On a Monday morning in May 2026, a CTO at a mid-sized company opened Anthropic’s billing panel and froze. The company had just switched to per-token billing. Their AI spend had jumped 7× overnight. His exact words: “We built a monster.”
This isn’t fiction. It’s a real scenario cited by Financial Times reporters Jamie John and others. Previously, the company paid $200 per user per month, and engineers called Claude without limits. Under per-token pricing, the same usage produced a dramatically larger bill. The CTO’s reaction was visceral: cut budgets, restrict access, reconsider whether every AI-generated line of code was actually worth the price.
He’s not alone. Over the past two months, from Fortune 200 enterprises to AI-first startups in Silicon Valley, CTOs and CFOs have suddenly started doing the same math: does AI’s output justify its bill?
The Subsidy Machine: 40× to 70× Burn
David Rosenthal (blogging as “dshr”) captures the AI platform business model in his June 23 piece AI’s Affordability Crisis as a “drug dealer algorithm” — the first taste is free, and the price hikes come once you’re hooked. The metaphor isn’t elegant, but the phenomenon it describes has data behind it.
SemiAnalysis ran an extreme test: how many tokens can a user burn under a $200/month subscription? The answer: Anthropic’s Claude can consume $8,000 worth; OpenAI’s ChatGPT can hit $14,000. That’s an implicit subsidy to enterprise customers of 40× and 70×, respectively.
The scale of subsidy can also be measured from another angle. OpenAI’s 2025 financials — disclosed by tech journalist Ed Zitron — show: $13.07 billion in revenue, $34 billion in total costs and expenses, an operating loss of $20.92 billion. A $41.55 billion non-cash loss came from the “fair value adjustment for the nonprofit-to-for-profit conversion,” but even stripping out non-cash items, the operating loss remains in the tens of billions.
An even more striking detail: OpenAI spent 44% of revenue ($5.73 billion) on sales and marketing — and at that spend level, enterprise adoption growth was still flattening.
This data supports two opposing interpretations. Pessimists say: your product can’t even gain traction when it’s practically free — what makes you think it’ll work at higher prices? Optimists — including some HN participants — argue that 44% marketing spend precisely indicates a market that still needs education; once the adoption curve crosses an inflection point, marketing as a percentage of revenue will naturally decline. Neither side has a definitive answer yet.
The Corporate Brake Pedal
HN user “burningChrome” provided a front-line perspective. He works at a Fortune 200 company that went through the standard AI adoption arc: three months of “Wild West” — every team freely using any LLM, some teams even cancelling multiple SaaS vendor contracts because they built their own AI tools and “thought the cost was zero.” Then the company signed enterprise agreements with Anthropic and Google. One month later, management saw that token consumption had blown past expectations and cut off access to Claude and Gemini entirely. Want access restored? Fill out multiple forms, pass multiple layers of approval, submit a solid business case — and before any of that, get in line behind thousands of other employees.
“The company is in damage-control mode. Someone saw the bill and decided to shut down the party.” His summary is concise and devastating.
This isn’t an isolated case. Multiple HN commenters described similar trajectories. One noted that IT departments started mass-emailing developers, educating them that “cheap models are good enough,” while imposing token or dollar caps on high-value model access. Another, who does client work for Fortune 100 companies, observed a common pattern: enterprises are giving developers a $500/month AI budget and requiring them to demonstrate productivity gains based on delivery output — not lines of code.
I won’t judge whether these practices are reasonable. But at the engineering level, one judgment is clear: when customer purchasing decisions shift from “let’s try it” to “ROI first,” pricing power is sliding from sellers to buyers.
The Other Side: Maybe AI Is Worth It
Before declaring “AI is too expensive,” you need a baseline. Several HN commenters offered compelling counter-arguments.
User “travisb” did a different kind of math: AI is the “ultimate contractor” — on-demand, zero idle time cost, zero hiring cycle, zero contract negotiation. A human engineer’s fully loaded cost (salary + benefits + office space + management overhead) in the US is roughly $95/hour. If AI can equivalently replace human output across a broad set of tasks, $200+/hour remains economically justified. “At that utilization level, AI vendors’ financials would look a lot better.”
User “qurren” was even more direct: “If an engineer costs $X in salary and AI helps them produce 3× the output, the company should happily pay up to $2X for AI.” Yet in reality, he observed the opposite — many companies start complaining when AI spend reaches 0.1X.
This asymmetric behavior suggests either: companies lack confidence in AI’s actual productivity contribution, or they’re fundamentally gaming the system — capturing AI’s productivity gains while hoping vendors keep burning cash on subsidies.
There’s also an important accounting clarification. HN user “raincole” pointed out that roughly $30 billion of OpenAI’s 2025’s $38.5 billion net loss came from the “one-time accounting treatment” of the nonprofit-to-profit conversion. Stripping that out, OpenAI’s core operating loss is far smaller than the headline number, and internal targets point toward profitability in 2026. This means dshr’s original citation of the $38.5 billion figure likely overstated the scale of ongoing losses.
Investor perspectives are also fragmenting. One HN user from wealth management observed that client conversations over recent months have shifted from “how do we get on the AI train” to “how do we preserve capital when AI crashes.” But another immediately questioned the source’s credibility — “Are you working in a wealth management office, or repeating someone else’s view?” — a question that itself exposes the blurry boundary between “narrative” and “facts” in current AI economics discussions.
