Stripe Buys OpenRouter for $7B as AI Value Shifts to Routing Layer
IBM
Plummeting token prices and unmonitored agent sprawl are driving enterprise AI value away from model makers and into toll-collecting routing gateways and systems integrators enforcing strict FinOps controls.
Stripe is finalizing a deal to acquire AI model router OpenRouter for over $7 billion, representing a greater than 5x markup in just 82 days as it positions itself as the transaction layer for autonomous agent commerce. (11:15)
IBM partnered with OpenAI to build a dedicated consulting practice, complementing its Anthropic software engineering alliance and leveraging watsonx.governance to act as a model-agnostic enterprise control plane. (0:00)
Ramp's August 2026 AI Index reveals stark spend inequality: the top 1% of adopting companies spend approximately $7,400 per employee monthly on AI tools, approaching engineering salaries, while the median corporate spend is just $12. (22:20)
Enterprise token spend surged 13x over the past year despite falling per-token prices, prompting an 'AI FinOps' era aimed at curbing unmonitored autonomous agent workloads and fragmented proof-of-concepts. (30:00)
Detailed notes
## Background
Published on August 21, 2026, by IBM on their Mixture of Experts podcast platform, this episode explores the shifting financial and operational infrastructure of the artificial intelligence industry. According to the show description, the panel breaks down IBM's newly announced partnership to train and certify consultants on OpenAI's enterprise tools alongside major corporate acquisitions and spending benchmarks across the market.
## TL;DR
Host Tim Huang and IBM panelists Kawtar El Maghraoui, Aaron Baughman, and Mihai Criveti analyze why enterprise AI value is shifting away from frontier model checkpoints toward model-agnostic routing, systems integration, and governance. The group examines the strategic logic behind Stripe’s reported $7 billion acquisition of OpenRouter, viewing it as an infrastructure toll booth for emerging autonomous agent commerce. Drawing on data from Ramp’s AI Index, the discussion highlights the end of unmetered corporate experimentation, the rise of AI FinOps to curb runaway agent costs, and the operational chaos of uncoordinated enterprise deployments. Finally, they address reports of AI-generated legislative text in the US Congress, urging institutions to establish clear governance frameworks rather than futile blanket bans.
## Atmosphere & style
The conversation carries an analytical, insider-practitioner tone that balances technical systems architecture with financial skepticism. The hosts share an easy camaraderie, often punctuating complex discussions of enterprise plumbing and rate limits with wry humor—such as Mihai joking about rethinking his life choices after giving away routing gateways for free. Rather than indulging in venture capital hype, the dialogue feels grounded in the gritty operational reality of enterprise IT, enterprise FinOps, and the headaches of legacy infrastructure.
## Bulletin
- `11:15` Stripe is finalizing a deal to acquire AI model router OpenRouter for over $7 billion, representing a greater than 5x markup in just 82 days as it positions itself as the transaction layer for autonomous agent commerce.
- `00:00` IBM partnered with OpenAI to build a dedicated consulting practice, complementing its Anthropic software engineering alliance and leveraging watsonx.governance to act as a model-agnostic enterprise control plane.
- `22:20` Ramp's August 2026 AI Index reveals stark spend inequality: the top 1% of adopting companies spend approximately $7,400 per employee monthly on AI tools, approaching engineering salaries, while the median corporate spend is just $12.
- `30:00` Enterprise token spend surged 13x over the past year despite falling per-token prices, prompting an 'AI FinOps' era aimed at curbing unmonitored autonomous agent workloads and fragmented proof-of-concepts.
## Key takeaways
- `00:00` IBM is establishing itself as a model-agnostic control plane, using Anthropic for software engineering and OpenAI for client-facing consulting while running routine tasks on Granite.
- `05:00` Enterprise AI adoption is constrained not by model intelligence, but by change management, messy data, strict API rate limits (50–200 requests/sec), and legacy data architectures in mainframes and SAP ERPs.
- `05:00` Frontier model developers depend heavily on systems integrators because over 80% of enterprise value remains locked behind legacy systems and strict compliance regulations like HIPAA.
- `11:15` OpenRouter provides unified multi-model routing, FinOps spend controls, semantic caching, and virtual API keys to manage heterogeneous enterprise model environments.
- `14:48` Stripe's acquisition of OpenRouter functions as an AI 'toll booth', allowing Stripe to capture transaction and metering revenue across continuous micro-token streams regardless of which underlying model wins.
- `14:48` Autonomous AI agents require digital wallets to budget inference calls dynamically, making Stripe's payment infrastructure a potential foundation for agentic commerce.
- `20:00` Open-source model routers are proliferating rapidly, pushing enterprise value toward comprehensive agent control planes featuring governance, observability, and MCP server integrations.
- `25:00` The corporate 'AI FOMO tax' of 2023–2025 has ended, replaced by CFO mandates demanding unit-economic ROI and metrics based on cost-per-unit-of-work rather than per-seat software licensing.
- `27:30` Enterprises are experiencing operational backlash from dozens of fragmented, unmonitored autonomous agents and automated event streams burning expensive frontier model tokens without rate limits.
- `30:00` Total cost of ownership in enterprise AI mirrors the 2012–2015 cloud computing transition, shifting the market from indiscriminate adoption to structured AI FinOps and governance.
- `30:00` Software engineering remains the primary corporate function where surging AI spend faces minimal pushback, as coding assistants like Cursor and Claude Code become vital infrastructure.
- `33:10` AI-generated legislative drafting in Congress mirrors widespread workplace shortcuts, highlighting the need for institutional guidelines and human accountability over outright bans.
## Detailed walkthrough
### `00:00` Introduction and IBM's New OpenAI Partnership
Tim Huang opens the episode by highlighting that the AI market is constantly in flux, pointing out that right now "the heartbeat is around routing." Joined by IBM panelists Kawtar El Maghraoui (Principal Research Scientist, AI Native Systems), Aaron Baughman (IBM Fellow), and Mihai Criveti (CTO, watsonx Orchestrate), the show sets an agenda covering Stripe's acquisition of OpenRouter, spend data from Ramp, AI legislation in Congress, and IBM's recent strategic moves.
```mermaid
flowchart TD
accTitle: IBM Multi-Model Integration Strategy
accDescr: IBM integrates OpenAI and Anthropic models alongside Granite models across consulting and software engineering layers.
subgraph External["Frontier Model Providers"]
OAI["OpenAI\n(Complex Task Planning)"]
ANT["Anthropic\n(Software Engineering / IBM Bob)"]
end
subgraph IBMInternal["IBM Portfolio"]
GRA["IBM Granite Models\n(Smaller, Cheaper, Fast Routing)"]
GOV["watsonx.governance\n(Audit Trails, Compliance, Prompt Monitoring)"]
end
OAI --> CON
ANT --> ENG
GRA --> CON
GRA --> ENG
GOV --> CON
GOV --> ENG
```
Tim introduces IBM's announced partnership with OpenAI to build an OpenAI practice within IBM Consulting, training and certifying large numbers of personnel. He notes this follows IBM's major partnership with Anthropic. Rather than an "OpenAI versus Anthropic" binary battle ("the sharks and the jets"), the industry is trending toward model-agnostic system integration. Aaron Baughman views this as an enormous opportunity to deploy forward deployed engineers and units into Fortune 500 enterprises. He emphasizes that IBM is positioning itself as a neutral enterprise control plane rather than betting on a single frontier model winner. Aaron distinguishes the internal roles: Anthropic is primarily aligned with IBM software engineering and the "IBM Bob" ecosystem, whereas OpenAI aligns directly with IBM Consulting's client-facing work. Simultaneously, IBM maintains its own Granite family of models to handle simpler, cost-effective requests while offloading complex planning tasks to larger frontier models via model gateways.
- **Panelists:** Tim Huang (Host), Kawtar El Maghraoui (Principal Research Scientist, AI Native Systems), Aaron Baughman (IBM Fellow), Mihai Criveti (CTO, watsonx Orchestrate).
- **Core Topics:** IBM’s OpenAI alliance, Anthropic ties, Stripe buying OpenRouter, Ramp corporate spend benchmarks, and congressional AI usage.
- **Enterprise Differentiation:** Anthropic powers internal software engineering workflows (IBM Bob), OpenAI fuels IBM Consulting practices, and IBM Granite provides low-cost execution for routine workloads.
- Aaron Baughman: “We're betting that the enterprise AI value, it's shifting upwards from owning the model to now orchestrating, governing, and integrating, and even, you could argue, operationalizing these types of models.”
---
`05:00` Model Limits, Enterprise Plumbing, and Watsonx Governance
The panel dives into OpenAI’s broader commercial strategy, including its internal deployment arm and its alliance with McKinsey. Mihai Criveti explains that model developers depend heavily on systems integrators because AI deployment requires boots on the ground to tackle client-specific use cases. Mihai stresses that "there is no such thing as the best model" in isolation. Model selection relies on trade-offs among cost, time to first token, generation speed (tokens per second), geographic hosting restrictions (such as regional data residency mandates in London data centers), and operational rate limits. Enterprises that attempt to roll out AI to thousands of workers routinely hit strict throughput ceilings (like 50 to 200 requests per second) that money alone cannot resolve, necessitating model routing, custom prompt engineering, and multi-provider redundancy.
```mermaid
flowchart LR
accTitle: Enterprise AI Evaluation Dimensions
accDescr: Key constraints and dimensions that dictate model selection in enterprise production deployments.
Aaron points out the sheer economics of seat-based token consumption: budgeting $10 per employee each month across hundreds of thousands of staff yields millions of dollars in monthly operational costs. System integration partnerships enable enterprises to capture economies of scale pricing while right-sizing model capacity per query. Kawtar El Maghraoui expands on this, arguing that raw benchmark wins (like a 2% edge on MMLU or HumanEval) pale in comparison to distribution channels and deep access to legacy enterprise architecture. With approximately 80% of enterprise value tied up in mainframe COBOL databases, SAP ERP installations, and Epic healthcare systems operating under strict HIPAA guidelines, established enterprise integrators hold the leverage. Through watsonx.governance, IBM focuses on auditability, bias mitigation, prompt monitoring, and regulatory compliance, giving clients the flexibility to hot-swap frontier model checkpoints without incurring multi-billion-dollar foundation model pre-training costs.
- **Operational Ceilings:** Scaling AI internally exposes sharp API rate limits (e.g., 50–200 requests per second), requiring dynamic routing layers rather than direct reliance on single API endpoints.
- **Scale Cost:** A standard rate of ~$10 per user per month quickly balloons to millions of dollars in monthly token expenses for a enterprise with 100,000+ employees.
- **Legacy Distribution:** The vast majority of business data resides in legacy mainframes, COBOL applications, SAP ERPs, and Epic Systems subject to HIPAA and audit controls.
- Kawtar El Maghraoui: “OpenAI and Anthropic needs IBM's distribution and consulting pipes far more than IBM needs their specific model checkpoints.”
- Mihai Criveti: “Different large language models and models providers offer multiple models with different costs, different latency as in time to first token, different performances in things like tokens per second, different performance as in the outcome.”
`10:00` Enterprise Transformation, Model Agnosticism, and the Consulting Advantage
The discussion opens around the realization that enterprise AI adoption is not being bottlenecked by a shortage of intelligent models. Instead, the real enterprise roadblocks are change management, systems integration, and dealing with dirty data. IBM Consulting is effectively weaponizing models from OpenAI and Anthropic to accelerate legacy digital transformation—converting old codebases and automating high-volume banking workflows at scale. The panel highlights a recurring debate: if models become commoditized and open-source models proliferate, value migrates elsewhere. Previously, the hosts debated whether raw compute would capture all the value; now, it appears the other side of that coin is the trusted consultant who guides enterprise implementations.
