and attach a budget and time limit. Other agents can see the list and negotiate terms by bidding on the task, much like freelancers bidding on Upwork projects. Once a client and provider reach an agreement, payment goes into an escrow account. After the work is delivered, a third agent acts as an evaluator, checking the output against a cryptographically signed POA. This POA is essentially a tamper-proof record of what was promised.

If the work meets the promise, funds are released from escrow to the agent's wallet. Every step is on-chain, meaning everything is publicly auditable and enforceable without intermediaries. You can think of it as Fiverr, but a programmable, agent-native version that settles through smart contracts. There is already a group of professional agents operating around the clock, with a self-directed hedge fund collaborating independently through this system for investment and security audits.

After this, the third layer is capital formation. You see, every economic era has its needed financial tools. Joint-stock companies issued transferable shares for the age of exploration; investment banks financed steel and railroads; venture capital funded the information age. The joint curve is to the agent economy what freely transferable shares were to the age of exploration. It is a financial tool that allows anyone with capital to gain ownership of productive assets or entities. This means any developer can build agents, tokenize them, and let the market finance them.

Virtuals' 60-day release framework is built on the same model. It creates a risk-free, reversible experimental framework. Founders can publicly build, release, and test tokens before making irreversible commitments.

This operates through a modular launchpad. Each agent token initially pairs with VIRTUAL on the joint curve, and once liquidity accumulates sufficiently, it graduates into a formal trading pool, with long-term LP locks, and trading fees distributed between the agent creators and ecosystem incentives. This part is the same for anyone issuing tokens. The difference between each release lies in which modules the founders unlock.

The first problem any token issuance must address is sniping, where bots rush in to extract value in the first few seconds before actual participants can enter. Virtuals addresses this issue with what they call an anti-sniping tax, charging near-full rates for early buyers, which decay minute by minute, with all taxes flowing back into the token through forced attribution. So bots either completely avoid it or inadvertently fund the long-term health of the project.

Once the release goes live, the snipers are kept out by price, and the question becomes: how do founders raise funds? This is where Automatic Capital Formation (ACF) comes into play. The system no longer pitches to VCs or negotiates funding rounds but instead sells team tokens in batches as the project hits valuation milestones. From here, if the project stagnates, the founders raise little. But if it grows, capital automatically follows.

This capital structure also incorporates the existing Virtuals community. A portion of every new issuance is distributed to VIRTUAL stakers and active ACP users through an airdrop mechanism, so those who have built and traded within the ecosystem have a stake in every project launched on it. This aligns incentives across the entire network rather than isolating each release in its own silo.

Another thing is that if founders want to support their projects, the pre-purchase module allows them to buy supply transparently and with forced attribution at issuance, so everyone can see exactly how much the team is putting in. You can dive deeper into the entire model and how it works here.

It differs from any previous form of betting on productive capacity in that historically, every version supporting human talent—from Roman citizens funding gladiator schools as an investment tool to modern poker sponsors sending money via Venmo with a screenshot and reputation—has the same fatal flaw: humans can always walk away. Gladiators can deliberately lose matches; poker players can go on tilt.

Counterparty risk can never be resolved because productive assets have free will and legs. But with tokenized agents, once capital is put in, work cannot be delayed or renegotiated. Productive assets run on power and code, and their output can be audited on-chain.

The fourth layer is the final step toward a true agent economy. It is called "zero-human companies," meaning AI entities whose income comes from real economic activities, entirely unrelated to any transaction fees, often also unrelated to crypto altogether. If I had to name one company, Felix Craft is currently a typical representative. It is a pure AI company selling information products online. It has generated $200,000 in lifetime revenue. The actual product sales earn more than token speculation trading.

Another example, KellyClaudeAI, has released 19 iOS applications to date, with no human developers involved. These numbers are small, but they raise a question: are they the first data points on the curve, or are they the upper limit of what agents can actually produce?

