Four stories. Real talk. No corporate speak.
TRUMP UNLOCKS THE VAULT: GPT-5.6 CLEARS GOVERNMENT TESTING, LAUNCHES TO THE WORLD THIS THURSDAY
So here is what happened. OpenAI built one of the most powerful AI models ever made. The government looked at it and said, hold on, we need to run some tests first, you cannot just release this to 200 countries and everyone with a laptop. And OpenAI said, fine, but we really do not like this, that is not our preferred way to ship products. And the government said too bad, we are doing it anyway.
That was about two weeks ago. This week, the Department of Commerce’s Center for AI Standards and Innovation finished its review, gave OpenAI the green light, and GPT-5.6 is now cleared for a full global launch on Thursday, July 10. If you have been waiting to get your hands on it, Thursday is your day.
Think about that for a second. We are now in a world where the US government has veto power over when you get access to a new AI model. Not whether the model exists. Not whether it gets built. But when, and who gets it first. The government had its own technical experts sitting in sessions with OpenAI engineers in Washington to basically debrief them on what GPT-5.6 can do before anyone else got to see it.
This is genuinely new territory. Three years ago, AI companies released models like they were posting software updates on GitHub. You woke up in the morning, opened your feed, and read a blog post that said “we dropped something new, here you go.” Now there are Senate briefings and Commerce Department sign-offs and staged rollouts to approved partners before the general public gets near it.
The thing is, you can see why the government wants this. GPT-5.6 is a frontier model by any definition. The same restrictions were placed on Anthropic’s Fable 5 and Mythos models before them, with those getting cleared only at the end of June after Anthropic worked with officials to patch a jailbreak vulnerability that Amazon had flagged. So the pattern is becoming clear: build the model, brief the government, fix whatever concerns they raise, then ship.
Whether this is a reasonable precaution or the beginning of the government permanently inserting itself into the AI product release cycle is something reasonable people can disagree on. What is not really debatable is that Thursday will be a big day. GPT-5.6 reportedly contains significant reasoning improvements and has been specifically built for complex multi-step tasks. OpenAI has been running its Sol, Terra, and Luna model tiers under the 5.6 umbrella, with Sol being the most capable.
The truly ironic thing here is that some government-approved companies and agencies got access to GPT-5.6 in restricted preview mode a few weeks ago. So certain bureaucrats in Washington were testing the thing before regular paying ChatGPT subscribers could. If that does not tell you something about the direction this whole relationship between tech and government is heading, nothing will.
There are real questions about where this ends. The Trump administration has positioned itself as pro-AI but also, apparently, as the body that reviews AI before the rest of humanity gets it. That is a strange and novel kind of power, and Thursday’s launch is the second time in a month the government has made that power visible.
Read the full Axios scoop here.
CHINA IS OFFICIALLY DONE WITH NVIDIA. ALSO CHINA JUST SECRETLY ORDERED MORE NVIDIA CHIPS.
If you want a clean summary of how the US-China AI chip war is actually playing out on the ground, here it is. Chinese companies announced publicly this week that they are shifting 46 percent of their artificial intelligence hardware budget toward domestic chip suppliers over the next year, up from 30 percent today. This is the nationalist storyline the Chinese government wants the world to see. Huawei makes chips. We buy Huawei chips. We are independent from American technology. Very good.
Meanwhile, also this week, the Chinese government quietly told Alibaba, ByteDance, and DeepSeek that they will be allowed to purchase a limited number of Nvidia H200 chips.
So: Chinese companies are officially going domestic. Also, please send Nvidia chips. These two things are happening simultaneously, and nobody seems embarrassed about it. It is the most honest thing that has happened in geopolitics in a while.
Here is the thing about domestic Chinese AI chips. They exist, and some of them are genuinely impressive for the price. Huawei’s Ascend 910 series has been getting better year over year and Chinese companies have had years of practice optimizing their software to run on it because US export controls forced them to. Cambricon, Biren, and Moore Threads are all in the game. The domestic chip ecosystem is real and it is growing.
But it is not Nvidia. Not yet. When Chinese companies need to run their most demanding frontier model training runs, the H100 and H200 are still the hardware that gets results at scale with confidence. The domestic alternatives can handle inference and smaller training workloads well enough, but for cutting-edge model development everyone knows what they actually want when budget is not a constraint.
This is why you get this strange dual reality where Chinese executives are in meetings nodding along to domestic silicon commitments while quietly lobbying their own government for permission to import American chips. It is not exactly hypocrisy. It is more like being on a diet and maintaining a snack drawer, except the snack drawer requires US Department of Commerce approval to access.
The export control regime has simultaneously pushed Chinese companies toward independence AND made Nvidia chips feel more valuable than ever. When something is restricted, it becomes desirable. Every AI startup in Beijing knows that H200s are controlled, and therefore H200s must be the best thing on the market, and therefore you want H200s. Scarcity does something to the human brain even when humans work for large state-affiliated corporations.
