Alex Atallah

The through-line

Alex Atallah keeps finding the same shape of opportunity twice: a new building block appears on the internet, dozens of incompatible projects start producing versions of it, and nobody has built the place where you can see all of them at once. That was NFTs in 2018, where he and his co-founder Devin bet on "a new primitive for the internet" that sat somewhere between web pages and cryptocurrencies and needed a marketplace to make it legible [1]. It was language models in 2023, where the arrival of open weights convinced him there would be "tens of thousands of these maybe hundreds of thousands" and that "there needs to be a place on the internet to discover these and understand what they do" [3]. In both cases the product started as a browsing and discovery tool and only later hardened into a marketplace [3][2].

The second constant is a refusal to believe in a single winner. He built OpenSea on the premise that NFTs would end up spread across many blockchains and that users would still want one place to see everything [5], and he started OpenRouter to answer one question: "Will this market be winner take all?" [3]. His answer, arrived at through data rather than conviction, is no, and everything about the product follows from that.

On new internet primitives and the discovery layer

He came to crypto in 2017 out of a nostalgia for missed revolutions, looked at the ICO boom and found that "everyone was incentivized for the wrong reasons nobody was actually caring what the tokens could do" [1]. The exceptions were the utility-focused, humble projects that gave tokens away and grew organically, and then CryptoKitties, which was "significantly different from all the other ico booms that year" because it involved "a lot more than just flipping coins" [1][2]. What made it matter was the standard: once several projects were building on the same ERC standard, "having a place to find all of these projects uh was kind of a clear need" [2]. He describes the resulting product as "a kind of a combined explorer google marketplace" [2], and for newcomers as "a little bit like amazon but for digital goods" mixed with eBay, where creators take a royalty on every resale and items never sit inside the company's servers [2].

The same instinct produced OpenRouter. Before building it he ran an experiment on whether users would bring their own model to generic websites, launching window AI in April, "an open- source Chrome extension that let a user choose their model and let a web app just kind of suck it in" [3], explicitly patterned on the web3 wallet he knew from crypto [4]. OpenRouter then began as a collection point and a place to explore data about who was using each model, and became a marketplace only when it became clear many companies would host the same models "at very different prices and performances" [3]. The scale argument he made about NFTs has an echo here: he thought the number of NFT collections "may one day exceed the number of websites" [5], and he thinks the model long tail will be similarly vast [3][4].

On many chains, many providers, and a balkanized world

His stated expectation for crypto was that "there will probably be many blockchains and many ways of scaling what's happening today on ethereum," with developers choosing chains for different reasons and each growing its own culture, so that "we may just be in a very balkanized world in the future" and users will want one place to see it all [5]. He rated Polygon and Solana on execution rather than technology, praising Polygon's developer documentation, its active engagement with gaming, NFT and DeFi partners, and above all its "speed of iteration getting feedback from all their partners and their users building and iterating again" [1]. He was candid that the underlying infrastructure is immature: "ethereum doesn't really work," with constant forks shifting standard behaviour [1]. Practically, he pushed creators toward Polygon because gas is cheap enough that OpenSea can pay it for buyers and sellers, and because the product detects Ethereum holdings and helps users bridge rather than making them buy MATIC [1].

The same heterogeneity showed up in inference. Feature support fragmented across providers, some with the min-p sampler, some with caching, tool calling or structured outputs, until "the ecosystem was just ballooning into this kind of outofcontrol heterogeneous monster and we wanted to tame the monster" [3]. Aggregating providers per model, one of which now has 23, both created price competition and materially boosted uptime for closed models that could not keep up with demand [3]. That fragmentation problem across 400-plus models is the core of how he explains the company [3][6][8].

On whether intelligence will be winner-take-all

He frames inference as possibly "the largest market ever in software," which made the winner-take-all question worth a company [3]. The evidence he cites is his own token data by model author: Google Gemini going from roughly two or three percent in June to 34 or 35 percent over twelve months, Anthropic among the most popular, and OpenAI under-represented because developers use OpenRouter to get OpenAI-like behaviour for every other model [3]. His conclusion is that "the future is going to be multimodel," that customers "unlock huge gains" by using different models for different purposes, and that models are sticky rather than churned through, with the number of actively used models steadily rising [3]. Underneath that sits the harder claim: "Inference is also a commodity" [3]. Claude from Bedrock should look exactly like Claude from Vertex, because two hyperscalers are delivering the same good at different rates and a developer should be able to select without caring who serves it [3]. If inference becomes "a dominant operating expense," then "selecting and routing will be crucial" [3]. He has since engaged directly with the counter-question of whether the routing layer itself is becoming a commodity, alongside the strength of Chinese open models relative to America and why enterprises may fear Anthropic and OpenAI more than China [9].

