The Real Threat to AI Isn't a Bubble. It's Free.

On June 12th the US Commerce Department ordered Anthropic to disable its two most powerful models (Fable 5 and Mythos 5) for all foreign nationals. The next day Z.AI released GLM-5.2 under an MIT license with no regional restrictions, available for anyone to download, modify, and run locally without asking permission.

OpenAI is valued at $852B and Anthropic at $965B. Together they represent nearly $2T in private market valuation built on a single assumption: frontier AI commands premium pricing permanently. GLM-5.2 performs at one-sixth the cost of Western equivalents on the benchmarks that matter to enterprise buyers. It cannot be geofenced or revoked, and for users with the hardware to run it locally, it costs nothing beyond the electricity to power it.

The Companies Behind the Valuations

Before the open-source argument, we need to break down who owns what. Microsoft owns ~27% of OpenAI. Google committed $43B to Anthropic and owns 14% (capped at 15%). Amazon owns an estimated mid-teens percentage of Anthropic, with up to $33B committed. Nvidia committed $30B to OpenAI and $10B to Anthropic before CEO Jensen Huang confirmed in March 2026 that these would be the company's final direct investments in major AI labs. Salesforce holds ~$5B in Anthropic, owning roughly 0.5%. The precise stakes become public when Anthropic's S-1 is unsealed ahead of its expected October IPO, this will be the first time the full ownership structure is disclosed.

In Q1 2026 Alphabet (Google's parent company) reported record profits of $62.6B, ~$28.7B of that (almost half) came from marking up the value of its Anthropic stake as Anthropic's funding round price rose. Amazon's Q1 net income included $16.8B in pre-tax Anthropic gains, which were more than half of their pre-tax income for the quarter.

These are paper profits and only exist because the most recent Anthropic funding round was priced higher than the previous one. They are not cash or revenue; they depend entirely on the assumption that AI premium pricing holds, and this rests on there being no free alternative.

But now there are free alternatives.

What GLM-5.2 Is

Z.AI is the brand of Zhipu AI, a Beijing-based AI lab founded in 2019 with backing from Tencent, Alibaba, and the Chinese government.

GLM-5.2 was released on June 16th. 753B total parameters using a mixture-of-experts architecture, meaning ~40B parameters activate per token, making it far more efficient than a dense 753B model, though still requiring professional-grade hardware to run locally. 1M token context window. MIT license with explicit language guaranteeing "no regional limits" and "technical access without borders."

API pricing: $1.40 per million input tokens, $4.40 per million output tokens. Claude Opus 4.8 costs $15 per million input tokens and $75 per million output tokens. The same task costs around one-sixth as much on GLM-5.2 via API.

On benchmarks: GLM-5.2 beats GPT-5.5 on multiple long-horizon coding evaluations, and it performs at/near Claude Opus 4.8 on maths reasoning. Cline IDE confirmed it is the first open model to cross 80% on Terminal-Bench. Z.AI published no official benchmark numbers at launch; these figures come from third-party evaluators.

The announcement for GLM-5.2 was on June 13th, the day after the US Commerce Department ordered Anthropic to disable Fable 5 and Mythos 5 for foreign nationals. The company's documentation explicitly describes the release as a response to "the geopolitical restriction of AI access."

GLM-5.2 is not an isolated release; it follows DeepSeek V4, Alibaba's Qwen family (1B+ downloads in January 2026), and Moonshot AI shipping a trillion-parameter open model in June. Epoch AI's analysis puts Chinese open-weight models at around seven months behind US frontier models on average, down from 14 months in 2023. The Stanford AI Index confirms the performance gap between the top-ranked model and the 10th fell from 11.9% to 5.4% in a single year.

The Open-Source Privacy Argument Removes the China Problem

The standard objection to using Chinese AI models is data sovereignty. If your queries go through Z.AI's servers, Chinese infrastructure processes them, which is a legitimate concern for regulated or sensitive industries. GLM-5.2 running locally eliminates that concern entirely.

An MIT license means anyone can download the model, run it on their own hardware, modify it and distribute it. No query leaves your building, none of your data touches Z.AI's servers, and the Chinese government can't access your prompts. The privacy objection to Chinese AI is an objection to using Chinese AI via API, not to running Chinese open-source models locally.

