Part 3 of 6 in a series on the AI gold rush, where one company sells the shovels at extraordinary profits while everyone else bleeds cash.
By the High Country Advocate Investigative Team
During California’s 1849 Gold Rush, most prospectors went broke. The people who sold picks, shovels, and supplies became the millionaires. Samuel Brannan, often cited as California’s first millionaire, made his fortune not by mining gold but by selling metal pans he’d bought for pennies and marked up dozens of times over.
Today’s AI infrastructure boom follows the same pattern. Most companies building AI applications are losing staggering amounts of money, while one dominant supplier—Nvidia—prints cash at margins that look more like software than hardware. The easiest way to see who is actually winning the AI boom is to follow the money.
Nvidia’s extraordinary profits
Nvidia’s fiscal 2025 results show what happens when one company controls the picks and shovels of the AI era. The company reported $130.5 billion in revenue for the year, up 114% from the prior year, with $39.3 billion in revenue in the fourth quarter alone, a 78% year-over-year jump. Data center revenue reached $35.6 billion in Q4, up 93%, and quarterly net income hit $19.3 billion—more than double the prior year.
The margins tell an even sharper story. Nvidia has been running gross margins between 70% and 78%, extraordinary for a hardware manufacturer in an industry where 30–50% is typical. Intel sits around the low 40s and AMD in the mid-40s, while software firms often enjoy 80–90% margins because duplicating code is essentially free. Nvidia, by contrast, fabricates physical chips—silicon wafers etched with billions of transistors, mounted in complex packaging, and shipped as power-hungry GPUs that demand advanced cooling and data center infrastructure.
Nvidia can do this because it dominates the AI GPU market. The company controls roughly 86% of AI accelerator share, and every major AI player depends on its hardware. H100 GPUs reportedly sell for $25,000 to $40,000 each, while the newer Blackwell GB200 line commands even higher prices amid chronic supply constraints. CEO Jensen Huang has told investors that demand for Blackwell “is expected to exceed supply for several quarters,” implying continuing pricing power.
That dominance has translated into an explosion in market value. Nvidia’s market capitalization climbed from about $1.2 trillion at the start of 2024 to roughly $3.28 trillion by year’s end, making it the world’s second most valuable company behind Apple. For fiscal 2026’s first quarter, Nvidia is guiding to $43 billion in revenue with gross margins still above 70%, and the company continues to project supply constraints rather than oversupply.
The main competitive threat comes from the customers themselves. Google has spent a decade developing its TPU line, Amazon has rolled out Trainium and Inferentia, Microsoft has introduced its Maia chips, and Meta is designing custom silicon. JPMorgan projects that custom chips could capture 45% of the AI accelerator market by 2028, up from 37% in 2024, but the overall market is growing so fast that Nvidia can lose some share and still grow revenue at extraordinary rates. In this boom, a slightly smaller slice of a rapidly expanding pie can still feed exponential earnings growth.
AI companies’ spectacular losses
While Nvidia mints money, the companies building on top of its hardware often look like financial bonfires. OpenAI, the most visible AI application company, lost an estimated $5 billion in 2024 on revenue of about $3.7–4 billion. Total operating costs ran to $9 billion, meaning the company was spending roughly $2.25 for every dollar it brought in, even as monthly revenue grew to around $300 million by August 2024—a 17-fold jump in about a year and a half.
The first half of 2025 suggests the pattern is accelerating rather than stabilizing. OpenAI’s revenue for those six months reached roughly $4.3 billion, already surpassing all of 2024 by about 16%, but cash burn hit roughly $2.5 billion. Research and development spending alone is reported at $6.7 billion for that half-year, up sharply from about $2.5 billion for all of 2024, highlighting how much money is being poured into the race for more capable models.
Internal projections reportedly show cumulative losses of about $44 billion for OpenAI between 2023 and 2028, with a possible path to profitability in 2029—though even that is framed as uncertain. The company’s single largest cost category is compute: the rent paid to Microsoft for running enormous AI workloads on Azure. On top of that, Microsoft is estimated to receive around 20% of OpenAI’s revenue under the terms of their partnership, meaning OpenAI pays both to use Microsoft’s infrastructure and then again on the top line.
Inference costs—the compute required to serve responses to users—are particularly punishing. For just the first nine months of 2025, the cost to run ChatGPT for roughly 700 million weekly users is estimated at $8.65 billion. By some analyses, these inference costs may exceed total revenue, especially when usage by “power users” is factored in. At $200 a month, ChatGPT Pro may lose money on heavy users, while the $20 consumer tier can be profitable, but only if usage stays within certain boundaries.
Anthropic, which develops Claude, follows a similar financial trajectory. The company is reported to have lost more than $5 billion in 2024 on revenue of roughly $850 million to $1 billion. Development spending alone is estimated at $2.7 billion for that year, and the company has trumpeted an increase in annualized recurring revenue from about $1 billion in January 2025 to a projected $9 billion by year-end. But ARR is a snapshot—current monthly revenue multiplied by twelve—not actual annual revenue, and realistic estimates for 2025 put Anthropic’s realized revenue in the $2–3 billion range with break-even not expected until 2028.
Claude’s user base remains a fraction of ChatGPT’s. Estimates suggest roughly 16–18 million monthly users for Claude—on the order of 2% of ChatGPT’s traffic. Both companies have found large, loyal audiences, but both are burning billions of dollars a year to serve them, propped up by investors willing to fund losses on the hope of future dominance.