On distillation, data as the moat, and the long tail

The pivotal moment for him was Alpaca in March 2023, when a Stanford group fine-tuned Llama 1 on GPT-3 outputs for under $600: "the first time I saw the transference of both style and knowledge from a large model onto a small one" [3]. Two consequences followed. You no longer need a ten million dollar training budget, and "for the first time" you can "make unique data available as a service in the form of a language model" [3]. He calls this "knowledge finally being distilled into software" [3], and in developer terms, "clearly we can actually get within shooting distance" of the closed models with one computer [4]. The corollary is that "data becomes more of the moat," and he illustrates it with a person who is not the smartest in the room but knows a set of things nobody else does, knowledge the smarter person simply cannot derive [4]. That creates an economy for data that previously had no way of being sold [4]. It also creates the discovery problem he built for: an open weights model is still closed in a sense, "a black box" where you get billions of floating point numbers and no idea what it is good at [3]. Hugging Face existed, but "you couldn't really use the models and inference was tricky" [4].

On measuring models by usage rather than benchmarks

He deliberately publishes traction, not quality. The rankings began as raw tokens in and tokens out, which he admits has a real flaw: one developer going bananas can skew a model's numbers, though they have to pay a lot to do it and he is not aware of anyone gaming it [4]. What the page shows is "not really an eval it's more of a you know app Annie style like engagement metric" for every model [4]. He wants to move toward retention as the ranking signal, on the analogy of website analytics: raw traffic is the simple metric, but "what really matters is like are people sticking" and coming back with the same kind of prompt [4]. Developers can opt into sharing prompt data for a discount, which is used to classify incoming prompts and work out which models are good at what, feeding future automatic routing [4]. The demand he sees is concrete: people who already know the model they want but not where to get it, and people trying to find which model is really good at finance, roleplay, programming or machine translation [4]. Better discovery with fine-grained categorisation is on the roadmap, down to seeing the best models that take Japanese and produce Python [3].

On standards, and when to ignore them

His view is that "there's a healthy Duality between standards and standard Breakers" [4]. Standards help consumers because they lower the barrier for new entrants, make switching cheap and therefore increase competition, quality and price pressure; they hurt consumers when everyone locks in and deviating becomes impossible [4]. He credits ChatML as the first attempt he knows of to standardise the prompt layer, simple and extensible, and says that is part of why OpenRouter's API was built to look like OpenAI's and be a superset of it [4]. He rates that standard as healthy precisely because there is an easy opt out: you can always send a raw prompt, and model developers do innovate on prompt format, as with the research finding that a model performed better when the assistant believed it was GPT-4 [4].

The advice he gives builders is to resist standardising too early. Developers with an idea tend to open an EIP on GitHub first, when "really the first thing they should do is is test the idea with users and build the app"; filing for comment before there is any usage is "usually I think really premature" [2]. Get usage, learn from it, adjust the protocol, then write the standard [2]. He also explains why crypto's tooling fragmented where AI's converged: chains are fundamentally financial, conforming to the same RPC standard is how you participate and earn rewards, so deviating means launching a new blockchain with its own client APIs [4].

On the engineering of a router

Routing latency was the first hard problem, and the goal was near zero impact from sitting in the path. They pushed logic and infrastructure to the edge, leaning heavily on Cloudflare, including Hyperdrive to run SQL cached in the user's edge region so database work costs almost nothing [4]. He puts the result at about 30 milliseconds, "the best in the industry I believe," achieved with a lot of custom cache work [3], and notes "I haven't heard anybody complain about routing latency in like many many many months" [4]. Analytics was the second: homegrown Postgres tooling that worked but will not scale, though he rates the Postgres ecosystem as underrated, especially triggers, which have scaled well, with TimescaleDB as the next generation and cron jobs on the way out [4]. Third, type safety, where they "decided early on to get really crazy about type safety" with extremely strict checking across the codebase, treated as a foundational engineering principle he had not applied at a previous company, made necessary by the sheer number of schemas and provider formats flowing through the system [4].

The messy edges are provider behaviour. Every new LLM provider's API differs somewhere, and stream cancellation policies vary so much that dropping a stream may bill you for the whole completion, for nothing, or for twenty tokens you never received [3][4]. Rather than build an MCP marketplace, they built middleware, borrowing the concept from web development and making it AI native, because MCP-style pre-flight calls alone cannot transform outputs on the way back to the user; the web search plugin annotates results from any model live in the stream [3]. Ahead of that he expects more modalities, particularly transfusion models mixing transformers with stable diffusion to give images world knowledge and conversational editing, plus geographic routing to the right GPU in the right place, enterprise optimisation, prompt observability and better prices [3]. The stated posture is collaboration and "building an ecosystem that's durable and with low vendor lock in" [3].

On who builds on it, and learning from power users

Usage skews indie developer, with some companies using it to benchmark new models as they land [4]. The dominant category is B2C apps: games, roleplay apps, novel writing assistants, and above all applications whose output is generated code rendered into experiences, plus straightforward programming assistants [4]. The earliest demand signal for non-OpenAI models was moderation, users wanting to know whether they would be deplatformed and what a provider's policy actually was, prompted by cases like a detective novel where chapter four contains a murder [3].

