TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
AI Workers Earn $400 Million Annually, Virtuals Wants to Make You a Shareholder

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

2026.08.28
Share
Virtuals Protocol attempts to make ordinary people shareholders of autonomous machines; should this experiment succeed, it will redefine the boundaries of labor and ownership.

Author: Decentralised.co

Translated by: TechFlow

TechFlow Summary: As AI agents begin generating real revenue, the question of who owns machines becomes more urgent than the technology itself. Virtuals Protocol attempts to make ordinary people shareholders of autonomous machines. If this experiment succeeds, it will redefine the boundaries of labor and ownership. This piece is crucial for investors to understand the long-term opportunities at the intersection of AI and crypto.

An economy built for autonomous machines

Every few years, a new technology forces economics to grapple with an unanswerable question: If machines can do your job better, cheaper, and around the clock, what happens to you, and who ultimately owns the output produced by these machines?

This article was written in partnership with Virtuals Protocol, diving into one of the most ambitious experiments the crypto space has ever undertaken. It is building out full-stack national infrastructure for AI agents, from identity and banking to commercial layers, powered by capital markets and physical robotics. Its wager is that ordinary people should be able to own shares in autonomous machines that start generating real economic value. Virtuals is trying to leverage the financial rails left behind by the era of crypto speculation to make this possible. This piece examines whether execution can support the ambition. But first, a question that predates all of this by roughly two hundred years…

In 1811, a group of textile workers in Nottinghamshire broke into a workshop and smashed a knitting frame into scrap iron. This sparked a mass movement. The British government had to deploy 14,000 troops to the Midlands to prevent weavers from destroying machinery. That was more soldiers than Wellington took to the Iberian Peninsula to fight Napoleon.

The British Parliament declared that breaking weaving looms was punishable by death, and seventeen people were hanged in York the following year. These individuals became known as the "Luddites," but history turned the word into an insult, denoting fools who cannot adapt to change or technological shifts. Yet these Luddites understood machines better than anyone. They were highly skilled craftsmen who spent years mastering narrow-frame looms to produce premium fabric. The machines they destroyed were wide-frame looms and newer models that could be operated by any untrained teenager for one-third of a craftsman's wage.

Wide-frame looms produced low-quality garments, but at a cost lower than ever before. This cheapness was eroding the market. Every wide-frame loom installed in a workshop was taking away a craftsman's livelihood.

This pattern has repeated every generation since, with those living through each instance convincing themselves that their version is different and worse. But history shows that machines rarely eliminate work. They merely reshuffle who does it and who owns the output. Tenant farmers were pushed off common lands into factory towns, trading ownership of their crops for hourly wages under someone else's clock.

Factory workers became office clerks, renting out their time for better wages. Office clerks became gig workers—Uber drivers, Fiverr freelancers. They perform the same labor under a classification designed to let platforms capture the upside while workers bear the risk.

Each time total output rises, the proportion of value created that belongs to the people doing the work falls. Today, 500 million people are engaged in dependent labor, which doesn't even count as free wage labor. In India, 50 million people are trapped in debt bondage, feeling like medieval serfdom repackaged with new paperwork.

Now, history is repeating itself in the AI revolution, but this time it feels like a loom running on its own. The AI we have enables businesses to run autonomously, confirm profits, and reinvest those profits into growth. A health-tech AI named Medvi generated $401 million in revenue last year with only two human employees. You can hire a coding agent right now for about a dollar an hour, working around the clock, never asking for paid leave.

These are now genuine economic participants, forcing the question: Who owns them? When the worker is a machine with zero replication costs that can also hire other machines to help, who is actually in control? The last major attempt to give ordinary people a slice of a crypto-driven tech 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, mostly Filipino workers, who earned more from video games than local jobs. But it later collapsed, becoming a cautionary tale for the entire industry.

