Why AI's Hardware Boom and Software Bust Are the Same Story

Two things happened in the AI industry this week. Apple sued OpenAI for stealing hardware trade secrets ^[1]. And a coalition of 17 news organizations asked a federal judge to sanction the company for destroying billions of ChatGPT conversation logs ^[2]. Neither case has anything to do with the price of chips. But both are symptoms of a split that has been widening for months — and that the market has already priced in. The AI industry has divided into two layers, and a single mechanism runs between them. The physical scarcity of advanced chips drives up the price of compute. That cost gets passed downstream through usage-based token pricing. And that makes the software layer economically unsustainable. The same shortage that guarantees hardware revenue is what makes software companies go broke. It is one story, and the connection runs in only one direction. On the hardware side, the numbers are extraordinary. Nvidia reported 85% revenue growth to $81.6 billion in its most recent quarter, and CEO Jensen Huang said demand has "gone parabolic" ^[3]. The hyperscalers — Amazon, Microsoft, Alphabet, Meta — committed roughly $700 to $740 billion in AI capital expenditure for 2026, a nearly 70% increase over 2025 ^[4][5]. The stock market has followed the money: in the first half of 2026, hardware and memory stocks dominated the Nasdaq-100, with Sandisk up 858%, Micron up 300%, and Intel up 280% ^[6][7]. The physical scarcity behind these numbers is real — AI memory chip shortages drove the first computer price increases since the 1980s, with prices rising more than 3% per month, a trend expected to persist through 2027 ^[8]. On the software side, the same scarcity shows up as a cost crisis. OpenAI is reportedly projected to lose approximately $38.5 billion in 2026 on roughly $13 billion in revenue, with losses expected to continue through at least 2029 ^[9]. Baidu's first-quarter results illustrate the pattern in miniature: AI revenue surged 49% and overtook legacy search as the company's primary revenue driver, but net profit plunged 55% as the cost of delivering that AI revenue consumed the gains ^[10]. The AI business is becoming the main event while simultaneously destroying profitability. The mechanism that connects these two pictures is token pricing. As OpenAI and Microsoft shifted from flat subscription fees to usage-based billing, enterprises hit a wall. PNC CEO Bill Demchak made the dynamic explicit.

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

Uber exhausted its full-year AI coding budget by April ^[11]. Major enterprises deploying Anthropic's Claude face monthly bills exceeding $500 million ^[12]. Multiple enterprises are now scaling back AI spending as operating costs exceed expectations and return on investment remains unclear ^[13]. The market was pricing software companies as worth less than the chips they run on, even as their revenue kept growing: Intuit fell 60%, Adobe dropped 37%, and Workday declined 37% in the first half of 2026, making them the worst performers in the Nasdaq-100 ^[6]. The software layer faces a second problem the hardware layer does not: a legal siege on multiple fronts. Nearly 400 local and regional newspapers sued OpenAI and Microsoft in June for systematically scraping copyrighted articles, including paywalled content, and stripping bylines and copyright notices ^[14]. Sam Altman had previously acknowledged that training today's leading AI models would be impossible without using copyrighted materials ^[14]. On July 9, the New York Times and 16 other news organizations sought sanctions against OpenAI for what they called a "campaign of deception" — deleting billions of ChatGPT conversation logs in violation of preservation orders and falsely claiming for two years it lacked the technical ability to search training datasets for copyrighted content ^[2]. On July 10, Apple sued OpenAI for theft of hardware trade secrets, alleging that OpenAI's Chief Hardware Officer pressured job candidates to bring physical Apple prototypes to interviews and that a former engineer exploited a security bug to download over a thousand pages of confidential files ^[1]. And in late June, Anthropic accused Alibaba of the largest known AI distillation attack — 25,000 fraudulent accounts generating 28.8 million exchanges with Claude to siphon software engineering and agentic reasoning capabilities ^[15]. The company cast the attack as a national-security problem.

Distillation attacks turn hundreds of billions of dollars in American investment and [research and development] into a massive subsidy for our geopolitical competitors — Anthropic

None of these lawsuits target the chipmakers. The hardware layer sells physical goods to paying customers; the software layer sells access to models trained on data whose ownership is now being litigated in courtrooms across the country. The hyperscalers are not passive victims of this dynamic. Amazon, Microsoft, and Alphabet are all building custom AI chips — Graviton and Trainium at Amazon, Maia 200 at Microsoft, eighth-generation TPUs at Alphabet — in an attempt to break Nvidia's pricing power ^[3]. TrendForce projects custom chip shipments growing 44.6% in 2026, compared to 16.1% for GPUs ^[3]. But Nvidia still reported 85% revenue growth, and Huang said demand has gone parabolic. The cost pressure on the software layer persists even as the hardware diversification effort accelerates. The "picks and shovels" framing that investors adopted during the AI boom is more literally apt than they realized. In a gold rush, the shovel-sellers get rich whether or not the miners strike gold. But the AI industry has added a twist: the scarcity that enriches the shovel-sellers is the same scarcity that makes the miners go broke, and the miners' claim to the gold itself is now being fought in court. The two layers cannot converge until chip supply catches demand. That is not expected before 2028 at the earliest ^[8].