How DeepSeek Is Changing AI Economics for Business

This article distills a recent LinkedIn Live event hosted by Marvin Labs. The panel featured Alex Hoffmann (Co-Founder & CEO, Marvin Labs), Mark Bolton (AI strategy and wealth management advisor, Transform Consulting Inc.), and James Yerkess (Former Global Head of Transaction Banking & FX, HSBC Wealth Management).

The discussion focused on where AI is impacting P&L today, how falling training and inference costs are changing adoption, and what signals analysts should track as the economics shift. Watch the full conversation below.

The DeepSeek effect: cost, not a technical leap

The panel agreed that DeepSeek’s launch was less about a technical breakthrough and more about a step-change in economics. By releasing strong models at materially lower prices per 1M tokens for inference, DeepSeek shifted what is viable to automate. That has immediate implications for unit economics across many AI-assisted workflows.

• Mix-of-experts architecture: DeepSeek leverages a sparse “mixture of experts” approach rather than purely dense training, which concentrates compute where it adds value.
• Pricing reset: The conversation highlighted how sub-scale or cost-constrained users can now access model quality that previously required meaningfully higher per-token pricing.

For almost every user of AI, cost is now variable. When model quality improves at a fraction of the prior price, use cases that were marginal quickly move into positive ROI territory.

Data locality, compliance, and the “China question”

Concerns about model provenance and data residency came up repeatedly. The panel’s practical takeaways for operating teams:

• Data location matters more than model origin. Many providers now offer US- or EU-hosted inference endpoints. If regulated data is in scope, require onshore processing, logs controls, and contractual restrictions in vendor addenda.
• Open weights are a safety valve. Where feasible, self-hosting open-weight models can reduce the data transfer surface area. DeepSeek’s repository and licenses are available on GitHub.
• Treat it as a standard vendor risk review. Apply the same diligence you would for any critical service: data classification, retention, auditability, and incident response. For broader social engineering risks as AI improves content quality, CISA’s guidance on phishing remains a useful baseline (CISA alerts).

CapEx: who should spend, and who benefits

Hyperscalers continue to signal large, multi-year capital plans for data centers and AI infrastructure. The panel separated two business models:

• Platform monetization: Cloud vendors and model labs spending heavily to sell training/inference capacity and platform features across enterprises.
• Proprietary ROI: Large platforms that do not primarily sell AI infrastructure, but embed AI to drive core KPIs, such as engagement and monetization.

Alex argued investors should be cautious about underwriting purely CapEx-driven theses while the cost curve is moving.

Be wary of investment cases predicated on ever-rising AI CapEx. If unit costs keep falling, parts of that infrastructure may earn lower returns than modeled.

What is actually driving the bottom line

The most consistent gains today are not from chatbots. They are from embedded AI inside mature workflows that directly tie to revenue or cost.

• Ads relevance and yield: Meta and Alphabet both highlighted AI-driven improvements in ad ranking and campaign automation in recent earnings communications. These include higher conversion rates and stronger return on ad spend for specific advertisers. See company investor relations for details (Meta IR, Alphabet IR).
• Operations and risk: In banking, AI is increasingly used in triage layers for AML, fraud, and sanctions screening to cut false positives before human review. The decision remains with humans, but first- and second-pass automation lowers cost and cycle time.
• Workflow tools: Industry-specific software has begun to embed AI for parsing, drafting, extraction, and routing. The front end does not look like “AI.” It looks like the same application completing steps faster and more accurately.

The best use cases don’t look like AI. They look like the same tools doing the job better — higher relevance, fewer manual steps, faster cycle times.

Variable costs change the adoption curve

As inference pricing falls, adoption depends less on upfront budgets and more on clear unit economics:

• Most near-term deployments can scale on variable cost only. This favors SMBs and mid-market teams that can test, iterate, and scale without heavy upfront investment.
• “Agentic” pipelines over fine-tuning. Many teams are moving from expensive custom training to well-structured prompts, tools, and retrieval that produce reliable outputs at lower cost per task.
• Decision support over decision autonomy. “Co-pilot” patterns remain the norm. AI does the draft and triage. Humans decide.

What to watch next: signals for analysts and investors

Analysts should focus on validated, data-backed indicators rather than generic AI claims.

• Adoption and engagement: Active users, session length, and task completion rates for AI-assisted features inside products.
• Unit economics: Cost per task or per 1M tokens versus measurable revenue lift or cost savings.
• Release velocity: Frequency and impact of model and product updates relative to peers.
• Data and compliance posture: Clear data residency, retention, and audit controls for enterprise buyers.
• Consolidation at the model-lab layer: Expect pressure on second-tier research labs that are lagging on quality, price, or cadence.

Why this matters now

DeepSeek’s pricing catalyzed a broader reset in expectations. Lower costs expand the viable frontier of AI-assisted tasks, especially in advertising, operations, and risk. The investment implications differ by position in the stack. Platform providers must prove returns on rising CapEx. Application providers with high-ROI use cases can compound gains with modest incremental spend.

Watch the full video to hear the complete discussion, audience Q&A, and how Marvin Labs is helping analysts cut through noise and focus on what matters.