DeepSeek Now Sells Compute by the Second—and That Changes Your Cost Model

What Just Launched

DeepSeek has opened a public elastic compute tier—internally branded DSec—that bills inference workloads by the second and scales active capacity to zero when no requests are in flight. For operators running DeepSeek models on their own infrastructure, or routing to DeepSeek's hosted API, this is a direct attack on the largest hidden cost in production AI: the GPU time you pay for while your model waits.

The announcement surfaced mid-2026 and has drawn immediate attention from the infrastructure community because it mirrors what cloud hyperscalers have offered for serverless functions for years—but applied to large language model inference, where cold-start latency has historically made true scale-to-zero impractical.

Why This Was Hard Before

Running a large language model isn't like running a web server. Loading a model with billions of parameters into GPU memory takes time—often several seconds. If you scale to zero and then receive a request, your user waits through that load before getting a single token. That latency penalty has pushed most operators to keep at least one instance warm around the clock, paying for capacity whether it's used or not.

DSec's approach, based on what DeepSeek has disclosed, uses pre-warmed model shards held in a shared pool rather than dedicated per-customer instances. When your workload needs capacity, it draws from that pool in milliseconds rather than loading from scratch. You pay only for the compute seconds your requests actually consume. The pool itself is DeepSeek's cost to carry—your idle time becomes their problem, not yours. This is the same economic logic behind cloud database serverless tiers, and it works at scale because aggregate demand across thousands of tenants is smoother than any single tenant's traffic.

The tradeoff is control. You are sharing infrastructure you cannot inspect, with isolation guarantees you must take on faith, administered by a company headquartered in China and subject to the data governance rules that come with that.

What Changes for Operators

For teams running bursty or unpredictable AI workloads—customer support queues, document processing pipelines, internal tools with irregular usage—the arithmetic shifts meaningfully. A workload that runs for three hours a day was previously anchored to 24-hour reservation costs. Under per-second billing with genuine scale-to-zero, that same workload could cost one-eighth as much. For startups or product teams burning reserved GPU budget on low-traffic features, this is real money.

The mistake would be treating this as a pure cost win and moving production workloads over without working through the trust and compliance questions first. Elastic, cheap, and opaque is a combination that requires due diligence, not enthusiasm.

Questions You Should Be Asking

• What are the data residency guarantees, in writing? Prompts and completions passing through a shared pool need explicit contractual commitments about where they are stored, for how long, and who can access them—not just a privacy policy paragraph.
• What does isolation actually mean here? Shared model pools reduce cost by sharing resources across tenants. Ask specifically whether request data can be observed by, or influence the outputs seen by, another tenant.
• How is cold-start latency guaranteed, and what happens during pool saturation? Millisecond warm starts depend on pool availability. Get the SLA in numbers, and ask what your p99 latency looks like when the pool is under peak load.
• Does your compliance posture permit this vendor? Depending on your industry and customer base, routing data through DeepSeek's infrastructure may conflict with existing regulatory obligations. This is a legal question, not a technical one—get it answered before the first production request goes out.
• What is the egress path if you need to leave? Low-cost elastic compute is a strong lock-in mechanism. Before you optimize your stack around DSec pricing, confirm you can migrate workloads without re-engineering your inference pipeline from scratch.

What To Watch Next

The signal that matters is whether Western cloud providers—AWS, Google, Azure—accelerate their own serverless LLM inference tiers in response. If Azure AI Foundry or Amazon Bedrock announce per-second billing with scale-to-zero within the next quarter, DSec's pricing pressure has landed. If they don't move, it likely means they've concluded the compliance risk keeps enterprise buyers away regardless of the cost advantage—and that tells you something important about how seriously to take this offer yourself.