The Vibe-to-Bankruptcy Pipeline
42% of enterprise AI projects were abandoned last year. The ones that survived cost more than human payroll.
For twenty years, software economics were simple: building the app was expensive, but scaling it was cheap. SaaS spread costs flat. Whether you had ten users or ten thousand, the marginal cost of software approached zero. Buyers learned to expect flat-rate scaling.
AI agents break this model completely.
Agents do not scale like software. They scale like human labor. Every action they take burns compute, energy and tokens. Your costs become strictly proportional to the work being done. And right now, the market is treating autonomous digital workers like a flat-rate SaaS subscription.
Which is why so many enterprise AI projects are driving straight off a cliff.
Vibe coding made AI demos genuinely cheap to build.
A working agent in a weekend.
Responses flowing, tasks completing, someone in the room saying we need to ship this.
Forty-eight hours, an API key and enough momentum to get a project approved.
The paradox is that the exact same toolchain that made starting cheap made scaling unexpectedly and brutally expensive. Not by accident. By design.
Jason Calacanis said it out loud on the All-In podcast last week. He described hitting $300 per day per agent on the Claude API. At 10 to 20% of productive capacity. That is roughly $100,000 per year per agent - before the system is doing anything close to full-load work.
Chamath Palihapitiya followed with the framing that hits harder for anyone managing a P&L:
You now have to assign your best developers a token budget, then decide whether their AI-assisted output justifies the cost of enabling them.
That is a new kind of management conversation. It used to be about software licenses. Now it is about metered inference.
We are seeing the fallout in the data. S&P Global Market Intelligence reported that 42% of enterprise AI initiatives were abandoned last year - up from a mere 17% in 2024. More than double. The demos worked. The pilots stalled. The projects were cancelled.
It is a failure of economics, not of technology.
The demo is the trap
Vibe coding does not break at the prototype stage. That is exactly why it is dangerous.
Demos are forgiving. Context windows are small. Tasks are narrow. The model does something impressive, the output looks clean and someone approves the next phase. The whole project gets funded on the strength of a forty-eight-hour prototype that ran on one developer’s laptop.
What nobody tells the executive committee is that the prototype was cheap because everything about it was minimal. One model. One context. One task at a time. No retries. No orchestration. No production reliability requirements.
Move to real automation and the architecture fundamentally changes. Agents do not complete a task once and stop. They retry when tools fail. They spawn subagents for parallel workstreams. They carry massive context payloads across sessions. They run in the background while your developer sleeps.
Anthropic's own documentation states it directly. Agent teams - multiple Claude Code sessions running in parallel - use approximately seven times more tokens than standard single-session work.
Seven times. Not 20% more. Seven times.
The vibe-coded demo you built in a weekend is now metered infrastructure running at 7x the cost you modeled when you made the case for the project.
This is where projects enter Pilot Purgatory. The term was popularized by McKinsey in 2018 for IoT startups that kept running pilots without ever shipping products.
I have actually lived in this specific purgatory.
In 2016, I launched an IoT pilot for a precision manufacturer across 50 machines on a single production line. At AWS, I advised Airbus teams on dragging shipment-tracking pilots out of the lab and into actual production and integration. Later, at a major retailer, I had to leverage a vendor’s investment program just to pivot a stalled IoT pilot into a live environment.
Scaling a pilot to a fully integrated production platform is hard work.
Now the term is back in 2026, but the mechanism is different. In 2018, pilots died because integration was hard. In 2026, the ROI was clear at the demo stage and completely wrong at the production stage. The math changed between the Friday pitch and the Monday bill.
The reality check in the boardroom
Developers think this is a technical hurdle. CFOs know it’s a margin killer.
If you think the conversation about token cost is just podcast chatter, you are ignoring the earnings calls.
Figma’s Q3 2025 earnings call is the perfect warning sign. CFO Praveer Melwani confirmed that subsidizing AI features dropped gross margins to 86%, but he framed this margin compression as a deliberate strategic investment rather than an unexpected failure. Figma is explicitly choosing to eat the inference cost today by delaying consumption limits to build workflow dependence tomorrow. It proves the structural hit to SaaS margins is real and the free ride for users expires quickly.
Deloitte sees the exact same trend, publishing a piece in the Wall Street Journal CIO section titled “AI Tokens: How to Navigate AI’s New Spend Dynamics,” noting that AI tokens are driving enterprise IT budgets up by 20% (full report in PDF).
The market response is predictable. Teams try to hide the cost under flat subscriptions. They buy the $200 per month Claude Max plan and assume they have capped their downside.