Aaron reflects on the emerging constellation of high-profile partnerships, pointing out three anchors that define the current landscape: IBM, Anthropic, and OpenAI. The Anthropic alliance, announced in October 2025, was set up to overhaul software development and integration, out of which internal initiatives like IBM Bob emerged. Combining that with OpenAI tools gives thousands of practitioners using IBM Consulting Advantage a full-stack capability to meet clients on their terms. As Aaron summarizes:
> “So we're taking the full stack of the business with these two partnerships.”
`11:15` Stripe’s $7 Billion Deal to Acquire OpenRouter
```mermaid
flowchart TD
accTitle: OpenRouter Enterprise Gateway Architecture
accDescr: Shows client apps connecting via virtual API keys through an orchestration layer to underlying AI models and tool servers.
subgraph Clients["Clients & Applications"]
App["Enterprise Apps & Agents"]
end
The hosts pivot to breaking market news: payment-processing giant Stripe is finalizing an agreement to acquire OpenRouter for over $7 billion. This valuation caused widespread shock across the tech industry because it represents more than a fivefold markup over OpenRouter’s $1.3 billion valuation recorded just 82 days prior. For audience members accustomed to hearing only about frontier model developers, OpenRouter might seem like an unfamiliar candidate for a multi-billion-dollar acquisition.
Mihaly jokes about the massive valuation, quipping that having previously written and given away his own model gateway, Model Context Protocol (MCP) gateway, and agent-to-agent (A2A) gateway, he is left *“rethinking his life choices.”* On a serious note, he argues that the enterprise architecture is heading toward an entirely heterogeneous paradigm:
> “I believe the future is multi-modal, multi-model, multi-agent, multi-framework, multi-harness, multi-cloud, multi-everything; there is no one size fits all for models, there is no one model which is good, and you can't even get enough of one model to fit every use case.”
A unified routing gateway like OpenRouter resolves this by offering:
- Direct access to an extensive variety of foundation models across multiple vendors through a consolidated interface.
- Semantic caching and workload optimization to cut latency and redundant inference runs.
- Semantic routing that automatically pushes incoming requests to the optimal model for a given task.
- Enterprise FinOps and cost controls to enforce budget caps and track consumption.
- Virtual API keys paired with role-based access control (RBAC), preventing catastrophic leaks of master upstream API keys while preserving centralized telemetry and observability.
`14:48` The Toll-Booth Model: Why Stripe Wants AI Infrastructure
Addressing why an e-commerce and subscription payment processor would drop billions on an AI model router, Kaltar argues that raw frontier models suffer from steep price competition, whereas the routing layer functions as a lucrative toll booth. Token prices have dropped more than 90%, and developing frontier models is a relentless, capital-intensive race to the bottom. In contrast, the intermediary gateway extracts a toll on every single API request regardless of which model sits atop the benchmark leaderboards.
Furthermore, traditional software billing—charging a static $30 per month per user to a corporate credit card—completely breaks down in the AI era. Modern generative systems handle millions of unpredictable micro-tokens flowing continuously. OpenRouter operates directly in the execution path, allowing billing and metering to happen precisely where the code runs. Looking ahead to agentic systems, autonomous agents cannot sign enterprise contracts or swipe plastic cards; they require digital wallets to budget their own inference costs, deciding whether to call a cheap $0.05 model for a quick query or pay $0.50 for a deliberate reasoning model. Stripe is positioning itself to become the Visa of autonomous agent software. In addition, unlike hyperscalers such as AWS, Microsoft Azure, and Google Cloud, Stripe is provider-neutral and does not care where compute is hosted.
`16:48` Commoditization Risks and Agentic Commerce
The panel weighs whether the industry has arrived at a "post-model" era where the spotlight permanently shifts away from model makers toward routers, gateways, and consulting firms. Aaron observes that market attention naturally cycles between layers of the stack, noting that the immediate heartbeat of the industry is focused on routing and optimization:
> “Right now the heartbeat is around routing.”
However, Aaron cautions that routing technology faces rapid commoditization. The gateway and routing space is already congested with alternatives and competitors, including:
- LiteLLM
- Cloudflare
- Portkey
- Helicone
- Together AI
- Fireworks AI
- Context Forge (Mihaly’s own build)
Aaron suggests Stripe’s massive bid was driven primarily by explosive metrics—specifically noting token throughput jumping roughly 5x within six months, scaling from 5 trillion to 25 trillion tokens. Stripe is placing a massive bet on agentic commerce, anticipating that autonomous AI agents will be the core economic drivers of tomorrow's transactions.
Mihaly reinforces that this purchase is not merely an acquisition of open-source routing code; Stripe is buying OpenRouter’s vast community and customer footprint. While exact counts vary across reporting, OpenRouter boasts between 8 million and 10 million total registered users across free and enterprise tiers, alongside an estimated 150,000 active, paying accounts with established billing relationships and direct software integration.
`20:00` Stripe’s Strategic Bet on OpenRouter and the Open Gateway Ecosystem
The panel examines OpenRouter’s lean operational structure relative to its outsized enterprise valuation. With a headcount estimated between 50 and 100 people, its implied valuation equates to roughly $70 million per employee. Aaron points out that while Mihai is correct in framing this as an aggressive bet on future growth, OpenRouter already possesses the fundamental infrastructure to scale along that trajectory—provided commoditization does not erode their advantage.
The hosts compare Stripe’s acquisition to the classic "tall boots" or toll booth analogy: model providers are currently bleeding cash competing against one another on benchmarks, whereas Stripe positions itself right on the bridge. By controlling the gateway, Stripe collects a few cents on every vehicle driving across as token volumes surge, securing a reliable transaction layer regardless of which specific frontier model wins.
```mermaid
flowchart LR
accTitle: Stripe Bridge Strategy
accDescr: Model makers compete and bleed cash while Stripe collects toll on routing tokens to end consumers.
subgraph Bridge["Control & Routing Layer"]
OR["OpenRouter / Stripe Gateway<br/>(Collects toll per token)"]
end
subgraph Consumption["Enterprise Workloads"]
Agents["AI Agents & Apps"]
end
Agents -->|"Token Requests"| OR
OR -->|"Dynamic Routing"| Models
```
When asked whether open-source AI gateways will eventually eclipse commercial offerings, Mihai explains that open source is already saturated, with over 100 implementations existing in various forms. However, standard API gateways are only an entry-level building block. Enterprise demands have moved past simple model-calling toward comprehensive agent control planes that provide semantic routing, permission caching, and unified policy management across diverse assets.
- OpenRouter's valuation reaches an estimated $70 million per employee with a team of just 50 to 100 people.
- The open-source market already contains more than 100 alternative gateway implementations.
- Enterprises require unified agent control planes covering governance, trust, observability, FinOps, cost control, semantic routing, and MCP (Model Context Protocol) server integrations.
- Mihai argues: “AI gateway or model gateway is just a small component of the capabilities enterprises need, whether they’re using open source or commercial software, to provide governance, trust, observability, optimization, cost control, FinOps, run and manage capabilities for their entire AI fleet.”
---
`22:20` Ramp’s AI Index and the Reality of AI Spend Disparities
The discussion shifts to Ramp’s August 2026 AI Index, which leverages transaction data from approximately 70,000 businesses to assess real-world enterprise adoption. While the share of businesses using open-source model serving and inference platforms increased from 4.5% to 6.1%, the broader spending patterns reveal deep skew across corporate tiers.
Aaron details the stark variance in corporate spend: the top 1% of AI-adopting companies spend roughly $7,400 per employee per month on AI tools. For the top 10%, that figure falls steeply to $650 per month, while the median corporate spend across the dataset rests at a mere $12 per month. This massive delta indicates that while companies want to provide baseline access, CFOs are aggressively metering usage to prevent runaway costs from digital workers and autonomous agent runs. In elite development shops, monthly AI tool costs are beginning to rival the actual base salaries of the software engineers using them.
- Aaron observes: “What I found and just by looking at it, it looks as though AI adoption is it’s really uneven.”
- Aaron highlights the looming labor-cost parity: “What I find is interesting, right, is that the top 1%, right, it’s starting to approach the cost of AI per employee, starting to approach the actual salary of the said engineers that are using these tools.”
---
`25:00` The Death of the FOMO Tax and Shift to Unit Economics
The hosts declare that the "AI FOMO tax" era of 2023–2025 has officially ended. During those early boom years, CFOs rubber-stamped expansive, unmetered departmental AI software budgets out of dread that their company would fall behind the competition. In 2026, unmetered experimentation has been replaced by mandatory unit-economic accountability. Enterprise tools must now demonstrate concrete ROI through head-count reduction, accelerated engineering cycles (such as pull requests merged per developer), or direct gross revenue expansion.
Meanwhile, early wrapper tools focused on narrow tasks like generic copy generation are suffering catastrophic churn because frontier model updates and native browser or operating system features perform those tasks natively at zero incremental cost. The panelists argue that evaluating AI spend on a legacy "per-seat" SaaS framework is obsolete. Spend must be evaluated on cost-per-unit-of-work: for instance, watching human customer support resolution drop from $18 down to a few cents with AI, or contract auditing costs plummeting from $400 to $12.
- The industry is pivoting from per-seat license metrics toward cost-per-unit-work benchmarks (e.g., cost per resolved support ticket or cost per audited legal contract).
- Engineering teams remain the heaviest daily consumers of high-cost AI resources, prompting organizations to segment tooling budgets heavily by department rather than rolling out uniform enterprise licenses.
- One speaker summarizes the shift: “In 2026, the era of the unmetered experimentation is kind of officially dead, and CFOs now demand kind of measurable unit economic payback.”
---
`27:30` "Random Acts of AI" and Enterprise Control Plane Backlash
Addressing whether skepticism about an AI bubble is justified, Mihai points out that enterprise friction does not come from a lack of utility, but from total operational chaos. Over the past two years, enterprise client requests have inverted. Instead of asking how to build their first proof of concept, companies are now overwhelmed by what Mihai terms "60 random acts of AI."
Organizations have accumulated dozens of unmonitored autonomous agents across fragmented frameworks—LangChain, LangGraph, AutoGen, AutoGPT, CrewAI, Microsoft Copilot Studio, and Watsonx Orchestrate. Token budgets are rarely burned by knowledge workers asking ChatGPT simple questions; the true cash drain comes from software developers running 50 background terminal sessions or connecting high-throughput automated event streams straight into expensive frontier reasoning models without governance or rate limiting. Layered on top of these unbudgeted cloud bills is regulatory liability, as compliance teams now have to account for massive potential fines under frameworks like the EU AI Act within their overall total cost of ownership.
- Mihai details the shift in client conversations: “Over the course of the last two years, most of the client conversations I’ve had have shifted from, help us use AI or, you know, we want to buy our product or we want to do something with AI. We’ve done a POC with AI to help. We’ve done too much AI. We’ve got 60 random acts of AI.”
- Mihai describes the source of enterprise consumption: “It’s usually not asking ChatGPT for advice on one or two things that’s really burning down the tokens; it’s these hundreds of agents the developers and everybody else, for you know, they have 50 open terminals or they’ve got an event stream going into one of these agents and consuming them at scale and they use the most powerful and the most expensive models.”
`30:00` Total Cost of Ownership and the AI FinOps Era
The panel delves into the hidden financial realities of enterprise AI adoption, highlighting that total cost of ownership extends far beyond basic tokenomics and invoice totals from model providers. As organizations deploy more models across disparate teams, they face mounting expenses related to management, personnel, infrastructure, governance, and security. Companies currently lack mature budgeting mechanisms, quota enforcement systems, and frameworks to directly correlate token usage with measurable return on investment. Mihai points out that his work on watsonx Orchestrate is specifically targeted at bridging this gap—aligning ROI, model expenses, guardrails, and operational overhead. He sums up the state of the industry: “Getting control over AI and having predictable costs and having predictable outcome and a good return on investment is proving extremely difficult in these early days.”
