OpenAI confirmed roughly $2 billion in monthly revenue by mid-2026, an annualized run rate of about $25 billion. Internal documents reportedly project a $14 billion loss for the year anyway, nearly triple 2025’s loss. The company burned $3.7 billion in the first quarter alone, more than half that quarter’s $5.7 billion in revenue. Cumulative cash burn is forecast at roughly $115 billion through 2029, and OpenAI does not expect to break even until after 2030.
Anthropic, the other frontier lab most often mentioned in the same breath, is not telling the same story. Its run-rate revenue crossed $47 billion in May 2026, up from $9 billion at the end of 2025, on training costs running roughly four times lower than OpenAI’s. Anthropic expects its first-ever operating profit, about $559 million, in the second quarter of 2026, with full profitability targeted for 2028 or 2029.
Two companies, built on the same underlying technology, are telling two very different financial stories at the same moment. This is the seventh chapter in an argument this publication has been making about AI generally: that it is not a technological revolution but a regime of extraction, taking language, art, labor, and legitimacy from the many and converting it into power for the few. The first chapter traced that extraction across the whole economy. The second traced it through YouTube, the third through software engineers, the fourth through the attention economy, the fifth through pollution and water, and the sixth through the price of a laptop and an electric bill. This chapter is about the industry’s own balance sheet, and about who is positioned to absorb the loss if the money-losing half of it turns out to be the norm rather than the exception.
Start with what “losing money” actually means at OpenAI’s scale. The company spent $2.37 for every $1 of revenue in 2024, $1.60 for every $1 in 2025, and $1.22 for every $1 in the first quarter of 2026, an improving ratio that is still deeply negative. Leaked financials reported by Fortune put 2025 losses at $21 billion against $13 billion in revenue. Compute is the primary driver of that gap, and it is not the kind of cost that compresses quickly. None of this makes OpenAI unique. It makes OpenAI the highest-profile example of an industry-wide pattern in which the two best-known companies in the field are, for now, on opposite sides of the same ledger.
Some of what looks like demand for that compute is not fully independent demand. Analysts estimate more than $800 billion in circular financing arrangements across the industry: chipmakers and cloud providers invest in AI companies, which then spend a share of that money buying the investors’ own chips or cloud capacity back. OpenAI’s roughly $300 billion cloud deal with Oracle is one example; Oracle in turn spends billions of that on Nvidia chips, and Nvidia is itself a major OpenAI backer. Nvidia separately agreed to buy $6.3 billion of cloud services from CoreWeave, which rents out Nvidia’s own chips; CoreWeave gave OpenAI $350 million in equity ahead of its IPO and later expanded its own cloud deals with OpenAI to as much as $22.4 billion. None of this is illegal or even unusual for a capital-intensive industry building out fast. But it means some of the revenue and demand figures the industry points to as proof of real appetite for AI are, in part, the same dollars changing hands among a small number of firms more than once.
The pricing customers actually see is not a reliable guide to the industry’s real costs either. Inference for GPT-4-class models has fallen roughly 95 percent in two years, but that price is not really the cost of running the model. OpenAI, Google, Anthropic, and Meta are all pricing inference below cost to win market share, funded by venture capital and hyperscaler cross-subsidy rather than by the transaction itself. Multiple industry analyses expect 30 to 50 percent API price increases within the next 12 to 24 months as that subsidy becomes harder to sustain. Businesses building products on today’s token prices are building on a foundation their own vendors do not expect to hold.
That foundation is already buckling under a different pressure than the one this chapter has described so far. Per-token prices have fallen roughly 98 percent, yet total enterprise AI spending has roughly tripled over the same period, driven by volume rather than price: agentic workflows that trigger repeated model calls per task, retrieval pipelines that inflate context windows, and reasoning models that spend far more compute “thinking” before they answer. Per-developer token usage is reportedly up roughly 18.6 times in under a year, and the average enterprise AI budget has grown from about $1.2 million a year in 2024 to $7 million in 2026. Uber burned through its entire 2026 AI coding budget by April, and Microsoft revoked most employees’ access to Anthropic’s Claude Code in May 2026 after per-engineer token costs reached $500 to $2,000 a month. Bryan Catanzaro, an Nvidia vice president who oversees deep learning at one of the companies profiting most directly from this spending, put it to Axios without qualification: “For my team, the cost of compute far exceeds the cost of employees.” That is not a story about AI failing to work. It is a story about AI working exactly as sold and still costing more than the jobs it was supposed to make cheaper.
Even if prices do eventually rise the way the industry is counting on, competition may not let them rise far enough. By May 2026, Chinese open-weight models accounted for roughly 61 percent of all tokens consumed on OpenRouter, the largest neutral model router, and four of the five most-used models overall were Chinese; Meta’s Llama, the open-weight leader two years earlier, had fallen out of the rankings entirely. DeepSeek’s latest flagship model lists at roughly a twelfth the price of a comparable OpenAI model at similar benchmark performance, and Alibaba’s Qwen scores within single digits of the closed frontier on coding at a tenth to a thirtieth of the cost per token. Cost, not capability, is the argument these models are winning on. That complicates the whole premise this chapter has been building toward. OpenAI’s plan to reach breakeven sometime after 2030 assumes it will eventually be able to charge enough to cover what compute actually costs. If a rival lab can offer comparable performance at a twelfth the price, that day may never arrive, no matter how much power and chip capacity OpenAI manages to secure. A financial model that depends on eventually raising prices does not survive contact with a competitor that is simultaneously cutting them.