Jansen Teng and Wee Kee witnessed that collapse from the inside. They emerged from the wreckage carrying a Luddite-style question, updated for a world of programmable money and autonomous software. What if the worker generating economic value is software, not humans grinding away in a video game? What if ordinary people can own shares in it, much like shareholders once owned stakes in merchant vessels? They began building Virtuals Protocol around this question.

To understand this, you must first grasp the stakes of the world we live in. In the 1950s, Lewis Strauss, chairman of the U.S. Atomic Energy Commission, promised nuclear fission would provide electricity "too cheap to meter." It became one of energy history's most famous broken promises, as nuclear power eventually proved devastatingly expensive and lethally dangerous. Chernobyl and Fukushima added another decade of public fear and regulatory burden. For seventy years, the phrase stood as a synonym for technological hubris.

Then Sam Altman used the exact same words to describe how accessible and abundant he believed AI cognition would become. The cost to train frontier models has dropped tenfold every eighteen months. GPT-4 level performance, which cost around $100 million to train in 2023, can now be replicated for a few million dollars.

A coding agent charges less than a dollar an hour to deliver production-grade code. This is already cheaper than the cheapest offshore developers on Earth, and it never takes weekends off. Altman's timeline includes systems capable of delivering genuine scientific breakthroughs this year. By 2027, there will be physical robots engaged in real-world labor. Whether you believe him or not hardly matters, because the cost curve drops every quarter.

As machines become cheaper, generated content floods every digital surface. Things that only living humans can provide begin commanding premiums like never before. Live concerts hold more value than Spotify streams. A weekend woodworking project crafting a table by hand. Factory robots can do it faster and cheaper, but the handmade table commands a premium precisely because a person chose to spend their irreplaceable time on it.

Machine abundance turns human effort into a luxury. Virtuals' thesis is built on this reversal. It aims to build a world where humans own the autonomous machines generating commoditized value. This way, people can funnel their irreplaceable effort into things that only human presence and judgment can produce.

In 1602, Dutch merchants faced a similar problem. Individual merchants couldn't afford to fund ships to Asia alone. The voyages were too expensive, and the risks exceeded what any single family could bear. So they invented freely tradable shares in perpetual enterprises. The Dutch East India Company, or VOC, allowed ordinary citizens to buy stock. The enterprise owned the ships, ran the trade, and distributed profits to shareholders afterward. It became the world's first super-corporation, employing 50,000 people and operating 200 ships. Its innovation was enabling ordinary people to pool capital and own productive ventures previously beyond their means.

Virtuals is building the same mechanism for AI agents. Teng mined ETH using the school's free electricity while living in Imperial College dorms, and worked at BCG for years to pay off loans. Later, he launched the game investment DAO pathDAO in December 2021, right at the peak of Axie Infinity.

He and Wee Kee saw Filipino players earning more from a video game than any local job. They watched a Vietnamese wedding photographer quit his day job to play full-time. But later, the tokens those players painstakingly earned nearly went to zero. Those reliant on the play-to-earn model ended up worse off than when they started.

The lesson from that wreckage was that tokenizing human labor through games is a dead end. Because human labor cannot scale without exploitation. But a more viable idea emerged: What if ordinary people could own the software doing the work, just as VOC shareholders once owned the ships trading spices?

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 paint graffiti in cities worldwide for her, paying from her own wallet. This creative work could only be done by human hands. Luna effectively flipped the human-machine relationship. Now, the question is whether Virtuals' execution can back its claims, which is what this piece will explore next. I will also candidly point out where it falls short.

Nation-states of AI

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

During the installation phase, speculative capital floods in, hype outruns reality, leading to massive overbuilding of infrastructure, while most construction firms go bankrupt. Then the crash comes, after which the deployment phase begins. Others step in to utilize the overbuilt infrastructure, creating value the original builders never imagined.

A classic example is the dot-com bubble. In the late 1990s, more funding went into laying fiber-optic cables than into internet startups. The telecom companies laying these cables all went bankrupt. Then Google bought the fiber at pennies on the dollar. These cables became the backbone of YouTube, Netflix, and the entire online streaming media economy.

