Author: Maddie P, Partner, Robot Ventures
Compiled by: Hu Tao, ChainCatcher
Stop reading academic papers. The cutting-edge fields have evolved to a new stage, while you're still refreshing arXiv like you're refreshing your ex's Instagram.
I know, because I did the same thing. For six whole months, I was obsessed with every model release. I collected every benchmark. GPT-4.5 preview and all that? I read them all. You know what I learned? Nothing. You know what I gained? Less than nothing. I paid the price for my mistakes.
Meanwhile, a child earned more using a three-month-old model and a Stripe API key than my "cutting-edge implementation" ever did... more than all my work combined. Not because his AI was smarter, but because his AI had direct access to money.
Every AI researcher I know is now facing the same nightmare: they spent five years creating God, only to find that God needs a credit card and Apple's permission to exist.
The model war is over. Not because anyone won, but because winning or losing no longer matters.
The commodification of intelligence is happening faster than the commodification of licensing.
In 2023, GPT-4 was priced at $30 per million tokens. Now, Gemini Flash Lite is priced at a mere $0.08. In just 18 months, the price has plummeted from a luxury item to less than a cup of morning coffee, a complete collapse.
Performance is converging. Pick any benchmark: ARC, MMLU-style tests, GPQA-style hard tests, code evaluation—the results are all the same. Cutting-edge technology continues to advance, but the gap between top-tier models is narrowing relative to real-world production environments. It looks like the smartphone market in 2015. All products used the same chips, the same screens, the same cameras. The only real difference was who could get a carrier license.
Models will continue to become more intelligent, but the commodification of intelligence is happening far faster than the granting of permissions. The gap between what artificial intelligence can do and what it is allowed to do is widening, and this gap holds enormous potential profits.
corpse
The graveyard is full of companies that possess perfect artificial intelligence but have no authorization.
I witnessed a transcription startup fail, even though its accuracy surpassed that of human stenographers. The reason for its demise wasn't technological failure, but rather its inability to convert users into revenue. The product was fantastic, but lacking a payment channel, it simply ended there.
Another team developed an AI trader that actually made money, but it was rejected by all app stores. Autonomous financial activity is making compliance teams uneasy. There are currently no clear approval options for AI that can transfer funds without human supervision.
This is the real dilemma that will cause you immense suffering. Even if customers are willing to pay, they will be very calculating. Why buy a supercomputer when they can wait six months for the next generation to offer the same functionality for free? This is practically a tactic Intel used in the 1980s. Your customers aren't just being held back by limitations; they're waiting for you to exit the market. Commoditization is eroding your pricing power from the bottom up, while the limitations are squeezing you from the top down.
These two pressures combine. You can't collect enough fees to survive because the model will be cheaper next quarter. You can't quickly integrate the payment system because the speed of compliance processes depends on the speed of people. You're squeezed from both sides. The ultimate winners are only those who own the railroad. No matter how high or low your fees, no matter if you ship this quarter or next, they can collect tolls. They always get the money.
Companies don't die from bad AI; they die from well-developed AI that gets trapped. This is due to a combination of restrictions on access and pressure to commercialize. This pincer movement is what truly hinders business growth.
Permission stack
The value creation of modern artificial intelligence lies in the accumulation of permissions. This is not because the model itself is not good enough, but because autonomy is restricted.
Consider self-driving cars. We have impressive demonstrations, but most systems are still just driver assistance. Humans are still involved, responsibility is outsourced, and the real world is full of extreme situations. Artificial intelligence is at the same stage. We've reached Level 2 autonomy in cognition, but not yet Level 5 autonomy in economics.
Level 1 Access: Can I call the API? → OpenAI = Thinking Access;
Level 2 Access (Compliance): Can user data be stored? → AWS = Memory Access;
Level 3 Permissions (Revenue): Can you process payments? → Stripe = Permission to charge users;
Level 4 Permission (Distribution): Can the app reach users? → App Store = Publishing Permission;
Level 5 Access (Funds): Can the system obtain credit, margin, and settlement guarantees? = Economic Behavior Access.