The growth these numbers assume also depends on power that may not arrive on schedule. Interconnection queues, the wait to actually connect a new facility to the grid, now run four to seven years in the country’s biggest data center markets, including Northern Virginia, Phoenix, and Dallas. A campus that joined Northern Virginia’s queue in May 2026 cannot realistically expect utility power before 2030 to 2033. Industry estimates suggest 30 to 50 percent of AI data centers scheduled to open in 2026 will be delayed or canceled, with only 5 gigawatts actually under construction against 16 gigawatts announced. Equipment lead times compound the delay: large power transformers now average 128 weeks to deliver and generator step-up transformers average 144 weeks, with switchgear sold out through 2028. Nationally, nearly 2,300 gigawatts of generation and storage capacity are stuck in U.S. interconnection queues, more than the country’s entire installed power capacity.
That delay is not a separate problem from the financial one. It is the same problem, compounded. A data center that cannot draw utility power until 2030 or later still requires land, construction, and equipment paid for now, capital committed years before it can generate a dollar of the revenue the industry’s own growth projections assume. Every year a facility sits waiting on an interconnection queue is a year of carrying cost added on top of losses that are already deeply negative, stretching the timeline for the breakeven OpenAI is not expecting until after 2030 in the first place. None of the financial projections in this chapter assume the power simply will not show up on time. Increasingly, on the timeline the spending plans require, it might not, and every year it doesn’t is a year the industry’s own math gets worse, not better.
Compute has its own version of the same bottleneck. The advanced packaging process required to assemble a modern AI accelerator, known as CoWoS, is now the binding constraint on chip supply, not the underlying silicon wafers themselves. TSMC’s CoWoS lines are sold out through 2027, with lead times running 52 to 78 weeks, against 2026 demand estimated near 1 million wafers, up from roughly 370,000 in 2024. Nvidia alone has locked in more than 70 percent of that packaging capacity, leaving AMD, Broadcom, and everyone else to split what remains. The memory that goes inside those chips is squeezed the same way: high bandwidth memory demand is growing 80 to 100 percent a year while supply grows only 50 to 60 percent, a gap not expected to close until 2028 or 2029 at current investment rates. Nvidia has already cut production of its consumer GeForce RTX 50-series GPUs by 30 to 40 percent in the first half of 2026 to redirect that scarce memory toward the AI accelerators data centers actually want. Samsung, SK Hynix, and Micron are investing more than $50 billion combined to expand HBM capacity, but a new memory fab takes 18 to 24 months to come online, an investment that cannot simply be accelerated by writing a bigger check.
That queue behaves exactly like the power queue, and compounds it. A GPU order placed today against packaging and memory capacity that does not exist until 2027 or later is capital committed now, against revenue that cannot start until the hardware physically arrives, arrives at a data center, and gets connected to power that may itself not be available until years after that. Money can buy a place in line. It cannot manufacture wafers, packaging capacity, memory fabs, or grid interconnections that do not yet exist, and every year that gap persists adds carrying cost to an industry already running on a schedule that does not reach breakeven until sometime after 2030.
Even where power and chips are both available, the community hosting the data center may simply refuse it. In the first quarter of 2026 alone, local backlash delayed or blocked at least 75 projects worth a cumulative $130 billion, with a separate estimate putting $64 billion in projects blocked or delayed by community opposition specifically. A Gallup poll found 71 percent of Americans oppose having a data center built near them, including 48 percent who oppose one strongly, a higher rate of strong opposition than Americans report toward a nearby nuclear plant. Water use is the single most cited concern, showing up in more than 40 percent of contested projects, the same resource strain this series has already documented in Memphis, Georgia, and Arizona, followed by energy consumption, electricity rate increases, and noise.
That opposition has started reaching state law. In June 2026, New York’s legislature passed the country’s first statewide data center moratorium, following an executive order from Governor Kathy Hochul barring new large-scale data centers of 50 megawatts or more for up to a year. More than 300 data-center-related bills were introduced nationally in early 2026, with roughly a dozen states weighing moratoriums of their own, and Senator Bernie Sanders introduced a federal AI Data Center Moratorium Act in March. By the end of June, 116 municipalities had passed their own local moratoriums, including a six-month pause in Lysander, New York, approved after more than 350 residents turned out to oppose a proposed 300-megawatt facility.
Unlike a supply chain delay, a moratorium does not resolve itself in a year or two of waiting. It can kill a project outright, or push it somewhere else entirely, and every relocation resets the clock on the power queue and the chip queue both. Every constraint in this chapter, financial, physical, and now political, points the same direction: the growth the industry’s valuations assume has to clear more obstacles, arriving later and later, than the money behind it was ever built to expect.
