During major technical platform shifts, technology inevitably moves faster than the infrastructure that supports it. In the early days, teams are left with inefficiencies, fragmentation, and coordination problems. These circumstances set the ground for generational infrastructure companies to be born.
Cloudflare was born in the days of the early web, when latency and security gaps made the internet barely fit for commerce. Stripe emerged for digital payments, unifying the friction of myriad payment rails into a single integration. Databricks unified the warehouses, lakes, and tools that big data had splintered across. CrowdStrike consolidated endpoint, identity, and workload protection once the corporate perimeter dissolved into the cloud.
AI is the next shift, and it’s already fragmenting in similar ways. Model and inference choices are exploding, forcing teams to manage multiple models across multiple providers and the complex decision process of when to use what.
Until OpenRouter.
OpenRouter is the control plane for inference, routing across models and inference providers for performance, latency, and cost. Developers build on and look to OpenRouter to see which models and apps are being used, and as a result model and app developers track their rankings on OpenRouter’s leaderboards closely.
OpenRouter’s leaderboards have become the industry standard for tracking AI model usage
OpenRouter is increasingly becoming the de facto place for enterprises and developers to work with AI. The easiest way to see this is in the data: OpenRouter is processing over 25 trillion tokens a week, up 5x in just the last 6 months.
Generational shifts breed generational companies
Our investment in OpenRouter was rooted in a few core beliefs:
1. Frontier models will continue to innovate
The frontier models will continue to push what is possible, and be able to charge for that service. That said, the intelligence frontier will be pushed by all foundation models, but in different ways. The true frontier will remain jagged: Developers will experience different performance from different models, causing them to need to use multiple models for best performance across all queries.
2. Open source models continue to follow closely
To date, open weight models have been roughly 3-6 months behind the frontier models in terms of capability, but at a fraction of the cost. In the last 12 months alone we have seen remarkable gains made by open weight models, followed by remarkable increases in usage. Teams now have more choice than ever when it comes not only to what model to use, but also what inference provider to use that model on.
3. The CFO reckoning is coming
As enterprises feel the pressure to adopt AI or get left behind by those who do, we have seen a tops down enterprise mandate to consume inference that has led to incredible efficiency in the enterprise, as well as behaviors that likely won’t last like tokenmaxxing. Goldman Sachs studies show that companies are overrunning their inference budgets by an order of magnitude. As enterprises start to look at inference consumption as an OpEx line item, they will also start to look at cost optimization.
2026 will mark the next phase of AI adoption
The confluence of these trends will come in 2026. To date, inference adoption has gone through three distinct phases:
2023-2024: Experimentation phase. AI budgets came from discretionary funds, “strategic initiatives” or innovation funds. Nobody was sure what worked yet, and nobody was counting tokens.
2025: Mandate phase. CEOs and boards demanded both an AI strategy and robust AI adoption. Inference use exploded in part because the mandate was simply “use it”. This is the tokenmaxxing era.
2026: Optimization phase. Inference moves from the innovation budget to OpEx and CFOs start asking about impact more than pure usage. Teams are still incentivized to leverage AI as much as possible, but also to do it responsibly.
This rationalization is inevitable. Anthropic, OpenAI, and Google Gemini are adding the equivalent of ~three Workdays in revenue a week. That level of new spend can’t go unchecked forever, no matter the promised impact.
Inference use, and spend, will continue to go up, both at the individual company level and in aggregate. But inference is becoming such a substantial line item for most companies that it will require the same governance and scrutiny as other large line items – on both cost and sourcing risk. In a world where model leadership shifts frequently, committing to a single provider on a mulit-year contract is a strategic risk for companies and a career risk for CIOs. In 2026, enterprises and individual developers will need to use frontier models to continue to innovate and address new use cases while also behaving responsibly toward costs and preserving optionality to switch between models as the frontier evolves. OpenRouter already gives every customer the routing, observability, sourcing flexibility, and provider arbitrage that CFOs and CIOs will demand in 2026 — built in from day one.
The “system of record” for inference
OpenRouter is quietly becoming the system of record for inference. Every request that flows through the platform–across 400+ models, 8 million users, and trillions of tokens per week–generates signal about how the world is using AI. This signal allows OpenRouter to intelligently route queries, and to build products that add value to every LLM call. Whether it is adding web search capabilities to any LLM, standardizing tool calling across LLMs, or enabling any LLM to generate an image from text, OpenRouter makes each LLM call not only more efficient but also more valuable.
As OpenRouter continues to see an increasing share of all inference, their ability to increase the value of every LLM call will grow. The more valuable their routing functionality is, the more inference they see. And the more inference they see, the more data they have to build value additive products on top. This is the compounding advantage of the foundational infrastructure companies have.
The team to build
A shift as seismic as this needs a team that is up to the task. Alex, Chris, and Louis are just that.
Alex Atallah is a brilliant technical mind capable of seeing where the trends are going and building for the future as well as the present, a skill that is deeply needed in a time as turbulent as this. Chris Clark is a remarkable operator who turns complex problems into simple solutions and beautifully balances long term strategy with near term pragmatism. Louis Vichy has consistently pioneered novel technical architectures, and is continuing to do just that at OpenRouter.
After every meeting with the team we felt that we just had to work together.
We feel so lucky that we get to.