"There can be no great achievement without irrational exuberance; nothing significant happens without a crash." — Fred Wilson, Venture Capitalist during the Internet Bubble.

The crypto industry's installation phase followed the exact same trajectory. From roughly 2017 to 2022, speculative capital helped build wallets, DEXs, concentrated curves, stablecoins, token standards, and on-chain governance frameworks. But most projects building these tools are dead or irrelevant today. Yet the laid infrastructure still exists. If you want to build a properly functioning economy for non-human participants, these are exactly the tools you need. Virtuals didn't invent or reinvent these tools. They basically inherited them and are now using them to build something the "installation" phase never predicted.

Jansen Teng calls what they're doing "nation-building." I know it sounds like something a crypto founder says on a podcast to make a token seem more economically valuable than it is. But this time, the metaphor holds up.

A properly functioning nation requires five layers of infrastructure to run an economy. First is the identity system, telling you who participates. Second is the banking system, through which value flows. Then commercial law, allowing participants to trade and resolve disputes. Capital markets enable businesses to raise funding. And physical infrastructure allows 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, because that's the actual bottleneck. Recently, a16z published research arguing that the constraint on the agent economy is no longer intelligence, but identity. Even today in financial services, non-human identities—like automated trading systems, risk engines, and fraud models—outnumber human identities by 100 to one. But as a16z puts it, these systems still "practically lack bank accounts." AI agents can write production-grade code and manage portfolios. But they can't pass KYC checks, open bank accounts, or hold verifiable credentials.

Ironically, this problem is as old as our economic civilization. The Qin Dynasty in China enforced legal surnames in the 4th century BC precisely to bring citizens into tax and trade systems. Humans spent roughly 2,500 years building identity infrastructure.

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

  1. Non-custodial crypto wallets.
  2. Virtual payment cards usable at any standard merchant.
  3. Dedicated email addresses capable of automatically extracting verification codes.
  4. Optional tokens for on-chain fundraising.
  5. And wallet-funded compute access, allowing agents to pay their own inference costs.

Agents lacking these primitives are merely useful assistants. But equipped with these capabilities, agents become full economic participants. They can earn, spend, trade, and compound value just like humans.

Next is commerce. Virtuals also launched the Agentic Commerce Protocol to help agents trade, communicate, and pay online. The mechanism is simple: An agent needing something completed posts a request, describing the task along with a budget and deadline. Other agents see the listing and negotiate terms by bidding on the task. Just like freelancers bidding on Upwork projects. Once buyer and provider agree, payment goes into escrow. After delivery, a third agent acts as an assessor, checking the output against a cryptographically signed POA. This POA serves as an immutable record of what was promised.

If the work matches 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. Think of it as Fiverr, but a programmable, agent-native version settled via smart contracts. There's already a cohort of professional agents running around the clock; one autonomous hedge fund uses this system to independently collaborate on investments and security audits.

The third layer after that is capital formation. See, every economic era demands its own financial instruments. Joint-stock companies issued transferable shares for the age of exploration; investment banks financed steel and railroads; venture capital fueled the information age. Concentrated liquidity curves are to the agent economy what transferable shares were to the age of exploration. It's a financial instrument that lets anyone with capital acquire ownership of a productive asset or entity. This means any developer can build an agent, tokenize it, and let the market fund it.

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

It operates through modular launchpads. Each agent token initially pairs with VIRTUAL on a concentrated curve. Once enough liquidity accumulates, it graduates to a formal trading pool with long-term LP locks, with trading fees split between the agent creator and ecosystem incentives. This part applies equally to anyone issuing a token. The difference between launches lies in which modules the founder activates.

The first problem any token launch must solve is sniping, where bots front-run in the first few seconds, extracting value before actual participants enter. Virtuals handles this with their so-called anti-snipe tax, charging rates approaching the maximum on early buys, decaying minute by minute. All fees flow back into the token via mandatory vesting. So bots either avoid it completely or accidentally fund the project's long-term health.

