DeepSeek has cut prices by 75%. The 100x problem remains

DeepSeek’s recent decision to drastically cut the prices of its V4-Pro model by 75% should have been unequivocally good news for enterprise AI vendors and developers. Instead, many are finding that cheaper models don’t automatically translate into healthier margins.

The reason is simple: as inference costs plummet, agent systems voraciously consume tokens faster than prices drop. For the past 2 decades, the software economy has been dictated by the same rule. Infra is getting cheaper every year as applications become more capable. AI was originally assumed to follow the same pattern. As marginal patterns improved and token prices fell, many assumed that mining would become a minor operating expense. That assumption began to break down exponentially.

A chatbot typically turns a user question into a mock call. An agent turns it into a chain of planning, retrieval, tooling, verification, summarization, and subsequent decisions. User sees one response. The seller pays for the cycle. This is the 100x problem: the same user-visible query can cost a lot more to serve as an agent workflow than as a chatbot or a Retrieval Augmented Generation (RAG) response. For longer workflows, the multiplier is higher. Falling model prices help, but they don’t fix a product architecture that turns one prompt into dozens of chargeable operations.

The scale of what is now at stake is clear in how model providers themselves price developer relationships. OpenAI’s proposed program to give each Y Combinator startup $2 million in API credits—a number that would have funded an entire seed round in any previous tech cycle and when the same cohort received several thousand dollars in AWS credits—is less a recruiting perk than a recognition of how much it now costs to run an AI-based company in its first year of product. For established businesses retrofitting agents into existing product lines, the absolute numbers are even higher.

What is token boost

In a single-turn chatbot, one user message produces approximately one model call. The input-invoice ratio is about 1:5.

In a multistage agent deployed in customer support, sales operations, finance, legal review, and engineering, this ratio routinely reaches 1:700 or higher. Each loop iteration carries forward the cumulative conversation, instrument outputs, and reasoning traces. Each step adds; nothing is omitted.

A "simple" querying an agent like “What did our lead customer ask last week?’ it usually touches seven price operations before returning a response:

1. User prompt (~50 tokens)

2. System prompt and tool definitions (~3000 tokens repeated on each call)

3. Fetch (~5000 characters per context)

4. Model Call #1 – Tool Selection (8000 in / 200 out)

5. Tool execution (~4000 tokens returned)

6. Model #2 Generalization – Generalization (12,000 Inputs / 400 Outputs)

7. Model Call #3 – Subsequent Solution (12,400 Inputs / 100 Outputs)

A single sentence charges approximately 35,000 input tokens. Somewhere between $0.10 and $0.40 per border model request. Multiply that by a million requests per month—the amount of stakes on the table for every enterprise B2B function—and the order is six figures.

Why this disrupts the existing AI business model

The dominant pricing story for enterprise AI is seat-based SaaS: Pay per user per month, provide agency capabilities, capture margin. This model assumes a reasonably limited cost per user.

Token amplification violates the assumption. A power user making 50 agent calls per day on a $40/seat plan may cost more in the bottom line than the plan charges. Token leverage disrupts the traditional SaaS pricing model. When a power user’s daily agent activity costs more than their monthly subscription fee, the provider’s gross profits become negative, a paradox that is amplified as customers deepen agent adoption, the very usage curve that sellers sell on their boards. Several vendors are now privately reporting negative gross margins for heavy users, mirroring recent cloud spending reports from Bessemer’s « Supernova » cohort, where the correlation between AI-agent adoption and gross margin contraction has shifted from a theoretical risk to a primary revenue headwind.

Visible symptoms began to leak into the public domain. This week, Bloomberg documented a growing gap between marketing demos of Salesforce’s Agentforce and the capabilities actually delivered to customers. It’s the kind of loophole that predictably opens up when the promised functionality is technically possible but uneconomical to service at the price the seat plan implies. Salesforce is the most watched case, not unique.

"For my team, the cost of computing far outweighs the cost of employees." — Bryan Catanzaro, Vice President of Applied Deep Learning, Nvidia

Strategic importance is not "AI is expensive." It’s that the dominant business model adopted by most companies’ AI-based plans does not survive contact with agent workloads.

A simple example

Consider an enterprise software vendor that charges $40 per user per month for an AI support assistant. A traditional chatbot can cost only a few cents per user per day in conclusion, leaving healthy gross margins.

Now replace that chatbot with a fully agent-based workflow capable of investigating tickets, querying internal systems, drafting responses, validating output, and escalating exceptions. If a heavy user is running 50 to 100 agent requests per day, the output consumption can increase by an order of magnitude. What was once a minor infrastructure expense is becoming a significant operational expense.

This creates an unusual dynamic: the customers who get the most value from the product are often the customers generating the highest output costs. In extreme cases, providers may find themselves with their most engaged users contributing the least profit. The result is a growing realization in enterprise software that agent adoption and margin expansion are no longer automatically synchronized.

Agent orchestration is the new moat

The technical answers are known and converging. They are not new, but they are critical to survival

• Cost-based routing: This technique involves a small classifier model that decides which level (equivalents to Haiku, Sonnet, Opus) handles each query. Well-tuned routers reduce port bills by about 60% without any degradation in quality

• Fast caching: Anthropic, OpenAI and Google now offer 75 to 90% discounts on cached prefixes.

• Context discipline: You can truncate tool output, trim logic traces, and limit tool depth to prevent your agent from going down a rabbit hole

• Speculative decoding: for self-hosted deployments, this technique guarantees 2 to 3 times the effective throughput on the same GPUs.

"Organizations using orchestration-driven management report stronger performance gains—a holistic orchestration layer is associated with six times greater performance impact than compliance-only approaches" — IBM

Companies that build this layer well are starting to look less like microservices operators and more like financial trading systems: Each routing solution with a cost, each path with its own profits and losses, each tenant with a metered budget.

What business leaders should actually do

Wellour moves separate the companies that will still have margin after 24 months from those that won’t:

1. Make inference costs a premium metric. Track it by feature, by tenant, by request class the same way cloud pricing was tracked in the mid-2010s.

2. Budget as a media buyer. Set CPM caps for each feature. Hats off to them. Exceeding alert. Engineering will not enforce this by itself.

3. Treat the router as a core infrastructure, not an optimization. This is the new load balancer.

4. Audit prompts quarterly. A system prompt of 4,000 tokens that grows organically over six months is a six-figure account in slow motion. Most teams have never read their own production guidelines cover to cover.

5. Negotiate the volume of engagements early. Frontier-model providers now offer reserved-instance-style prepaid commitments at significant discounts. List price is the worst price any business will ever pay.

The next 24 months

Structural change under agent AI is not that it is expensive. As DeepSeek’s price cut today highlights, the unit cost of frontier inferences is falling roughly 3x per year, and the curve isn’t slowing down.

That’s the shift amplification outpaces price reductions. Cutting costs per token by 75% doesn’t help a company whose agents make 700x more tokens per user request than its pricing model suggests. For the first time since the dawn of the cloud era, architectural decisions are once again real-time financial decisions. Rapid redesign is a margin event. A poorly connected agent cycle is a break with a credit card attached.

The companies that survive the next 24 months of AI infrastructure pricing won’t be the ones using the cheapest model. They will be the ones whose agents are smart and you know how much they cost to think.

This is the 100X problem. And it’s arriving faster than price cuts can hide it.

Maitreyi Chatterjee is a senior software engineer at a large technology company.

Devansh Agarwal works as an ML engineer in a leading technology company.

Orchestration,DataDecisionMakers

#DeepSeek #cut #prices #100x #problem #remains