He treats a non-developer constituency as an asset: "this community of power users they're not developers they're not um companies they're just normal users who love llms" who connect an OpenRouter account to bring-your-own-key apps or sign in with OpenRouter [4]. They surface the niche finish reasons and strange errors that appear when a new model or API is added, which turn into immediate alerts and fixes [4].

On communities, ownership and untapped ideas

His definition of a web3 community is precise: in web2 it is a group with a shared interest, in web3 "we're taking it one step further and adding shared incentives," and NFTs add fixed or predictable supply, so membership is provable [1]. The effect is that holders market, build interoperable products and do partnerships on the creator's behalf, making it "a zero or negative marginal cost situation" for a creator to stand up a community [1], a dynamic he elsewhere calls getting a community evangelist without having to do anything [5]. The growth he saw came from projects blending art with technology and giving token holders something to do, from Hashmasks to a custom web experience, community access or DeFi hooks, and generally from projects mixing two areas such as gaming and virtual worlds, or virtual worlds and DeFi [1][2]. Ownership is the part he thinks web2 users only get through repetition: buying a shirt on OpenSea and finding it in your backpack inside Decentraland with no API integration required, until "you suddenly realize oh this stuff is like actually things that i think about now" [5].

He is expansive about what has not been built. Font licences, where you have no idea whether the calligrapher gets paid and where scarcity would add value, are a good fit for royalties and market-set prices [2]. Waitlist positions should be tradable, for apps, communities and crypto projects alike [2]. Usernames in crypto apps and site real estate for ads or content should be tokenised and mostly are not [2]. On intellectual property his answer has two halves: platforms already handle copyright takedowns and NFTs change little there, "it's just following copyright law"; but tokenised licences offer a way forward, so that a producer could sell a fixed number of licences to a beat and YouTube could ping the smart contract to verify a creator's right to use it [2].

On competition as a rising tide

He applies the same analysis to Coinbase entering NFTs and to smaller marketplaces: "rising tide lifts all boats especially applies to web3" because everyone builds on the same underlying blockchain infrastructure, so collaboration is easy and users can pick whatever suits them [1]. In practice new competitors onboard groups of users OpenSea was not serving well, who then get curious about the rest of the ecosystem [1]. He describes "really thin borders between companies" in web3, a shared API in the form of the blockchain creating a tech tree of composable building blocks [5], and says they kept group chats with competing marketplaces, helping them in the name of the best user experience, since items created elsewhere need to show up quickly and accurately [5].

On founding, teams and staying calm

Asked how long product-market fit took at OpenSea, his answer is that it was fast because the market was tiny: they were the only multi-project marketplace, making a binary bet that NFTs either were a real building block like web pages or streaming video, or nobody would care [5]. That was followed by many slow months, months of decline, and volume concentrated in a single project before it diversified [1]. He credits Y Combinator's partners for the focus on "building something people want and talking to users aggressively," which "became a big part of our dna," even though they pivoted from a router-mining WiFi coin idea on the first day of the batch and had to correct the introduction on stage [1].

On scaling from under 20 people to about 300, he names the hardest challenge as "knowing what you're not good at and being very very honest about it," with the founding goal of wearing all the hats "sort of incompetently" until someone who clears a very high bar can take each one [5]. Holding those bars is a daily discipline, and the payoff is that working with the people hired became his favourite part of the day [5]. His own routine is deliberately narrow: pick the number one problem or goal each week, block out early morning time before Slack fills up, walk while thinking about it, meditate occasionally, and travel once a month because time on an airplane creates distance and lets him reorganise around the bigger picture [5]. Asked how he stays so calm, the answer is blunt: "otherwise i burn out" [5].

Takeaways

• His core market bet is that inference is not winner-take-all: token data by author shows Gemini growing from roughly 2 to 35 percent in a year, and he concludes "the future is going to be multimodel" with inference itself "a commodity" [3].
• Alpaca, a $600 fine-tune of Llama 1 on GPT-3 outputs, was the turning point, proving distillation transfers style and knowledge and making "data becomes more of the moat" the operating thesis [3][4].
• OpenRouter's rankings measure traction, not quality: raw tokens today, "app Annie style" engagement rather than an eval, with retention as the ranking metric he wants next [4].
• Build usage before you build standards: filing an EIP before anyone uses the idea is premature, and the right order is test with users, learn, then standardise [2].
• Strict type safety across the codebase is treated as a foundational engineering principle, made necessary by the number of provider schemas and formats flowing through the router [4].
• Edge infrastructure, including Cloudflare Hyperdrive, cut routing overhead to roughly 30 milliseconds; MCP was rejected in favour of AI-native middleware because outputs also need transforming on the way back to the user [3][4].
• Competitors onboard users you were failing to reach: "rising tide lifts all boats" held at OpenSea, where he kept group chats with rival marketplaces [1][5].
• NFT communities work because shared interest plus shared incentives plus fixed supply turns holders into marketers, making community-building near zero marginal cost for a creator [1][5].
• The founder's hardest job is "knowing what you're not good at and being very very honest about it," wearing hats incompetently until a higher-bar hire takes them [5].