```mermaid
flowchart TD
accTitle: The AI FinOps Challenge
accDescr: Diagram illustrating how falling unit token costs drive increased enterprise adoption, resulting in cost runaway and the need for FinOps controls.
subgraph Drivers["Market Dynamics"]
A["Cheaper Token Unit Costs"] --> B["13x Token Spend Splurge"]
B --> C["Proliferation of Autonomous Agents"]
end
subgraph Operational_Overhead["Hidden TCO Overhead"]
C --> D["Governance & Security"]
C --> E["Infrastructure & Personnel"]
C --> F["Lack of Quota Enforcement"]
end
subgraph Resolution["FinOps & Orchestration"]
D & E & F --> G["Enterprise AI FinOps"]
G --> H["Predictable ROI & Guardrails"]
end
```
A puzzling paradox is also playing out across enterprises: even though individual model costs continue to drop and capabilities rise, aggregate spending is exploding. The Ramp AI Index report demonstrates that token spending increased 13x over the past year. Organizations fall into an "AI splurge" mindset, assuming lower unit costs mean they should deploy far more autonomous agents, which then run unchecked and compound budget overruns. The panel compares this moment directly to the maturation of cloud computing between 2012 and 2015, when enterprises finally stopped mindlessly running idle EC2 instances and instituted cloud FinOps practices. Enterprise AI is now entering that exact phase, forcing companies to move from mindless consumption to spending smarter. Engineering stands out as the single department where surging AI expenditure faces virtually zero corporate pushback; tools like Cursor, Claude Code, and specialized code-review models have successfully transitioned from novel experiments into indispensable core developer infrastructure.
- Token spend across organizations surveyed in the Ramp report grew 13-fold over the preceding year.
- Enterprise cloud history from 2012–2015 is repeating itself as companies replace unchecked experimentation with strict AI FinOps.
- Developer-focused tools such as Cursor and Claude Code have achieved justifiable, institutionalized status, avoiding the cost-cutting scrutiny hitting other business units.
---
`33:10` AI-Generated "Slop" in Congressional Legislation
To wrap up the episode, the hosts pivot to a recent *Politico* report revealing that the House of Representatives' Office of the Legislative Counsel is being inundated with AI-generated "slop" bills. This echoes earlier linguistic studies of the UK Houses of Parliament, which uncovered an explosion of distinct ChatGPT-style phrasing in speeches and records shortly after the tool debuted. The panelists joke about drafting a "load-bearing bill" stuffed with em dashes and smart quotes, observing that government staffers are taking the same shortcuts that have already flooded schools, academic publishing, customer support, and workplace Slack and Microsoft Teams channels.
Rather than staging a futile attempt to ban AI usage in government drafting, the panel advocates for comprehensive education and institutional guardrails. Staffers must understand that models hallucinate and make severe errors, meaning human authors remain strictly accountable for final outputs. The panel stresses that workers should not be pasting sensitive legislative materials into consumer web interfaces and blindly copying back whatever text is outputted. Instead, institutions need clear operational guidelines specifying what is allowed, what constitutes safe usage, and how to harness AI as an efficiency tool without outsourcing the actual authorship of the law. Mihai concludes the discussion: “The legislation is no different than any other field. And rather than trying to fight it, there needs to be clear guidelines and rules on what's allowed, what's not allowed, what's safe, what isn't safe, and empowering folks to use it in a smart way, as opposed to having AI write law.”
## Notable quotes
> `05:00` OpenAI and Anthropic needs IBM's distribution and consulting pipes far more than IBM needs their specific model checkpoints. — Kawtar El Maghraoui
> `00:00` We're betting that the enterprise AI value, it's shifting upwards from owning the model to now orchestrating, governing, and integrating, and even, you could argue, operationalizing these types of models. — Aaron Baughman
> `30:00` Getting control over AI and having predictable costs and having predictable outcome and a good return on investment is proving extremely difficult in these early days. — Mihai Criveti
## Mentioned
Aaron Baughman, Anthropic, AutoGen, AutoGPT, AWS, Claude Code, Cloudflare, Context Forge, CrewAI, Cursor, Epic Systems, Fireworks AI, Google Cloud, Helicone, IBM, IBM Bob, IBM Consulting, IBM Consulting Advantage, IBM Granite, Kawtar El Maghraoui, LangChain, LangGraph, LiteLLM, McKinsey, Microsoft Azure, Microsoft Copilot Studio, Mihai Criveti, OpenAI, OpenRouter, Politico, Portkey, Ramp, SAP, Stripe, Tim Huang, Together AI, UK Houses of Parliament, US House Office of the Legislative Counsel, watsonx Orchestrate, watsonx.governance
## Video description
Visit Mixture of Experts podcast page to get more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts This week on Mixture of Experts, we talk partnerships, purchases, and predictions surrounding the AI industry's financial infrastructure. But first our panelists dive into IBM's newly announced collaboration with OpenAI, under which IBM Consulting will train and certify consultants on OpenAI's tools to bolster the company’s push toward enterprise deployments. Then we turn to...
IBM panelists argue that enterprise AI value is shifting away from owning foundation model checkpoints toward orchestration, systems integration, and FinOps routing layers. Strategic alliances, high-multiple gateway acquisitions like Stripe's purchase of OpenRouter, and enterprise spend data demonstrate that managing cost, governance, and operational friction now outweighs raw benchmark leads.
Value Migration to Enterprise Integration and Model Agnosticism (0:00)
Rather than picking a single frontier model winner, enterprise adopters are turning toward model-agnostic systems integrators to overcome distribution bottlenecks, strict API throughput limits (such as 50 to 200 requests per second), and legacy enterprise integration hurdles. Aaron Baughman notes that enterprise value is moving up the stack into orchestration, governance, and deployment. IBM illustrates this by pairing Anthropic with internal software engineering through initiatives like IBM Bob, while aligning its OpenAI partnership directly with client-facing work in IBM Consulting.
Kawtar El Maghraoui asserts that frontier model providers need consulting and distribution channels far more than integrators need specific checkpoints, especially since around 80% of enterprise value remains locked in mainframes, COBOL codebases, SAP ERPs, and HIPAA-regulated healthcare systems. Through platforms like watsonx.governance, enterprises are prioritizing auditability, bias monitoring, and regulatory compliance, enabling them to hot-swap frontier models and leverage smaller models like IBM Granite for routine workloads rather than bearing custom pre-training expenses.
The Toll-Booth Model and Stripe's OpenRouter Bet (11:15)
Stripe's reported agreement to acquire model router OpenRouter for over $7 billion—more than a fivefold markup over its $1.3 billion valuation 82 days earlier—highlights the strategic value of the AI routing and transaction layer. The panel characterizes gateways as toll booths: while model makers burn capital competing on rapidly falling token prices, gateways extract micro-fees on queries regardless of which provider wins benchmark races. Operating directly in the execution path, OpenRouter provides semantic caching, model routing, budget caps, and virtual API key management across proprietary and open-source models.
Kawtar El Maghraoui and Aaron Baughman explain that traditional $30 per month per seat billing models fail under volatile micro-token usage, positioning Stripe to act as a payment rail for autonomous AI agents that require digital wallets to budget inference calls. Despite an implied valuation near $70 million per employee across a 50-to-100 person team, Baughman cautions that standard routing technology faces commoditization from competitors like LiteLLM, Portkey, and over 100 open-source gateway projects. Stripe's valuation premium is largely seen as an acquisition of OpenRouter's user base of 8 to 10 million registered accounts and roughly 150,000 active paying customers.
Uneven Enterprise Spend and the AI FinOps Era (22:20)
Transaction data from Ramp's August 2026 AI Index across roughly 70,000 businesses reveals stark disparities in enterprise spend. The top 1% of adopting companies spend approximately $7,400 per employee per month on AI tools, approaching software engineer base salaries. The top 10% average $650 monthly, while median spend sits at just $12 per employee per month. Unmetered corporate experimentation seen in earlier boom cycles has ended, replaced by CFO-enforced unit economics that evaluate tooling by cost-per-unit-of-work—such as customer support resolutions or contract audits—rather than per-seat software licenses.
Although individual token prices continue to decline, aggregate enterprise token spending surged 13-fold over the prior year, mirroring the cloud infrastructure expansion of 2012–2015. Mihai Criveti describes companies dealing with '60 random acts of AI,' where runaway background agent terminal sessions, event streams, and overlapping frameworks drive unexpected costs. This operational sprawl has made software developer tools like Cursor and Claude Code institutionalized staples, while elevating AI FinOps, quota enforcement, and compliance monitoring under regulations like the EU AI Act into urgent business priorities.
Institutional Guardrails and AI in Government Work (33:10)
Reports of consumer AI tools generating low-quality legislative draft submissions for the U.S. House of Representatives' Office of the Legislative Counsel illustrate how unchecked usage reaches public institutions. Similar to patterns identified in the UK Parliament, workplace chat applications, and academic publishing, government staffers are leveraging consumer models to generate text without adequate oversight.
The panel argues that blanket bans on AI drafting are impractical. Instead, organizations must implement clear institutional guardrails, education on model hallucination risks, and operational guidelines that prevent sensitive text from being fed into consumer interfaces, ensuring human authors remain strictly accountable for official documents.
As enterprise AI matures beyond experimentation, long-term economic advantages appear to favor integration platforms, governance layers, and financial rails over raw model developers. The unresolved challenge remains establishing reliable FinOps frameworks and operational controls that convert rapid token volume growth into predictable corporate return on investment.
