OpenAI’s compute spending surges, on track to reach $600 billion by 2030

OpenAI’s current compute outlays already dwarf the $10‑million‑a‑year cloud budgets that funded its early GPT‑2 experiments, according to a PYMNTS.com analysis that projects the firm’s annualized spend will climb to roughly $60 billion by 2027 before accelerating to $120 billion in 2029, putting the cumulative total on a path to $600 billion by the end of the decade. The model assumes a compound‑annual growth rate of 45 percent, driven primarily by two forces: the scaling of large language model (LLM) training runs to ever‑larger parameter counts, and the expansion of inference workloads across an estimated 2 million enterprise customers that the company disclosed last year. PYMNTS notes that OpenAI’s “compute‑heavy” product mix—ChatGPT Plus subscriptions, API calls, and bespoke fine‑tuning services—requires both high‑throughput GPU clusters for training and massive inference fleets to meet latency targets for real‑time user interactions.

The report breaks down the projected spend by hardware class, showing that Nvidia’s H100 tensor cores will dominate OpenAI’s procurement strategy. PYMNTS estimates that by 2025 roughly 70 percent of the firm’s compute budget will be allocated to H100‑based servers, a share that rises to 85 percent by 2029 as the company retires older A100 and V100 generations. The analysis also highlights a parallel investment in custom ASICs, citing OpenAI’s recent partnership with a leading chipmaker to develop purpose‑built inference accelerators that promise a 2‑3× improvement in token‑per‑joule efficiency. If those ASICs achieve the projected performance gains, the PYMNTS model suggests they could shave up to $15 billion off the total 2030 spend, but even with that discount the trajectory remains firmly on a $600‑billion course.

OpenAI’s CFO Sarah Friar has publicly argued that the scale of this compute demand will outstrip private financing capacity, urging the U.S. government to “backstop chip financing” in a Wall Street Journal‑hosted event, as reported by The Information. While the PYMNTS piece does not quantify the fiscal gap, it infers that the firm’s capital expenditures will exceed $30 billion annually by 2028, a level that would likely require external financing mechanisms beyond traditional venture debt. The report therefore flags a potential policy lever: federal loan guarantees or direct subsidies for AI‑focused semiconductor fabs could lower the effective cost of capital for OpenAI’s hardware purchases, thereby moderating the steep spend curve.

From a technical standpoint, the compute surge is underpinned by OpenAI’s shift toward “sparse‑mixture‑of‑experts” (MoE) architectures, which enable models to activate only a subset of their parameters per token, reducing per‑inference compute while preserving overall model capacity. PYMNTS cites internal engineering estimates that MoE‑based models can deliver the same quality of output as dense 175‑billion‑parameter GPT‑3 models while using roughly 30 percent less GPU time per request. However, the training phase for MoE models is more resource‑intensive, requiring larger batch sizes and longer convergence periods, which explains the bulk of the projected spend growth. The report also points out that OpenAI’s data ingestion pipeline—now augmented by a newly released web crawler, as noted by The Information—feeds petabytes of fresh text into the training loop, further inflating the compute budget needed to keep models up‑to‑date.

Finally, the PYMNTS analysis warns that OpenAI’s compute trajectory could reshape the broader AI ecosystem. At a $600‑billion spend level, the firm would command roughly 15 percent of the global AI‑focused GPU market, according to the report’s market‑share calculations. This concentration could pressure suppliers to prioritize OpenAI’s custom ASIC programs over competing customers, potentially accelerating the industry’s move away from commodity GPUs toward proprietary accelerators. The report concludes that unless alternative financing models or more efficient hardware architectures emerge, OpenAI’s compute spend will continue its exponential climb, cementing its role as the single largest consumer of AI compute resources by the end of the decade.