I. Tokens Are Too Expensive
On March 7, at an All-In Podcast live event, Chamath Palihapitiya made a statement that shocked the industry to his three friends who were also betting on AI:
"My costs triple every three months. My revenue does not."
Chamath is the founder of Social Capital and a former early executive at Facebook.
In 2024, he launched a project called 8090, meaning "achieve 80% of functionality with AI to save 90% of costs."
The goal was simple: rewrite the world's legacy software.
Today, the company named "save 90% costs" sees its own AI bill triple, approaching $10 million (including AWS compute rentals, Cursor subscription fees, Anthropic API, etc.).
Chamath's post on X, whether mocking himself or the industry:
"Thanks to VCs for footing the bill for this Token buffet with massive investments."
This is likely the most ironic scene in Silicon Valley in 2026. A top VC publicly mocks other VCs for underwriting his company's Token bills.
Dax Raad, founder of OpenCode, also revealed on X that CFOs are getting restless.
"CFO: What do you mean, an extra $2,000 per engineer per month in Token costs?"
In the past, corporate technology spending was clear: servers, cloud computing, SaaS subscriptions. One annual budget sufficed. Now, there is an uncontrollable Token bill that keeps rising.
The last time I saw such uncontrollable costs leading to astronomical fees was during the 4G era when 30MB cost $0.70. People joked then that they could lose their house overnight. In 2026, Token buyers are about to say the same thing.
Why is using AI becoming more expensive?
To improve intelligence, we constantly seek better solutions, and long-chain reasoning has become the technical winner of 2025.
Training Costs: Each generation of model training costs increase by about 10 times, while performance improves by only 2 times. GPT-3 training cost was about $4.6 million, GPT-4 exceeded $60 million, and GPT-5 has surpassed $500 million. A 10x input yields only a 2x return.
Agentic Cost Explosion: The cost formula for Agentic AI is k × O(N²), where k is the number of tool calls per round. Every tool call requires reprocessing the entire conversation context. The cost of a 40-round conversation is approximately 4 times that of a 20-round conversation, not 2 times.
Whether training costs or usage costs are rising, consider this example:
GPT-5.4 Pro had an extreme case: someone typed "Hi", waited 5 minutes and 18 seconds, and spent $80. The reasoning chain expanded to an absurd degree.
The Cost of Intelligence Is Paying More
Chamath also encountered the epic "Ralph Wiggum" plugin. This plugin forces the AI Agent to continuously loop to determine if the task goal is met. If not, it continues looping.
Ralph Wiggum is the "stupid but persistent" character from The Simpsons.
This plugin causes the Agent to repeatedly retry the same problem. It might solve it, or it might solve nothing. You could lose your house overnight.
This is exactly what is happening with OpenClaw. The biggest problem in deploying OpenClaw is that behind every seemingly high-quality completion lies top-tier models and continuous improvement, continuous monitoring, where every action burns RMB.
II. You Don't Want to Burn Cash, But the Market Forces You
The market is quietly changing. The traditional software business model is clear: huge upfront investment, with marginal costs near zero later. Whether you sell the 10,000th or the 1,000,000th license, the cost is almost the same. The more you sell, the more you earn.
AI applications work in reverse. Every user interaction incurs a direct cost. Every Token must be paid for. The more you use, the more you lose. The more autonomous you want it to be, the more expensive it gets.
In traditional SaaS, all user costs are basically the same, approaching zero. In the AI era, it's reversed: your most loyal users are your most expensive users, while low-usage customers might switch to a competitor tomorrow.
Faster growth means more losses. The classic VC script of "burn cash first, monetize later" has become a highway to bankruptcy in the AI era.
Competitive structures force everyone to a new level of intensity.
Many AI entrepreneurs do not choose the low-cost "Human + Code + AI" model, using AI only at key nodes and leaving the rest to humans or processes? Because large model manufacturers no longer offer this option.
OpenAI, Anthropic, and Google start with Agents from the very first frame. Claude Code is directly an autonomous coding agent; Codex is directly an end-to-end coding agent. These compilation tools are almost killing Cursor. If upstream manufacturers' products are fully autonomous, how can you still be doing "Human + AI collaboration" without looking foolish?
Competition intensifies: pressure forces entrepreneurs to adopt the Agent model → The Agent model is the most Token-intensive model (k × O(N²)) → Token costs skyrocket → Profits are eaten away.
Entrepreneurs don't want to burn cash, but the market forces them to compete using the most expensive methods. Previously in tech, you could choose to "stay light," winning with fewer resources and smarter ways. In the Token economy, this is no longer possible. "Staying light" means your product capabilities are inferior to big tech's free tools. Entrepreneurs are caught in the middle,进退两难 (at a loss).
However, if the value created is significant enough, burning cash can be a good strategy. But many entrepreneurs are creating things that are no longer valuable, especially traditional Apps.
Once Coding Agents pass the capability singularity, the cost of software production approaches zero. But not everyone wins.
"After coding passes the singularity, the software world will not become fairer, only more crowded. It's not 'everyone wins,' but 'just building it no longer constitutes an advantage.'"
You use AI to write an app in minutes. But your competitor also uses AI to write the same app in minutes. Copy costs approach zero, and the time difference approaches zero.