Subscriptions cap cost. They also kill observability.
When you hit the rate limit - and with any real agentic workload you absolutely will - the agent does not fail gracefully. It stalls. The work stops.
The token burn from developers running background processes 24/7 became so severe last year that Anthropic had to impose strict weekly usage caps, even on their highest $200-a-month tier. Within weeks, they were forced to roll out new enterprise administrative controls allowing IT teams to set hard organizational spend limits and monitor token usage per seat, just to stop developers from exhausting company quotas.
Paper Compute published the sharpest version of this argument last week. Today’s AI pricing functions as a subsidy designed to create workflow dependence. The correction does not arrive as a massive price increase. It arrives as friction. Rate limits, throttling and capacity constraints accumulate quietly until the workflow you built around the tool becomes a liability.
YOU BUILT A BUSINESS PROCESS ON BORROWED INFRASTRUCTURE.
That is not a criticism. It is what happened to cloud computing. To SaaS. To every infrastructure layer before this one.
The difference is that this layer is metered at the level of thought.
The system revealed
Let me put the pieces together:
Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027 due to rising costs and unclear business value. Industry veterans are already warning that 40% is a highly optimistic floor, given that standard digital transformation failure rates sit closer to 70%.
Open-source projects like ccusage are being built solely to track Claude Code spend across developer machines because official tooling lacks visibility.
New infrastructure is emerging solely to solve the token burn problem. In November 2025, Anthropic’s own engineering team published data showing that by forcing agents to write code and filter data in a sandbox before the payload hits the LLM, they could drop a 150,000-token enterprise workflow down to just 2,000 tokens. This is a 98.7% reduction.
This is the market building its own immune response.
Projects die in the pilot phase because no one modeled the burn rate.
The projects that survive cannibalize budgets meant for human headcount.
The entire ecosystem is now pivoting to patch the economics before the budgets run dry.
When token spend outpaces salaries, AI stops being a software license. It becomes metered infrastructure. You pay for it exactly the way you pay for cloud compute. You optimize it exactly the way you optimize digital ad spend. The companies that survive this transition are not writing better prompts. They are treating token allocation as a supply chain problem.
We saw this exact cycle with cloud infrastructure. Cheap pilots created workflow dependence. That dependence triggered massive bills. Those bills forced the creation of FinOps and internal chargebacks. The companies that won that era built their governance before the invoice arrived.
Agentic AI is on the exact same arc. It is just moving ten times faster.
The 90-day collapse points
If AI token cost is truly shifting from a developer problem to a boardroom governance problem, the market will adapt its business models.
Here is what will prove or disprove this within 90 days.
By May 31 2026 - A major enterprise SaaS vendor abandons flat-seat pricing for AI agents in favor of a hybrid consumption model.
The collapse point hits when traditional B2B software admits they are losing money on power users and forces the token bill back onto the customer.
By June 30 2026 - A Tier-1 cloud provider acquires an independent AI routing or context compression startup. If inference is the new cloud compute, hyperscalers must natively integrate token FinOps to protect their margins.
By June 30 2026 - Two Fortune 500 companies publicly mandate human-in-the-loop constraints for internal AI agents strictly as a financial control to cap autonomous retry loops.
This proves organizations are using operational friction as a budget governance tool.
By June 30 2026 - S&P Global or Gartner downgrades near-term enterprise AI ROI citing unmodeled agentic operating costs. The 42% abandonment rate was a backward look.
The test is when analyst firms formally bake token burn into forward-looking models.
If none of these fire, my thesis is wrong. The noise is just early-adopter friction and enterprise AI economics will stabilize quietly.
But the demo always works. The meter always starts.
Where are you in this pipeline?
Again, here’s my ask.
It takes me six to seven hours of deep research and synthesis to write one of these Signals. I do not outsource the thinking.
If you aren’t subscribed, subscribe. It takes two seconds. It costs nothing. It separates the signal from the noise.
If you have a voice, restack it. Let your network see the signal before the invoice arrives.
And if this resonated with you, if it gave you the language to name the problem, pledge your support.
It is the only way to support the weekly Signals and influence the next deep-dive Analysis.
Excellent article!
Vibe coding has hythnotized business leaders. Many of the AI failed projects you mentioned are probably not Vice coding related. Just companies being too early and believing in X/Y/Z SI or vendor. The problem IMHO is that everyone started to sell the AI dream back in 2024 when they could barely spell AI internally. The AI tech was immature. It's improving fast. Vibe coding is an atrocity for engineers, an opportunity for business users (with proper security guardrail).