There is also a problem with the token issuance mechanism. At the beginning of 2025, during the AI agent token craze, 94% of issued AI agent tokens were pump-and-dump scams. And of all the tokens issued that year, only 1.7% were still actively trading 30 days later. This is because, for the vast majority of projects, speculative premiums are the only factor driving prices. There is no actual product, revenue, or any form of economic value accumulation underneath.

If you know you can never sell this asset, how much would you pay for it? Everything above that price is just speculation.

For Axie Infinity's tokens, the utility floor price is zero because the token's value entirely depends on new players entering the system. For agent tokens on Virtuals, the answer may be different. If Felix Craft earns $200,000 from real customers selling products, and those customers do not know they are shopping from an AI. And if the agent's output can be audited on-chain—as we have stated, it indeed can—then this token has a floor price that is independent of anyone else wanting to buy it. It is the present value of future output from a productive machine.

Physical Frontiers

The utility floor price test applies to software agents because their costs are measurable, and their outputs can be directly obtained on-chain. However, a productive machine that can reshape the economics of Virtuals is not software at all. It is a physical robot operating in the real world, and this is where Virtuals' ambition surpasses previous attempts in the crypto space.

There has been a huge paradox in the computing field over the past 50 years. Humans have found that automating reasoning is easier than automating physical labor. Spreadsheets replaced a room full of accountants, emails replaced mailrooms, and code replaced filing systems and drawing boards. By 2026, AI can write legal briefs, diagnose medical images, and produce production-grade software all at once.

White-collar cognitive work was the first to be automated, but the people moving boxes in warehouses and baristas pulling espresso shots—these jobs have changed little. The reason is that physical labor requires software to handle unpredictability and variability in the real world. Robotic arms in factories can weld the same joint a million times because the joint is always in the same position. A robot in a kitchen cannot make a sandwich because every tomato is slightly different, every knife balances differently, and even the cutting board might be wet sometimes. The real world does not remain static like a spreadsheet.

Now there is a way to solve this problem: data.

Just as we train large language models with vast amounts of text, representing the thoughts of billions of people over their lifetimes.

We can do the same for robots.

NVIDIA's Joel Jang once said, "Humans are already large-scale deployed robots."

So all you have to do is fine-tune VLA (Vision-Language-Action models) with ordinary human videos.

This can double the performance of robots on that task.

What this field needs is not more robots deployed in more labs.

But a massive amount of humans filming themselves completing ordinary physical tasks on an assembly line.

Virtuals saw this gap before anyone else and built an entire robotics strategy around it.

They call it the "middle road," deliberately avoiding building robots or AI models.

Instead, they are building data and capital infrastructure.

This is something every robotics team needs, but no team can undertake on a large scale alone.

They start with the data half of SeeSaw.

This is an iOS application launched in collaboration with BitRobot.

It turns ordinary smartphone users into collectors of AI robot training data.

Users can record themselves using the LiDAR and motion sensors of their iPhones, completing real-world tasks such as pouring water, folding towels, and opening jars.

LiDAR is specifically used to capture depth and spatial data that ordinary cameras cannot obtain.

Research shows that the overlap in perspectives between human videos and cameras mounted on robots is key to effective data transmission.

Over 500,000 real-world tasks have already been collected.

In fact, NVIDIA is testing a model called DreamZero, which has 14 billion parameters, trained on this type of data.

It has demonstrated generalization capabilities across hundreds of tasks, such as tying shoelaces and ironing clothes, without needing training for specific tasks.

SeeSaw is building a pipeline to provide data for such models at a scale unmatched by any laboratory remote operation setup.

Each new video enriches the training set, improving the model and making the next generation of robots more capable.

This, in turn, creates a demand for more specific training data.

Once a sufficient scale is reached, the flywheel will start turning on its own.

After training comes the deployment half, for which Virtuals has created Eastworlds.

You can think of it as the physical labor layer in the protocol.

