[Editor's Note: this piece previously said "Blackstone" instead of "Blackrock," which has now been fixed.]

I've been struggling to think about what to write this week, if only because I've written so much recently and because, if I'm honest, things aren't really making a lot of sense.

NVIDIA claims to have shipped six million Blackwell GPUs in the last four quarters — as I went into in my last premium piece — working out to somewhere between 10GW and 12GW of power (based on the power draw of B100 and B200 GPUs and GB200 and GB300 racks), which...does not make sense based on the amount of actual data center capacity brought online.

Similarly, Anthropic claims to be approaching $10 billion in annualized revenue — so around $833 million in a month — which would make it competitive with OpenAI's projected $13 billion in revenue, though I should add that based on my reporting extrapolating OpenAI's revenues from Microsoft's revenue share, I estimate the company will miss that projection by several billion dollars, especially now that Google's Gemini 3 launch has put OpenAI on a "Code Red," shortly after an internal memo revealed that Gemini 3 could “create some temporary economic headwinds for [OpenAI]."

Which leads me to another question: why?

Gemini 3 is "better," in the same way that every single new AI model is some indeterminate level of "better." Nano Banana Pro is, to Simon Willison, "the best available image generation model." But I can't find a clear, definitive answer as to why A) this is "so much better," B) why everybody is freaking out about Gemini 3, and C) why this would have created "headwinds" for OpenAI, headwinds so severe that it has had to rush out a model called Garlic "as soon as possible" according to The Information:

Last week, OpenAI’s chief research officer Mark Chen told some colleagues about the new model, which was performing well on the company’s evaluations, at least when compared to Gemini 3 and Anthropic’s Opus 4.5 in tasks involving coding and reasoning, according to a person with knowledge of the remarks.

But Garlic may be a bigger deal. Chen said OpenAI is looking to release a version of Garlic as soon as possible, which we think means people shouldn’t be surprised to see GPT-5.2 or GPT-5.5 release by early next year.

Garlic is a different model from Shallotpeat, a new large language model under development which Altman told staff in October would help OpenAI challenge Gemini 3. Garlic incorporates bug fixes that the company used in developing Shallotpeat during the pretraining process, the first stage of model training in which an LLM is shown data from the web and other sources so it can learn connections between them.

Right, sure, cool, another model. Again, why is Gemini 3 so much better and making OpenAI worried about "economic headwinds"? Could this simply be a convenient excuse to cover over, as Alex Heath reported a few weeks ago, ChatGPT's slowing download and usage growth?

Experts I've talked to arrived at two conclusions:

Gemini 3 is good/better at the stuff tested on benchmarks compared to what OpenAI has.
OpenAI's growth and usage was decelerating before this happened, and this just allows OpenAI to point to something.

I don't know about garlic or shallotpeat or whatever, but one has to wonder at some point what it is that OpenAI is doing all day:

Altman said Monday in an internal Slack memo that he was directing more employees to focus on improving features of ChatGPT, such as personalizing the chatbot for the more than 800 million people who use it weekly, including letting each of those people customize the way it interacts with them.

Altman also said other key priorities covered by the code red included Imagegen, the image-generating AI that allows ChatGPT users to create anything from interior-design mockups to turning real-life photos into animated ones. Last month, Google released its own image generation model, Nano Banana Pro, to strong reviews.

Altman said other priorities consisted of improving “model behavior” so that people prefer the AI models that powers ChatGPT more than models from competitors, including in public rankings such as LMArena; boosting ChatGPT’s speed and reliability; and minimizing overrefusals, a term that refers to when the chatbot refuses to answer a benign question.

So, OpenAI's big plan is to improve ChatGPT, make the image generation better, make people like the models better, improve rankings, make it faster, and make it answer more stuff.

I think it's fair to ask: what the fuck has OpenAI been doing this whole time if it isn't "make the model better" and "make people like ChatGPT more"? I guess the company shoved Sora 2 out the door — which is already off the top 30 free Android apps in the US and at 17 on the US free iPhone apps rankings as of writing this sentence after everybody freaked out about it hitting number one. All that attention, and for what?

