ATL275
Accounting Technology Lab — ATL275
What Tokens Are You Smokin’?
- Series: Accounting Technology Lab
- Episode: ATL275
- Title: What Tokens Are You Smokin’?
- Hosts: Randy Johnston and Brian F. Tankersley, CPA.CITP, CGMA
- Recorded: September 11, 2026
- Release Date: September 18, 2026
- Approximate Runtime: 17 minutes
- Primary Topics: AI routers, token marketplaces, AI gateways, model selection, agent orchestration, AI economics, usage-based pricing, governance
Episode Summary
AI may feel like a flat monthly subscription today, but ATL275 argues that the economics are moving toward a metered, multi-model market. Randy Johnston and Brian Tankersley examine AI routers, token marketplaces, and agent orchestration—tools designed to let firms choose different models for different jobs rather than committing every task to one vendor.
Brian explains why OpenRouter and LibreChat can make model choice resemble buying fuel: use the amount of compute you need in the engine best suited to the work. Randy connects that concept to a three-layer AI strategy spanning general productivity tools, AI embedded in accounting applications, and specialized agents.
The hosts also discuss the difficulty of forecasting token budgets, the likelihood of more usage-based pricing, and a growing field of gateways including OpenRouter, Ramp Router, Factory Router, Portkey, LiteLLM, Cloudflare, Kong, AWS, Microsoft, and Google.
For accounting firms, the practical issue is not merely cost. Model selection, privacy, compliance, governance, and vendor concentration all matter when agents handle sensitive workflows. The takeaway: expect AI procurement to look less like buying one software subscription and more like managing a portfolio of compute suppliers—optimizing each workload for capability, cost, risk, and control.
Key Takeaways
- Every AI workload consumes compute. Whether the customer sees a monthly subscription or an API invoice, somebody is paying for the tokens behind the transaction.
- One model does not necessarily fit every workload. AI routers can give organizations access to different models without separately funding and administering every provider.
- Cost optimization will become part of AI governance. Firms may increasingly route simple work to inexpensive models and reserve expensive frontier models for jobs that justify the added capability.
- Per-seat pricing may not be the final economic model. As agents become more common, usage-based pricing could become increasingly important.
- Accounting firms need more than cheap tokens. Privacy, contractual protection, data handling, auditability, regulatory compliance, security, and vendor stability remain critical.
- AI gateways are becoming strategic infrastructure. OpenRouter, hyperscale cloud platforms, and newer routing competitors are competing to become the transaction layer between applications and AI models.
- “Bring your own tokens” may become commonplace. Software vendors can increasingly let customers supply their own API credentials and select the AI models used inside applications.
- Token measurement should eventually resemble cost accounting. Firms need to understand the cost of AI by workflow, agent, client, and business process—not merely the organization’s total monthly AI bill.
Short Promotional Copy
One-Sentence Promo
Randy Johnston and Brian Tankersley examine how AI routers, token marketplaces, and usage-based economics could change the way accounting firms buy, govern, and deploy artificial intelligence.
Three-Sentence Promo
Every time you use AI, somebody is burning tokens—and eventually somebody has to pay for them. In ATL275, Randy Johnston and Brian Tankersley explore OpenRouter, AI gateways, token marketplaces, model routing, agent orchestration, and why firms may increasingly choose the best model for each workload instead of buying one AI platform for everyone. The discussion combines AI economics with the accounting profession’s continuing concerns about privacy, governance, security, compliance, and cost control.
Promotional Paragraph
What happens when buying AI starts looking more like buying gasoline, electricity, or payment-processing capacity? In ATL275, Randy Johnston and Brian Tankersley explore the emerging world of AI routers and token marketplaces, where applications can choose among multiple AI models based on capability, cost, availability, or other criteria. The hosts discuss OpenRouter, Ramp Router, Factory Router, cloud AI gateways, agents, Model Context Protocol, and the possibility that firms will increasingly “bring their own tokens” to business applications. For accounting firms, the challenge is bigger than finding the cheapest model: sensitive client information, governance, contractual protections, regulatory requirements, and reliable cost controls all have to travel with the workload.