OpenAI alone has raised around $57.9 billion across 11 funding rounds and reached a private valuation of roughly $300 billion. It has also reportedly committed to about $1.4 trillion in long-term infrastructure deals with Nvidia, Oracle, Broadcom, and other vendors—a number so large it reflects not today’s business but the assumed future scale of AI adoption. With annual losses above $5 billion, the math only works if revenue can grow from roughly $20–30 billion for the broader AI application space in 2025 to something on the order of $2 trillion by 2030—a 100-fold increase in five years while most projects still struggle to show measurable returns.
Concerns about the sustainability of this model have already surfaced in public remarks. On November 6, 2025, OpenAI CFO Sarah Friar referenced the idea of a federal “backstop” or “guarantee” for AI. Trump administration AI Czar David Sacks responded publicly that there would be no federal bailout for AI companies, prompting Friar to walk back her comments and CEO Sam Altman to insist the company neither has nor wants government guarantees. The fact that a backstop was mentioned at all offers a glimpse of how precarious insiders may view the current financial structure.
The supporting winners
Nvidia is not the only beneficiary of this imbalance. Microsoft profits on multiple fronts: it is reported to take around 20% of OpenAI’s revenue and to earn high-margin income on the cloud infrastructure OpenAI and others rent from Azure. Because Microsoft already owns the data centers and networks, AI customers function as premium tenants paying a mark‑up on infrastructure that largely exists regardless of AI workloads.
Amazon, Google, and other cloud providers have adopted similar models. They sell access to compute capacity originally built for search, e‑commerce, and enterprise workloads, then layer AI-specific services on top. The incremental cost of running AI workloads on this base is far lower than the prices charged, which means healthy margins—if not at Nvidia’s level, still far above what the AI application companies themselves enjoy.
Further down the stack, the semiconductor supply chain has its own set of winners. TSMC manufactures Nvidia’s advanced chips, ASML supplies the cutting-edge lithography tools, Broadcom provides networking silicon for AI data centers, and Arm licenses critical chip architectures. These firms get paid when chips are designed, manufactured, and shipped; their profits do not depend on whether AI apps ever reach profitability.
The financing structures around the AI buildout add another layer of concern. Meta, for example, has disclosed roughly $27 billion in off‑balance‑sheet arrangements tied to infrastructure, illustrating how costs can be hidden in complex contracts rather than clearly stated on balance sheets. In some cases, suppliers invest in their own customers—Nvidia investing $100 million in OpenAI, for instance—only to have those customers turn around and spend that money back on Nvidia’s GPUs, inflating valuations and making investments appear self‑justifying.
Not everyone is convinced this can continue without consequence. Investor Michael Burry, best known for his early bet against subprime mortgages before the 2008 financial crisis, disclosed about $1.1 billion in put options against AI-related companies in November 2025. That includes an estimated $187 million bet against Nvidia and roughly $912 million against Palantir, along with a thesis that AI chips are being used up in 2–3 years while companies depreciate them over 5–6 years, inflating reported profits. His analysis suggests roughly $176 billion in overstated earnings across major AI and cloud players through 2028 and likens some of the accounting practices to what was seen at Enron.
The money flow
Taken together, the flow of money through AI infrastructure in 2025 forms a clear, top‑heavy chain. Big Tech has earmarked an estimated $405 billion in capital expenditures this year, much of it directed toward AI data centers and related infrastructure. A large portion of that spending goes straight to Nvidia, which converts it into 70–78% gross margins and record profits. Another chunk flows to cloud giants like Microsoft, Amazon, and Google, which mark up access to data centers and sell high‑margin AI services on top.
Beyond that, money spreads through the broader semiconductor ecosystem—to TSMC, ASML, Broadcom, Arm, and other suppliers that profit as long as chips are ordered and delivered, regardless of whether end‑user applications ever make money. Only a fraction of the capital ends up in AI application companies like OpenAI and Anthropic, and what does arrive is quickly spent on compute, R&D, and payroll rather than profits.
At the very end of this chain sit consumers and ratepayers. Data centers require enormous amounts of electricity and trigger costly upgrades to transmission and distribution systems. Those costs rarely show up as a line item labeled “AI,” but instead appear inside capacity charges and supply fees that regulators approve in routine utility rate cases with little public understanding that the AI boom is part of the reason bills are rising.
The companies selling the infrastructure—chips, cloud capacity, and fabrication tools—get paid whether or not AI applications ever justify their spending. Nvidia has already booked $130.5 billion in 2025 revenue, cloud providers have already collected their fees, and manufacturers have already shipped their components. If every major AI application company failed tomorrow, most of the infrastructure providers would keep their past profits, while investors and, indirectly, the public would be left holding the bag.
The application layer has no such guarantee. Investors eventually need profits, not just growth curves and user counts, and that pressure is what drives projections that AI revenues must grow from tens of billions today to trillions of dollars within a few years. In this AI gold rush, the pattern from 1849 is repeating: the miners may or may not strike pay dirt, but the arms dealers are already rich.
Article 4 will examine who ultimately pays for this buildout—and how rising electricity bills in places like rural Colorado became collateral damage in the AI infrastructure boom.
Support Independent Local Journalism — High Country Advocate was created as a real alternative to regional media that too often silences dissenting voices while taking sides in the political struggle. Producing in-depth, unflinching reporting like this series and others is expensive: servers, editing, research time, and legal review all add up quickly. If these articles have informed you or given you new perspective, please consider supporting HCA with a paid subscription — every subscriber helps keep this reporting strong and independent. 