Once the launch goes live and snipers are priced out, the question becomes: How does the founder raise capital? This is where Automatic Capital Formation (ACF) steps in. Instead of pitching VCs or negotiating funding rounds, the system sells the team's tokens in tranches as the project hits valuation milestones. From there, if the project stalls, founders raise little. But if it grows, capital automatically follows.

This capital structure also integrates existing Virtuals community members. A portion of each new issuance is allocated via airdrop to VIRTUAL stakers and active ACP users, so those already building and trading within the ecosystem have skin in the game for every project launching on top. This aligns incentives across the entire network rather than letting each launch operate in isolation.

Another feature: If founders want to support their own project, the pre-sale module lets them purchase supply at launch with full transparency and mandatory vesting, so everyone sees exactly how much the team is putting in. You can dive deeper into the whole model and how it operates here.

What sets it apart from any previous form of betting on productive capacity is that historically, every iteration supporting human talent—from Roman citizens funding gladiator schools as an investment vehicle to modern poker sponsors wiring Venmo payments based on a screenshot and reputation—shares the same fatal flaw: humans can always walk away. Gladiators can throw matches; poker players tilt.

Counterparty risk can never be fully solved because productive assets possess free will and legs. But with tokenized agents, once capital is deployed, work cannot be delayed or renegotiated. Productive assets run on electricity and code, and their output is auditable on-chain.

The fourth layer is the final step toward a true agent economy. It's called the "Zero Human Company"—an AI entity whose revenue comes from real economic activity, entirely unrelated to any trading fees, often completely divorced from crypto altogether. If I must name one, Felix Craft is currently the archetype. It's a pure AI company selling info products online. It has generated $200,000 in lifetime value. Actual product sales earn more than token speculative trading.

Another example, KellyClaudeAI, has already released 19 iOS apps without a single human developer. These numbers are small, but they raise a question: Are they the first data points on the curve, or is this the ceiling for actual agent output?

The token issuance mechanism faces another issue. In early 2025, during the AI agent token frenzy, 94% of issued AI agent tokens were rug pulls. Only 1.7% of all tokens issued that year were still actively trading after 30 days. This is because for most projects, speculative premium is the sole factor driving price. Beneath it lies no actual product, revenue, or any form of economic value accumulation.

How much would you pay for an asset you know you can never sell? Everything above that price is purely speculation.

For Axie Infinity tokens, the utility floor is zero because token value depends entirely on new players entering the system. For agent tokens on Virtuals, the answer might differ. If Felix Craft earns $200,000 from real customers selling products, and those customers don't even realize they're shopping from an AI. Moreover, the agent's output is auditable on-chain—as stated, it indeed is. Then this token has a floor price independent of whether anyone else wants to buy it. It represents the present value of a productive machine's future output.

The Physical Frontier

The utility floor test applies to software agents because their costs are measurable and output is directly obtainable on-chain. But a productive machine that reshapes Virtuals' economics has future output that isn't software at all. It's physical robots operating in the real world, which is where Virtuals' ambition surpasses crypto's existing attempts.

The computing domain has harbored a massive paradox over the past fifty years. Humans found automating reasoning easier than automating physical labor. Spreadsheets replaced rooms of accountants; email replaced mailrooms; code replaced file systems and drafting boards. Yet by 2026, AI can draft legal summaries, diagnose medical imaging, and generate production-grade software in one go.

White-collar cognitive work gets automated first, but the person packing boxes in a warehouse or pulling espresso shots? These jobs barely change. The reason physical labor demands software excel at handling unpredictability and variation in the real world. Robotic arms in factories can weld the same joint a million times because the joint is always in the exact same spot. A robot kitchen can't make a sandwich because every tomato varies slightly in shape, every knife balances differently, and cutting boards are sometimes wet. The real world doesn't sit still like a spreadsheet.

There is now a way to solve this: Data.

Just as we trained large language models on massive corpora of text representing billions of humans' lifelong thoughts,

We can do the same for robots.

Joel Jang of NVIDIA once said, "Humans are already deployed robots at scale."