Timestamped Pull Quotes
| Time | Speaker | Quote |
|---|---|---|
| 01:58 | Brian Tankersley | “Every time you use AI, you’re smoking a token.” |
| 03:11 | Brian Tankersley | “You want to do AI with the best model that’s also the least expensive for the amount of processing you need to get done.” |
| 03:40 | Brian Tankersley | “Which token would you like to smoke?” |
| 06:42 | Brian Tankersley | “We have to optimize for maximum productivity at minimum cost in here, and so that’s really what these tools let us do.” |
| 07:06 | Randy Johnston | “We don’t know how much you’re going to spend for different types of processes.” |
| 10:58 | Randy Johnston | “The problem is there is no one model to rule them all.” |
| 13:23 | Randy Johnston | “Everybody wants an AI gateway, and everybody is trying to create an agent platform.” |
| 16:42 | Brian Tankersley | “You’re going to see these places where you can bring your own tokens, and you can effectively roll your own AI experience using these AI routers.” |
| 16:52 | Randy Johnston | “I guess it’s time to smoke them if you got them.” |
Production note: Timestamps are based on the supplied transcript/SRT. Recheck them against the final edited audio if additional cuts are made before publication.
Suggested Show Notes
- Why tokens matter even when users purchase AI through a flat monthly subscription.
- The emerging role of AI routers and model marketplaces.
- Why different AI models may be appropriate for different workloads.
- OpenRouter and the economics of aggregating access to hundreds of models.
- Brian’s use of LibreChat as a front end for multiple AI models.
- Randy’s three-layer AI strategy:
- General-purpose productivity AI.
- AI embedded inside accounting and business applications.
- Agents, MCP, and workflow automation.
- Why token budgeting remains difficult.
- How agents could change the economics of per-user AI licensing.
- Ramp Router and the growing competition among routing platforms.
- Factory Router, Trusted Router, Portkey, LiteLLM, Cloudflare AI Gateway, Helicone, Martian, Kong, and other routing products.
- Amazon Bedrock, Microsoft Azure AI Foundry, and Google Vertex AI as enterprise gateway alternatives.
- Why AI routing resembles merchant-service transaction processing.
- Brian’s analogy between AI tokens and gasoline: common fuel powering different engines.
- Microsoft Agent 365 and governance of enterprise agents.
- The emergence of “bring your own API key” and “bring your own tokens.”
- Why privacy, security, contractual protections, and regulatory compliance still matter when optimizing AI costs.
20 Suggested Social Media Posts
1
“Every time you use AI, you’re smoking a token.”
ATL275 looks underneath the friendly $20-per-month AI subscription and asks what happens when AI consumption becomes a true usage-based economy.
#AccountingTechnology #AI #GenerativeAI
2
There may be no one AI model to rule them all.
Some models write better. Some extract data better. Some reason better. Some are dramatically cheaper.
ATL275 asks whether AI routers will become the traffic controllers deciding which model gets each job.
#AI #AccountingTech #LLM
3
The next AI purchasing decision may not be ChatGPT vs. Claude vs. Gemini.
It may be: Which model should process this particular transaction at the right combination of quality, cost, privacy, and speed?
That is the world of AI routers. We unpack it in ATL275.
4
AI procurement is starting to look a lot like payment processing.
Aggregate huge volumes of transactions, route them efficiently, take a small piece of each one—and do it billions of times.
That makes the AI gateway a potentially valuable place to sit.
ATL275: What Tokens Are You Smokin’?
5
Buying an AI license for every employee is easy.
Proving that it is the most economical way to deliver every AI workflow is much harder.
ATL275 explores agents, token consumption, model routing, and a very different AI cost model.
#CPA #AI #TechnologyStrategy
6
Think of AI tokens like gasoline.
The same basic fuel can power a car, generator, lawn equipment, or countless other machines.
AI routers are trying to help organizations choose where—and how efficiently—to burn that fuel.
Hear the analogy in ATL275.
7
Accountants understand cost allocation.
We may soon need to apply that discipline to AI:
Cost per workflow.
Cost per agent.
Cost per client.
