DeepSeek calculates a $38 maximum per GPU card-hour under optimal conditions, establishing a theoretical ceiling for token factories that contrasts sharply with realized market economics. When we examine the mechanics of spot market pricing, the divergence between advertised hardware capabilities and actual cash flow generation becomes stark. The spot market operates on highly volatile short-term clearing prices that fluctuate based on regional grid loads, data center utilization rates, and sudden shifts in training versus inference workloads. Unlike long-term reserved instances or multi-year enterprise contracts that guarantee baseline utilization, spot capacity exposes operators to severe revenue compression when token demand dips or when competitors dump excess capacity at marginal cost. Contract structures thus dictate whether an operator secures a predictable margin or absorbs the full brunt of market volatility, transforming theoretical card-hour ceilings into highly unstable cash flows.
The Revenue Gap
The SpaceX example illustrates both the contract price of $5.27 and the total realized revenue factoring in operational costs, training, and unbillable workload. According to SpaceX CFO Bret Johnsen on the Q3 earnings call, at current pricing, new compute deployments pay for themselves in less than a year. Assuming these figures reflect fully loaded costs, the inference business demonstrates structural health at scale, yet actual cash flows depend heavily on contract structures rather than list prices. These contractual frameworks—ranging from pay-as-you-go spot agreements to multi-year committed capacity commitments—determine how effectively a provider can hedge against the relentless collapse of intelligence prices.
GPU spot prices have shown extreme volatility while the price of intelligence collapses across the board. OpenAI cut Luna prices by 80% in July and another 50% in September. A benchmark threshold that cost $0.55 eighteen months ago now clears at $0.0015, representing a 377-fold reduction. These opposing vectors mean gross profit per card hour relies entirely on throughput gains outpacing token deflation, forcing operators to constantly optimize their spot market bidding strategies and contract tiers just to maintain positive unit economics.
Infrastructure Bottlenecks
"Some of the companies at the forefront of AI are telling Qualcomm, by 2028, I need to make sure that I have the ability to run at least a hundred billion parameter model into a phone."
As Qualcomm CEO Cristiano Amon noted on the Sources podcast, this push toward running large models directly on handsets bypasses cloud inference entirely, shifting compute demands away from centralized data centers. At the same time, physical infrastructure struggles to keep pace with enterprise demand. According to JPMorgan's estimate, more than 60% of the data center capacity planned for 2027 is running behind schedule.
This severe mismatch between planned facilities and operational readiness creates profound operational friction for enterprise data centers. Hardware idle times—where expensive silicon sits uninstalled or unpowered on warehouse floors awaiting grid connections—directly erode the capital efficiency of data center operators. Power connection delays extend the period during which capital is tied up without generating revenue, forcing operators to factor these holding costs into their long-term contract pricing and spot market offerings. With physical hardware idling on warehouse floors awaiting power connections, whether deflationary pressures will eventually overwhelm surging enterprise demand remains an open question for operators and buyers alike.
Editorial analysis
How the story developed
In June 2026, Google agreed to pay SpaceX $920 million per month for access to AI compute, helping drive SpaceX's third-quarter revenue to $14.6 billion by October. Now DeepSeek highlights the unit economics behind such capacity, calculating an ideal ceiling of $38 in revenue per GPU card-hour that contrasts directly with volatile spot rates across regional power grids and data centers. The gap between DeepSeek's theoretical $38 per GPU-hour and volatile spot pricing highlights the tension between ideal token factory yields and real cash generation.
Sources: our stories of 13.06.2026, 21.10.2026