DeepSeek cut prices 75%. The 100x problem remains

DeepSeek cut pricing on its V4-Pro model by 75%, but the article argues enterprise AI vendors may still see margin pressure because agentic workflows “amplify” token usage—often from ~1:5 (chatbot) to ~1:700+ (agents), turning modest per-token cost declines into still-material per-query expenses. Using an example of a $40/user/month support assistant, 50–100 agent requests/day could increase inference costs by ~1 order of magnitude, potentially compressing or even driving negative gross margins on heavy users. The write-up cites widening gaps between AI-agent demos and shipped customer capability (e.g., Salesforce Agentforce) as a symptom of uneconomical execution at seat-plan pricing.

Analysis

The immediate loser is any enterprise software name that is trying to package agentic capability into a flat subscription price. That model turns usage into COGS leakage: the more the customer likes the feature, the worse the vendor’s gross margin profile can become. CRM is the cleanest public proxy because its AI narrative is tightly tied to workflow automation; if adoption rises before monetization is redesigned, the market should start discounting lower software gross margin and slower multiple expansion.

Second-order winners are less about model providers and more about the control plane around them. IBM stands out as a beneficiary if orchestration, governance, routing, and cost controls become budget line items rather than afterthoughts; that is higher-margin services/software than raw inference. GOOGL and AMZN can still win on volume, but only if they own the metering layer and can absorb price compression with enough query growth; otherwise, token efficiency and self-hosting can leak demand away from their high-value API mix. NVDA is not the clean short here: more agent activity still means more compute, but the market may be overestimating how much of that demand survives routing, caching, and smaller-model substitution.

The real catalyst window is the next 1-3 earnings cycles, when investors will start pressing for AI gross margin disclosure, per-feature unit economics, and evidence that “AI attach” is not just revenue with hidden costs. Over 6-18 months, the industry likely migrates from seat-based pricing to usage caps, action-based pricing, and internal cost governors; vendors that cannot reprice will have to throttle features. The contrarian miss is that the bullish AI consensus is treating lower model prices as if they automatically improve software margins, when in practice cheaper tokens often just enable customers to do more expensive work per request. The thesis breaks if vendors prove they can charge per action, per workflow, or per outcome fast enough to keep incremental gross margin stable.

Key Decisions for Investors

• Short CRM into strength or buy a 3-6 month put spread ahead of the next earnings print; thesis is that AI feature adoption will pressure gross margin before it meaningfully lifts ARR. Falsifier: disclosed AI gross margin/attach shows stable or improving economics.
• Pair long IBM / short CRM for 1-3 months as a relative-value expression on orchestration and governance winning over bundled-agent pricing. Risk/reward improves if enterprise CIOs start talking about token budgets and routing as mandatory controls.
• Do not short NVDA outright on this theme; if anything, keep it as a conditional long on weakness only if hyperscaler capex and inference volume continue to inflect. Falsifier: cloud vendors guide to materially slower AI spend growth despite usage growth.
• Watch AMZN and GOOGL for evidence that AI consumption is monetizing through metered infra rather than margin dilution; if not, avoid assuming their cloud mix benefits automatically. Alert item: any commentary on AI-related gross margin pressure or customer self-hosting moves.