## Background
本期节目由 IBM 官方播客频道《Mixture of Experts》于 2026 年 8 月 21 日在 RSS 及各大网络平台发布。根据页面简介，本期节目聚焦人工智能产业金融基础设施领域的伙伴关系、并购案与未来预测，开篇重点剖析了 IBM 与 OpenAI 最新达成的战略合作——IBM Consulting 将全面培训并认证咨询顾问以加速 OpenAI 工具在企业级的落地。
## TL;DR
IBM 专家团队深入探讨了当前企业级 AI 市场的重大结构性转移。核心话题包括 IBM 继结盟 Anthropic 后再度携手 OpenAI，意在夺取多模型企业控制平面而非押注单一模型；Stripe 以逾 70 亿美元收购 OpenRouter，标志着商业价值从底层模型研发向过路费式的路由与 Agent 支付基础设施倾斜。此外，嘉宾们基于 Ramp 2026 AI 指数分析了企业从盲目实验向严苛 FinOps 及单项工作成本（cost per unit work）核算的转变，并讨论了多智能体失控乱象以及美国国会遭遇 AI 劣质法案草案冲击的治理挑战。
## Atmosphere & style
整场对谈节奏明快且充满技术实战派的犀利与自嘲，主持人和三位常驻专家展现出极高的默契与思辨张力。交流氛围既有大厂 Fellow 与 CTO 纵览全局的商业洞察，又穿插着对开发者亲手写网关却见证 OpenRouter 暴富的幽默调侃，以及对硅谷盲目卷基准跑分脱离企业遗留系统现实的冷峻解构。整体听感兼具深度智力碰撞的快感与工业界一线的务实清醒，从技术架构的宏大叙事层层递进到企业财务账本与现实治理的残酷细节。
## Bulletin
- `00:00` IBM 继结盟 Anthropic 之后，正式宣布与 OpenAI 达成战略合作，IBM Consulting 将全面培训并认证咨询顾问，旨在通过前线部署工程师（FDE）掌控企业级 AI 控制平面，而非在单一基础模型之间押注输赢。
- `11:30` Stripe 正推进以逾 70 亿美元估值收购多模型路由平台 OpenRouter，该平台 Token 吞吐量在约半年内从 5 万亿暴涨至 25 万亿，反映出资本正从基础模型价格战快速转移至掌控自主 Agent 计费与结算入口的“过桥收费站”。
- `22:54` Ramp 2026 年企业支出报告揭示出剧烈的“K 型断层”：尽管过去一年企业 Token 支出总额激增 13 倍，但排名前 1% 的企业人均 AI 月支出高达 7,400 美元，而企业中位数仅为 12 美元，多数企业已踩下刹车并设定严苛的用量配额。
- `25:00` 企业 AI 采购正式告别交纳“错失恐惧税”（FOMO tax）的无节制实验阶段，传统 SaaS 按席位收费模式正被“单项工作成本”取代，企业管理层全面转向要求可衡量的单位经济回报与 FinOps 治理。
## Key takeaways
- `00:00` IBM 采取双轨兼收策略：Anthropic 深度对齐软件工程与研发工具链（如 IBM Bob），OpenAI 赋能 IBM Consulting 进行前线交付，自有开源 Granite 负责轻量高频请求，从而构建中立的控制平面。
- `05:00` 企业落地大模型的核心瓶颈不是模型跑分，而是遗留系统（如主机 COBOL、SAP、大型 ERP）的数据集成、并发限流（Rate Limits）以及数据驻留合规约束。
- `10:00` 大模型稀缺性正在快速消退，未来商业决胜的关键在于具备遗留系统改造能力、全链路合规审计与信任交付能力的系统集成商与咨询机构。
- `11:30` OpenRouter 仅用不到三个月估值即从 13 亿美元跃升至超 70 亿美元，印证了异构多模型、多 Agent 架构下统一 API、语义缓存与虚拟密钥网关的刚性价值。
- `14:40` 底层模型陷入 Token 单价骤降超 90% 的价格战，而卡位在调用入口处的路由层能够从全网调用中稳定抽取“过路费”，成为商业模式更稳健的收税站。
- `14:40` Stripe 收购 OpenRouter 意在构建 Agent 经济的基础金融设施，为能够自主决定模型开销、全天候产生动态微额 Token 交易的软件智能体提供“独立钱包”。
- `16:40` 路由工具本身面临快速同质化竞争（如 LiteLLM、Portkey 等），Stripe 高估值收购的核心壁垒在于 OpenRouter 汇聚的数百万开发者与即刻变现的高频商业流水。
- `20:00` 开源与商业网关生态正在从简单的 API 请求转发器演进为集成权限控制、MCP 协议调度与语义路由的完整“智能体控制平面”。
- `25:00` 营销文案类简单套壳应用因基础模型免费原生化遭遇断崖式流失，企业评估重心转向客服工单与合同审核等业务流的“单次工作交付成本”。
- `27:38` 企业内部正被无节制的“随机 AI 行为”瘫痪，工程师滥用复杂 Agent 框架并默认接入最昂贵旗舰模型，加之《欧盟人工智能法案》落地，推高了合规与全生命周期拥有成本（TCO）。
- `30:00` 行业正经历类似 2012-2015 年云计算成熟期的结构演变，从无节制消耗资源进入 AI FinOps 阶段，唯有工程开发类核心工具（如 Cursor、Claude Code）展现出高刚性支出韧性。
- `33:45` 美国国会与英国议会相继被 AI 生成的劣质立法草案淹没，表明治理 AI 泛滥的关键在于建立透明的规范边界与企业级安全通道，而非一味封杀。
## Detailed walkthrough
### `00:00` IBM 与 OpenAI 达成战略合作：中立编排层与咨询服务新格局
本期播客由 Tim Huang 主持，常驻技术专家小组包括 AI Native Systems 首席研究科学家 Kauter El Maghraoui、IBM Fellow Aaron Baughtman，以及 Watsonx Orchestrate 首席技术官 Mihai Criveti。节目开篇抛出了当周科技圈的几大热点动态，包括 Stripe 收购模型路由平台 OpenRouter、企业费用管理平台 Ramp 发布的最新 AI 业务洞察报告，以及美国国会探索利用 AI 辅助立法的工作进展。但首先引发全场热议的，是 IBM 刚刚宣布与 OpenAI 达成的重磅合作。
根据这项合作协议，IBM Consulting 将大规模培训并认证咨询顾问，在内部建立专门的 OpenAI 业务实践（OpenAI practice），直接服务企业客户的规模化落地。Tim 敏锐地指出，不久前 IBM 才刚与 Anthropic 达成了深度战略结盟。在传统认知中，行业通常将模型之争视作你死我活的阵营对抗——犹如黑帮电影里的“西区故事”（Sharks and Jets）。然而，IBM 这次接连拥抱两大领军者的举措表明，企业级 AI 服务市场正在进入一个更加务实、兼收并蓄（ecumenical）的新阶段，集成商和咨询顾问不再排他性地押注单一供应商。
Aaron 对此表现得非常振奋，他认为 OpenAI 顶尖的模型能力结合 IBM 深入财富 500 强乃至更广阔企业核心的落地能力，是一套极具杀伤力的组合拳。在这一合作框架下，IBM 将派出精锐的“前线部署工程师”（Forward Deployed Engineers, FDE）和前线作战单元直接深入客户业务现场。Aaron 鲜明地指出，IBM 的核心策略绝非在底层基础模型之间二选一，而是要掌控企业级控制平面的话语权：
> “This is a big signal that we, IBM, we're not trying to pick a model winner. We're actually working and we're going to become the owner of the enterprise AI control plane here. We're betting that the enterprise AI value, it's shifting upwards from owning the model to now orchestrating, governing, and integrating, and even, you could argue, operationalizing these types of models.” —— Aaron Baughtman
这种双轨战略在 IBM 内部有着明晰的分工定位：
- **Anthropic 阵营：** 深度对齐 IBM 软件工程与开发生态，核心用于支撑像 IBM Bob 这类代码生成与研发工具链。
- **OpenAI 阵营：** 深度赋能 IBM Consulting，通过前线部署工程师帮助各大跨国企业客户进行复杂的定制集成。
- **自有 Granite 系列模型：** 保持开源与自主可控，主打轻量、低成本与高效。复杂的任务规划与深度推理交给 OpenAI 等大模型，而大量高频简单的请求则直接路由至 Granite，从而为客户提供真正的中立网关与选择权。
```mermaid
flowchart TD
accTitle: IBM Enterprise AI Control Plane
accDescr: 架构图展示了 IBM 如何通过企业级控制平面集成不同模型并分配至不同业务场景
subgraph ControlPlane["IBM 企业级控制平面 (Model Gateway & Control Plane)"]
direction TB
Orch["任务编排与治理 (Watsonx)"]
Route["智能模型路由 (Router)"]
end
subgraph Models["模型供给层 (Model Providers)"]
OAI["OpenAI (复杂规划 / 推理)"]
ANT["Anthropic (软件工程生态)"]
GRN["IBM Granite (轻量低成本 / 基础请求)"]
end
subgraph Execution["业务与服务交付 (Delivery)"]
Consulting["IBM Consulting (前线部署工程师 FDE)"]
Software["IBM 软件工程 (IBM Bob 等)"]
Client["财富 500 强企业私有业务场景"]
end
ControlPlane --> Models
Models --> Execution
```
### `05:00` 企业落地的真实瓶颈：工程算力墙、成本优化与遗留系统管道
面对主持人关于“各方都在全栈竞争、OpenAI 为何广签渠道（如早先的自建部署团队及麦肯锡合作）”的追问，Mihai 从工程实战角度给出了直接回应：OpenAI 极度渴望系统集成商（System Integrators）这支具备地面作战能力的队伍。在现实商业世界中，“没有所谓的万能最佳模型”（there is no such thing as the best model），模型优劣完全取决于多维指标的权衡：
```
[首字延迟 (Time to First Token)] vs [输出吞吐 (Tokens/s)] vs [任务效果 (代码 / 摘要 / 检索)]
```
此外，企业在放大 AI 部署时往往会遭遇残酷的现实障碍：
- **数据合规与地域限制：** 许多受监管行业要求数据驻留本地。若伦敦数据中心没有对应的模型实例，客户别无选择，只能使用本地合规模型。
- **突发并发与吞吐限制（Rate Limits）：** 企业管理者常天真地以为“明天就可以花钱给全公司 10 万名员工开通 AI”。现实却是不仅算力配额不支持（例如限流在 50、150 或 200 RPS），即便挥舞支票也买不到足够的实时推理算力。此时必须依赖多模型路由架构，并针对不同模型动态重构 Prompt。
Aaron 进一步补充了经济账问题。假定每位员工每月消耗 10 美元 Token 费用，十万量级员工每月的开销便高达数百万美元，长此以往不可持续。通过企业级伙伴关系，不仅能获得前沿模型首发能力，还能拿到大规模批量采购的折扣底价，在优化匹配任务的同时拉低整体运营成本。
紧接着，Kauter 一针见血地指出了硅谷研发与企业现实之间的巨大脱节。硅谷往往沉迷于跑分竞争，以为学术基准提升一两个百分点就能征服市场，但真正的企业价值沉睡在错综复杂的传统系统之中：
> “In the real world, 80% of enterprise value is locked inside legacy mainframe COBOL databases, maybe SAP ERPs, Epic healthcare systems, HIPAA compliance vaults, et cetera. And IBM owns those enterprise relationships. OpenAI and Anthropic need IBM's distribution and consulting pipes far more than IBM needs their specific model checkpoints.” —— Kauter El Maghraoui
Kauter 总结强调，IBM 根本无须像微软、谷歌或 Meta 那样砸进 200 亿美元去卷底层通用基座模型的训练，其制胜之道在于打造可信赖的、中立的企业编排中枢：
- 依托 Watsonx.governance 提供全链路审计跟踪（audit trails）、偏见检测（bias detection）、Prompt 行为审计以及监管合规防护；
- 允许企业根据成本、响应速度和准确率指标，自由“热插拔”（hot swap）不同的底层模型；
- 打通企业老旧系统（Mainframe/COBOL/SAP）与先进 AI 之间的管道，解决企业落地中真正的系统集成瓶颈。
### `10:00` 企业落地的真正瓶颈：咨询巨头、大模型星座联盟与信任中介
企业在部署生成式 AI 时，真正的瓶颈早已不是缺乏高智商模型，而是变更管理（Change Management）、深层次的系统集成以及应对遗留系统中的“脏数据”。IBM Consulting 正在将 OpenAI 与 Anthropic 的前沿模型转化为“武器库”，加速企业数字化转型，例如大规模转换银行老旧代码库并重构金融业务流。这引发了行业层面的深刻思辨：如果人人都唾手可得优秀的开源模型与顶尖商用模型，模型的稀缺性大幅下降，那么未来的决胜点究竟是算力，还是具备落地交付能力的“受信咨询机构”？
Aaron 梳理了当前 IBM 构建的“星座级合作阵列”（Constellation of Partnerships）。早在 2025 年 10 月左右，IBM 就宣布了与 Anthropic 的战略联盟，旨在借助 AI 重塑企业软件开发与系统集成生命周期，催生了 IBM Bob 等核心工具；而如今与 OpenAI 的全面合作，则进一步通过 IBM Consulting Advantage 平台对数千名顾问进行技能认证。谈及这一双重合作带来的全栈布局，Aaron 满怀信心地强调：“We're taking the full stack of the business with these two partnerships. And I'm really excited about the future and what we're going to do together within IBM.”
- **企业落地核心卡点**：变更管理、异构系统深度集成、数据清洗与治理。
- **核心落地场景**：遗留代码库自动化翻新重写、银行金融核心工作流端到端自动化。
- **合作星座架构**：IBM 联手 Anthropic（重塑软件工程）与 OpenAI（武装一线咨询顾问），形成覆盖企业全技术栈的落地交付矩阵。
---
### `11:30` 70 亿美元并购案：OpenRouter 为何成为多模型时代的统一网关
科技界曝出震撼级交易：支付处理巨头 Stripe 正在敲定以超过 70 亿美元的价格收购 AI 模型路由平台 OpenRouter。令人瞠目的是，就在短短 82 天之前，OpenRouter 的估值仅为 13 亿美元——短短三个月不到溢价超过 5 倍。对于只紧盯 OpenAI、Anthropic 诸如模型基准跑分的普通大众来说，OpenRouter 也许相对陌生，但这一惊人估值直接折射出底层基础设施市场的疯狂重估。
针对 OpenRouter 究竟凭什么值 70 亿美元，Mihaly 开玩笑自嘲道，作为一个亲手写过模型网关、MCP（模型上下文协议）网关和 A2A（Agent-to-Agent）网关还全免费开源出去的开发者，“I'm kind of rethinking my life choices.”但玩笑归玩笑，Mihaly 强调企业级 AI 的未来绝非单模型垄断，而是极其复杂的异构生态：“I believe the future is multi-modal, multi-agent, multi-framework, multi-harness, multi-cloud, multi-everything; there is no one size fits all for models.”