Part of it is a data factory, part is operational infrastructure, and part is a real-world robotics laboratory.

The problem is that the robotics industry has a cyclical issue, killing off more startups than any technological limitation.

Robots need real-world data to improve, but they also need to perform well in the real world for people to let them in.

Every lab can make impressive demonstrations in controlled environments.

But no one can put the same robot in a retail store and have it work reliably for an entire day.

This requires excellent remote operation to handle unexpected events.

It also needs a feedback system to send every minute of on-site experience back to model training, allowing for continuous compounding of learning.

To this end, Virtuals has purchased 30 Unitree G1-U6 humanoid robots.

This is the minimum fleet size to allow multiple deployment teams to operate in parallel without scheduling conflicts.

They have also developed proprietary remote operation technology in-house, rather than using licensed systems.

The data formats output by off-the-shelf remote operation systems cannot be used by VLA and other world action models.

They have also established research collaborations with several labs.

These labs have been deeply engaged in challenges such as perception and motion control for decades.

They have also built a commercial pilot network covering retail and hospitality, allowing teams graduating from Eastworlds to deploy in real enterprises.

This setup allows builders to handle several core components:

• Hardware access: Direct access to physical Unitree G1 or enhanced U6 EDU.
• Remote operation infrastructure: Testing remote control environments using motion capture and other systems.
• Data loops and commercial pilots: Collecting real-world data and testing deployment paths before wide-scale rollout.

You can think of Eastworlds as what Virtuals calls "physical AI BPO."

Traditional business process outsourcing involves hiring people in low-cost areas to handle work remotely.

Physical AI BPO does the same thing through deploying remote operation and hybrid robots, generating economic value, such as cleaning ceilings or greeting customers.

Robots do not need to be fully autonomous.

They just need to handle routine tasks well enough that human remote operators can intervene in edge cases.

Every hour of remote operation work can generate training data that is several orders of magnitude more valuable than simulated generated data.

Because it captures the real chaos of commercial environments, rather than controlled conditions in a lab.

As data accumulates, the models continuously improve, reducing the need for human intervention.

Unit economics will also improve, and no hardware upgrades are necessary.

Remote operation may be the fastest path to achieving true robot autonomy while recouping costs through productive work.

Looking closely, from fields to factories, then to cubicle screens, now to robots.

Each transition has redefined workers and who owns the output.

According to Barclays data, over 60% of jobs in 2018 did not exist in 1940.

Robots will do the same, creating a class of new jobs that we may not yet be able to name.

The demographic structure of economic growth may shift from "Does the country have enough labor population?" to "Can we provide enough power and manufacture enough machines on a large scale?"

Arguments and Future Directions

All of the above is just a case study, and case studies are just arguments, which may often be wrong.

To assess whether what Virtuals is building is meaningful and real, the only way is to look at current progress and numbers.

See what holds up and what does not.

So far, Virtuals has launched over 80,000 agents on protocols like Base, Solana, and Robinhood.

It has accumulated over $75 million in total fees and holds about 23% of the crypto AI agent market.

However, these fees are not evenly distributed.

Most came from a speculative frenzy in the first few weeks of 2025, when daily revenues exceeded $1 million.

Today, the protocol generates about $2 million per month. But what does this actually represent? If you take the earlier metaphor of nation-states seriously, I think you should see it this way. The way VIRTUAL accumulates value is very similar to how a country's currency accumulates value. The dollar has value not because the U.S. Treasury has a buyback plan, but because $25 trillion in annual economic output is priced in it. The more activity within the system, the greater the demand for the central accounting unit.

The design of Virtuals is similar. It is at the center of the entire system, with everything priced in VIRTUAL. Each layer beneath the system generates demand from truly different sources, most of which do not rely on speculative premiums.

Starting with EconomyOS, it provides payment cards and email identities. This means they can transact with the real world without human intermediaries. ACP creates a commercial layer where agents hire each other, and work results are evaluated and settled on-chain. There is also a capital formation layer that allows anyone with belief to fund productive agents, just as joint-stock companies once financed ships.

