Most strategy games guide you through menus. Build this fortress, research that technology, wait twelve minutes. Pax Historia asks you to type what you want—"annex Belgium," say—and an AI decides if your neighbors declare war in response.
That shift, from clicking predetermined buttons to issuing commands in plain English, has drawn 35,000 daily users to the YC-backed startup's browser-based alternate history game in roughly two months. The founders, Eli Bullock-Papa and Ryan Zhang, met as freshmen at Virginia Tech. Their pitch: large language models can replace the rigid logic trees that have governed grand strategy titles for decades.
Players pick a country, choose a time period, and start typing. The AI simulates how rival nations react, redraws borders, even lets you negotiate with algorithmically-controlled heads of state. Want to test whether Napoleon wins at Waterloo with different tactics? Whether the Ottoman Empire endures into 2025? The model will iterate on your premise.
Prompt, Consequence, Repeat
The core loop is straightforward. Describe your move; the AI narrates the fallout. No tech trees. No scripted events. A scenario editor has enabled players to publish more than 4,000 custom historical setups—each reportedly played at least 50 times by launch. Time jumps let you compress campaigns: skip ahead a week, a month, six months, a year. The what-if experiments pile up fast.
The game runs entirely in a browser, currently tagged "Alpha," and it's free to start. But LLM-native products come with LLM-sized compute bills. Players receive a small daily allotment of free tokens and can purchase more, with consumption tied to whichever AI model they select. A "Pax Patron" subscription unlocks premium models without per-token charges. One reviewer estimated that a dollar buys 15 to 20 minutes on a mid-tier model like Gemini Flash—a friction point that surfaces repeatedly in community threads.
The Infrastructure Burden
Two months in, the numbers tell a story of ambition colliding with operational reality. Around the time of its January 2026 YC announcement, Pax processed more than 100 billion tokens in a single week. OpenRouter, which routes the company's inference requests across providers, logged 4.87 million requests the prior month and access to 28 models from seven vendors. Job postings name backing from Y Combinator, Pace Capital, and Z Fellows, and claim nearly 20 million rounds played to date.
Running a game on inference instead of pre-rendered assets introduces overhead traditional studios never encounter. A case study from infrastructure vendor Infron AI—take the source accordingly—says Pax integrated over 30 AI providers and adopted real-time health monitoring across 60 endpoints. The company reported a 60 percent reduction in AI costs and 99.9 percent reliability after implementing those systems, though those figures come from promotional material.
Growing Pains and Grinding Gears
Growth has arrived with complications. Community threads from late March surface complaints of "ongoing lag" and "critical bugs." Players note odd map behaviors: frontlines freeze, nations dissolve unexpectedly after updates. A recurring worry: the AI seems to degrade over time, especially under load or following silent patches. Reddit threads debate whether certain models have been swapped out or throttled without announcement.
Monetization remains unsettled. Early adopters recall more generous free-token allocations; recent posts suggest the runway has tightened. Regional pricing proves difficult when upstream API providers charge uniform global rates. Creators reportedly earn a 10 percent commission on token purchases made during plays of their scenarios—according to a user-generated wiki entry—but that economy hinges on sustained player spending. A harder sell when sessions burn through credits in minutes.
The studio has also weathered occasional outages. One authentication incident in February lasted over four hours. The Terms of Service were updated as recently as February 5 to add export control and regulatory language—signs of a product maturing under pressure in real time.
What the Model Can't Yet Simulate
Multiplayer is in development, no public date. A Steam release remains a community wish. The team size sits somewhere between three employees, per the YC directory, and seven, per LinkedIn—a small crew managing a product that processes billions of tokens weekly and serves tens of thousands of concurrent users.
Pax Historia demonstrates that you can build a grand strategy game on large language models instead of pre-scripted logic. Whether that architecture scales profitably—whether players will pay for the privilege of prompting their way through history—is the question the founders are still answering. For now, the traction suggests an audience exists for AI-native experiences. The infrastructure to support them? Still being assembled, one inference request at a time.