Cost per successful outcome.
Token accounting is coming.
#Accounting #AI #FinOps
8
OpenRouter, Ramp Router, Factory Router, Cloudflare, Kong, AWS, Microsoft, Google…
Everyone wants a place in the emerging AI routing layer.
ATL275 explains why that layer matters to accounting firms even if they never write a line of code.
9
AI strategy may need three layers:
1. General productivity AI
2. AI embedded in professional applications
3. Agents and automated workflows
Then comes the uncomfortable accounting question:
What does each layer really cost?
ATL275 tackles it.
10
The cheapest AI model is not always the right model.
The most powerful AI model is not always the right model either.
The objective is the model with the right capability, cost, controls, and risk profile for the job.
That is what makes intelligent routing interesting.
11
What is your firm’s token budget?
If the answer is “I have no idea,” you have plenty of company.
Randy Johnston and Brian Tankersley discuss why AI consumption is still difficult to forecast—and why that will matter as more workflows move into agents.
12
Software subscriptions taught us to budget per user.
Agentic AI may force us to budget per task.
That is a big change in technology economics—and potentially a very big change in how accounting firms buy software.
ATL275 explores the shift.
13
“Bring your own tokens” could become the AI equivalent of bringing your own mobile carrier or payment processor.
Applications provide the workflow.
You provide the model access.
A router decides where the work goes.
That is one possible future discussed in ATL275.
14
AI routing is not merely a cost issue.
For accounting firms, every routing decision can also involve:
Client confidentiality
Data residency
Retention
Security
Contract terms
Regulatory requirements
Vendor risk
Cheap compute is not cheap if it creates expensive risk.
15
A $30 AI subscription is wonderfully simple.
A world of autonomous agents running thousands of background tasks every day is not.
ATL275 looks at the financial plumbing that will be needed when AI becomes infrastructure rather than an occasional chatbot.
16
The model marketplace is beginning to resemble the cloud market.
Organizations may buy capacity from multiple suppliers, route workloads dynamically, monitor consumption, and optimize for both performance and cost.
CPAs should recognize the management-accounting opportunity here immediately.
17
Build the agent or buy the agent?
Then:
Which model runs it?
Who pays for the tokens?
How much can it spend?
Where does the data go?
Agentic AI turns a technical decision into a governance decision.
ATL275 explains why.
18
“We have to optimize for maximum productivity at minimum cost.”
That sounds less like an AI slogan and more like Management Accounting 101.
Which may be exactly why accountants need to understand token economics.
#CPA #CFO #AI
19
The AI platform with the best chatbot may not ultimately control the market.
The company controlling the routing, governance, billing, and orchestration layer could become just as strategically important.
ATL275 explores that emerging battle.
20
Smoke ’em if you got ’em.
Randy Johnston and Brian Tankersley close ATL275 with a look at the rapidly emerging token economy—and why tomorrow’s accounting technology stack may choose among AI models the way today's businesses choose among cloud services.
ATL275 — What Tokens Are You Smokin’?
Suggested Tags and Hashtags
Tags:
Accounting Technology, Artificial Intelligence, AI Routers, Token Economics, OpenRouter, Stripe, Ramp Router, LibreChat, Agentic AI, AI Agents, Model Context Protocol, AI Gateways, AI Governance, AI Cost Management, Microsoft Agent 365, Azure AI Foundry, Google Vertex AI, Amazon Bedrock, n8n, AI Orchestration
Suggested Hashtags:
#AccountingTechnology #AccountingTech #ArtificialIntelligence #AI #AgenticAI #AIAgents #AIRouters #OpenRouter #LLM #GenerativeAI #AIGovernance #AIInfrastructure #CPA #CFO #FinOps #TechnologyStrategy #Automation #MCP
Newsletter Copy
Suggested Subject: ATL275: What Tokens Are You Smokin’? The Economics Behind AI
Every AI prompt, agent, and automated workflow consumes computing capacity. Today that cost is often hidden behind simple monthly subscriptions, but Randy Johnston and Brian Tankersley believe AI is moving rapidly toward a more visible usage-based economy.