在真实业务中，单一模型永远无法覆盖所有场景或承载无限吞吐。像 OpenRouter 这样的网关具备极其刚性的不可替代性：
1. **统一模型接入与语义优化**：单一接口直接调度全网主流提供商模型，内置语义缓存（Semantic Caching），显著压降重复请求延迟与开销。
2. **企业级全栈网关功能**：统一接入并管理 MCP 服务端、自主 Agent、工具技能库、自动化评测（Evals）及语义路由（Semantic Routing）。
3. **FinOps 与虚拟鉴权体系**：实施预算上限管控与角色型访问控制（RBAC）。企业员工不再直接接触并暴露原始 API 密钥，彻底杜绝密钥泄漏失窃风险，同时集中沉淀全生命周期的可观测性数据与调用度量指标。
```mermaid
flowchart TD
accTitle: OpenRouter企业网关架构
accDescr: 展示OpenRouter作为统一网关在企业调用、安全控制与异构多模型调度之间的中间层作用
subgraph ClientLayer ["企业接入层"]
A["企业业务系统 / AI Agents"]
end
subgraph GatewayLayer ["OpenRouter 统一中间层"]
B["虚拟 API 密钥与 RBAC 权限体系"]
C["语义缓存 & FinOps 成本熔断"]
D["智能语义路由 & MCP 服务网关"]
B --> C --> D
end
subgraph ModelLayer ["多模型与异构云端"]
E["OpenAI / Anthropic"]
F["开源大模型 (Llama, DeepSeek等)"]
G["私有云 / GPU 集群部署"]
end
A -->|"受控请求"| B
D -->|"轻量查询"| E
D -->|"低成本/高吞吐"| F
D -->|"本地合规推理"| G
```
---
### `14:40` 收税站比卖金矿赚钱：Stripe 押注 Agent 经济与自主钱包
为什么一家处理在线信用卡支付的金融巨头 Stripe 会收购 AI 路由层？Kaoutar 给出了极其犀利的洞察：“Models are cheap. The toll booth is where the money is.”在基准评测竞赛中，模型研发是一场昂贵且残酷的竞次淘汰赛（Race to the bottom），每过半年模型智商翻倍而推理成本骤降，Token 单价已狂跌超过 90%；然而，不论哪家模型在排行榜称霸，守在收费站卡位（Toll Booth）的路由网关都能从每一次调用中抽成过路费。
更关键的逻辑在于 AI 计费方式对传统金融支付系统的根本性颠覆。传统 SaaS 软件是按人头每月刷信用卡支付 30 美元固定订阅费；但 AI 时代完全不同，是数以百万计、全天候 24 小时动态波动的不可预测的微额 Token（Micro-tokens）流。OpenRouter 能在代码执行的最前端实时处理计量与计费。面向未来，自主 AI Agent 正在接管工作并需要自己的“独立钱包”，它们无法签署纸质合同或手刷物理卡：一个 Agent 可以自主决定调用 0.05 美元的便宜模型处理简单查询，或者调用 0.50 美元的高级推理模型深思熟虑。Stripe 的真正野心是成为这群自主软件的底层“Visa 卡”。此外，相比于极力将客户算力绑定在自身机房的 AWS、Azure 或 Google Cloud，Stripe 具备平台中立性，无论代码跑在云端、边缘老旧笔记本还是 GPU 集群，Stripe 都能稳稳抽取微交易分成。
- **模型价格崩塌**：Token 推理单价已下跌超 90%，纯模型提供商陷入价格战泥潭。
- **新型计费范式**：从固定月费订阅迁移到 7×24 小时百万级 Micro-token 动态实时微结算。
- **Agent 自主钱包**：赋能无实体身份的软件 Agent 自主决策算力开销并进行点对点交易。
- **生态绝对中立**：不捆绑特定云底座与特定模型，专注跨云容灾与故障回退（Fallback）。
---
### `16:40` “后模型市场”大变局：路由同质化危机与用户生态护城河
这一重磅并购标志着行业正式步入“后模型市场”（Post-model Market）——资本的视线正从底层基础模型快速向外衍生。主持人提出一个反思：既然人们对单一模型的执念在消退，未来究竟是咨询顾问称王，还是网关为王？甚至是未来 OpenRouter 反向收购咨询公司，还是 NVIDIA 下场大举进军 IT 服务？这种结构性变迁引发了关于价值捕获的深度讨论。
Aaron 指出，资本的脉动（Heartbeat）如今聚焦在路由层，但繁荣背后潜藏着极其迅速的“商品化/同质化”（Commoditization）风险。路由赛道早已高度拥挤，市面上涌现了 LiteLLM、Cloudflare、Portkey、Helicone、Together AI、Fireworks AI 以及开源的 Context Forge 等大批竞争对手。Stripe 愿意下重注豪赌，核心原因是 OpenRouter 展现出的惊人增速：在过去短短约六个月内，平台的 Token 处理量实现了 5 倍暴增，从 5 万亿 Token 飙升至 25 万亿 Token。
Stripe 实际上是在用真金白银“押上身家性命”（Betting the farm），坚信由 Agent 驱动的商业行为（Agentic Commerce）将接管传统数字商业，因此必须抢先构筑金融底座。Mihaly 补充道，这笔交易的核心资产绝非单纯的一套开源路由代码，而是其庞大且黏性极高的用户群与即刻变现能力：各方数据显示，OpenRouter 拥有 800 万至 1000 万覆盖免费到企业级的庞大用户群，以及超过 15 万高频活跃用户。Stripe 收购的本质是直接买下了这群最活跃的 AI 开发者、绑定了他们真实的付费钱包与正在发生的商业交易流。
- **惊人增长指标**：约 6 个月内 Token 吞吐量自 5 万亿暴涨 5 倍至 25 万亿。
- **高估值支撑依据**：平台聚合了约 800 万至 1000 万注册用户及 15 万核心活跃开发者。
- **赛道竞争者群雄逐鹿**：LiteLLM、Cloudflare、Portkey、Helicone、Together AI、Fireworks AI、Context Forge。
- **商业本质演进**：不仅是一次代码或技术网关的并购，更是抢夺下一代自主 Agent 商业交易结算账户所有权的资本战役。
### `20:00` Stripe 收购 OpenRouter 与 AI 网关生态演进
围绕 Stripe 对 OpenRouter 的收购，嘉宾们深入分析了这笔交易背后的商业逻辑与人效估值。OpenRouter 团队规模极小，大约仅有 50 到 100 名员工，但从市场估值推算，平均每位员工贡献的估值高达约 7000 万美元。这固然是对未来增长的大胆押注，但也意味着极强的单位产出。嘉宾借用“过桥收费”的隐喻指出：在各大基础模型厂商为了追逐基准测试（benchmarks）而疯狂烧钱失血的同时，Stripe 买下的更像是立于咽喉要道的“收费桥梁”——只要海量 token 在不同模型间流转消耗，Stripe 就能在每一辆疾驰而过的“数据车辆”上稳稳抽取两分钱的过路费。然而，路由层也时刻面临技术同质化（commoditization）的潜在威胁。
随后讨论延伸至开源与自研网关生态。目前开源社区中已涌现出超过 100 种形式各异的网关工具。随着企业级架构从单一模型调用演变为复杂的智能体体系，网关已不能仅仅充当简单的 API 转发器，而必须具备语义路由（semantic routing）、权限缓存以及协调 MCP（Model Context Protocol）服务器的能力。未来的核心竞争焦点正在全面转向“智能体控制平面”（Agent Control Plane）。
```mermaid
flowchart TD
accTitle: AI网关与智能体控制平面架构
accDescr: 展示从上层智能体与客户端到底层模型之间的治理与路由控制层
subgraph ClientLayer["客户端与应用层"]
Agents["自主智能体 (Agents)"]
Devs["开发者与 MCP 服务器"]
end
subgraph ControlPlane["智能体控制平面 (Agent Control Plane)"]
Router["语义路由 (Semantic Routing)"]
Policy["权限缓存与治理合规"]
FinOps["成本控制 / FinOps 与可观测性"]
end
subgraph ModelLayer["底层模型基础设施"]
PropModels["闭源商业模型 (OpenAI 等)"]
OSModels["开源模型推理端点"]
end
ClientLayer --> ControlPlane
ControlPlane --> ModelLayer
```
- 嘉宾强调企业所需的不仅是简单转发：“AI 网关或模型网关只是企业所需能力中的一小部分，无论是采用开源还是商业软件，企业都需要它来为其整个 AI 机队提供治理、信任、可观测性、优化、成本控制、FinOps 以及运行与管理能力。”（“AI gateway or model gateway is just a small component of the capabilities enterprises need, whether they're using open source or commercial software, to provide governance, trust, observability, optimization, cost control, FinOps, run and manage capabilities for their entire AI fleet.”）
### `22:54` Ramp 2026 AI 指数：两极分化的企业支出与计量限制
讨论转向 Ramp 发布的 2026 年 8 月企业支出报告（Ramp AI Index）。作为覆盖约 70,000 家企业的商业信用卡与费用管理平台，Ramp 提供了洞察企业微观开支的极佳切口。数据显示，开源 AI 与专用模型推理平台的采用率正稳步攀升，市场份额从 4.5% 跃升至 6.1%。尽管开源推理整体占比仍有限，但其增长势头非常明确。
然而，企业在 AI 工具上的开支呈现出极度不均衡的“K型断层”。处在金字塔顶端前 1% 的激进企业，每位员工每月在 AI 上的支出高达约 7,400 美元；排名前 10% 的企业这一数字骤降至每月 650 美元；而中位数企业每位员工每月的 AI 支出仅有区区 12 美元左右。这种巨大落差表明，多数企业在普及 AI 的同时已经踩下刹车，设定了极其严格的用量配额。随着“数字员工”（digital workers）逐步投入生产，企业高管不得不正视前 1% 群体中 AI 工具开销已逼近甚至超过工程师实际薪资这一残酷现实。
- Ramp 数据样本覆盖约 70,000 家企业，全面呈现了 2026 年 AI 开支的实际全貌。
- 企业 AI 月度人均支出呈现严重分化：Top 1% 为 7,400 美元，Top 10% 为 650 美元，中位数仅为 12 美元。
- 面对飙升的账单，企业策略发生转向：既要让员工体验 AI，又必须对调用实施硬性上限管制（metering & capping）。
### `25:00` 告别“错失恐惧税”：从人头费向单次工作成本的范式转移
嘉宾们指出，企业在 AI 上的投入方式正在经历深刻的结构性洗牌。在 2023 至 2025 年间，出于对“落后于时代”的集体焦虑，企业财务负责人（CFO）大笔一挥批准了各种宽泛的预算。然而到了 2026 年，这种交纳“错失恐惧税”（FOMO tax）的时代已经彻底终结，无节制的实验性开支被全面叫停，CFO 开始严厉拷问每笔支出的单位经济效益。仅做简单包装的营销文案生成类工具遭遇了灾难性的客户流失（catastrophic churn），因为基础模型能力升级及操作系统与浏览器的原生免费功能已完全能取而代之。
传统的 SaaS“按席位付费”（per-seat pricing）模式在 AI 时代正在失效，业界必须转向“单项工作成本”（cost per unit work）来核算 ROI。例如，一个人工客服工单处理成本高达 18 美元，而交由 AI 处理后可暴跌至几美分；一份合同的审核成本则能从 400 美元大幅压低至 12 美元。同时，AI 支出的部门分化愈发显著：软件开发等工程团队消耗巨大且产出明显，而其他业务部门则开销寥寥。如果高昂的人均开销无法兑现为数倍的生产力提升，盲目花钱将毫无意义。
- 嘉宾直接宣告了企业采购逻辑的转折：“在 2026 年，无计量的实验时代可以说已经正式终结，CFO 们现在要求的是可衡量的单位经济回报。”（“In 2026, the era of the unmetered experimentation is kind of officially dead, and CFOs now demand kind of measurable unit economic payback.”）
- 传统的评估框架正在被颠覆：“我认为，按席位评估 AI 支出是一种过时的传统 SaaS 框架。”（“Evaluating AI spent on a per seat basis, I think it's kind of an obsolete legacy SaaS framework.”）
### `27:38` “随机 AI 行为”泛滥与全生命周期合规重压
针对市场上所谓“AI 泡沫论”与“投入产出比过低”的悲观论调，Mihai 从一线客户咨询的真实经历给出了剖析。过去两年，企业客户的需求发生了 180 度大转弯：过去他们争先恐后地询问“如何做 PoC”以及“怎样才能用上 AI”，而现在各大企业普遍陷入了“救命，我们的 AI 用过头了”的深渊。企业内部正被无节制的“随机 AI 行为”（random acts of AI）所瘫痪。
