Dharma AI, the Hugging Face-based lab known for its specialized OCR models, argues that utilization, not model quality or fleet size, has become the number that decides who wins in enterprise AI. The reasoning is a matter of accounting rather than opinion. A GPU bills by the calendar hour, covering depreciation, power draw, cooling, and financing costs, no matter whether a job is running on it, while it only earns money by the hour it actually computes. Two companies with identical GPU budgets can post sharply different returns. What separates them is the share of that hardware actually computing at a given moment, not the size of the budget itself.
The scale of the underlying commitments has grown even as the constraint shifted. In 2020, Microsoft assembled a dedicated supercomputing cluster for OpenAI: more than 10,000 GPUs alongside 285,000 CPU cores, ranked per contemporary reporting among the five largest computing systems then in existence. By 2026, according to Dharma AI, Anthropic was running concurrent multi-gigawatt compute deals across four hardware vendors, AMD, Amazon, Microsoft, and Google, struck within months of each other, while Meta signed a similarly sized agreement of its own. A buyer locking in deals with four separate suppliers inside the same few months, despite near-unlimited capital, still cannot secure enough compute from any single one.
Owning the hardware only converts the problem. It does not solve it. The same rack today might run a training job, a quantization pass, batched inference, low-latency real-time inference, embedding generation, fine-tuning, and a model evaluation suite, sometimes for the same organization and even the same model. No two of those workloads make the same demands on the chip, which is why a scheduler tuned for one will misallocate the rest.
Utilization dashboards can look strong even when the fit is wrong. Several queued jobs may simply be waiting on a GPU shaped for a different task that happens to be occupied elsewhere. Dharma AI borrows an aircraft analogy to frame the stakes, then undercuts it: a grounded plane can be sent to almost any other route, while a grounded GPU can serve only the jobs whose memory and latency needs it happens to fit.
The lab has a direct stake in half of its own argument. Dharma AI builds and sells specialized models, including DharmaOCR, and this piece frames specialization, alongside orchestration, as one of two levers needed to close the utilization gap. Nothing in the essay names a GPU-orchestration product Dharma AI sells, so the commercial interest runs through the specialization half of the case rather than the management layer itself.
AI Insiders covered an argument two days ago that compute could get ten times more expensive. If that holds, utilization stops being an optimization and becomes the whole margin.
Enterprises planning a GPU cluster this quarter should budget for an orchestration layer alongside the hardware order, not after it arrives. A rack of idle accelerators depreciates at the same rate as a busy one, and the invoice does not care which.
Dharma AI published this analysis, credited to Erick Lachmann, Gabriel Pimenta de Freitas Cardoso, and Gustavo Lucchetti, on the Hugging Face blog on April 17, 2026.