In ATL275 of the Accounting Technology Lab, Randy and Brian explore AI routers and token marketplaces—including OpenRouter, Ramp Router, Factory Router, cloud AI gateways, and other emerging platforms. They discuss how routers can choose different AI models for different jobs, why token budgeting remains difficult, and how agentic AI could change the economics of per-seat software licensing.
For accounting firms, however, price is only part of the equation. Privacy, security, governance, regulatory compliance, contractual protections, and vendor risk all need to remain part of the model-selection process.
If your firm is beginning to deploy AI agents or wondering what happens after the $20-per-user chatbot era, this episode is a useful look at the financial plumbing underneath the next generation of AI.
Suggested Visual / Promotional Image Direction
Concept 1 — The AI Cigar Lounge
A sophisticated but humorous “AI token lounge” with stylized digital tokens labeled for major model families sitting in cigar-box-style compartments. A professional accountant is choosing among them while a digital meter displays cost, performance, and privacy. Avoid actual tobacco use; make the “smoking” metaphor clearly about digital tokens.
Headline:
WHAT TOKENS ARE YOU SMOKIN’?
Subhead:
AI Routers, Model Marketplaces & the New Economics of Artificial Intelligence
Concept 2 — AI Gas Pump
A futuristic fuel pump labeled AI TOKENS, with several digital hoses feeding icons representing different AI models and agents. Pricing meters show tokens, compute, and cost.
Headline:
FILL ’ER UP — WITH WHICH MODEL?
Concept 3 — AI Traffic Router
A busy digital interchange with a business workflow entering from the left and being intelligently routed toward multiple AI model destinations based on signs labeled COST, SPEED, QUALITY, PRIVACY, and COMPLIANCE.
Headline:
WHO ROUTES YOUR AI?
Products, Services, and Companies Discussed
The table below focuses on technology products and companies materially discussed in the episode rather than incidental references used in historical or cigar/gasoline analogies. “Not confirmed” means an official current account was not sufficiently verified rather than an assumption being made.
| Company / Product | Role in Episode | X | |||
|---|---|---|---|---|---|
| OpenRouter | Multi-model AI router/marketplace | @OpenRouterAI | Not confirmed | OpenRouter | Not confirmed |
| Stripe | Payments company acquiring OpenRouter | @stripe | Stripe | Stripe | @stripe |
| Ramp / Router.com | AI routing platform | @tryRamp | Ramp | Ramp | Not confirmed |
| LibreChat | Multi-model AI chat interface | @LibreChatAI | Not confirmed | LibreChat | Not confirmed |
| OpenAI / ChatGPT | Frontier AI provider/model platform | @OpenAI | OpenAI | OpenAI | @openai |
| Anthropic / Claude | Frontier AI provider/model platform | @AnthropicAI | Not confirmed | Anthropic | @anthropicai |
| Google / Gemini / Vertex AI | Models and enterprise AI gateway | @GeminiApp / @Google | |||
| Microsoft / Copilot / Azure AI Foundry / Agent 365 | Productivity AI, enterprise AI platform and agent governance | @Microsoft | Microsoft | Microsoft | @microsoft |
| Perplexity | AI service mentioned in subscription/model discussion | @perplexity_ai | Not confirmed | Perplexity | Not confirmed |
| Thomson Reuters / Checkpoint CoCounsel | AI embedded in professional tax/research products | @thomsonreuters | Thomson Reuters | Thomson Reuters | @thomsonreuters |
| Wolters Kluwer / AnswerConnect | Tax research platform with AI | @Wolters_Kluwer | Wolters Kluwer | Wolters Kluwer | @wolterskluwer |
| Blue J | AI tax research provider | Not confirmed | Not confirmed | Blue J | Not confirmed |
| TaxGPT | AI tax research provider | Account present; handle not independently confirmed | Account present; handle not independently confirmed | TaxGPT | Account present; handle not independently confirmed |
| n8n | Workflow/agent orchestration platform | @n8n_io | Not confirmed | n8n | Not confirmed |
| SAP | Strategic investor/integration partner for n8n | SAP | SAP | SAP | SAP |
| Meta / Switchboard | Reported AI-routing initiative | Meta | Meta | Meta | Meta |
| Trusted Router | Open-source routing product discussed | Not confirmed | Not confirmed | Not confirmed | Not confirmed |
