Welcome to Ignition, Catalyst Investors’ briefing on what we’re seeing at the intersection of AI, robotics, software, and growth equity, combining the best of human and AI expertise.
In this issue:
- The investment implications of the shift toward open-weight models
01 — THE SIGNAL
The Open-Weight Squeeze: What the Server Wars Tell Us About Frontier Lab Valuations
As Anthropic and OpenAI move toward IPOs reportedly valued north of a trillion dollars, we keep circling the same question from the investor seat. How durable is the frontier-model advantage? For the past several months we have been watching open-weight models, and the Chinese ones in particular (DeepSeek, Qwen, GLM, Kimi), to see what they do to the economics of the premium labs. If the frontier commoditizes, it does not just reprice OpenAI and Anthropic. It reprices the entire stack of infrastructure and application companies that we invest across.
We ran the numbers. We plotted model capability against estimated training cost across generations and labs, using Claude as our research engine. The result is below. The US frontier labs climb a steep, expensive curve. The Chinese labs track a roughly parallel curve well to the left of it, reaching comparable capability at a fraction of the spend. The Chinese labs achieve comparable performance levels at about a three-month lag, according to the underlying data.
CHART 1 – Capability vs. estimated training cost, by lab
A word on the cost data: no lab discloses training costs for its 2025 and 2026 models, so every dollar figure here is an estimate. The Chinese numbers are the softest of all and are almost certainly understated. DeepSeek’s headline figures exclude R&D, failed runs, and the full cost of infrastructure. Its latest model reportedly trained on Huawei silicon rather than Nvidia, which breaks any estimate built by analogy to Western hardware. There are also credible accusations, some from Anthropic itself, that a portion of the apparent efficiency is unauthorized distillation of US models rather than home-grown engineering.
Discount all of it as heavily as you like. The conclusion survives. The Chinese models are fast-following the US frontier, at a small fraction of the cost. The magnitude is in dispute. The direction is not.
The usage data says the same thing more bluntly. On OpenRouter, the neutral marketplace where developers route traffic across models with no lock-in, the US-model share of tokens has fallen from about 70% to about 30% in twelve months.
CHART 2 – OpenRouter token share, US vs. China
Two caveats. First, the “Other” slice is large and growing, and much of it is other Chinese models (Qwen, GLM, Kimi) that sit outside the chart’s named China group. The chart understates Chinese share, it does not inflate it. Second, this is routed inference volume, not the whole market. First-party ChatGPT, Gemini, and Claude traffic dwarfs it, and the premium labs still capture most of the revenue. Anthropic holds roughly 12% of tokens on the platform but close to 46% of the revenue. Volume and revenue have come apart.
Still, it is hard to see how this arrangement holds in its current form. AI application developers are the ones that largely route traffic to the most efficient model. Will developers lock into a premium tier priced at ten to thirty times the alternative, against a good-enough open tier that has already taken the volume? Something gives. The question is what.
The model wars in historical context – servers and PCs
For that, it helps to look at the last time a premium technology met a cheap-and-open one. In the early internet era, servers ran on premium proprietary Unix (names like Solaris, AIX, and HP-UX), sold on reliability and support, locked to specific hardware. Then two challengers arrived at once. Open-source Linux, free and running on commodity boxes. And Microsoft’s Windows NT, cheaper proprietary software that undercut Unix without being free. By 2000 the public web-server layer was roughly a dead heat. Linux eventually went on to dominate internet infrastructure. Windows NT carved out a large, durable enterprise server business. And premium Unix was pushed into a shrinking, high-value niche where reliability was worth paying for. In that story, the premium frontier labs look like Unix, or perhaps like Windows NT. The cheap open-weight models, and especially the Chinese ones, look like Linux.
There is a more optimistic precedent, however. On the PC desktop, in an earlier era, Microsoft Windows became the default operating system and stayed there for decades. Not because it was the cheapest or the best, but because of lock-in: OEM bundling, the Win32 application ecosystem, network effects that compounded. The moat was never the operating system. It was everything that was built on top of it. If a frontier lab becomes the platform that all AI applications are built on and depend on, it captures that kind of durable economics no matter how cheap the alternatives.
So, which is it? Three scenarios, where Microsoft plays two of the three roles, in two different markets.