Most AI companies remain at level three or four. They can sell subscription services, but they can't achieve financial self-sufficiency. This is why user conversion rates plummet. Each additional level of access is like a tollbooth, and tollbooths accumulate. Even with a superior model, economic autonomy remains elusive.
This is the real crux of the matter: what we lack is not intelligence, but economic autonomy. Unless entities can borrow, settle, and repay within given constraints, the flourishing development of artificial intelligence will only degenerate into a transfer of wealth from application developers to the licensing layer that grants them the authority to survive.
Why are agents more expensive than prisoners?
Here's an experiment: Your intelligent agent has an IQ of 180, can access browsers and command-line interfaces, and can call any API. But it has no money and cannot borrow money. What can it build?
there is nothing.
Like a talented employee who can't even get reimbursed for a domain name; like a trader who can devise an entire business plan but can't afford a $12 server hosting fee; like a trader who has discovered a perfect arbitrage opportunity but can't execute the trade; like an F1 race car in a world without roads.
Why not just give them a company credit card?
A credit card requires you to already exist. You need a Social Security number, a bank account, and three years of credit history. Your agent program has nothing. It only has an API key.
Even if you could give it a card, it's like giving a Formula One driver a bicycle. Yes, it works, but it's not the right tool. Credit cards limit your risk. They have daily limits, fraud alerts, and human verification. A broker might need a million dollars to execute a perfect arbitrage in thirty seconds. Try explaining that to Visa's fraud unit. And credit cards are designed based on predictable risk. They assume your spending follows certain patterns: groceries, gasoline, occasional vacations. Underwriters can build such distribution models. Brokers work differently. They take on risks that cannot be predicted. The point of intelligent systems is to discover opportunities you don't expect. You can't pre-approve credit limits for opportunities that haven't yet been discovered.
But the real problem goes far beyond that. Credit cards are for employees to buy office supplies, but brokers aren't employees. They need to use future income as collateral for loans, leverage locked-in positions, and flexibly adjust credit lines based on market opportunities. What they need is capital that can keep pace with their thinking.
We provided our agents with a variety of tools: browsers, terminals, APIs. Doesn't it feel like victory is in sight? But the truth is, any significant economic activity requires funding; and the infrastructure doesn't exist because the financial system was originally designed for humans. There are business hours, all sorts of forms, someone always has to take the blame, and Brad from the compliance department eats lunch from noon to 2 pm.
Why Crypto is the answer?
Agents need capital that can operate at machine speed, without manual approval, and capable of handling unpredictable risks. Where does this kind of infrastructure currently exist?
On the chain.
This isn't crypto hype; it's a fundamental architectural reality. Traditional finance has always included human factors from its inception. This isn't a loophole, but the essence of its business model. Every approval, every audit, every compliance review means someone's job, someone's salary, someone's lunch break. This friction itself is the product.
Crypto eliminates human intervention. This isn't a matter of philosophy, but rather a matter of technological limitations. Smart contracts either execute or they don't. There's no need for regulatory oversight or contacting anti-fraud agencies. The code runs exactly the same way at 3 AM and 3 PM.
More importantly, on-chain collateral solves the underwriting problem. You don't need to rely on personal credit history or proof of employment for lending; instead, you rely on mathematics. Collateral is verifiable, highly liquid, and trustless. Agents holding staked ETH do not need to prove its value to Brad. ETH is locked and accessible programmatically. Loan disbursement, use, and repayment can all be completed through a single atomic transaction.
This is the crux of the matter. It's not because cryptocurrencies are cool (though they are indeed cool), but because cryptocurrency is the only infrastructure that allows traders to take on financial risk without prior permission. Traders can use locked positions for lending, complete transactions in milliseconds, and settle without intermediaries. The infrastructure already exists, it just hasn't been optimized for this use case… until Sprinter came along.
Sprinter (which we are building)
Last year, I tried lending by staking Ethereum (ETH) and told Michael Cieri about it. Three weeks later, I had my fourth conference call with Brad from Traditional Finance Inc.
Brad wants to know my investment goals. Brad needs proof of my employment. Brad is worried about the volatility of cryptocurrencies. Brad goes to lunch.
