DeepSeek Ledger: Demand vs Price in AI Ramps

Before you underwrite an AI raise, you need to know whether the run rate is demand or a price reset. DeepSeek’s annualized revenue run rate was near $1 billion in the September 24, 2026 reporting, after the pace was below $500 million several months earlier.

The acceleration was partly linked to model prices being raised by about 2.3 to 4.5 times. The same revenue line can support very different valuations, so the first job is to split the growth into the part that survives a price reset and the part that would not.

The run-rate delta is the first number

Write down the gap between the two reported paces. That delta is the first number in the ledger, and it should be treated as a point-in-time comparison, not a forecast, because a run rate is not a year of booked revenue. It tells you how much the pace changed, not where it will land.

The price multiplier tells you how much of it is price

Compare the run-rate delta with the price multiplier. An AI price increase can make a flat or shrinking user base look like growth, so the ratio needs usage data behind it. Put the two together before you compare the growth to any benchmark.

• Revenue outpaces price: the extra growth is demand.
• Revenue matches price: the growth is price-led.
• Revenue lags price: volume fell.

Divide the run-rate delta by the price multiplier to estimate implied demand growth. A high price multiplier with a modest run-rate delta points to price-led growth. A modest price multiplier with a large run-rate delta points to demand-led growth. The first case is the one that should slow a raise.

Ask for the price schedule and the usage data behind the run rate. The price schedule shows when the price move happened; usage data shows whether customers stayed on. If the lab cannot show usage, the ratio is only half the answer.

Margin and compute test whether demand is real

Revenue for January through July 2026 was about 475 million yuan, compared with about 47.5 million yuan for all of 2025. That gap is useful, but it does not prove demand. Margin tells you whether the model is actually selling. API gross margin for the first seven months of 2026 was about 82.9%, while company-wide gross margin was about 44.6%. High API margin can mean the lab is charging more, not that more customers are buying.

Ask for the inference load behind the margin. Rising usage makes the margin a signal of scale. Flat usage points to pricing power. Falling usage is a warning.

Liang Wenfeng, DeepSeek’s CEO, reportedly said more than 70% of computing capacity goes to training new models, leaving under 30% for inference, the serving side of the business. If most compute is reserved for training, inference capacity may not scale with the revenue line. Price can lift revenue, but demand must show up in usage, inference load, and repeat spend.

The price-led portion needs a lower multiple

DeepSeek is seeking a second outside round of about 50 billion yuan, or about $7.45 billion, at a valuation of about 500 billion yuan, with completion targeted for the end of October 2026. Liang Wenfeng, DeepSeek’s CEO, told investors about the $1 billion annualized revenue run rate in a recent investor meeting. Before you set a multiple, split the revenue into price-led and demand-led buckets.

The price-led bucket is the part of revenue that would disappear if the price reset. The demand-led bucket is the part that survives a price reset. Apply a lower multiple to the price-led bucket, and apply the higher multiple only to the demand-led bucket. If the lab cannot show usage, inference load, or repeat spend, keep the discount larger.

Present the split to your investment committee as a bridge: show the reported run rate, the price-explained portion, and the demand-led remainder. Pair that with the margin and compute checks that support the demand-led remainder, and let the lower multiple follow the price-led portion.