None of this has escaped people whose job is to price risk. Investor Michael Burry, who bet against subprime mortgages before the 2008 crash, said the current environment feels like “the last months of the 1999-2000 bubble,” citing Apollo chief economist Torsten Slok’s data that 87 percent of venture capital funding now goes to AI-related companies, with AI-linked borrowers accounting for nearly half of investment-grade bond issuance and roughly 38 percent of high-yield debt issuance, compared with internet companies making up under 40 percent of VC funding at the actual 1999 peak. Slok has separately warned that AI valuations could face a “painful repricing” if returns outside the technology sector take longer to materialize than markets currently expect, and flagged the industry’s growing focus on “token optimization,” using fewer tokens to do the same work, as an early sign that adoption may be slower than assumed. The Bank for International Settlements, the institution that serves as a central bank to the world’s central banks, used its 2026 annual report to warn about AI capital spending increasingly financed with debt, widening credit spreads, and the risk that circular investment could spread losses further than any one company’s balance sheet if a downturn hits.
Burry’s skepticism runs deeper than valuation multiples. He has separately accused the hyperscalers running this buildout of an accounting choice that flatters their own numbers: Meta, Amazon, Microsoft, Google, and Oracle depreciate their Nvidia GPUs over five to six years, Microsoft having stretched its own schedule from four years to six and Meta using five and a half, when Nvidia’s annual release cadence means a chip’s real competitive life runs closer to two or three years. By Burry’s estimate, that mismatch understates depreciation and overstates profits by roughly $176 billion between 2026 and 2028, with reported operating income at companies like Oracle and Meta potentially more than 20 percent above what the underlying economics support. If he is right, some of the profitability propping up confidence in this buildout is an accounting assumption, not a fact on the ground.
Slok’s warning about profit margins failing to expand outside the technology sector already has early, concrete evidence behind it. MIT’s widely cited 2025 research found that 95 percent of generative AI pilots at companies fail to deliver any measurable financial return, with researchers pointing to integration and workflow problems rather than model quality as the reason. The gap between claimed and real returns shows up even at companies making the loudest claims about them. JPMorgan Chase has said its AI tools deliver roughly 10 percent in developer productivity savings, yet the bank’s compensation expense grew about 6 percent in 2025 and headcount kept rising. Bank of America has cited 2,000 full-time positions avoided through AI, while its own compensation expense climbed $2.9 billion over the same period. None of this proves the tools do nothing. It does suggest that the productivity gains large customers are willing to publicize have not yet shown up in the one place that would make them real: their own expenses.
None of these warnings prove the industry is a bubble that is about to pop. Whether it is remains genuinely disputed among people paid to have an informed opinion about it, and by at least one measure the buildout has room left to run: AI capital spending currently equals about 0.8 percent of GDP, well under the roughly 1.5 percent of GDP that prior tech-investment booms reached at their peak. What is not disputed is the concentration this has already produced. The so-called Magnificent Seven, Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla, now make up roughly 34 percent of the S&P 500, matching the concentration of the top seven stocks at the peak of the dot-com bubble. JPMorgan estimates more than $6 trillion in funding will be needed through 2030 for AI data centers, power, and supply chain buildout, much of it financed with debt in an industry where, as of right now, one of its two most prominent companies loses $1.22 for every dollar it brings in.
If that concentration ever unwinds badly, the people most exposed to it did not choose to be. A draft U.S. Treasury Department report warned that an AI bubble could put millions of Americans’ retirement savings at risk in a serious downturn, because public markets are now dominated by the same handful of AI-linked companies whose swings have outsized influence on 401(k) balances. Most savers did not pick that exposure deliberately. Since the Pension Protection Act of 2006 encouraged employers to auto-enroll workers into retirement plans, the overwhelming majority of 401(k) participants default into target-date funds, and most of those investors keep their entire retirement account in a single such fund. Because those funds track the broader market, a standard target-date or index fund now carries 30 to 40 percent of its equity exposure in AI-linked stocks, a concentration bet nobody sat down and made on purpose.
So yes, say this plainly too. It is economic colonialism when the gains from an industry concentrate among founders, early investors, and a handful of chipmakers and cloud providers trading the same capital back and forth, while the downside, if the underlying economics do not hold, lands on people who were auto-enrolled into that exposure through a retirement plan default they never actively chose. It does not require the bubble question to be settled to see the shape of the bet. OpenAI is losing more than a dollar for every dollar it earns, on a schedule that does not reach breakeven until sometime after 2030, in an industry where credentialed economists at Apollo and the Bank for International Settlements are on the record warning about debt-financed concentration matching the dot-com peak. Anthropic’s leaner path shows the losses are not inevitable, which is exactly why they are a choice, not a law of nature. Whoever is right about whether this ends in a crash, the people who stand to lose the most from being wrong were never asked whether they wanted to place the bet at all.