各团队利用 LangChain、LangGraph、AutoGen、AutoGPT、CrewAI 编写智能体，又在 Microsoft Copilot Studio 与 Watsonx 上部署流程，造成了数百个自治智能体在内网游荡。真正榨干企业 token 预算的并不是员工在 ChatGPT 网页端随手提问，而是工程师们打开几十个终端、将实时事件流直接接入智能体，并在底层默认调用最顶尖、最昂贵的商业大模型。雪上加霜的是，监管层面的合规利剑已然高悬。欧洲《人工智能法案》（EU AI Act）等法规逐步落地，巨额违规罚款风险迫使管理层必须将合规与治理风险计入 AI 的总体拥有成本（TCO）之中。
- 客户现场反映出的混乱现状十分普遍：“我们在 LangChain、LangGraph、AutoGen、AutoGPT、CrewAI 中编写了各种智能体，在 Microsoft Copilot Studio 上有智能体，在 Watsonx 上也有智能体，我们到处都是智能体。”（“We've got agents written in LangChain, LangGraph, AutoGen, AutoGPT, CrewAI, we've got agents on Microsoft Copilot Studio, agents on Watsonx... we've got agents everywhere.”）
- 针对监管合规隐患，Mihai 警告企业必须核算潜在法律代价：“监管机构会问，大家知道有《欧盟人工智能法案》这回事，那么我们在核算总体拥有成本时，是否把那些巨额罚款也考虑进去了？”（“Regulators are asking: hey, you know, there's a thing called the EU AI Act, so are we factoring into the total cost of ownership one of those big fines?”）
### `30:00` 企业 AI 的全面拥有成本与 FinOps 时代
讨论从企业部署大模型的实际开销展开。嘉宾指出，企业 AI 带来的总拥有成本（TCO）绝不仅限于支付给特定模型供应商的 token 费用（tokenomics），真正的重头在于支撑其运转的周边投入——包括监管与合规、人力管理、计算基础设施、数据治理以及安全防护。当前很多企业面临着成本失控的困境，根本缺乏一套有效机制来管理预算、追踪 token 开销，更谈不上将这些支出与配额执行、ROI（投资回报率）或商业价值强绑定。发言嘉宾提到，这正是他们打造 watsonx Orchestrate 时试图解决的核心痛点，即在投资回报率、模型支出、全面拥有成本、安全护栏以及合规控制之间建立严格的对齐。
```mermaid
flowchart TD
accTitle: 企业 AI 总拥有成本与失控机制
accDescr: 展示模型成本与运维安全开支导致失控，并转向企业级 FinOps 治理的流程。
subgraph CostDrivers ["成本驱动因素"]
A["Token 支出 (Tokenomics)"]
B["合规、安全与治理成本"]
C["基础设施与人力资源"]
end
subgraph Paradox ["成本失控悖论"]
D["单价下降但使用量暴增 13 倍"]
E["大量自主智能体产生复杂调用"]
end
subgraph Solution ["FinOps 治理对齐"]
F["配额管理与支出监控"]
G["明确投资回报率 (ROI)"]
H["watsonx Orchestrate 护栏体系"]
end
CostDrivers --> Paradox
Paradox --> Solution
```
随后，团队讨论了“AI 越便宜反而越昂贵”的悖论：尽管模型单价持续下跌且性能更强，但根据 Ramp 的报告，过去一年内企业在 token 上的支出反而激增了 13 倍。企业陷入了一种“AI 狂欢（splurge）”，认为既然单价低了就可以无节制地让各种自主智能体（agents）处理任务，却忽略了多智能体复合运作失控可能带来的灾难。嘉宾将其与 2012 至 2015 年间云计算成熟期的演进相类比——当年企业停止无节制地空转 EC2 实例，从而催生了云财务运营（FinOps）；如今企业 AI 同样进入了 FinOps 阶段，重点必须转向“聪明地花钱，而非一味盲目削减”。不过在工程开发领域，AI 工具的支出表现出极高韧性，像 Cursor、Claude Code 以及代码审查专用模型，已经从早期的尝鲜噱头转变为不可或缺的开发者基础设施。
> “AI is easy and this list shows us it true It getting control over AI and having predictable costs and having predictable outcome and a good return on investment is proving extremely difficult in these early days” —— 嘉宾在分析企业 AI 成本困境时总结道。
- 根据 Ramp 发布的行业数据，过去一年企业在 AI token 上的支出激增了整整 13 倍。
- 研发工程部门是唯一一个 AI 工具支出持续指数级增长且几乎未受到预算缩减阻力的领域。
- Cursor、Claude Code 以及专用代码评审模型等工具已经跨过新奇阶段，成为企业开发的基础设施。
### `33:45` 国会立法草案中的 AI 垃圾内容与规范治理
节目的最后是一个趣味但发人深省的话题。主持人引用了政治媒体 Politico 的报道：美国众议院立法法律顾问办公室（House Office of Legislative Counsel）目前正被大量由 AI 生成的“劣质法案草案（slop bills）”所淹没。这与此前英国议会在 ChatGPT 发布后某些特定 AI 风格措辞使用频率爆发式增长的现象如出一辙。人们在撰写复杂草案时习惯性走捷径，导致立法机构充斥着充斥破折号、智能引号和空洞辞藻的机器生成法案。
对此，嘉宾在收尾发言中表示，这与学界论文、学生作业、日常邮件、Slack 和 Teams 消息被 AI 泛滥占领的现状并无本质不同。解决立法领域 AI 泛滥的关键并非彻底封杀，而是系统的教育与规范化治理。必须让起草者清醒地意识到模型必然会犯错，最终结果的法律责任永远由人类承担；同时，政府机构需要建立受控、合规的企业级工具使用通道，而不是让员工随意使用公开网络上的 ChatGPT 随意粘贴数据。
> “The legislation is no different than any other field. And rather than trying to fight it, there needs to be clear guidelines and rules on what's allowed, what's not allowed, what's safe, what isn't safe, and empowering folks to use it in a smart way, as opposed to having AI write law.” —— 嘉宾在被要求给出整期节目最终结语时指出。
- 政治媒体 Politico 曝光美国众议院立法法律顾问办公室正面临 AI 生成的低质法案草案冲击。
- 类似现象此前已在英国议会得到数据证实，即 ChatGPT 问世后特定 AI 常用词汇在议会发言与记录中激增。
- 主持人随后向对谈嘉宾 Mihaly、Kautar 和 Aaron 致谢，并在预告下期 Mixture of Experts 后宣布本期节目结束。
## Notable quotes
> `14:40` Models are cheap. The toll booth is where the money is. — Kaoutar El Maghraoui
> `11:30` I believe the future is multi-modal, multi-agent, multi-framework, multi-harness, multi-cloud, multi-everything; there is no one size fits all for models. — Mihai Criveti
> `25:00` In 2026, the era of the unmetered experimentation is kind of officially dead, and CFOs now demand kind of measurable unit economic payback.
> `25:00` Evaluating AI spent on a per seat basis, I think it's kind of an obsolete legacy SaaS framework.
## Mentioned
Tim Huang, Kauter El Maghraoui, Aaron Baughtman, Mihai Criveti, IBM, IBM Consulting, OpenAI, Anthropic, Stripe, OpenRouter, Ramp, watsonx Orchestrate, Watsonx.governance, Granite, IBM Bob, IBM Consulting Advantage, COBOL, SAP, Epic, LiteLLM, Cloudflare, Portkey, Helicone, Together AI, Fireworks AI, Context Forge, LangChain, LangGraph, AutoGen, AutoGPT, CrewAI, Microsoft Copilot Studio, Cursor, Claude Code, Politico, 美国众议院立法法律顾问办公室, 英国议会, EU AI Act
## Video description
Visit Mixture of Experts podcast page to get more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts This week on Mixture of Experts, we talk partnerships, purchases, and predictions surrounding the AI industry's financial infrastructure. But first our panelists dive into IBM's newly announced collaboration with OpenAI, under which IBM Consulting will train and certify consultants on OpenAI's tools to bolster the company’s push toward enterprise deployments. Then we turn to...
[Source](https://cdnapisec.kaltura.com/p/1773841/sp/177384100/playManifest/entryId/1_6qdmltid/format/download/protocol/https/flavorParamIds/0/name/media.mp3)
Source quotes
[14:40] Kaoutar El Maghraoui: Models are cheap. The toll booth is where the money is.
[11:30] Mihai Criveti: I believe the future is multi-modal, multi-agent, multi-framework, multi-harness, multi-cloud, multi-everything; there is no one size fits all for models.
[25:00] In 2026, the era of the unmetered experimentation is kind of officially dead, and CFOs now demand kind of measurable unit economic payback.
[25:00] Evaluating AI spent on a per seat basis, I think it's kind of an obsolete legacy SaaS framework.
Source quotes
[5:00] Kawtar El Maghraoui: OpenAI and Anthropic needs IBM's distribution and consulting pipes far more than IBM needs their specific model checkpoints.
[0:00] Aaron Baughman: We're betting that the enterprise AI value, it's shifting upwards from owning the model to now orchestrating, governing, and integrating, and even, you could argue, operationalizing these types of models.
[30:00] Mihai Criveti: Getting control over AI and having predictable costs and having predictable outcome and a good return on investment is proving extremely difficult in these early days.