Open weights also mean the model can be modified. Companies can fine-tune open-source models on proprietary data that never leaves their infrastructure, strip out behaviours they don't want, or distil it into smaller specialised models they own outright. OpenAI and Anthropic offer fine-tuning only through their APIs, within their constraints, on their servers.

This issue matters because it removes the primary argument for paying premium prices for Western AI. Data sovereignty, keeping sensitive information away from external servers, is achievable with local deployment. Running GLM-5.2 locally gives you data sovereignty against China, the US government, Anthropic, OpenAI, and every other third party simultaneously.

OpenAI updated its US privacy policy in April 2026, formalising advertising in ChatGPT, contextual ads in the free tier, purchase data received from advertisers, and user information shared with marketing partners for ad targeting. 94.5% of ChatGPT's 800M weekly users are on the free tier and OpenAI is burning $27B per year, and advertising is the only mechanism for monetising that user base at scale. The conversations those 800M users have with ChatGPT train the model by default (unless they opt out), inform the ad targeting, and compound the value of the platform.

A local model running on your own hardware generates none of that data; the user gets the AI and the AI company gets nothing.

What Your Average PC Can Run Today

The Steam Hardware Survey from April 2026 (the most comprehensive real-world PC hardware dataset available) shows what average consumer hardware looks like.

The most common set-up: RTX 3060 GPU (3.99%), 12GB VRAM, 16GB RAM (40.72%), and 6-core processor.

On that hardware today, the average person can run 7B to 13B parameter models at full quality. These include Llama 3.3 8B, Gemma 3 9B, Mistral 7B, and Phi-3 medium. These models handle coding assistance, summarisation, writing, question answering, and general chat. They are not frontier models, but for most everyday tasks the difference is not perceptible.

Quantised 27B parameter models also run on 12GB VRAM with modest quality reduction. Qwen3.6-27B runs on that hardware and delivers what MindStudio describes as "strong SWE-Bench performance" on coding benchmarks.

What the average Steam user cannot run: GLM-5.2 at 753B parameters, as it requires multiple professional-grade GPUs with 80GB or more VRAM combined; that is server hardware, not a gaming PC.

The average person today gets capable local AI (not frontier), and the gap only matters for complex reasoning and tasks at the frontier of what models can do. It is more than enough for writing emails, summarising a document, debugging code, or answering questions, which is what most people actually use AI for.

The Landscape a Year From Now

For hardware, the RTX 50 series ships 16GB VRAM on midrange cards as standard vs. 8GB on equivalent cards two years ago; by mid-2027 the average consumer GPU will have 16 to 24GB VRAM. A 70B parameter model quantised to run on 24GB VRAM is already possible today with an RTX 4090; by mid-2027 that hardware will be midrange. Google's TurboQuant research, published at ICLR 2026, reduces the memory overhead required for long-context models on consumer hardware and is already being integrated into open source tools.

For models, the lag between open-source and frontier is around seven months today; in 2023 the lag was 14 months. The gap is reducing by roughly 1-2 months per year; by mid-2027 the lag is likely 4-5 months. At that pace, the average consumer PC in mid-2027 runs models that are today considered competitive midrange (roughly GPT-4o). Still not frontier, but good enough for 90% of tasks the average user actually performs.

Z.AI has said its target is a Mythos-level model within six months, meaning a model performing at frontier level, available free, by the end of the year. Whether that timeline holds true is unknown, but the trajectory of GLM releases (GLM-4 in 2024, GLM-5.1 in early 2026, GLM-5.2 in June 2026) is consistent with a company closing the gap fast.

The Two-Tier AI the US government Is Creating

The Commerce Department's June 12th order made frontier AI legally concentrated among US entities and approved partners, while everyone else has access to capable but not frontier open-source Chinese alternatives.

The strategy works if the quality gap between the restricted Western frontier and freely available open source remains large enough to matter for real applications. It fails if the gap narrows to the point where open-source is good enough for most purposes.

For most everyday tasks that point may already be here. For genuinely difficult reasoning, novel scientific problems, and complex agentic workflows, the frontier is still in the lead. The question is how many users actually need frontier models and whether that number justifies an almost $2T valuation.

