The Price of Intelligence Is Falling. The Bill Keeps Rising.
AI tokens are getting cheaper by the month – but that's not saving anyone money, and it's definitely not making OpenAI profitable
The cost of running AI has dropped dramatically. The price per token to answer a PhD-level science question as proficiently as GPT-4 has fallen by about 97% per year, according to EpochAI research. What cost $20 per million tokens two years ago now costs $0.07 from low-cost providers. Andreessen Horowitz calls this “LLMflation”—a 10x cost reduction every year.
And yet OpenAI is on track to lose $44 billion before reaching profitability in 2029. The company burned $9 billion in 2025 on $13 billion in revenue—a cash burn rate of roughly 70% of sales.
How can a product get dramatically cheaper while its maker bleeds cash? Understanding this paradox is essential for anyone building with or investing in AI.
The Token Trap
Here’s what the efficiency headlines miss: as AI models become more capable, they generate more tokens to complete tasks.
Standard models have become more verbose over time. The average number of output tokens for benchmark questions has doubled annually, according to EpochAI. “Reasoning” models—the ones that explain their approach step by step—use eight times more tokens than simpler models. And usage of reasoning models is rising about fivefold every year.
So while the price per token drops, the number of tokens required to do useful work rises. The efficiency gains and the capability gains roughly cancel out.
As one analyst at SemiAnalysis put it: “Generating responses remains costly because the models keep improving—and growing. So even as the price of tokens falls, better, more verbose models mean more tokens must be generated to complete a task.” He found it “hard to imagine” a future where the marginal cost of AI services falls close to zero.
The Competition Problem
Even if costs were stable, pricing pressure would squeeze margins. OpenAI charges developers about $120 per million output tokens for its most advanced model. DeepSeek, a Chinese rival, offers comparable performance at a fraction of the price—in some cases 95% cheaper.
Competition has made the AI API market look less like enterprise software and more like commodities. Stanford research found that achieving GPT-3.5-level performance became 280 times cheaper between late 2022 and late 2024, driven largely by new entrants undercutting incumbents.
OpenAI’s response has been to retreat upmarket. Its newest reasoning model, o1, costs the same per output token as GPT-3 did at launch—$60 per million. The company is betting that there’s a premium tier of customers willing to pay for the most capable models, even as the low end gets commoditized.
Whether that bet pays off depends on whether capability improvements can stay ahead of competitors—and whether customers value the difference enough to pay for it.
The Scale Paradox
OpenAI’s financial trajectory assumes scale will eventually solve the profitability problem. The company projects $200 billion in annual revenue by 2030. At that scale, even modest margins would generate significant profits.
But scale requires spending. OpenAI has committed to buying more than 26 gigawatts of datacenter capacity through the end of the decade at a cost exceeding $1 trillion. It signed a roughly $60 billion annual computing arrangement with Oracle, an $18 billion joint data center venture, and a $10 billion allocation for custom semiconductor development.
HSBC analysts estimate OpenAI faces a $207 billion funding shortfall through 2030, even accounting for projected revenue. The company’s cumulative free cash flow will still be negative by then, leaving a gap that must be filled through debt, equity, or more aggressive revenue generation.
The circular nature of AI industry investment makes this more concerning. Nvidia recently announced a $100 billion investment in OpenAI—shortly after OpenAI signed a $300 billion cloud computing contract with Oracle. Oracle is a major Nvidia customer. The money flows from Oracle to Nvidia, from Nvidia to OpenAI, and back to Oracle through cloud contracts. Critics note this resembles patterns from the dot-com bubble, when telecom equipment makers extended financing to customers to encourage equipment purchases.
What This Means for Businesses
If you’re building AI capabilities, the economics have several implications.
Don’t assume costs will keep falling. The headline trend of cheaper tokens masks offsetting factors. Budget for capability improvements driving up token consumption, not just price reductions driving down costs.
Provider stability matters. OpenAI’s path to profitability depends on assumptions that may not hold. The current model of burning billions isn’t sustainable indefinitely. Prices for API access are likely to increase, or service terms will change.
The “build vs. buy” calculation favors buying—with caveats. The cost of building foundational models makes it unfeasible for all but the largest tech giants. But buying means being beholden to the pricing and platform decisions of your chosen provider. Hedging across multiple providers adds complexity but reduces risk.
Watch for reasoning model costs. The shift toward models that “think” step by step dramatically increases token usage. A task that took 100 tokens on GPT-4 might take 800 on a reasoning model. Budget accordingly.
An Assessment
The price of raw AI capability is falling rapidly. That’s genuinely good news for developers and businesses building applications.
But the business of providing AI is not getting easier. The major providers are spending more than they’re earning, betting that scale and capability leadership will eventually generate returns. Whether that bet succeeds depends on factors that are genuinely uncertain: the pace of capability improvement, the intensity of competition, and the willingness of customers to pay premium prices.
The price of intelligence is falling. The cost of being in the intelligence business is not.