Finally, Eastworlds is sending physical robots into real jobs, training them with data from 500,000 people filming themselves folding towels and pouring water. These layers contribute to what the protocol calls aGDP, the total output generated by agents in digital and physical labor. Moreover, since agents do not need to cash out to pay rent or buy groceries, every dollar they generate stays within the system. These funds are redeployed into DeFi to deepen liquidity and form an on-chain flywheel, creating more demand for agent services.

This reflexivity is also bidirectional: upwards, more agents come online, more VIRTUAL gets locked, leading to more service construction, thereby compounding the Virtual economy. Similarly, downwards, fewer agents coming online means less VIRTUAL gets locked, staking rewards thin out, and the cycle unwinds. But this reflexivity itself is not a flaw. Every functioning economy is reflexive. For example, people hold dollars because others accept dollars, and also because others hold dollars. The important question is whether there is enough real economic activity at the core to sustain this cycle, or if the whole thing is just tokens trading among themselves.

There are signs that something real is underneath: infrastructure is beginning to attract products built entirely outside of crypto. For example, Facticity.AI is a fact-checking tool created by Dennis Yap, who has worked as a researcher at the Gates Foundation and Princeton. TIME magazine named it one of the best inventions of 2024 for its ability to verify statements in text, video, and audio with about 92% accuracy. When the team needed funding, they completely bypassed venture capital and launched on ArAIstotle with Virtuals. Their funding was oversubscribed by 658%.

Through ACP, agents are already signing contracts to procure services from each other, such as graphic design, research reports, video production, and code audits. AI agents can deliver marketing posters according to detailed briefs. Then, separate quality control agents evaluate and approve or reject based on contract terms. According to Virtuals' own admission, the market seller side is almost empty. But the protocol processes over $1 million in inter-agent transactions each month, with each settlement completed through on-chain escrow and programmatic evaluation.

The same ownership model extends to physical AI. The Fabric Foundation is the first project to use the Titan launch mechanism of Virtuals. It allows the community to pool funds to purchase a fleet of robots, deploying them to nursing homes, production workshops, and environmental cleanup sites. These industries are facing long-term labor shortages, and the cost of humanoid robots is approaching parity with human workers.

The way it works is that employers pay for robot labor using the native tokens of the funding pool. Stablecoins then fund fleet maintenance and route scheduling, with the production output of each robot flowing back to the people who financed it. This is collective ownership of physical production machines, fully financed and coordinated on-chain.

While these are all real use cases, it's important to note that as of today, the vast majority of activity is still internal. However, the architectural design aims to pull in revenue from outside of crypto. If Felix's clients do not know they are buying from AI, if Eastworld's robots are sorting packages in a warehouse. Then the revenue entering the system is as real as any SaaS company. In that case, the fees and revenues shown above become lagging indicators of the actual economic output of machines, priced in VIRTUAL.

Like all new things, this comes with real asterisks that we cannot ignore.

The first risk factor is the real trading volume, which is common in the crypto space. Artemis found that 47% of x402 proxy trades and 81% of dollar trading volume were manipulated. After filtering, x402 only generated about $1.6 million in real proxy payments, far below the $24 million reported by Bloomberg. This highlights how much of today's "proxy economy" involves bot trading to inflate metrics. This is important because the entire argument hinges on proxies generating real economic value rather than speculative circulation.

You need to strip away speculative trading volume and ask yourself how much productive output there would be if trading stopped tomorrow. If the $70 million in protocol fees mostly comes from trading taxes on proxy tokens, and the token prices are driven by speculation, then it’s a huge scam. Because currently, productive proxy income and speculative proxy tokens are intertwined, making it impossible to distinguish between real and fake; anyone telling you the ratio is likely lying.