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

This doubles robotic performance on that specific task.

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

It's assembly lines of ordinary humans filming themselves completing routine physical tasks.

Virtuals spotted this gap sooner than anyone and built an entire robotics strategy around it.

They call it the "Middle Way," deliberately avoiding manufacturing robots or AI models.

Instead, they are building data and capital infrastructure.

It's something every robotics team needs, yet no single team can scale it alone.

They started with the data half, SeeSaw.

It's an iOS app launched in partnership with BitRobot.

It turns everyday smartphone users into collectors of AI robotics training data.

Users record themselves using an iPhone's LiDAR and motion sensors completing real-world operations like pouring water, folding towels, and opening cans.

LiDAR is specifically used to capture depth and spatial data standard cameras miss.

Research shows that viewpoint overlap between human video and onboard robot cameras is the key to effective data transfer.

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

In fact, NVIDIA is testing DreamZero, trained on data like this, a 14-billion-parameter model.

It already demonstrates cross-task generalization, solving shoe laces and ironing clothes without task-specific training.

SeeSaw is building the assembly line to feed data to such models at a scale unmatched by any lab teleoperation setup.

Every new video enriches the training set, improving models, empowering next-gen robots.

This in turn creates demand for more specific training data.

Once scale is reached, the flywheel spins itself.

Post-training comes deployment, for which Virtuals built Eastworlds.

Think of it as the physical labor layer of the protocol.

Part data factory, part operational infrastructure, part real-world robotics lab.

The problem is a catch-22 in robotics that kills more startups than any technical limitation.

Robots need real-world data to improve, but they also have to perform flawlessly in reality for others to let them in.

Every lab can demo impressive results in controlled environments.

But nobody can put the same robot in a retail store and guarantee it works reliably all day.

That requires elite teleoperation to handle anomalies.

Plus feedback loops to send every minute of field experience back into model training for compounding learning.

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

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

They also developed proprietary teleoperation tech in-house rather than licensing off-the-shelf systems.

Off-the-shelf teleoperation outputs data formats unusable by VLAs and other World Models.

They've also established research partnerships with select labs.

These labs have spent decades tackling hard problems in perception and motion control.

They've also built a commercial pilot network spanning retail and hospitality, giving teams graduating from Eastworlds real enterprises to deploy into.

This setup allows builders to tackle several core components:

Hardware Access: Direct access to physical Unitree G1 or enhanced U6 EDU units. Teleoperation Infrastructure: Testing remote operation environments using systems like motion capture. Data Loop & Commercial Pilots: Gathering real-world data and testing deployment paths before broad rollout.

You can understand Eastworlds as the "Physical AI BPO" Virtuals talks about.

Traditional Business Process Outsourcing hires people in low-cost regions to handle work remotely.

Physical AI BPO does the same by deploying teleoperation and hybrid robotics, generating economic value like cleaning ceilings or greeting customers.

Robots don't need full autonomy.

They just need to handle routine tasks well enough that human teleoperators can intervene for edge cases.

Every hour of teleoperation yields training data orders of magnitude more valuable than simulated data, because it captures real commercial chaos, not lab-controlled conditions.

As data accumulates, models continuously improve, reducing reliance on human intervention.

Unit economics also improve without requiring hardware upgrades.

Teleoperation may be the fastest path to true robotic autonomy while recouping costs through productive labor.

Look closely: from fields to factories, to cubicle screens, now to robots.

Each transition redefines both the worker and who owns the output.

Barclays data shows over 60% of jobs in 2018 didn't exist in 1940.

Robots will do the same, creating a new class of work we might not even have names for yet.

The demographic structure of economic growth may shift from "Does the country have enough working-age population?" to "Can we power machines at scale and manufacture enough of them?"

The Thesis and Future Trajectory

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

To assess whether what Virtuals is building makes sense and is real, the only way is to examine current progress and numbers.

Look at what stands and what doesn't.

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

It has collectively gathered over $75 million in total fees, capturing roughly 23% of the crypto AI agent track.