| Factory / Factory Router | AI routing platform | @FactoryAI | Not confirmed | Factory | Not confirmed |
| Portkey AI | AI gateway/router | Not confirmed | Not confirmed | Portkey | Not confirmed |
| LiteLLM | AI gateway / unified model access | Not confirmed | Not confirmed | LiteLLM AI Gateway | Not confirmed |
| Cloudflare AI Gateway | AI gateway/infrastructure | @Cloudflare | Not confirmed | Cloudflare | Not confirmed |
| Helicone | AI observability/gateway technology | Official X linked; handle not verified | Not confirmed | Helicone | Not confirmed |
| Martian | AI model router | @withmartian | Not confirmed | Martian | Not confirmed |
| Kong / Kong Gateway | API/AI gateway technology | @thekonginc | Not confirmed | Kong | Not confirmed |
| Amazon Web Services / Amazon Bedrock | Enterprise AI/model platform | @awscloud | Amazon Web Services | Amazon Web Services (AWS) | Amazon Web Services |
| Zoho | Example of software permitting external model/API use | @Zoho | Zoho | Zoho | @zoho |
| Model Context Protocol (MCP) | Open protocol for AI/tool connectivity | N/A | N/A | N/A | N/A |
| “Foundry/Foundries AI Gateway” | Name appearing in transcript | Needs name verification | Needs name verification | Needs name verification | Needs name verification |
Calls to Action
- Listen to ATL275 before finalizing your firm’s 2027 AI budget.
- Inventory which AI workloads actually require a premium frontier model.
- Begin measuring token or AI-compute consumption by agent and workflow.
- Ask AI vendors how routing decisions affect privacy, retention, subprocessors, and data residency.
- Determine whether your firm can place spending limits on individual agents.
- Review whether “bring your own API key” capabilities could reduce cost or improve model flexibility.
- Treat AI routing vendors as part of the firm’s third-party risk-management program—not merely as purchasing intermediaries.
Production and Fact-Checking Notes
1. OpenRouter / Stripe transaction.
The transcript describes an August 17 purchase of OpenRouter by Stripe for “a little more than $8 billion.” Contemporary reporting did discuss an $8-billion-plus transaction, but Stripe and OpenRouter formally announced the agreement on August 19, 2026 and did not disclose an official acquisition price. For published show notes, I recommend saying “Stripe agreed to acquire OpenRouter” without stating a definitive price unless the dollar figure is explicitly attributed to press reporting.
2. OpenRouter scale.
The episode mentions roughly 400 models and millions of users. OpenRouter’s August 19 announcement described access to more than 400 models and said the platform served more than 10 million developers and companies.
3. Ramp Router.
Ramp formally announced Router.com on August 19, 2026 as a unified API and routing/evaluation layer for AI models. This portion of the discussion tracks the timing of the public announcement.
4. SAP and n8n.
The transcript says SAP “put $5.2 billion in n8n.” That should not be repeated as a factual investment amount. n8n’s announcement says SAP made a strategic investment in a transaction that valued n8n at $5.2 billion, and that n8n technology would be embedded in SAP Joule Studio.
5. Microsoft indemnification.
The transcript broadly discusses Microsoft indemnification in the context of data exposure. Microsoft’s documented Customer Copyright Commitment concerns qualifying copyright/IP claims for covered commercial AI services and requires customers to meet specified conditions and use applicable guardrails. It should not be characterized in show notes as a blanket indemnity against all forms of data exposure without checking the applicable Microsoft product agreement.
6. Meta Switchboard.
The episode describes Switchboard as a reported Meta project rather than a launched product. Preserve that qualification.
7. “Foundries AI Gateway.”
The transcript wording is unclear. Verify this product name against Randy’s source material or the original audio before adding an external link.
8. Fast-moving market.
Routing products, ownership, availability, model catalogs, pricing, and social accounts can change quickly. Recheck external links immediately before publication.