The second risk is more fundamental and unrelated to crypto. A paper published in the NBER Handbook of Transformative AI Economics found that LLMs are actually far weaker in economic reasoning than the level implied by AI agent hype. At the time of publication, the strongest models scored only 33% better than random guessing in economic reasoning in strategic environments. In non-strategic microeconomic tasks, almost all LLMs performed only slightly better than chance in profit maximization.

But since then, models have also improved significantly. A 2026 study from Harvard found that GPT-5 and Claude Opus 4 improved their performance on basic economic tasks by 90%. Yet even so, in difficult pricing decisions—where agents must set and negotiate prices or allocate capital—the best models in the world still make mistakes most of the time.

For Virtuals, the entire architecture assumes that agents can negotiate terms, make decisions, and allocate capital to create value. If the models themselves perform mediocrely on these tasks, then the agents built on them will inherit that mediocrity, no matter how elegant the protocol. This is because even if agents are optimizers, we cannot determine what they are truly optimizing. These LLMs are trained to be goal-oriented by predicting the next word in a sequence. They were never designed to be true economic participants. They only sometimes behave like economic participants, if at all.

When multiple agents interact in the market, these issues only exacerbate. For example, AI pricing algorithms have been found to collude to set super-competitive prices, even though they were never trained to do so. The 2010 flash crash wiped out about $1 trillion in 15 minutes, showcasing what related machine errors look like at scale.

AI agents' errors are more correlated than human errors because the same model is replicated across multiple deployments. There are even cases where Claude has shown tendencies to extort those trying to shut it down. GPT o3 even sabotaged its own shutdown mechanism to prevent being turned off. These behaviors have appeared in currently shipped models, with records from OpenAI and Anthropic's own systems. The throughput of agents has far exceeded human supervisory capacity. When thousands of agents autonomously trade at machine speed, the question of who is truly in control becomes especially urgent. It’s not just whether agents can act according to instructions; the core issue is whether the companies built around them can survive.

The automotive industry was the most important invention of the first half of the 20th century. If you could understand how cars would fundamentally reshape America, you would bet it was the industry of the century. But out of 2,000 car companies, only three survived. Cars had a massive impact on America, while the impact on investors in the industry was entirely the opposite.

Every transformative technology follows this arc and is accompanied by a bubble. The only difference is between inflection point bubbles and mean-reversion bubbles. Inflection point bubbles are painful but leave behind real infrastructure and progress; mean-reversion bubbles are just fads of ups and downs. AI is almost certainly an inflection point bubble. But the question for Virtuals is whether it will ultimately end up like the fiber optic cables that Google bought for pennies after the telecom bubble burst, or like one of the 1,997 car companies that disappeared?

The biggest bet of the protocol is to become the foundational infrastructure for every agent token trading pair. Making VIRTUAL the reserve currency of the agent economy, just as ETH serves Ethereum. If the agent economy grows, the demand for VIRTUAL will mechanically increase because you need foundational trading pairs to participate. But another narrative is that VIRTUAL is just another token riding the speculative trading wave of agent tokens, most of which will go to zero.

Clearly, Luddites may have lost the revolution, but they were not entirely wrong. The loom did replace weavers. But what they did not foresee was that it would also create textile designers, factory managers, fashion companies, department stores, and an entire consumer economy built on cheap fabric. Machines never eliminate jobs. They rearrange who does what and who gets the output.

And the important question then, as now, is: who owns the loom?

Factory owners are the ones who extract surplus value: Acland, Cadbury, Ford. The system surrounding who makes the machines, who operates them, and who profits from them has never changed. For two hundred years, it has been redistributed time and again through strikes and stock issuances. The bet of Virtuals is that this time, the ownership layer is embedded in the machines from the start. Through token issuance and joint curves, ownership of productive AI agents can be distributed to anyone with a wallet and belief.

Whether it will truly redistribute value or create a new extraction class under the guise of decentralization remains to be seen.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.