The export control regime accelerates exactly the dynamic it is trying to prevent. Restricting Western AI for foreign nationals removes the argument for buying Western AI at premium prices and gives Z.AI a direct commercial and political reason to keep releasing capable open-source models. Every restriction on US AI companies is an advertisement for open source models.

On June 30th, the Trump administration lifted all export controls on Mythos 5 and Fable 5, with access restoring July 1st. US tech executives warned the restriction was gifting China a competitive window, and the government reversed under that pressure 18 days later.

The Circular AI Economy

Google committed $43B to Anthropic. Anthropic then uses Google Cloud as its primary cloud provider, paying Google for compute. Google marks up its Anthropic stake as valuations rise, reporting record profits on paper gains for a company that pays Google for its infrastructure. Amazon committed $33B to Anthropic. Anthropic runs its largest deployments on AWS, Amazon reports $16.8B in Anthropic gains in a single quarter. Microsoft owns 27% of OpenAI. OpenAI pays Microsoft for Azure compute; Microsoft's AI revenue includes OpenAI workloads running on its infrastructure.

The cloud giants are simultaneously the investors in, the infrastructure providers for, and the reported beneficiaries of AI valuations. When Anthropic raises a round at a higher price, Google and Amazon report higher profits. When OpenAI's revenue grows, Microsoft's Azure revenue grows.

Nvidia confirmed in March 2026 it will make no further direct equity investments in major AI labs. CEO Jensen Huang's stated reason was the coming IPOs, though the announcement followed months of criticism that deals like Nvidia's $100B OpenAI commitment (where OpenAI would spend the money on Nvidia chips) amounted to circular revenue. The $700B in planned AI capital expenditure for 2026 is flowing through traditional capital expenditure channels rather than equity investments.

The valuation of Anthropic at $965B rests on its revenue (currently a $47B annualised run rate) continuing to compound at current rates. That revenue rests on enterprises paying $15 per million tokens for Claude when an MIT-licensed alternative charges $1.40 for comparable performance. Companies like Uber have already confirmed they burned through entire AI budgets in months at current token pricing. The search for cheaper alternatives is not hypothetical; those discussions are happening right now.

If open-source models reach 90% of Western frontier performance in the next 12 months, the pricing premium that justifies the almost $2T valuation of Anthropic and OpenAI compresses. None of this happens overnight; revenue compounds, enterprise contracts are locked in for set periods, and the first-to-market advantage is real.

OpenAI is already responding to pricing pressure, weighing token price cuts to fend off Anthropic. This would compress already-thin gross margins at a time when inference costs are rising from $8.4B in 2025 to $14.1B in 2026. A sustained price war between the two largest Western AI labs, driven by competition from each other and from open source alternatives, pushes cash-flow breakeven further out with OpenAI's current estimate at 2030.

My View

The standard AI bubble argument is demand destruction: AI does not generate enough returns to justify the capex. That may prove correct, but it is not the argument I have outlined.

The supply normalisation thesis is different. Not that AI demand collapses, but that the pricing premium sustaining $2T in OpenAI and Anthropic erodes as open-source alternatives improve. Demand for AI is not going anywhere; the question is whether users pay $15 per million tokens or host it themselves on their own hardware.

OpenAI operating loss hit $20B in 2025 and is projected to reach $27B in 2026. Anthropic projected its first operating profit for Q2 2026 ($559M on $10.9B quarterly revenue). Both of these companies depend on premium pricing holding. GLM-5.2 is another straw on the camel's back, we saw the market reaction to hyperscalers when DeepSeek released a free model in January 2025. Multiple Chinese AI labs are releasing capable open-source models at a fraction of Western API costs.

The US government's response of restricting access to frontier Western AI for foreign nationals (then removing this restriction 18 days later) concentrates the most powerful tools among the fewest users while giving the rest of the world a commercial and political argument for developing alternatives.

A $2T valuation built on permanent AI premium pricing is not going to collapse overnight because one Chinese lab released an open-source model. But it might compress significantly if the pattern holds for another 12 months and the capability gap continues to narrow.

Ticker Thoughts is independent analysis and not financial advice. No position held in any of the tickers mentioned at the time of publication. All open and closed positions are detailed on the positions page.