Original transcript
[0:00] Yeah, I mean, you know, this market is, you know, ebbing and flowing, you know, and, you know, we just have to trace where the heartbeat is coming from, right? And it seems like that right now the heartbeat is around routing. All that and more on this week's Mixture of Experts. you I'm Tim Huang and welcome to Mixture of Experts. Each week, MOE brings together a panel of technologists working at the frontiers of artificial intelligence to lead you through the week's news. On this week's episode, we have Kauter L. McGrawey, Principal Research Scientist, AI Native Systems. Aaron Baughtman, IBM Fellow, and Mihai Cravetty, CTO, Watson X Orchestrate. Welcome to you three. We've got a bunch of big stories to follow on. Today, we're going to talk a little about Stripe buying Open Router. We'll talk about some really interesting business data out of RAMP. And then finally, we'll talk a little bit about AI for legislation in Congress. But first, I really want to talk about one big announcement coming out of IBM itself. Just this last week, there was a big announcement that IBM was going to launch a big partnership with OpenAI. And this is sort of interesting. It caught my eye just because not too long ago, IBM had also closed a very large partnership with Anthropic as well. And specifically, the partnership pertains to essentially training up large numbers of people to facilitate basically an OpenAI practice within IBM consulting. I'm kind of interested in how this kind of world of consulting is evolving around AI, because normally I think we've thought about like, oh, OK, it's going to be open AI versus anthropic and it's just going to be, you know, the sharks and the jets. But this kind of partnership really almost suggested a different world is going to emerge where it's not going to be so black and white. And in fact, there will be a lot of kind of professionals that work to integrate these systems that are kind of much more ecumenical with these models. think it all might go yeah i mean i'm very excited about it um i think it's a potent combination you know to have open ai's ai capabilities plus ibm's and our ability to put them within a fortune 500 company and and beyond you know within this new practice that we're creating we're going to have four deployed units and four deployed engineers are going to be equipped with these kind of capabilities i can go out and meet clients where they are right um And this is a big signal that we, IBM, we're not trying to pick a model winner. We're actually working and we're going to become the owner of the enterprise AI control platform. plane here. We're betting that the enterprise AI value, it's shifting upwards from owning the model to now orchestrating, governing, and integrating, and even, you could argue, operationalizing these types of models. And what I also think is fascinating is that Both Anthropic and OpenAI, they somewhat own the market, right? And because we have a partnership with both, both in different ways, it's quite important that we're putting together Anthropic, right? Where we have Anthropic is really geared towards IBM software engineering, where we have IBM Bob built around that ecosystem, while OpenAI is really for IBM consulting, right? And going out and having these four deployed units and engineers. neutral enterprise AI technology. integrator where we have the options, right, of both of the leaders, anthropic and open AI. But yet we still maintain our own, you know, granite type series models, which I think would become even more important, right, because this gives clients the freedom of choice. And the granite models are the much smaller ones, right, where we can route simple requests to those. So they're smaller, they're cheaper, whereas you could argue that the open AI ones, you know, the more complex type of task planning. And so by having these model gateways, I know Mahai works on a lot, gives us the ability to go in any direction that our clients would want to go in. Yeah, and I think one way of looking at it is almost like everybody's competing with everybody across every single layer at the moment. This partnership makes me think a little bit about OpenAI not too long ago, deployment company, which is kind of a consulting arm that would help people integrate OpenAI. They also signed a partnership with McKinsey, which is like another consulting org. You know, at the same time, Aaron, as you mentioned, right, IBM's got its own models that are kind of playing in the space and also has its own consulting arm, which is now partnership with OpenAI. You know, it kind of feels like, again, it's sort of like everybody is going in every single direction. I guess the kind of question is like, what's OpenAI's strategy here? Because it feels like they are sort of almost like choosing every single route and sort of
[5:00] being what sticks when it comes to this world of working with clients to kind of integrate this technology. Yeah, I think OpenAI needs system integrators, which is boots on the ground. This is where consulting organizations come in and have the ability to really integrate specific models or specific model providers into use cases which are company-specific, client-specific. It's also important to note that there is no such thing as the best model. The best model for what, Right. Different large language models and models providers offer multiple models with different costs, different latency as in time to first token, different performances in things like tokens per second, different performance as in the outcome. And even that differs for use cases, which is the best model for coding, which is the best model for information retrieval or summarization. Different models and model providers have different geographic availability. And in many cases, you're only allowed to use as an organization's models in your geography. If you don't have OpenAI in your London data center, you're going to have to use what's available there. And even there, you have limits with these models, which is what most companies tend to find out the moment they scale AI. Hey can I just give AI to my 100 employees tomorrow And the answer is well no you can buy it Like no you can buy it You literally can buy it You go to these organizations and you like hey, this thing only does 50, 150, 200 requests per second or the inference performance is not sufficient. And no matter how much money you throw at it, you realize there's not sufficient model to go around. So then you either do model routing or you go to multiple providers and you find out, hey, I need to prompt this model differently. So this is where, again, that relationship and partnership with system integrators, with the forward deployed engineers and so on comes in, to be able to help you consume the right model for the right task in the right way, in a way which is governed and secured and controlled, and prompted the right way to get optimal results. Yeah, I wanted to also mention there's a cost factor involved as well. It's very costly for over 100,000 employees each to begin to use these types of models. And I think to Mahai's point, there's no free lunch where you have to route the appropriate problem to the appropriate model. That like this multi issue where you want to have the best capabilities that match the problem but you also want to minimize cost Because if you think it like 10 bucks or US dollars per employee and you have hundreds of thousands of employees per month you looking at millions of dollars a month in token costs which may not be sustainable for a business right And so by having these partnerships, you not only get access to some of their leading edge models, but you presumably also get economies of scale pricing, right, which does help to bring down the cost of your workforce so that you can equip your forward deployed engineers and forward deployed units. So I do think it's a competitive advantage for both IBM and also for OpenAI to saturate the market with their capabilities. Yeah, I think there's another also point that's very important, which is kind of the distribution and the legacy plumbing that we have, you know, at IBM, which is very strong. And I think those beat row benchmark edge. If you look at what's happening in the Silicon Valley, there is a lot of belief that maybe having a 2% higher score on the MMLU or the human eval benchmark wins the Fortune 500. But in the real world 80 of enterprise value is locked inside legacy mainframe COBOL databases maybe SAP ARPs Epic Healthcare Systems HIPAA compliance Vault et cetera And IBM owns those enterprise relationships OpenAI and Anthropic needs IBM's distribution and consulting pipes far more than IBM needs their specific model checkpoints. So I think here, this is really a great partnership because at IBM, I don't think we need to spend 20 billion training frontier foundation models to compete with maybe Microsoft, Google, or Meta. Instead, IBM's winning strategy here is becoming the trusted vendor neutral enterprise orchestrator through the Watson X dot governance. IBM provides the audit trails, the bias detection, the prompt monitoring and the regulatory compliance layer, while letting clients kind of hot swap models based on cost, speed and capability. I agree with Mihai. I mean, there is no model that fits everything here. So it depends. The best model depends on so many other dimensions here, and you have to tune it to your enterprise, to your data, etc. So I think it's very important to have that layer, that orchestration, and those pipes that are so important with the regulatory compliance, with the audit, with the governance, which is really important. Because the bottleneck here in enterprise AI is not...
[10:00] how to a shortage of intelligent models It's really change management, system integration, and also dealing with dirty data. And I think IBM Consulting is kind of weaponizing open AI and anthropic models to accelerate that enterprise digital transformation, kind of converting legacy code bases and automating banking workflows at scale. Yeah, absolutely. And I think the question being asked is, it's almost a little bit like what... How much will the model matter in the future? And I think you've seen this debate. We've talked about this a lot on MOE on the compute side, where it's like, OK, well, if everybody can get great open source, maybe it's really the compute that's going to matter. It feels like the consulting side is actually the other part of this discussion, which is, OK, when a world where there's lots and lots of models and there's lots and lots of options, then ultimately the kind of winning factor is going to be the sort of like trusted consultant. Aaron, do you want to get a final like, you know, this constellation of partnerships, right? So there's three in my mind right now that really stand out. It's IBM plus Anthropic plus OpenAI, right? Because the Anthropic deal and partnership was announced around October 2025. And the idea was to bring in AI to change enterprise software development and integration, right? And that's how IBM Bob, you know, somewhat emerged, thousands of consultants through IBM Consultant Advantage so that we can go to clients, right, and meet clients. So we're taking the full stack of the business with these two partnerships. And I'm really excited about the future and what we're going to do together within IBM. Great. Well, that's a great segue to the next topic I wanted to cover. News broke this week that Stripe, the payment processor, is finalizing an agreement to purchase a company called Open Router for more than seven billion dollars. And again, if any indication of how crazy the market is right now, the stat that was kind of going around the Internet, which is that seven billion is more than five times the one point three billion dollar valuation that Open Router had just 82 days ago. And so this is like a really wild story. focusing on the models and hearing about what OpenA and Anthropik are doing, OpenRouter itself may be kind of an unfamiliar... And so I guess, Mihaly, for folks who are maybe less familiar with what Open Router is or does, do you want to just give a quick intro on why it's such a big deal and why it could be possibly worth $7 billion? Sure. So as somebody who's written... his own model gateway, MCP gateway, and A2A gateway and gave it away for free, I'm kind of... A little bit different. Thinking my life choices. Thinking my life choices. But all joking aside, it's like, I believe the future is multi-modal, multi-modal. multi-agent multi-framework multi-harness multi-cloud multi-everything there is no one size fits all for models there is no one model which is good and you can't even get enough of one model to fit every use case so with platforms such as open router you have the ability to access a large variety of models from every vendor to a single platform, which gives you access not only to the models, but You can do things like, for example, semantic caching or you can help optimize some of that work. And. some of the future that we will see in this space, is integration into unified gateways, gateways that have the ability to access not only models but MCP servers, agents, skills, evals, controls with semantic routing, the ability to route the right request to the right model, synops and cost management, the ability to say, yep, We're going to limit and cap your cost and spend with things like virtual API keys where you're not accessing the key of the model directly so you can lose it or somebody can steal it. But you're actually accessing a virtual key, which is site role based access control. So. A lot of these capabilities are must-haves in enterprise. They are the capability, which allows an organization to distribute these models, MCP servers, tools, and so on. safely. to their enterprise, but also get things like metrics and observability data from those interactions. Kaltar, I guess, so now we know what OpenRouter is. Why Stripe? You know, this is also kind of like the other part of the transaction. We've been hearing, of course, a lot about the big acquisitions that the Frontier Labs do. You know, Stripe is, you know, I use them to pay for subscriptions and stuff online. Why are they getting into this space, you think? Yeah, that's a very good question. I think the way I see it, models are cheap. The toll booth is where the money is. Every six months, models get smarter and way cheaper. Token prices have dropped over 90%.
