In 1994, Netscape launched its browser Navigator, which cost around fifty dollars at the time. For a pretty long moment it worked and led to a very successful IPO, because Netscape was the web and the web was Netscape. Then Microsoft bundled Internet Explorer into Windows and gave it away for free. The paid browser was dead inside two years. Netscape open-sourced what was left, and that code became Mozilla, and Mozilla became Firefox. Years later Google shipped Chrome, free, faster, everywhere, and won.
Essentially, the browser became a commodity. But the web did not become worthless when the browser became free – rather, the value moved to another part of the stack. It went up, to the layer built on top. Google, Amazon, eBay, Shopify all monetised the long tail. All of the value diffused to the application layer and it has never looked back.
The ultimate winner of this era is undoubtedly Google, but even Google did not make money on Chrome. Chrome was a free distribution move to protect the thing that actually printed money, the best ad business in history.
I believe the foundational model layer will follow this trend.
The cost of intelligence is collapsing at an alarming rate. On OpenRouter, the largest neutral marketplace routing developer traffic across hundreds of models, US-origin closed models fell from around 70% of token volume in June 2025 to roughly 30% in June 2026. Open-weight models, most of them Chinese (DeepSeek, Qwen, Kimi, etc.) now run close to a third of all tokens on Vercel’s enterprise gateway, up from a ninth two months earlier. On OpenRouter, DeepSeek is now the single largest provider by token volume. DeepSeek alone processes more tokens than Google, Anthropic or OpenAI. These models are 60-90% cheaper, and for the workloads that burn the most tokens they are already good enough.
To see where this argument actually comes from, look at one model. In April, Moonshot AI released Kimi K2.6, open-weight, self-hostable, free to download. On SWE-Bench Pro, the benchmark that best mirrors real engineering work, it scored 58.6%, edging ahead of both GPT-5.4 and Claude’s then-current flagship. An open model had beaten the closed labs on serious coding. And it did it on price: roughly $0.95 per million input tokens and $4 per million output, against $5 and $25 for Claude Opus 5-6 times cheaper on the API, and cheaper still once you self-host the weights. DeepSeek undercuts even that. The frontier answered within weeks and retook the coding lead, as it usually does. But that is the pattern, the gap opens, an open model closes it inside a single release cycle, and the price of good enough keeps falling through the floor.
There is no doubt the frontier is still ahead. Claude is genuinely the best thing in the world at code and I use it every day. Codex is also gaining market share. On that same Vercel gateway, open models run about a third of the volume, yet on under 4% of the spend. The dollars still sit with the frontier, because the high-value work still routes to the best model.
So the question is not whether intelligence commoditises, it seems inevitable it will on a volume basis. The question is how large and how durable the frontier is, and whether it is worth a combined $2 trillion. A narrow, shrinking premium on a commoditising input does not hold a trillion-dollar mark.
Combined, OpenAI and Anthropic are worth about $1.8 trillion in the private market today, and both are pointed at public listings that would put the pair near $2 trillion.
Anthropic has an estimated current run rate of ~$70bn, and that run-rate has a concentration risk sitting inside it. Marc Benioff recently outlined that Salesforce will spend around $300 million on Anthropic this year, most of it on coding. But in the same breath he called for an intermediary layer that routes routine work to cheaper models and saves the frontier for the hard problems. The question of durability is real: when an open-source model does that same coding at 90% of the performance for a fraction of the price, how long before Salesforce’s own shareholders start asking why $300 million is buying a frontier premium the workload no longer needs?
If those valuations are a bet on selling intelligence – the raw provision of tokens, the model as the product – then I think it’s difficult. The path to the valuation runs through the application layer, exactly like it did for Google. What it does not run through is the sale of intelligence itself.
When the middle commoditises, the margin leaves it in two directions. It goes down, to infrastructure (the compute, the cloud, etc.) that everyone needs no matter which model wins. And it goes up, to the applications, being the companies that take cheap intelligence and integrate it into workflows, datasets, a distribution channel and the like.
At Quadri, we have always believed this is where the true value would sit, and have constructed our portfolio with this inevitability in mind. For example:
- Below the model is emma, the compute layer itself. A sovereign, multi-cloud control plane that sits over raw GPU capacity and manages the cost, complexity and jurisdiction of everything running on top. It gets more valuable the more compute the world burns, not less.
- Above it are Allstacks and Corgi. Allstacks points models at the proprietary graph of how a company actually ships software, and turns that into decisions engineering leaders pay for. Corgi is rebuilding commercial insurance from the ground up – AI-native underwriting in a multi-trillion-dollar market modern software had barely touched, and one of the fastest revenue growth trajectories in the world. It sits above the model, not below it. It consumes intelligence; it does not sell it.
- Further, also above the model we have Trase, an Ai agent operating system for secure and regulated industries. What it manages is not the intelligence but the right to use it, I.e., governance, audit ability and sovereignty in enterprise and government environments.
For these companies, falling model cost is a tailwind for the layer that owns something the model does not.