Cracks: Technical or Financial?
A key data point in dshr’s article comes from the Financial Times and Panmure Liberum: under the most optimistic “zero cost” assumption — measuring only revenue returns against capital expenditure — the implied AI investment returns for the five major hyperscalers are: Microsoft -9.2%, Alphabet -15.7%, Amazon +7.2%, Meta -28.8%, Oracle -35.6%. Only Amazon is barely positive.
This data needs two layers of context. First, it assumes zero operating cost, significantly underestimating the actual depth of losses. Second, it measures “returns on sunk investment” against “current revenue” — if future revenue grows dramatically (whether from model capability breakthroughs or price increases), these numbers could be rewritten substantially. Which assumption holds depends on whether you believe the revenue curve can catch up to the investment curve’s steepness.
Will Lockett did a highly simplified calculation: assume the AI industry accumulates roughly $3 trillion in debt over the coming years, at 3% interest over 10 years — that’s $309 billion in annual debt service alone. Optimistically assuming AI achieves 10% profit margins, cost parity with human labor, and can perform most job functions — each replaced role contributes roughly $6,600 in annual profit to AI companies. At that rate, debt service alone requires replacing 46.8 million US jobs, approximately 27% of current US employment.
Two engineering-level corrections: first, human labor’s total employer cost includes not just salary but payroll taxes, health insurance, office space, etc. — per BLS data, benefits cost roughly 30.1% of employer cost, so the equivalent profit per role is roughly $9,500, reducing required replacements to ~32.5 million. Second, this calculation assumes AI can achieve human-equivalent capability — and a 2024 MIT study showed that in 77% of scenarios, using humans still outperforms AI. These two uncertainties point in opposite directions and don’t cancel each other out.
Two Possible Exits: Open Source and Price Cuts
Two variables emerged in the HN discussion that could weaken the crisis narrative.
The first is the impact of open-source models. “tacone” pointed out that the OpenAI/Anthropic duopoly naturally lacks price competition pressure, while Chinese models and open-source models are engaged in genuine price competition. Models like GLM 5.2 are approaching frontier model performance at extremely low cost. One user posed a simple question: why pay $8,000/month for Claude when you could spend one month’s fee on an AMD machine or Mac Mini and run an equivalent open-source model?
The blind spot in this logic is latency and throughput. As “wqaatwt” noted: cloud batch inference efficiency far exceeds single-machine local inference — beyond hardware cost, latency and throughput are equally critical. For latency-sensitive agent applications, local deployment isn’t necessarily economical.
The second is the possibility of proactive platform price cuts. dshr’s original piece cited Sam Altman saying costs have become a “huge issue” for customers and that OpenAI is considering “significant” price cuts to counter Anthropic’s enterprise market lead. Meanwhile, Anthropic announced in June a “pause” on its Claude Agent SDK’s per-token billing changes — hitting the brakes before the price hike took effect. But there’s a logical tension here: if price cuts are commercially viable, why did both companies wait until the eve of an IPO to consider them? If they aren’t viable, this looks more like a short-term concession to “maintain the growth narrative until the IPO closes.”
The Third Way: It’s Not About Cost
HN user “woeirua” offered a framework that bypasses the “technical cost” debate entirely: “This is fundamentally a financial viability question. Models themselves are getting cheaper extremely fast — by this time next year, Fable 5 will cost less than today’s Sonnet. That’s not the problem. The problem is that many companies will find they simply can’t extract ROI from AI. Faster code output doesn’t equal more profit. Most business ideas are just bad ideas — implementing bad ideas faster with AI doesn’t create profit growth.”
This perspective moves the debate entirely from the “technical side” to the “application side.” It suggests that even if inference costs dropped to zero, AI’s economic sustainability remains in question — because the bottleneck on value extraction is the quality of demand itself.
User “gexla“‘s confession reinforces this suspicion: “Every time I see the cost indicator in a tool and remember I might be building something useless, I realize — everyone is probably doing the same thing. Spending imaginary money building imaginary value. Then I open social media and see walls of AI-generated content, all talking about skills, systems, agents, and ‘Karpathy wiki systems’ to produce more useless things.”
This is existential unease. But it’s worth acknowledging that this sentiment may be survivorship bias — the people genuinely creating value may not be on HN discussing cost problems.
As for the future, the data pool is full of contradictory signals. Hyperscaler AI infrastructure investment for 2026 is projected at $725 billion, roughly 36% year-over-year growth. Meanwhile, enterprise budget controls have already kicked in, shifting from unlimited access to ROI-based allocation. Both trends can’t continue simultaneously — either the investment proves worthwhile and revenue catches up, or we face an abrupt price discovery.
Who you believe depends on how you answer one core question: when the subsidies stop, when enterprises shift from “fear of missing out” to “fear of losing money” — does the AI industry leave behind a revolutionary productivity tool, or a classic case of capital misallocation?
I can’t answer that question. But it’s the question everyone watching this space needs to keep asking.
This analysis is based on currently available public information and community discussions. If you have a different perspective or additional information, discussion is welcome.