The article opens with two data points — SpaceX‘s third-quarter report and DeepSeek‘s inference post — as evidence that inference businesses can be healthy at current prices and efficiencies, and that there is considerable upside if demand persists at higher prices per card-hour. It lays out a four-rung GPU revenue ladder from a theoretical maximum down to realized revenue. DeepSeek calculates a maximum of $38 per GPU card-hour for a GPU under optimal conditions: fully utilized for external customers, no internal work, 100% of output usable and billable, sold at full list price. Actual revenue may be lower because V3 is priced below R1 and because prices vary between peak and off-peak hours. The next rung is the B200 spot price of $8.77 per card-hour, what it costs to rent a card on the open market. SpaceX provides both a negotiated contract price of $5.27 and realized revenue of $2.44 per card-hour, with the latter including operations costs, training, and unusable or unbillable work.

On SpaceX‘s Q3 earnings call, CFO Bret Johnsen said, “At current pricing, new compute deployments are generating a payback of less than a year.” If that measures payback after tax on fully loaded costs, the article says the inference business can be a very good one. But the case rests on demand persisting and on efficiency outrunning price cuts.

Two forces push the other way. GPU rental prices doubled in six months, from $4.40 to $8.08 per card-hour, while the price of intelligence fell much faster. OpenAI cut Luna 80% in July and another 50% in September. A benchmark threshold that cost $0.55 eighteen months ago now clears at $0.0015, a 377-fold reduction. If tokens deflate that quickly while cards cost more, gross profit per card-hour depends entirely on throughput gains keeping pace.

The second force is inference moving off the datacenter entirely. AI companies are asking Qualcomm to run large models on the handset. Qualcomm CEO Cristiano Amon said on the Sources podcast that companies at the forefront of AI are telling Qualcomm that by 2028 they need the ability to run at least a 100-billion-parameter model in a phone. Inference on laptops and phones never sends tokens to the cloud.

Still, the cloud may not be big enough to support demand. New data centers are behind schedule, with GPUs idle on warehouse floors waiting for power. JPMorgan finds more than 60% of data center capacity planned for 2027 is not under construction, and another 7% is delayed. With demand continuing to drive record revenue growth for AI, deflationary forces may lose to too much demand. The article concludes that large-scale inference is healthy at scale, and token factories have more rungs to climb on the pricing ladder.