Anthropic files for IPO at a near trillion dollar valuation
Anthropic confidentially filed a draft IPO with SEC on June 1, 2026. No price, ticker, or IPO date has been officially set yet, but clearly, the AI IPO race is no longer theoretical (OpenAI also working on a similar project).
Anthropic’s revenue run rate reached around $47 billion in May 2026, up from roughly $10 billion a year earlier. Its latest funding round (Series H) pushed the company to a valuation of around $965 billion. This would make a potential listing one of the largest tech IPOs ever.
The narrative is interesting. OpenAI has the consumer brand and massive ChatGPT distribution. Anthropic has the enterprise story: Claude Code, strong traction with developers, safety positioning, and now cybersecurity through Mythos and Project Glasswing (See next section).
But there is a big question behind the excitement: margins. Frontier AI is still extremely expensive to run. Compute contracts, inference costs, data center access, and energy are not small operational details. They are becoming the core economics of the business. As example, Anthropic is paying SpaceX $1.25 billion per month through May 2029 for compute. That is $15 billion per year to a single vendor. It is a remarkable number that reflects both the cost of running frontier AI at scale and the degree of infrastructure concentration the industry has developed
My view: Investors are no longer only betting on model quality. They are betting on whether AI labs can convert usage into durable and forecastable revenue, and revenue into profit. The full equation is still to be solved.
Claude Fable 5 and Mythos: the new dual use AI template
Anthropic also launched Claude Fable 5 for general use and Claude Mythos for selected cyberdefenders, infrastructure providers, and later some life science researchers. The key detail is that both are based on the same underlying model.
Fable 5 is the public version, with conservative safeguards. Mythos is the restricted version, with some safeguards lifted for vetted users. Anthropic says the model is especially strong in software engineering, research, vision, long context work, cybersecurity, and some scientific tasks. Pricing is set at $10 per million input tokens and $50 per million output tokens, which is less expensive than the earlier Mythos preview.
This dual release is important because it shows how frontier labs may increasingly ship the same capability in different access layers. One version for the general market. Another for trusted users. Another perhaps for governments. The model is the same, but the permissions around it are different.
My take: this is an interesting turn on probably a future of "dual" use AI. The most powerful models will not be released in a simple public or private binary. They will come with access tiers, monitoring, retention rules, invisible controls, (and government involvement?). That may be necessary for safety. But it also creates a transparency problem, and from the moment a model is released, even within a limited population, can it really be controlled ? I do not think so. So the risk of those "dangerous" models going out of hand is a risk not yet fully accounted for.
Microsoft goes model independent with seven MAI models
Microsoft introduced seven in house models under the MAI brand, including MAI Thinking 1, MAI Code 1 Flash, MAI Image 2.5, MAI Transcribe 1.5, MAI Voice 2, and additional reasoning models for Azure AI Foundry. It is also slowingly moving out from its previous position of ChatGPT/OpenAI distributor channel. The most strategic one is probably MAI Code 1 Flash, a Microsoft built coding model integrated into GitHub Copilot and Visual Studio Code.
Microsoft has spent the last years benefiting massively from its OpenAI partnership. But dependence on one external model provider is risky, especially when that provider is preparing for its own public market story (andpossibly IPO) and building more direct enterprise relationships.
Copilot is becoming a model independent enterprise platform where Microsoft can mix OpenAI, its own models, and possibly other providers depending on cost, latency, privacy, and use case. And by the way, the model dispatcher capability is becoming a real solution given the variance in token prices vs accuracy vs needs.
My view: Microsoft is building the AI control plane for enterprises. The model is becoming one component of the stack, not the whole strategy. Also, over the years, we have seen multiple cases where Microsoft was lagging behind, but gradually taking the lead. Microsofot consistency overtime still surprises me. Everytime we think they might be behind including the AI race, they manage to catch up, and in some cases, lead.
DeepSeek returns: cheap AI starts eating token volume
DeepSeek had a very strong month. The Chinese AI startup is reportedly preparing a first funding round of around $7.4 billion, with investors including Tencent and CATL. At the same time, DeepSeek V4 models are gaining real production usage because of their aggressive pricing.
The Vercel AI Gateway data is probably the most interesting signal here. DeepSeek’s share of tokens jumped from less than 1 percent to 17 percent in one month, while its share of spend stayed near 1 percent. DeepSeek is absorbing a large amount of usage without absorbing much of the money.
My take: This will push more companies toward a smarter model mix. A cheap model can draft, classify, summarize, route, or propose. A stronger model can verify, reason, review, or make the final recommendation. In other words, companies might increasingly use low cost models as workers and frontier models as reviewers or advisors. Another option, would be to have smarter models "plan", and lower ones, "execute".
But, this is also one of the biggest threats to frontier AI pricing power. If enterprises learn to route most token volume to cheaper models, OpenAI and Anthropic may still dominate prestige and high stakes work, but not necessarily every dollar of AI usage. Cheap AI does not need to be better than frontier AI. It just needs to be good enough for enough tasks.
US AI governance: softer rules, stronger politics
The US AI governance story remains unclear. Trump first postponed an AI executive order in May, saying he did not want to weaken the US position against China. A narrower version was then signed in June, based on voluntary cybersecurity testing of frontier models before public release.
David Sacks has been one of the loudest voices against heavy AI regulation, warning that an FDA style approach could slow the US and help China. That argument is now shaping the policy tone: avoid mandatory licensing, avoid slowing model releases, and keep regulation light unless a clear problem is proven.
At the same time, Congress is discussing a large federal AI bill that could freeze some state level AI rules for three years. Colorado’s AI law is also approaching enforcement, which creates a direct collision between state regulation and federal preemption.
My view: the real issue is not regulation versus innovation. It is governance quality. Bad regulation can slow useful innovation. But weak governance can also create legal risk, trust issues, and public backlash.
