OpenAI’s alleged “top 30” enterprise users: what this leak hints at, and why it matters
A table has been doing the rounds on Reddit, claiming to list OpenAI’s 30 biggest customers by token usage – reportedly more than 1 trillion tokens processed. OpenAI hasn’t confirmed the list, but if it’s even directionally right, it offers a useful snapshot of how enterprise AI adoption is taking shape.
Before we get carried away: the source is unverified, and several details are not disclosed. Still, the mix of companies named and the categories they fall into align with what we’re seeing in the market across the UK and beyond.
Source: Reddit discussion.
Who’s reportedly in the “top 30”, and what patterns stand out
The list blends AI-native startups (Perplexity, Cognition, Sider AI), vertical specialists (Abridge in healthcare, Tiger Analytics in services), developer infrastructure (JetBrains, Warp.dev, Datadog), and large SaaS platforms (Salesforce, Shopify, Zendesk, Notion). There are also consumer-scale brands like Duolingo, Canva, WHOOP, and a major telco (T-Mobile).
The Reddit post groups them into four archetypes. Here’s how that maps out:
| Archetype | What it means | Examples (from the list) |
|---|---|---|
| AI-Native Builders | Products fundamentally designed around reasoning and LLM-first workflows. | Cognition, Perplexity, Sider AI |
| AI Integrators | Established platforms embedding AI into existing customer workflows. | Shopify, Salesforce, Notion, Zendesk, HubSpot |
| AI Infrastructure | Developer tools and platforms enabling building, routing, and observability. | OpenRouter, Warp.dev, JetBrains, Datadog |
| Vertical AI Solutions | Domain-specific apps optimised for a single industry or task. | Abridge (clinical notes), WHOOP (health), Tiger Analytics (services) |
That spread matters. It suggests AI is not just a feature; it’s becoming a stack. From routing and DevOps, to product-layer assistants, to specialist vertical tools, value is accruing across the pipeline – not just at the chatbot front-end.
About that “token war”: economics and scale
Tokens are the basic billing unit for most large language models (LLMs). Roughly, 1,000 tokens equals 750 words of text. If the leak is accurate, these firms are consuming tokens at extreme scale, either in product features (e.g. AI search, meeting notes) or behind-the-scenes automations.
“Whoever compounds reasoning the fastest shapes the next decade of software.”
Token economics drive every design decision: prompt length, context window, retrieval pipelines, and model choice. If you’re rolling out AI at work, monitoring tokens is non-negotiable. Pricing varies by model and tier; see OpenAI pricing for current rates, and plan for variability as models change.
One claim in the Reddit post is that “over 70% of ChatGPT usage is non-work” – advice, planning, personal writing. That’s not confirmed by OpenAI here, so treat with care. It does align with observed behaviour: consumer-scale habits often precede enterprise adoption, and those habits inform the product surface area that integrators then bring into organisations.
Why UK teams should care: risk, compliance, and opportunity
Data protection and residency
UK organisations need clarity on where data goes, how it’s stored, and who can access it. If you require regional controls, consider Azure OpenAI Service, which offers enterprise-grade compliance and data residency options via Microsoft’s cloud regions (including UK regions). For direct OpenAI use, review data usage and retention settings carefully and ensure a data processing agreement is in place.
Procurement and lock-in
Heavy token usage can lead to vendor concentration risk. Build abstraction layers where possible (e.g. model routers, internal APIs) so you can trial alternatives without rewriting your entire stack. The presence of OpenRouter and Datadog on the list reflects the importance of routing, monitoring, and observability as token volumes grow.
Model reliability and safety
Even the strongest LLMs hallucinate. For regulated workflows (healthcare, legal, finance), combine models with retrieval-augmented generation (RAG) and clear human-in-the-loop checkpoints. Track failure modes, run A/B tests, and establish model performance baselines before scaling.
Practical steps for UK developers and product teams
- Pick your archetype: are you integrating AI into an established workflow, or building an AI-native product? The build-versus-integrate choice sets your architecture and cost profile.
- Instrument token spend early: log prompts, completions, errors, latency, and per-feature costs. Tie token consumption to business metrics (leads, resolved tickets, reduced minutes per task).
- Start with contained, high-ROI workflows: meeting notes, support triage, sales email drafts, or code review assistance. Keep humans in the loop until you have confidence in outputs.
- Use governance by design: prompt templates, content filters, role-based access, red-team tests, and clear user messaging on limitations and data handling.
- Meet users where they are: simple connectors into spreadsheets, docs, CRM, and ticketing systems beat flashy demos. If you’re experimenting, here’s a practical guide to wire AI into Sheets: How to connect ChatGPT and Google Sheets.
What we don’t know (yet)
- Verification: OpenAI has not confirmed the table. Treat it as indicative, not definitive.
- Token totals: It’s unclear whether “over 1 trillion tokens” applies per customer or in aggregate. Not disclosed.
- Timeframe: No dates are provided. We don’t know over what period the usage was measured. Not disclosed.
- Model mix: No breakdown by model family (e.g. GPT-4o variants) or modality (text, image, audio). Not disclosed.
The competitive landscape: where value may accrue
If this leak is even half right, two themes stand out. First, integrators with distribution (Salesforce, Shopify, Zendesk, Notion) can ship AI into existing workflows at pace. Second, AI-native builders focused on reasoning (Perplexity’s search, Cognition’s coding agents) are pushing the frontier and creating new usage patterns that everyone else will copy.
For the UK, the take-away is simple: AI capability is becoming a capability advantage. Whether you buy it (integrate), build it (AI-native), or run it (infrastructure), the teams that instrument cost, quality, and safety from day one will be better placed to scale when the proof-of-concept glow fades.
Final word
The token war has already started.
Even if the list is wrong in places, the direction of travel is right. Track your tokens, pick the right architecture for your goals, and build the governance to match your ambition. The rest is execution.
References
- Reddit discussion: OpenAI might have just accidentally leaked the top 30 customers
- OpenAI pricing and token costs: openai.com/pricing
- Azure OpenAI Service (enterprise and residency options): learn.microsoft.com/azure/ai-services/openai/