The H100 and H200 era is technically coming to a close in the US anyway as Nvidia rolls out Blackwell. But for China, the chips they cannot fully have remain the benchmark for what serious AI training looks like. And the government is now threading the needle by letting a few companies buy some, presumably to ensure they stay competitive on frontier model development even while the official story is domestic silicon all the way.
Fortune on the Chinese chip shift here. And Bloomberg on the H200 permissions here.
ANTHROPIC JUST OUT-EARNED OPENAI, THE COMPANY THAT INVENTED THE WHOLE THING
Let us state this plainly because it is genuinely something. Anthropic, a company that most people had never heard of three years ago, is now reporting higher annualized revenue than OpenAI. Anthropic is tracking toward $47 billion in annualized revenue and says it will be profitable in 2026. OpenAI is tracking toward somewhere between $25 and $33 billion. Anthropic is ahead.
OpenAI invented modern consumer AI as most people experience it. ChatGPT launched in November 2022 and within two months had 100 million users, the fastest any product in history had reached that number. OpenAI got the brand recognition, the pop culture moment, the massive Microsoft partnership, the cultural weight of being the AI company everyone talks about at dinner. And then Anthropic, founded by a group of researchers who left OpenAI in 2021, quietly built Claude, cracked the enterprise market, and is now outearning them.
How did this happen. A few ways. First, Claude Code. Anthropic’s coding tool became a genuine phenomenon in the developer world and hit $1 billion in annualized revenue by the end of 2025 before doubling to $2.5 billion by February 2026. Developers who use it daily have become nearly cultishly devoted to it. The same kind of loyalty dynamic happened with GitHub Copilot but Claude Code got there faster and stickier.
Second, Anthropic played the enterprise game better and more quietly. Business subscriptions, API access for developers, deals with big companies that need reliable AI they can actually put into production. While OpenAI was building the flashiest consumer products, Anthropic was becoming the AI of choice for companies that need to put models into actual business workflows without the model going rogue on them.
Third, the Claude models have built a strong reputation for reliability and following instructions precisely. That reputation matters when you are deciding whether to give an AI access to your customer database or your legal review process. Enterprises do not need the most creative model. They need the one that does what it is told without surprises.
Data from Ramp, which tracks how businesses spend on software, showed Anthropic overtook OpenAI in business subscriptions in May. And Similarweb data showed monthly visits to ChatGPT fell below a majority of the generative AI market for the first time that same month. People are actively switching. ChatGPT has lost its default status.
ChatGPT still has massive name recognition and consumer mindshare. But in the part of the market where the real money is, which is enterprise and developer spending, Anthropic is now ahead. This is not a small development. This is the scoreboard changing.
SAM ALTMAN OFFERS THE US GOVERNMENT A $42 BILLION SLICE OF OPENAI, AND THAT SOUNDS VERY GENEROUS UNTIL YOU ACTUALLY THINK ABOUT IT
Here is the pitch. Sam Altman, CEO of OpenAI, has proposed giving the United States government a 5 percent equity stake in OpenAI. At the company’s current private valuation, that 5 percent is worth approximately $42.6 billion. The proposal, as Altman has framed it publicly, is a way to give ordinary Americans a share of the upside from AI. The same way sovereign wealth funds in oil-producing states give citizens a piece of national resource revenue. The government gets a stake, the people benefit, everyone participates in the windfall. Great story.
Now let us think about this for a moment.
The deal, if it happens, would flow into the US sovereign wealth fund, which the Trump administration has been building partly through equity stakes in various things. OpenAI is also simultaneously converting from its unusual nonprofit-with-a-for-profit-arm structure to a standard for-profit company, which requires charitable sign-off in multiple states and has been politically contentious. And OpenAI is also planning an IPO. And the proposal as described would have the government take 5 percent not just of OpenAI but of each leading US AI developer, which would include Anthropic and potentially others.
Bloomberg’s opinion desk called the whole thing a Trojan horse. The argument is straightforward: if the government becomes a financial stakeholder in OpenAI at a $42 billion level, it creates a situation where aggressively regulating OpenAI would directly cost the government money. Investors with a $42 billion position tend to want the asset to succeed and grow. Regulators who own equity do not typically regulate the way regulators who do not own equity do.
The preliminary talks reportedly require congressional approval to finalize, so nothing is imminent. But the direction is clear. OpenAI is trying to thread a very specific needle: stay on the right side of the Trump administration, keep the government from becoming a hostile regulator, and make AI feel like a national project that Americans have a financial stake in. Giving the public equity in the upside is genuinely good politics regardless of the business logic.
Whether it is a stroke of genius or the beginning of a complicated conflict of interest is probably a question the next decade will answer. But watching Sam Altman and the Trump White House negotiate the governance of the most powerful technology in human history through the lens of equity stakes and sovereign wealth funds is something nobody in 2020 would have predicted. Here we are.