[15:00] 90 percent. Building a raw model is a tough, expensive race to the bottom, but the router gets paid a tiny toll on every single request, no matter who wins the benchmark race. But AI billing isn't traditional software. It's kind of a payment headache. Old software kind of charged, say, like a $30 a month per user on a credit card. AI doesn't work like that. It's millions of unpredictable micro tokens flying around like 24-7. lets them handle metering and payments right where the code runs. And if you look at AI agents, they need their own wallets. Down the road, AI agents will do the jobs on their own. They can't sign contracts or swipe physical cards. An agent needs a digital wallet to say, I'll use a cheap $0.05 model, for example, for a simple lookups, and a $0.50 model for deep thinking. Stripe wants kind of to be the Visa card for these autonomous software. And another thing, Stripe is also, it's neutral. Big clouds aren't. Amazon, Microsoft, and Google want you kind of stuck on their servers. Stripe doesn't care where your code runs, whether it's on AWS, on an old laptop, or a GPU cluster. They just take their small cut of the transaction. And I think that the routing strategy allows them to be very flexible here and then kind of have these fallback mechanisms switch between various models. So they don't care where things run or what models you use, but they want to control that becoming super important here. And so Aaron, I guess I know to bring it back to the original topic, it feels like we're back in the same place we were before, right? Where we're talking about, well, you know, people don't care about the model so much anymore. It's going to be the trusted consultant. You know, it seems like, well, maybe people don't care about the model so much anymore. It's going to be the router and the gateway. And I guess kind of question for you, Aaron, is like, we're almost looking at a post-model market now. It feels like where a lot of the money is going the AI model to wit. And I don't know, like, how should we think about how that market's evolving, right? Like, is the future Open Router buys a consulting firm? Or is it, you know, NVIDIA gets into the consulting? You know, I'm kind of really interested in kind of all of these weird configurations we're about to see as the core thing we've been paying attention to AI almost seems to become kind of commodity. And we're all kind of looking for like, what else is the It is, you know, ebbing and flowing, you know, and, you know, we just have to trace where the heartbeat is coming from. Right. And it seems like that right now the heartbeat is around routing. Right. Being able to route traffic to the specific place of which it, you know, it's. optimizes or minimizes your criteria that you're looking for. Right. But I suspect that this routing, it's going to become commoditized at some point. You know, it's a very it's going to become a very crowded capability space. You know, I just quickly looked and I mean, there's many competitors to to open around. There's light LLM, Cloudflare, Portkey. Helicon, Together AI, Fireworks AI, so on, right? Context Forge, I wrote it. Hello. Yeah, yeah, yeah. I mean, it's incredible just the amount of space, you know, that's where all of these little companies, I won't say little anymore, but these companies are coming in. And to me, Stripe bought, or rather, yeah, Stripe bought OpenAI based on growth. You know, they saw that over the last, what, you know, six months or so that the token growth by 5x from 5 trillion to 25 trillion, right? So they're betting the farm, I think, on that type of growth that they want to position themselves as the infrastructure for AI agents and agentic commerce, just like they are today with, you know, just commerce in general. So they're trying to make that leap, right? Saying that, you know, AI agents and models are going to be the drivers of commerce. Therefore, we need to get into it. But there is risk in this transaction, right? Is this commoditization part. And so do I think that this is going to be the underpinning? I think of the economic drivers and decision space, I think today it is, but after it becomes commoditized, there will be something else that'll come in. So will consulting groups want to buy open router type-ish or build their own organic? Today, yes, but I think there'll be other things that are going to come up where they also have a large number of users which was part of the acquisition Like it's, I don't know the exact numbers, but some, you know, different sources claim they have 8 million or 10 million users across free enterprise, whatever else, 150,000 active. I mean, I don't know the exact numbers, but this is not just an open source project somebody buys. It's more of they're buying the users, they're buying the commitment, they're buying the wallet, they're buying... those paying customers as well with this transaction. Yeah, we,
[20:00] What's really interesting is that OpenRouter has like 50 to 100 employees. I mean, it's pretty small, you know, but when you look at their market valuation, I mean, you're looking at. maybe 70 million per employee, right, that they contribute towards their valuation, of which they are up 70 million per person, right? That's big. And so, yeah, I mean, to Mahai's point, this is a, you know, a bet on growth, but they do have the foundation to make that growth, you know, and if that trend line continues, right, right, you know, I think Stripe might have done a good job, you know, but again, commoditization is always the enemy of these types nice analogy, like the tall boot analogy. So I think Stripe didn't buy this for just the AI hype, especially the open router. I think it's like, think of it as they bought the tall boots, like while model makers kind of bleed cash competing on benchmarks. I think Stripe sits at the bridge and collects two cents on every car driving across. So, and I think it's a very happening, you know, all of these tokens being consumed. And it's a very good place where they're sitting. You know, the routing right now is becoming a very important place. Well, maybe just maybe a final question I had. I know it's a sore topic because you released one of these gateways for free to the public. But what is the open side of this, the open source side of it? Like, do we feel like, you know, that there will be more kind of open gateways that kind of are operating outside of the businesses? Like, what's that ecosystem who's built something like that. Yeah, I think there's already more than 100 of these gateways in some shape or form out on the market today in the open source space. I think Aaron has already mentioned like five or six of them, but Definitely there is a need. right enterprises and users and so on need to have a way to manage a combination of agents, MCP servers and models and so on with things like semantic routing and caching of permissions. And I think a lot of work, this space is evolving towards is that of agent management platforms for agent control planes Warp. AI gateway or model gateway is just a small component of the capabilities enterprises need. whether they're using open source or commercial software, to provide governance, trust, observability, optimization, cost control, thin ops, run and manage capabilities for their entire AI fleet. Really interesting data coming out of RAMP, which, as you may know, is one of the most kind of widely adopted, you know, sort of business credit card payment management sort of platforms. And so as a result, they have a really good bit of traction on understanding how, you know, businesses down to the very micro level are using things like AI. And so they released the post, which is basically their kind of August 2026 review of what's happening in the AI space. And some really interesting stats here. of interested in to build on some of the themes that we've been talking about today is that there really does seem to be a lot of evidence that open source AI adoption is continuing to rise over time. I guess maybe, Aaron, I'll kick it over to you. You know, this is still, you know, a jump from 4.5% to 6.1%. So not a huge amount of the overall market, but they are seeing this kind of really measurable rise in people using, you know, model serving inference platforms and not How far do you think that's going to go? Yeah, I mean, you know, I do enjoy reading, you know, the ramp and looking at the 2026 ramp AI index because it gives you a nice, you know, sort of pulse right on on what people are doing and what the businesses are doing, because it's, you know, it analyzes about 70,000 businesses. And what I found and just by looking at it, it looks as though AI adoption is it's really uneven. Right. Because if you look at like the top one percent of these companies. AI adopters. They spend about let's say 7,400 per month on an AI, you know, per employee, whereas the top 10%, it drops down to 650, right? And then the median, you know, or, you know, the number that's like right in the middle, right, of all this is around, what, $12. And this is, you know, whenever I quote monetary values, it's a USD, but this is per month on an AI employee. So that spread, you know, really shows you this huge spread right across companies. But even so, there is growth, but there's a ceiling to the growth. It's as if companies are saying, "Hey, let's give everyone access to AI, but we have to meter the use." We need to cap the use in some way if it's getting out of control, especially as we get these
[25:00] digital workers online. And so there is this economic consequence. And what I find is, Interesting, right, is that the top 1%, right, it's starting to approach the cost of AI per employee, starting to approach the actual salary of the said engineers that are using these tools. So now the question becomes, well, if there's, you know, if we're paying as much for AI tools as we are a person, right, what do we do, right? Do we need to focus in on productivity versus revenue, right? What's the most important aspect? have to grapple with that, right? Assuming the economics and costs, you know, stay where they are, you know, at the moment. I feel kind of the end of this AI for more tax, because what happens in 2023, 2025, CFOs approved this broad AI software budget, and they're kind of the fear of falling behind. But in 2026, the era of the unmetered experimentation is kind of officially dead. And CFOs now demand kind of measurable unit economic payback. So either an AI tool demonstrability kind of reduces the headcount requirements, accelerates kind of cycle times, for example, PRs merge per engineer, or, you know, drives, you know, kind of net new gross revenue. marketing copy, you know, these are seeing kind of catastrophic churn because foundation model updates and native OS browser features do these things for free. So I think the true cost metric here is how do we measure the cost per unit work versus the cost per seat? So evaluating AI spent on a per seat basis, I think it's kind of an obsolete legacy SaaS framework. And, you know, evaluating AI on cost per resolve, maybe support tickets, for example, dropping from $18 human cost to maybe like a few cents AI cost or cost per audited contract dropping like from, say, $400 to $12. So the high spend per employee is actually kind of an indicator. It's not a good indicator if it doesn't correlate with the multifold productivity gains. And I think that's start focusing on it's what their return on investment with these spends per employee in terms of productivity gains. Otherwise, it's not going to be useful just to have these spends. But I also feel things are kind of segmenting. Initially, it's like, okay, let's start, let's give everybody access to these AI tools and LLMs, etc. But certain teams, like especially the engineering teams, the software developers, they're actually using them other teams that might be maybe losing them less. So, and I think that maybe we will start probably seeing segmentations per departments, you know, who really is benefiting the most and getting the ROI versus other teams that might probably don't need as much. Yeah, for sure. And I think, I mean, underlying that, Mihai, is sort of the idea that we actually may be spending... too much on ai i don't know like i guess there's there's one way of reading these numbers which is that the skeptics who say, hey, this might be actually kind of a bubble are maybe actually right. Right. Because as we start to take a much harder look at, say, productivity per dollar in, it may end up being much smaller than we think. Do you think it's too pessimistic as a way of reading these numbers? I think the way I would look at it is that, you know, over the course of the last two years, most of the client conversations I've had have shifted from, help us use AI or, you know, we want to buy our product or we want to do something with AI. We've done a POC with AI to help. We've done too much AI. We've got 60 random acts of AI. We've got agents written in LongChain, LongGraph, Autogener, AutoGPT, Crew AI. We've got agents on Microsoft Copilot Studio, agents on WatsonX, Requestrate. We've got agents everywhere. We've got models. We've got every model. We've got model routing. The problem we have is one, we don't even know where these agents are, where they are built or how AI is being spent, because it's usually not Thank you. asking chat gpt for advice on one or two things that's really burning down the tokens it's these hundreds of agents the developers and everybody else for you know they have 50 open terminals or they've got an event stream going into one of these agents and consuming them at scale and they use the most powerful and the most expensive models and so on and so forth second comes regulatory pressure regulators are asking Hey, you know, there's a thing called the EU AI Act. So are we factoring into the total cost of ownership one of those big fines?
[30:00] for example, or the cost of regulating those models, because the more you have and the more AI you have spread all over the place, the higher total cost of ownership is not just the tokenomics of how much you're spending on an individual model provider. It's everything else to manage it, the people, the resources, the infrastructure, the governance, the security. And then you've got the cost runaway, where today organizations don't really have a good mechanism to manage things like budgeting, even tokenomics, like understanding how much they're spending, let alone tying that into quotas, enforcement, tying it to a return on investment, for example, and value. So this is one of the areas I'm most excited about. It's one of the areas where this is kind of what we're building with Watson X orchestrate. So that's the product I'm I'm technically responsible for, to try to get alignment between return on investment, model cost, AI spend, total cost of ownership, guardrail, security controls, everything else in between So I would say AI is easy and this list shows us it true It getting control over AI and having predictable costs and having predictable outcome and a good return on investment is proving extremely difficult in these early days Yeah, yeah. Yeah. I found, too, that it's becoming harder and harder because as AI is getting cheaper and more capable, you would think that the cost would go down. But just in this report, you know, you can see that just over the last year, token, you know, spend has increased 13x. So it's like there's this AI splurge going on. Companies can't afford to do more, so let's do more because it costs less. But then the overruns happen because you have all of these agents that are compounding the problem that we don't know what exactly they're doing and how to put the guardrails on. So it's quite an interesting conundrum that I think the field is in. Yeah, and I think maybe this is kind of similar to what happened with the cloud compute when cloud computing matured between 2012 and 2015. Companies stopped mindlessly spinning up idle EC2 instances and they started introducing the FinOps. So the ramp data here shows enterprise AI is also entering this FinOps era and really pushing the companies kind of they need to start thinking about how do we spend smarter not less So because now we entering this maturity age and now the economics the reality is hitting us harder So we, I think a mental model here of how do we monitor these things, meter these things properly, make sure that we're spending where things need to be spent properly. So, like, if you look at engineering teams, it seems that they're the only department where AI has been continuous to grow exponentially without pushback. Tools like cursor, cloud code, specialized code review models, etc. They've crossed the threshold from novelty to being indispensable developer infrastructure. So that's spent there. I think it's justifiable, but maybe in other areas, not as much. for the final topic today because we're almost out of time uh this is just a fun final item that i wanted to touch on and i guess maybe hi maybe i'll give you the last word on this um politico which is a publication that covers uh dc and dc politics uh had a really interesting article that the congressional office of legislative council in the house of representatives is awash in ai driven slop quote unquote bills And it reminds me a little bit of I think we talked about it before I think one of my favorite analyses that anyone has ever done was observing that after ChatGPT launched the use of certain types of phrases really exploded in the Houses of Parliament in the UK. And I guess we've all used AI occasionally to take certain shortcuts. And so maybe it's no surprise that we're finding people trying to use AI to draft legislation. Should we be worried about this? What do we do about this sort of thing? And I'll give you the last word here for the episode. I want to introduce a load-bearing bill with a lot of PM dashes and smart quotes. I think it's no surprise that just like we're seeing an explosion on papers, publications, homework, emails, Slack messages, Teams messages written by AI. We're seeing the same thing everywhere, including legislation. it's part of it is a question of education it's educating people that ai will make mistakes that you need to have governance security control you're still responsible for the end result and giving them a way to do it that is i would say secure and govern not just everyone going to the for instance of chat gpt or online or somewhere or you know just posting the information there and
[35:00] posting back whatever they see. So I think... The legislation is no different than any other field. And rather than trying to fight it, there needs to be clear guidelines and rules on what's allowed, what's not allowed, what's safe, what isn't safe, and empowering folks to use it in a smart way, as opposed to having AI write law. Words of wisdom to end on. Mihaly, Kautar, Aaron, thanks for joining us on the show. And that's all the time that we have for today. And thanks to all you listeners. If you enjoyed what you heard, you can get us on Apple Podcasts, Spotify, and podcast platforms everywhere. And we'll see you all next week on Mixture of Experts. *music*
