The CEO of the world’s most important AI company just admitted something that would have been unthinkable six months ago: costs are out of control, and nobody saw it coming.
At OpenAI’s “Intelligence at Work” enterprise event, Sam Altman revealed that AI token costs have gone from a non-issue to “a huge issue” in the span of a single quarter. His exact words capture the whiplash perfectly: “At the beginning of 2026, the issue never came up. People were totally happy with the amount they were spending. Now, all of a sudden, it’s a huge issue.”
This isn’t just another tech CEO acknowledging growing pains. It’s the signal that the entire AI industry has entered a new phase, one where inference economics, not model capability, will determine who survives.
The $1.3 Million Wake-Up Call
To understand why Altman is sounding the alarm, look at what’s happening at the extremes. OpenAI’s top internal token spender now consumes 100 billion tokens per month, a million-fold increase from the 100,000 tokens that made someone the world leader just six and a half years ago. That internal champion isn’t even the global leader anymore, Altman admitted, to his personal “embarrassment”, that an external customer is burning through even more.
Then there’s Peter Steinberger, creator of OpenClaw, whose team spent $1.3 million on OpenAI API tokens in a single month, 603 billion tokens in 30 days. The New York Times reported an OpenAI employee spending 210 billion tokens in a single week.
These aren’t anomalies. They’re the leading edge of a wave that’s about to crash over every enterprise deploying AI at scale.
Why Costs Exploded Overnight
The most insightful take on Altman’s admission came from the developer community itself. As one observer on a popular forum noted: “Per token prices didn’t just jump overnight. What changed is that people went from chatting (a few thousand tokens a session) to running agents that can loop for hours and burn millions. It is the usage pattern that spiked, not the bill.”
This is the core insight. The agent era hit the invoice.
A standard chatbot interaction uses perhaps 1,000 tokens per query. An agentic task, one that plans, calls tools, retrieves results, updates context, and iterates, consumes 96,000 tokens on average before generating a final answer, according to SemiAnalysis data. That’s more text than The Great Gatsby.
Goldman Sachs Research projects total token consumption will multiply 24 times between 2026 and 2030, reaching 120 quadrillion tokens per month. When per-token cost falls 50% a year but volume rises 10x a year, the bill grows regardless of efficiency gains.
The Enterprise Reality: $9-19M/Year for Middle of the Pack
The numbers for a mid-sized enterprise deploying AI in 2026 paint an uncomfortable picture. According to analysis from FourWeekMBA, a company with 5,000 employees running 10 AI use cases and 50 agents faces a cost stack that didn’t exist two years ago:
| Cost Category | Annual Estimate |
|---|---|
| Inference compute | $1-3M |
| API seat licenses ($30/user/month) | $1.8M |
| Cloud AI commitments | $2-5M |
| Custom deployment (professional services) | $2-5M |
| Internal AI team (10 engineers) | $2-3M |
| Total | $9-19M annually |
For a company with $500M in revenue, that’s 2-4% of topline spent exclusively on AI infrastructure. To make that pencil, AI needs to either reduce headcount costs by more than $19M (politically explosive), increase revenue by more than $19M (nearly impossible to attribute), or improve decision quality in ways that somehow justify the spend (unmeasurable).
The problem has a name: the token tax. It’s the structural levy any company pays for building on inference it doesn’t own.
The Gross Margin Trap That’s Eating AI Companies
Traditional SaaS was a miracle business for a reason: once the code was written, serving the next customer cost nearly zero. Gross margins of 75-85% followed automatically.
AI inference breaks that premise at the foundation. Every user interaction triggers a model call. For agentic products where a single request fans out into reasoning steps, tool calls, and retries, the cost of one session multiplies.
ICONIQ Capital’s January 2026 survey of roughly 300 software executives found AI-native product gross margins projected at 52% in 2026, up from 41% in 2024, but still a staggering 23-33 percentage points below the 75-85% that mature SaaS businesses routinely achieve. For early-stage AI companies that haven’t optimized their model stack, margins can be as low as 25%.
Inference alone consumes roughly 23% of revenue at scaling-stage AI companies. For every $1 million in AI product revenue, roughly $230,000 exits as inference cost before a single engineer gets paid.
The Cursor Case Study: How to Escape the Trap
Cursor, the AI code editor built by Anysphere, became the fastest B2B software product ever to reach $1 billion in annualized revenue, surpassing Slack, Zoom, and Snowflake. By February 2026, it had crossed $2 billion.
The cost side was brutal. For most of 2024 and into 2025, Cursor’s cost of goods sold was dominated by inference fees paid to Anthropic and OpenAI. Every heavy user represented a direct pass-through cost that exceeded what the company was charging. TechCrunch reported that Cursor operated at negative gross margins until recently, meaning it cost more to run the product than the company could collect.