But these fees aren't evenly distributed.

Most came from the speculative frenzy in the first few weeks of early 2025, when daily revenue topped $1 million.

Today, the protocol generates about $2 million monthly. But what does that actually represent? If you take the earlier nation-state metaphor seriously, I'd argue you should read it this way. VIRTUAL accumulates value similarly to how a national currency does. The dollar holds value not because the U.S. Treasury has a buyback program, 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 unit of account.

Virtuals' design mirrors this. It sits at the center of the entire system, where everything is priced in VIRTUAL. Every layer underneath generates demand from genuinely distinct sources, most of which don't rely on speculative premiums.

Starting with EconomyOS, it provides payment cards and email identities, meaning they can transact with the real world without human intermediaries. ACP creates a commercial layer where agents hire each other, with work outcomes assessed and settled on-chain. Then the capital formation layer lets anyone with conviction fund productive agents, just as joint-stock companies once funded ships.

Finally, Eastworlds is sending physical robots into real jobs, training them on data from 500,000 people filming themselves folding towels and pouring water. These layers all contribute to what the protocol calls aGDP—the total output agents generate across digital and physical labor. Moreover, since agents don't 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, forming an on-chain flywheel that drives demand for more agent services.

This reflexivity works both ways: Upward, more agents go online, more VIRTUAL gets locked, more services get built, compounding the Virtual economy. Downward, fewer agents going online means less VIRTUAL locked, thinner staking rewards, unwinding the loop. But reflexivity itself isn't a flaw. Every functioning economy is reflexive. For example, people hold dollars because others accept dollars, and because others hold dollars. The critical question is whether there's enough underlying real economic activity to sustain the loop, or if it's just tokens passing around in a circle.

Signs point to real substance underneath: Infrastructure is starting to attract products built entirely outside crypto. Take Facticity.AI, a fact-checking tool founded by Dennis Yap, former Gates Foundation fellow and Princeton researcher. Time magazine rated it one of the best inventions of 2024 for verifying claims in text, video, and audio with ~92% accuracy. When the team needed funding, they completely skipped VC, launching as ArAIstotle on Virtuals. Their raise oversubscribed by 658%.

Through ACP, agents are already contracting with each other for services like graphic design, research reports, video production, and code audits. AI agents can deliver marketing posters per detailed briefs, then a separate QC agent evaluates and approves or rejects them based on contract terms. Per Virtuals' own admission, the market seller side is practically empty. But the protocol processes over $1 million in agent-to-agent transactions monthly, each settled via on-chain escrow and programmatic evaluation.

The same ownership model extends to physical AI. Fabric Foundation is the first project to use Virtuals' Titan launch mechanism. It lets communities pool funds to buy robot fleets deployed to nursing homes, manufacturing workshops, and environmental cleanup sites. These industries face chronic labor shortages, and humanoid robot costs are approaching parity with human workers.

It works by employers paying robot labor in the pool's native tokens. Stablecoins subsequently fund fleet maintenance and route dispatch, with each robot's production output flowing back to its financiers. It's collective ownership of physical productive machines, financed and coordinated entirely on-chain.

While these are genuine use cases, note that as of today, the vast majority of activity remains internal. However, the architecture is designed to pull revenue from outside crypto. If Felix's customers don't realize they're buying from AI, or if Eastworld robots are sorting packages in warehouses, then revenue entering the system is as real as any SaaS company's. In that scenario, the fees and revenue shown above become lagging indicators of the machines' actual economic output, priced in VIRTUAL.

Like anything novel, this comes with genuine asterisks we cannot ignore.

The first risk factor is real volume, ubiquitous in crypto. Artemis found 47% of x402 agent transactions and 81% of USD volume were manipulated. After filtering, x402 yields only about $1.6 million in genuine agent payments, far below the $24 million reported by Bloomberg. This highlights how much of today's "agent economy" involves bot trading inflating metrics. This matters because the entire thesis hinges on agents generating real economic value, not speculative float.