Inputs are inflating (Tokens are getting more expensive), outputs are depreciating (software is becoming less valuable). Double squeeze.
"Building it is not valuable" does not mean nothing is valuable. In the Token economy, what is truly valuable has changed.
Previously, the scarce resource in the software industry was "ability to build." Being able to write a usable SaaS product was a barrier in itself; hiring engineers was expensive, development cycles were long, and technical barriers were high. Now AI has driven the cost of "ability to build" to near zero. The scarce resource has become "knowing what to build." Judgment, industry know-how, and deep understanding of user needs are things Tokens cannot buy.
When WeChat Official Accounts first rose, the scarce resource was "ability to write": daily updates, layout, finding materials. Later, everyone learned this, and the scarce resource became "what to write": topic judgment, industry insight, unique perspective. Now AI does even the "ability to write" for you. The only thing it cannot do for you is "know what readers need." The Token economy and the content industry share the same underlying logic: production costs go to zero, strategy and judgment become the entire barrier.
IV. The New AI Industry Chain is Also Being Reconstructed
Chamath moved from Cursor to Claude Code. The surface reason is "too expensive." But in reality, for intermediaries like Cursor, AI Agents have become so powerful they no longer need them. Traditional software was for humans; the new AI interaction landscape is quietly unfolding.
In 2025, Cursor was still an AI IDE unicorn, but by 2026, they reached a life-or-death situation, even holding a "wartime mobilization" all-staff meeting. They stated that the coding capabilities of AI models had reached a critical point: Developers can directly issue natural language instructions to Agents, skipping the "IDE" middle layer.
New interactions replace old interactions. Say two sentences, and the task is handled; you don't even need to open the software.
Software barriers are gone. Software without proprietary industry data or process barriers no longer exists. You can see 10,000 people making RSS summary daily report Agent versions, and 100,000 people making quantitative model Agent versions. There are no barriers; using AI computing power to create miracles means catching up to your product is just a 3-month task. The "Model Manufacturers → Tool Providers → Users" three-layer structure is gradually transitioning to "Model Manufacturers → Users."
SaaS not swallowed by large models are now working for model manufacturers. Although Mike from a16z (former Salesforce executive) says "It is too early to say SaaS is dead. In the real business world, tacit knowledge is deeply embedded in the underlying code of large SaaS systems. Special clauses in Indiana maternity leave laws, differentiated local tax rules—these cannot be replicated with a few prompts. Vibe Coding is a long-tail supplement, not a core replacement."
Part of this is correct. But he missed the other crucial half. SaaS will not die, but its profit structure will be eaten by the Token economy. Traditional SaaS profits come from "zero marginal cost." But once you stuff AI features into SaaS, every user interaction generates Token costs, and that "zero marginal cost" curve bends. SaaS probably won't die from external substitution; the AI features it added itself are eating its own profits.
What if you say you won't add AI features? Then vertical competitors are eagerly waiting for you to exit the market.
One last data point: Traditional SaaS gross margins are above 75%. AI product gross margins are generally between 50-60%, with some early companies even as low as 25%.
AI product COGS (Cost of Goods Sold) looks more like infrastructure services than software. You think you are selling software, but you are actually selling computing; you are just working for large model manufacturers.
So, Is AI Still Worth Doing?
Yes, but.
The problem with 8090 is that what it tries to do with AI—rewriting legacy software—is being devalued by AI itself. What you build with Tokens, others can also build with Tokens. The Tokens you spend only buy a ticket to the entry; the ticket itself is never a barrier.
The problem with Cursor is that it is in a track with no obvious competitive barriers, 80% dependent on AI, leaving it vulnerable to any large model manufacturer kicking it. Do not choose any track that is strongly dependent on AI without proprietary data support; victories here are temporary.
I have a simple judgment. When your input costs are exactly the same as your competitors, and your output is also exactly the same as your competitors, the business is reduced to a game of luck. Now, doing AI is about Token cost-performance ratio. But if your Token efficiency has no essential gap with competitors, you are just participating in a war of attrition.
In the Token economy, the survivors are not the companies that burn the most Tokens, but the companies that get the most irreplaceable value per Token.
Only those who have figured out "what one dollar of Token actually gets back" have a chance to survive until the day Tokens become cheaper.
Finally,
I also want to say that just figuring it out is not enough. In the Token economy, we are witnessing a cruel intensification: more computing power is invested, but the output value may become less obvious. Entrepreneurs are trapped in a dilemma: without Agents, the product is a toy; with Agents, the company is consumable material for model manufacturers.
But historical turning points often happen in this "extreme intensification." Perhaps the breakthrough will not appear in tracks that continue to pile up Tokens, but in places where they "dare to put a period to Tokens." They know when to turn off that persistent loop like Ralph Wiggum. They know to delegate 90% of mundane problems to cheap, deterministic code, leaving only the remaining 10% of core decisions to expensive intelligence.
In an intelligent era where everything tends to be free, the most expensive thing is actually the judgment of "knowing when to stop thinking and start acting."
Figuring out whether your one dollar of Token is to buy an infinite loop or to buy an irreplaceable moment of decision is not just a matter of survival.