DeepSeek's decision to cut pricing on its V3-Pro model by 75% looked like unambiguous good news for developers and enterprise AI vendors. In practice, many are finding that cheaper inference doesn't automatically translate into healthier margins.
The reason comes down to one structural tension: while token prices are falling, agentic systems are consuming tokens faster than those prices decline.
The 100x Problem Explained
For two decades, software economics followed a reliable pattern — infrastructure got cheaper, applications got more capable. Early AI was expected to follow the same curve. As frontier models improved and token prices dropped, inference was assumed to become a negligible operating cost.
That assumption is now breaking down.
A chatbot converts one user question into roughly one model call. An agent converts it into a chain of planning, retrieval, tool use, verification, summarization, and follow-up decisions. The user sees one answer. The vendor pays for the entire loop.
This is the 100x problem: the same user-visible request can cost orders of magnitude more to serve as an agentic workflow than as a simple chatbot or RAG response.
What Token Amplification Actually Looks Like
In a single-turn chatbot, the input-to-billed ratio is roughly 1:5. In a multi-step agent deployed across customer support, sales ops, finance, or legal review, that ratio routinely hits 1:700 or higher.
A seemingly simple agent query — "What did our top customer ask about last week?" — typically touches seven priced operations:
- User prompt (~50 tokens)
- System prompt and tool definitions (~3,000 tokens, repeated on every call)
- Retrieval (~5,000 tokens of context)
- Model call #1 — tool selection (8,000 in / 200 out)
- Tool execution (~4,000 tokens returned)
- Model call #2 — summarization (12,000 in / 400 out)
- Model call #3 — follow-up decision (12,400 in / 100 out)
One sentence in, roughly 35,000 input tokens billed — somewhere between $0.10 and $0.40 per query on a frontier model. At a million queries a month, that's a six-figure line item.
Why This Breaks the SaaS Business Model
The dominant enterprise AI pricing model has been seat-based SaaS: charge per user per month, deliver agent capability, capture margin. That model assumes a reasonably bounded cost per user.
Token amplification destroys that assumption. A power user running 50 agent invocations a day on a $40/seat plan can cost more in inference than the plan charges — flipping vendor gross margins negative. This is a paradox that compounds exactly as customers deepen their agent adoption, the very usage curve vendors are pitching to their boards.
"For my team, the cost of compute is far beyond the costs of the employees." — Bryan Catanzaro, VP of Applied Deep Learning, Nvidia
Several vendors are now privately reporting negative gross margins on heavy users, mirroring findings from Bessemer's 'Supernova' cohort, where AI-agent adoption and gross margin contraction have moved from theoretical risk to a primary P&L headwind.
The gap is surfacing publicly too. Bloomberg recently documented a widening chasm between Salesforce's Agentforce marketing demos and the capabilities actually shipping to customers — the predictable result when promised functionality is technically possible but uneconomical to serve at the seat-plan price point.
The Technical Responses
The engineering toolkit for managing token costs is known and converging. None of these are novel, but all are critical:
- Cost-aware routing — A small classifier decides which model tier handles each query. Well-tuned routers cut inference bills by ~60% with no measurable quality degradation.
- Prompt caching — Anthropic, OpenAI, and Google now offer 75–90% discounts on cached prefixes.
- Context discipline — Truncating tool outputs, pruning reasoning traces, and capping tool depth prevents agents from running expensive rabbit holes.
- Speculative decoding — For self-hosted deployments, this technique delivers 2–3× effective throughput on the same GPU hardware.
"Organizations using orchestration-led governance report stronger productivity gains — a holistic orchestration layer is associated with six times greater productivity impact than compliance-only approaches." — IBM
Companies building this orchestration layer well are starting to resemble financial trading systems more than microservice operators: every routing decision priced, every path with its own P&L, every tenant on a metered budget.
What Enterprise Leaders Should Do Now
Four moves separate the companies that will still have margin in 24 months from those that won't:
- Make inference cost a first-class metric. Track it per-feature, per-tenant, per-query class — the same discipline applied to cloud cost in the mid-2010s.
- Budget like a media buyer. Set cost-per-thousand-queries ceilings per feature. Cap them. Alert on overruns. Engineering won't enforce this on its own.
- Treat the router as core infrastructure. It is the new load balancer — not an optimization to revisit later.
- Audit prompts quarterly. A 4,000-token system prompt that grew organically over six months is a six-figure bill in slow motion. Most teams have never read their own production prompts end to end.
One additional lever: negotiate volume commits early. Frontier-model vendors now offer reserved-instance-style prepaid commits at substantial discounts. List price is the worst price available.
The strategic implication isn't simply that AI is expensive. It's that the dominant business model assumed by most AI-native company plans does not survive contact with agentic workloads — and price cuts from DeepSeek or anyone else don't change that underlying architecture.
