DeepSeek's DeepInfer could make the GPU market bigger

DeepSeek opens its serving playbook

On October 20, DeepSeek released DeepInfer under Apache 2.0. The framework supports DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. It exposes serving optimizations including speculative decoding and asynchronous scheduling. The public release provides offline inference; operators still need service components such as cache management and load balancing.

The commercial significance is the spread of operational know-how. More developers can examine and adapt the software that makes a model practical to run. Our thesis is that making useful intelligence cheaper will create many more occasions to use it.

Source: DeepSeek DeepInfer release README

What the 80% margin story actually requires

DeepSeek reports $38 of revenue per GPU-hour for its own system at off-peak pricing. This is an issuer-reported operating figure, not an audited profit measure or a result every cloud can reproduce. The documentation does not establish a universal 80% margin.

Here is the arithmetic: margin = (revenue − cost) / revenue. At $38 revenue, an 80% margin requires cost of $7.60 per GPU-hour. A margin above 80% requires cost below $7.60. We have not established a matched all-in cost number. Storage, networking, idle capacity, service operation and the chosen accounting treatment all affect the comparison.

Opening code can improve another operator's economics. That does not make its traffic, selling price, hardware or costs identical to DeepSeek's. We exclude a universal margin promise from this report.

Source: DeepSeek V4.1 Flash inference deployment Q&A

The Jevons trade

The exciting investment mechanism is demand expansion. A coding agent can try more solutions before returning its answer. A software company can make model calls throughout a workflow instead of reserving them for one expensive feature. Businesses can analyze documents or conversations that previously cost too much to process.

Those examples describe potential uses, not measured adoption following today's release. They explain why a cheaper unit of intelligence can enlarge the market. Developers gain more economically viable things to build, and users can consume more inference through every application.

Our forecast is that this expansion will be large enough to keep raising total accelerator demand. Calling that the Jevons trade identifies the mechanism; it does not turn every efficiency improvement into an automatic law of higher resource consumption.

A worked example: efficiency and demand can rise together

Consider a deliberately hypothetical workload requiring 100 equivalent GPUs at baseline. Suppose software doubles throughput per GPU while customer workload triples. Required capacity becomes 100 × 3 ÷ 2 = 150 GPUs, a 50% increase. Both improved efficiency and a larger hardware market are present.

At two times throughput and only two times workload, the requirement stays at 100 GPUs. At unchanged workload it falls to 50. These are scenarios, not DeepInfer performance measurements, demand forecasts or a conversion from token price into GPU cost.

The threshold matters: total workload must grow faster than throughput per GPU for required GPU capacity to rise, holding workload mix and other assumptions constant. Our bullish call is that AI has enough valuable work still waiting to clear that threshold.

What investors can own

NVIDIA provides direct listed exposure to accelerator hardware and software. Microsoft, Amazon and Alphabet offer public exposure to cloud infrastructure and AI distribution. These are business exposures, not estimates of revenue attributable to DeepInfer.

We expect efficient serving to improve the earning potential of useful installed capacity and strengthen the incentive to expand it. If adoption outruns new supply, utilization and rental pricing can remain firm. That is the route from cheaper inference to valuable, scarce compute.

Hardware competition matters within that bullish aggregate view: this release supports Huawei as well as NVIDIA. A larger accelerator market does not assign every incremental dollar to one supplier. Nor does stronger utilization guarantee appreciation of every older GPU or a particular stock.

Source: DeepSeek DeepInfer release README

What would change our view

The central counterargument is that savings exceed the demand response. Then providers can handle more work with a smaller fleet. Aggressive price competition could also pass the gains to customers faster than operators build durable profits. Hardware supply and architecture changes can change which capacity remains economically useful.

The evidence to follow is workload growth alongside throughput, realized cloud pricing, utilization and capacity additions. We would also want reproducible third-party end-to-end serving economics before treating DeepSeek's operating figures as another provider's achievable return.

Our conclusion remains bullish: more efficient AI makes many more applications worth building. We expect DeepInfer to contribute to a much larger inference market, with productive compute capacity remaining an important constraint.

AI-assisted research and writing. Primary documentation retrieved October 20, 2026 at DeepInfer commit 774ce244b87cc0eb8d4b94ab9f5de2ab22049f01. Calculations are MacroShed scenarios; forward-looking conclusions are editorial judgments.

Source: DeepSeek DeepInfer release READMESource: DeepSeek V4.1 Flash inference deployment Q&A