OpenAI's AI Gambit: Custom Chips to Curb $150B Compute Hunger

OpenAI has revealed that its projected spending could reach $115 billion by 2029, a significant increase from earlier estimates, according to a report by The Information. The organization, which has long operated in the high-cost space of generative AI, is now anticipating an additional $80 billion in expenditures. This shift in financial outlook reflects growing challenges in managing the astronomical costs of large language model development and deployment. OpenAI previously estimated spending would rise to over $8 billion in 2024, an increase of about $1.5 billion from its earlier forecasts [3].

A major contributing factor to this financial strain is the company’s heavy reliance on expensive compute resources for both training and inference operations. According to internal data cited in the report, OpenAI has already spent $9 billion in 2024 to lose $5 billion and is projected to lose upwards of $8 billion in 2025. The firm expects to spend over $150 billion on computing costs alone from 2025 through 2030, underscoring the unsustainable economics of current generative AI practices [1]. These figures highlight a systemic issue across the AI industry, where even the providers of core model technologies are struggling to maintain profitability.

To mitigate these costs, OpenAI is reportedly developing its own data center server chips and infrastructure. This strategic move is intended to reduce dependency on external cloud providers such as Microsoft Azure and to better control server rental expenses. The company has already taken steps toward this goal by commissioning a $10 billion order from Broadcom for custom-designed AI accelerators. These chips, tailored for specific workloads, are expected to be delivered in Q3 2026 and will be integrated into large-scale AI rack platforms [2]. The shift to in-house silicon design is a critical component of OpenAI’s long-term plan to optimize its inference workloads and reduce financial exposure to third-party vendors.

The broader AI industry faces similar challenges, with many companies sending 100% or more of their revenue directly to cloud providers or model developers. For example, Perplexity spent 164% of its 2024 revenue on compute from OpenAI, Anthropic, and AWS. OpenAI’s own inference costs alone accounted for 50% of its revenue, and when combined with training expenses, that figure rose to 75%. These numbers illustrate the structural inefficiencies of the current AI business model, where companies essentially fund the expansion of competitors or infrastructure providers rather than achieving sustainable growth [1].

Despite the rising costs, OpenAI and other model developers continue to pursue aggressive innovation strategies, including the development of increasingly complex models. The company’s CEO, Sam Altman, has previously suggested that inference is becoming “profitable” when excluding indirect costs such as training and R&D. However, critics argue that such metrics obscure the broader financial realities, as OpenAI burned at least $3 billion in cash on salaries in 2024 alone [1]. The company’s ability to secure large-scale hardware commitments—like the Broadcom deal—may offer a temporary reprieve, but the underlying issue of unprofitable AI operations remains unresolved.

The financial trajectory of OpenAI highlights a broader industry-wide dilemma. While the firm is investing heavily in custom infrastructure and silicon, the fundamental economics of AI deployment remain fragile. The increasing scale of compute requirements, combined with the limited ability to pass these costs onto users, suggests that profitability in generative AI is still elusive. OpenAI’s spending projections and hardware investments may indicate a commitment to long-term technological leadership, but they also underscore the immense financial risk associated with the current AI landscape [2].

Source: [1] Why Everybody Is Losing Money On AI (https://www.wheresyoured.at/why-everybody-is-losing-money-on-ai/) [2] OpenAI widely thought to be Broadcom's mystery $10 ... (https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-widely-thought-to-be-broadcoms-mystery-usd10-billion-custom-ai-processor-customer-order-could-be-for-millions-of-ai-processors) [3] OpenAI says spending to rise to $115 billion through 2029 (https://fortune.com/2025/09/06/openai-spending-outlook-115-billion-through-2029-data-center-server-chips/)