A Trillion Tokens and The Decision Most Don’t Understand
On March 24, 2026, Portkey announced that its AI Gateway would become fully open-source. Immediate coverage focused on scalability numbers: over a trillion tokens processed daily, 120 million requests per day, and 180 million dollars spent on AI annually across 24,000 organizations. Quite impressive figures that inevitably resulted in an easy headline.
Yet the real strategic move is not in the numbers. It lies in the decision to release precisely what has historically been protected: the infrastructure layer connecting businesses to over 250 language models under a unified interface, along with the governance, observability, authentication, and cost control capabilities that previously required a separate paid subscription.
For any company operating in this segment, it amounts to throwing the lock on the floor and saying, "The value wasn’t here." The question worth analyzing is not whether this is generous, but what portfolio architecture is backing that decision, and how solid it actually is.
The Portfolio Logic Behind Opening the Code
Portkey is not in the packaged software business. It exists to be the control layer between enterprises and the proliferation of AI models. That distinction completely alters how this move is evaluated.
When a company releases its core infrastructure as open source, it’s typically pursuing one of two strategies: either the core product has ceased to be differentiating and needs community support to survive, or the company has identified that the true revenue engine lies at a higher layer and needs maximum distribution at the lower level to feed it. Clearly, Portkey resides in the latter.
The open-source gateway operates as a massive acquisition engine. Every engineering team that adopts it because it’s free, self-hosted, and flexible becomes a qualified candidate for the paid tiers: managed infrastructure with a 99.99% SLA, log storage and indexing at scale, semantic caching, advanced prompt management, and, above all, compliance certifications like SOC2, HIPAA, and ISO 27001 that no regulated company can build internally in mere weeks.
This is intelligent exploitation of the present business funding the exploration of the future business. The current revenue engine, the managed platform with its paid tiers, remains untouched. What opens up is the entry funnel. The marginal cost of acquiring a new technical user falls to zero while the value Portkey captures in the enterprise segment does not decrease at all.
One detail that merits specific attention: the announcement is not, strictly speaking, the first time Portkey has released open code. The gateway was already operating under that model in previous versions. What’s announced in March 2026 is the unification of separate components and the inclusion of the MCP Gateway as general availability. The narrative of "it is now fully open" also serves as a declaration of product maturity: there are no longer scattered pieces forcing the user to pay for cohesion.
The MCP Gateway as an Exploration Bet
The most revealing element of the announcement isn’t the open code itself, but the incorporation of the MCP Gateway, built on the Model Context Protocol. This protocol enables AI agents to use external tools, access databases, and execute actions in enterprise systems in a controlled manner.
Here is where Portkey’s bimodal logic is most clearly seen. The unified gateway for LLMs is today’s business: mature, with impressive scale metrics, perfectly defensible with the argument of migration costs. The MCP Gateway is the bet on tomorrow’s business: governance of autonomous agents in enterprise environments, a market that is still defining its standards in 2026.
Launching the MCP Gateway as general availability simultaneously with opening the code is a deliberate positioning maneuver. Portkey communicates two things at once: "We’re the most mature in an existing problem" and "We’re the first in the upcoming problem." The risk of this position lies in simultaneous execution. Maintaining operational quality of a system that processes a trillion tokens daily while building and scaling a completely different capability, agent orchestration, requires an organizational design that protects both lines of work without one cannibalizing the resources of the other.
Direct competitors, particularly LiteLLM in the self-hosted high availability segment and TrueFoundry in agent governance, are already operating in these areas. Portkey’s advantage isn’t isolated technology; it’s the installed base of 24,000 organizations that already trust its control layer for AI expenditure. Turning that trust into adoption of the MCP Gateway is an account expansion problem, not a new market conquest, which significantly reduces the risk of the bet.
The Sobriety That Scaling Numbers Conceal
There is a structural tension that the announcement doesn’t explicitly resolve: latency. The measured overhead of 20 to 40 milliseconds that Portkey introduces when advanced observability and guardrail functions are activated is no minor technical detail. For agent workloads that chain dozens of model calls in sequence, that overhead accumulates and can alter the economic viability of an entire workflow.
This matters because the promise of the MCP Gateway is precisely to enable complex agent workflows. If the cost of governance is accumulated latency, Portkey will need to demonstrate with production metrics that control doesn’t destroy performance. Competitors like Bifrost, built on Go and explicitly designed to minimize overhead, are poised to attack that argument.
The self-hosting model also deserves a cold reading. While the open code is available, deployment in isolated or air-gapped environments is only available at the Enterprise level. This means that organizations with strict data residency requirements, a segment that grows with each new AI regulation, must commit to the highest tier before assessing if the product meets their technical standards. LiteLLM offers full self-hosting with no licensing fees, a simple and powerful sales argument for teams with DevOps capability.
Open Source as the Infrastructure of the Revenue Model
Portkey is executing a portfolio model where the open and free layer does not compete with the paid business; it feeds it. Opening the code is the investment in acquisition; the managed platform with compliance certifications and SLAs is the return. It’s a structure that has already proven valid in data infrastructure, development tools, and databases, and is now being transitioned to enterprise AI middleware.
The long-term viability of this model depends on two variables that current numbers cannot confirm: the conversion rate of free gateway users to paid plans and the speed at which the MCP Gateway captures adoption before the agent governance market consolidates around a different standard. With 24,000 organizations already within the perimeter and a trillion tokens daily as a signal of operational retention, the conversion base exists. Execution on the agent segment will be the variable that determines whether Portkey’s portfolio strengthens or fragments over the next eighteen months.