Scenario one, the desktop-Windows outcome. A frontier lab becomes the default AI platform, with ecosystem lock-in and durable pricing power. This is the bull case, and it is the one the trillion-dollar valuations quietly assume.
Scenario two, the Unix outcome. The labs are relegated to a small, high-margin niche: the customers who genuinely need the absolute frontier and will pay for it. A real business, but a narrow one.
Scenario three, the Windows-NT-on-servers outcome. A large, broadly adopted, respectable revenue line whose pricing power is permanently capped because a cheaper, good-enough alternative sits right next to it. NT won meaningful server share and made real money. It just never got to charge Unix prices, because Linux was always the fallback.
We lean toward the third scenario. It is the scenario that fits the facts we can already see, including the one that would otherwise sink the bear case: Anthropic is reportedly running near a $47 billion revenue run-rate while losing token share. Scenario three holds both truths at once. The labs keep a substantial, growing business, but the constant presence of near-frontier cheap alternatives caps what they can charge, and therefore potentially limits the valuation multiple.
The desktop-Windows outcome looks hard to reach from here. The lock-in that made Windows durable is mostly absent at the model layer. Capable open-weight substitutes already exist. Switching between models is often a single line of code. The challengers are months, not years, behind. And application developers have every reason to stay portable rather than build their business on a single vendor like Anthropic or OpenAI that could later decide to compete with them directly, up the stack, for the same customers.
The honest counter is that AI does have lock-in vectors that server software never had. Distribution through ChatGPT and Claude Code. Consumer-facing businesses for ChatGPT and Gemini can be monetized with advertising. Enterprise trust and compliance, where a regulated buyer will pay up for a name they can defend. Accumulated memory and context. Agentic workflows that embed a model deep into how work gets done. None of these are as strong as Win32 was. But they aren’t nothing, and all are growing. We think they are real, but we think they are insufficient to produce the desktop-Windows outcome. Reasonable people can disagree, however, and it is the crux of the whole debate.
What also keeps this from collapsing into the pure Unix outcome is a disanalogy that runs the other way. Updating Linux is nearly free at the margin. It is distributed, incremental, and requires no single expensive event. A fast-follower AI model is not like that. To stay three months behind the frontier, a challenger still must run a tens-of-millions-of-dollars training run every generation. Much cheaper than the frontier, yes. Cheap, no.
That single fact cuts in the labs’ favor twice. It means scale and capital matter far more in AI than they ever did in server software, which is exactly why the open-weight winners are well-funded labs (DeepSeek’s hedge-fund parent, Alibaba, Zhipu, Moonshot) and not a distributed hobbyist community. The frontier labs’ resource advantage is more durable than Unix’s ever was against Linux. And it means the cheap tier cannot fall to zero, because fast-following is expensive enough that it consolidates around a handful of capitalized players rather than becoming truly free. That puts a higher floor under frontier pricing than the Linux analogy alone would suggest.
Scale and capital keep the labs alive, which rules out the Unix niche. Near-frontier cheap alternatives cap their pricing power, which rules out the Windows PC platform. What is left is the Windows NT outcome: a good business, permanently squeezed.
For a company priced on durable premium pricing power, that is the problem. We are not arguing that the frontier labs lack a viable business. They clearly have one, and a large one. We are arguing that the current valuations embed platform economics that the server wars suggest are very hard to hold when a capable, cheaper alternative is always one release behind. A highly competitive, capital-intensive business squeezed on price may not be a trillion-dollar business.
Which is why, for our own capital, we keep coming back to the layers around the model rather than the model itself. The infrastructure and compute-access layer, where the spending lands regardless of which model wins. And the application and workflow layer, where switching costs are real and the moat, as ever, is not the model. In the AI Game of Thrones, the throne may be the least defensible seat in the room.
*Sources and method: Capability-vs-cost chart built by Catalyst from public benchmark data (GPQA Diamond, SWE-bench Verified) and directional training-cost estimates; 2025–26 cost figures are unofficial and, for Chinese labs, likely understated. OpenRouter token-share chart: Bloomberg / Exponential View, OpenRouter data, second week of June 2026. Revenue and token-share figures: OpenRouter, company reporting, and press accounts as of July 2026. This is analysis, not investment advice.*
Forward to a founder, operator, or investor who is navigating AI adoption.
ir@catalyst.com
261 5th Ave
Suite 1102
New York NY 10016