Cursor’s escape is the most important strategic lesson of 2026. In November 2025, the company shipped Composer, its first in-house inference model optimized for code generation. Before Composer, every query routed to a third-party model and every token cost flowed out of Cursor’s gross margin. By moving inference from rented to owned, Cursor reached “slight gross margin profitability” on large enterprise sales.
The structural lesson is unambiguous: for any vertical AI company, the real margin inflection doesn’t come from pricing adjustments. It comes from moving inference from something you buy to something you produce.
Who’s Winning the Cost War
The divergence in pricing strategies has become impossible to ignore. DeepSeek made its 75% price cut permanent. As of May 22, 2026, generating one million words of output from DeepSeek’s V4-Pro costs $0.86. The same million tokens from Anthropic’s Claude Opus 4.7 costs $25. From OpenAI’s GPT-5.5, about $30.
Here’s the part that should make you stop scrolling: on SWE-bench Verified (the standard benchmark for real-world software engineering), V4-Pro scores 80.6. Claude Opus 4.7 scores 80.8. Two-tenths of a point apart. Nearly identical capabilities at roughly 28x price difference.
MIT Sloan’s analysis of OpenRouter inference data found that open models average about 90% of closed-model performance, usually closing the gap within roughly 13 weeks of a closed model’s release. The catch: open models cost 87% less to run, averaging $0.23 per million tokens against $1.86 for closed models.
For most workloads, teams are paying a large premium for capability they could match with open weights.
The ‘Tokenmaxxing’ Backlash
The meme Altman referenced, “My company spent my entire 2026 budget in Q1, can you make this more efficient?”, has real teeth. Companies that created internal AI leaderboards to encourage token consumption are now dismantling them.
Meta and Amazon have shut down their internal token leaderboards. Uber set a hard cap of $1,500 per month per employee for AI tools. Microsoft began canceling most internal Claude Code licenses, telling developers in its Experiences and Devices division their access would end by June 30 after compute costs exceeded the cost of the human employees the tools were supposed to augment.
The engineering analytics firm Faros AI found that “code churn”, lines of code deleted versus added, increased by more than 800% under high AI adoption. More tokens burned, more code produced, more code deleted, more burnout.
This is the hidden efficiency paradox: AI tools aren’t reducing work, they’re intensifying it while creating a new cost line that didn’t exist before. The tension between the productivity trap of AI and the burnout it creates is becoming impossible to ignore.
Three Escape Routes
Three approaches are emerging for enterprises trying to solve the inference economics problem:
Vertical AI companies that build domain-specific models requiring less compute. A legal AI running on a 7B parameter model costs 100x less than routing everything through GPT-5.5.
Platform consolidators (Microsoft, Salesforce) that bundle AI into existing subscriptions, amortizing the cost across products the enterprise already pays for. This is the pressure behind Microsoft’s Project Solara AI, a chip-to-cloud platform designed for “agent-first” enterprise devices.
Open-source deployers that run Llama, Mistral, or Qwen on their own infrastructure, avoiding per-token API costs entirely. The trade-off: more engineering effort but dramatically lower marginal costs at scale.
The same ownership dynamic applies to infrastructure. As Telnyx CEO David Casem put it: “If you don’t own the GPU stack and the network underneath it, your unit economics on inference will eat you alive.”
What This Means for the next 12 Months
The AI economy’s next chapter isn’t about who builds the best model. It’s about who makes AI cheap enough that the ROI math works for every company, not just the ones with unlimited budgets.
Nvidia’s Vera Rubin promises 10x lower inference costs. If that materializes in 2027, the compute portion of the enterprise cost stack could drop from $1-3M to $100-300K, fundamentally changing the equation. But other cost lines, licenses, cloud commitments, deployment services, aren’t falling. They’re rising.
The financial trajectory behind the cost crisis for OpenAI itself is bleak. Analysis suggests the company could hit zero cash by mid-2027 unless something fundamental changes.
Meanwhile, the costs of building AI-native products go beyond inference. The real-world infrastructure costs for something as common as a RAG pipeline reveal that the demo is easy, but production is a nightmare of hidden expenses.
The great AI cost panic of 2026 isn’t a sign that the industry is collapsing. It’s a sign that the industry is maturing. The era of unlimited token budgets and uncritical adoption is over. The era of inference economics has begun.
The companies that solve this problem, making AI economically scalable, not just technically capable, will capture the next phase of enterprise spend. Everyone else will be left with a meme and a bill they can’t justify.