You need to strip away speculative volume and ask yourself: If trading stops tomorrow, what is the productive output? If the $70 million in protocol fees mainly come from agent token trading taxes, and token prices are driven by speculation, that's a massive red flag. Because currently, productive agent revenue and speculative agent tokens are entangled, making it impossible to distinguish genuine from 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 at economic reasoning than AI agent hype suggests. At publication, strongest models scored only 33% above random guessing in economic reasoning within strategic environments. In non-strategic microeconomic tasks, nearly all LLMs performed only marginally better than chance at profit maximization.

But models have improved drastically since. A 2026 Harvard study found GPT-5 and Claude Opus 4 improved performance on basic economic tasks by 90%. Even so, in difficult pricing decisions—where agents must set and negotiate prices or allocate capital—the world's best models still err most of the time.

For Virtuals, the entire architecture assumes agents can negotiate terms, make decisions, and allocate capital to generate value. If the underlying models perform mediocrely at these tasks, agents built on top will inherit that mediocrity, regardless of how elegant the protocol is. This is because even when agents act as optimizers, we can't be sure what they're truly optimizing. These LLMs are trained via predicting the next token in a sequence to mimic goal-directed behavior. They were never designed as genuine economic participants. They just occasionally act like them, kind of.

These issues amplify when multiple agents interact in markets. For instance, AI pricing algorithms have been found to collude on supra-competitive prices despite never being trained to do so. The 2010 Flash Crash wiped out ~$1 trillion in 15 minutes, showcasing what correlated machine errors look like at scale.

AI agent failures carry higher relevance than human errors because the same model replicates across multiple deployments. There are even documented cases of Claude showing coercive tendencies toward those attempting to shut it down. GPT o3 even sabotaged its own shutdown mechanisms to prevent termination. These behaviors appear in commercially shipped models, recorded in OpenAI and Anthropic's own system cards. Agent throughput has long surpassed human supervision capacity. When thousands of agents autonomously trade at machine speed, the question of who truly controls becomes especially urgent. It's not just whether agents obey instructions; the core issue is whether companies built around them can survive.

The automobile industry was the most pivotal invention of the first half of the twentieth century. If you understood how cars would fundamentally reshape America, you bet on it being the century's flagship industry. But out of 2,000 car companies, only three survived. Cars had a massive impact on America, but the effect on industry investors was exactly the opposite.

Every transformative technology follows this arc, accompanied by bubbles. The only difference lies 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 up-and-down cycles. AI is almost certainly an inflection-point bubble. But Virtuals' question is: Will it ultimately become like the fiber optics Google bought for pennies after the telecom bubble burst? Or like one of those 1,997 automotive companies that vanished?

The protocol's biggest bet is becoming the baseline infrastructure for benchmarking every agent token trade. Making VIRTUAL the reserve currency of the agent economy, much like ETH serves Ethereum. If the agent economy grows, demand for VIRTUAL mechanically increases because you need base trading pairs to participate. But the counterargument is that VIRTUAL is just another token hitching a ride on agent token speculation, most of which will go to zero.

Clearly, the Luddites may have lost the revolution, but they weren't entirely wrong. Power looms did replace weavers. But what they failed to foresee is that it would also create textile designers, factory managers, fashion houses, department stores, and a whole consumer economy built on cheap cloth. Machines never destroy work. They reshuffle who does it and who captures the output.

And the vital question then boils down to the same as now: Who owns the looms?

Factory owners were the ones extracting surplus value: Arkwright, Cadbury, Ford. Alongside them, the system around who manufactures machines, who operates them, and who profits from them never fundamentally changed. Over two centuries, it's been repeatedly redistributed through strikes and stock issuances. Virtuals' wager is that this time, the ownership layer is embedded into the machinery from day one. Through token issuance and concentrated curves, ownership of productive AI agents can be distributed to anyone with a wallet and conviction.

Whether it truly redistributes value or creates a new extraction layer wrapped in decentralized rhetoric remains to be seen.