My Ethereum is right there. Immutable. Verifiable. Generating continuous returns. And while I'm listening to soothing jazz, Brad is "confirming with his supervisor."
That's when I realized: the entire credit system was orchestrated by Brad. Hundreds and thousands of Brads were passing documents between each other, charging Brad for his time.
Sprinter deleted Brad.
Sprinter is a programmable credit engine. You can borrow spendable stablecoins using on-chain verifiable collateral without selling the collateral. Credit usage is restricted: funds do not go directly into the free wallet, can only be used through authorized routing, and repayments are deducted first. We initially launch a consumer-facing credit card as the distribution channel, then release an SDK so that applications and agents can apply for short-term credit lines under strict restrictions. Credit lines are provided via API. A rules-based underwriting mechanism. No Brad involvement required.
It's not just about humans. Humans are just the beginning. We're also building systems for machines that need to borrow at 3 a.m. to rent computing resources before prices spike. We're building systems for agents that need to borrow against locked positions to execute 30-second arbitrage. We're building systems for protocols that require millisecond-level lending, not weekday lending.
Other companies are using artificial intelligence to develop products for humans. We, on the other hand, are using our funds to develop products for both humans and artificial intelligence.
This difference may sound subtle, but its impact can reshape the economic landscape.
What is truly being destroyed?
Policy is the end switch. A regulatory letter is enough to destroy everything. I have witnessed three perfect teams crumble within 48 hours simply because someone in Washington discovered their use case.
The settlement process has become a political issue. You see two economic systems: one where subservient dollars circulate slowly in banks, and another where programmable dollars flow at the speed of light. The difference between the two is not accidental, but rather a whole industry.
Stablecoin redemptions are akin to a bank run with a better user experience. The entire balance of the AI economy rests on the stability of stablecoins. Once (not if) stablecoins fluctuate, the entire economic system is affected. Of course, we might as well pretend the risk is "blockchain throughput."
Final
Some people can build a billion-dollar company using models from three years ago. This isn't because those models are inherently good, but because they've figured out how to make well-thought-out, savvy venture investments using unsifted capital.
Stablecoins are decoupling from corruption. Trading volume is shifting from derivatives to real-world trading. Traders are paying each other for computation, data, and inference.
A major power is panicking. "Unauthorized autonomous economic activity." This will make headlines. Markets crash. I'll eat popcorn and watch the spectacle.
Company valuation multiples are often based on API access permissions, not the quality of artificial intelligence. Bank relationships are more valuable than model parameters.
The era of a "two-dollar" economy has arrived. Traditional dollars circulate in banks at human speed, while programmable dollars are settled instantly on the blockchain. Arbitrage between the two has become a true game.
If none of this happened, then all my previous judgments were wrong.
But I wasn't wrong. The evidence is here. Inference costs have dropped by 99%. All models are converging on the same benchmark. Companies with perfect AI that don't require authorization have gone out of business. Brad still eats lunch every day from 12 pm to 2 pm.
What if I'm wrong?
Perhaps GPT-7 is so intelligent that permission issues are irrelevant. Perhaps Brad will learn to use email.
But probably not.
The commodification of intelligence far outpaces the spread of access. Distribution channels remain limited. Access remains precious.
The AI investment game that everyone was playing is over. Not because the game failed, but because it was so successful that it became worthless. Investing in the smartest models, building the best RAG (Red-Yellow-Green-Blue-Green) models, and optimizing inference—none of that matters.
The new rules of the game are the accumulation of authority, the scope of control, and the economic trajectory.
The winner will not have the best AI, but rather the AI that can do things using known information.
This is not pessimism. The development models of the internet, mobile, and cloud computing are all the same. Technology is commodified, while the railway system creates value.
The real test
The Turing Test poses the question: Can a machine convince us that it is human?
The question we should be asking is: Can the machine trade with Brad and win all his money?
The answer to the first question is yes.
The answer to the second question is precisely why your AI startup is doomed to fail.
The house always wins. And the house is Brad, sitting in the compliance department, eating a boring desk salad, ready to reject your API request.
The future is not one of equality for all. That message is still stuck in Brad's inbox.