Indeed, signs seem to be pointing towards reduced demand for these services. As The Information reported a few days ago...

Multiple Microsoft divisions, for instance, have lowered how much salespeople are supposed to grow their sales of certain AI products after many of them missed sales-growth goals in the fiscal year that ended in June, according to two salespeople in Microsoft’s Azure cloud unit.

Microsoft, of course, disputed this, and said...

A Microsoft spokesperson said “aggregate sales quotas for AI products have not been lowered” but declined to comment specifically on the lowered growth targets. The spokesperson pointed to growth in the company’s overall cloud business, which has been lifted by rentals of AI servers by OpenAI and other AI developers.

Well, I don't think Microsoft has any problems selling compute to OpenAI — which paid it $8.67 billion just for inference between January and September — as I doubt there is any "sales team" having to sell compute to OpenAI.

But I also want to be clear that Microsoft added a word: "aggregate." The Information never used that word, and indeed nobody seems to have bothered to ask what "aggregate" means. I do, however, know that Microsoft has had trouble selling stuff. As I reported a few months ago, in August 2025 Redmond only had 8 million active paying licenses for Microsoft 365 Copilot out of the more-than-440 million people paying for Microsoft 365.

In fact, here's a rundown of how well AI is going for Microsoft:

Its chips effort is falling behind, with its "Maya" AI chip delayed to 2026, and according to The Information, "when it finally goes into mass production next year, it’s expected to fall well short of the performance of Nvidia’s flagship Blackwell chip."
According to The Information in late October 2025, "more customers have been using Microsoft’s suite of AI copilots, but many of them aren’t paying for it."
In October, Australian's Competition and Consumer Commission sued Microsoft for "allegedly misleading 2.7 million Australians over Microsoft 365 subscriptions," by making it seem like they had to pay extra and integrate Copilot into their subscription rather than buy the, and I quote, "undisclosed third option, the Microsoft 365 Personal or Family Classic plans, which allowed subscribers to retain the features of their existing plan, without Copilot, at the previous lower price."
- This is what a company does when it can't sell shit. Google did the same thing with its workspace accounts earlier in the year. This should be illegal!
According to The Information in September 2025, Microsoft had to "partly" replace OpenAI's models with Anthropic's for some of its Copilot software. Microsoft has, at this point, sunk over ten billion dollars into OpenAI, and part of its return for doing so was exclusively being able to use its models. Cool!
According to The Information in September 2025, Microsoft has had to push discounts for Office 365 Copilot as customers had "found Copilot adoption slow due to high cost and unproven ROI."
In late 2024, customers had paused purchasing further Copilot assistants due to performance and cost issues.

Yet things are getting weird. Remember that OpenAI-NVIDIA deal? The supposedly "sealed" one where NVIDIA would invest $100 billion in OpenAI, with each tranche of $10 billion gated behind a gigawatt of compute? The one that never really seemed to have any fundament to it, but people reported as closed anyway? Well, per NVIDIA's most-recent 10-Q (emphasis mine):

Investment commitments are $6.5 billion as of October 26, 2025, including $5 billion in Intel Corporation which is subject to regulatory approval. In the third quarter of fiscal year 2026, we entered into a letter of intent with an opportunity to invest in OpenAI.

A letter of intent "with an opportunity" means jack diddly squat. My evidence? NVIDIA's follow-up mention of its investment in Anthropic:

In November 2025, we entered into an agreement, subject to certain closing conditions, to invest up to $10 billion in Anthropic.

This deal, as ever, was reported as effectively done, with NVIDIA investing $10 billion and Microsoft $5 billion, saying the word "will" as if the money had been wired, despite the "closing conditions" and the words "up to" suggesting NVIDIA hasn't really agreed how much it will really invest. A few weeks later, the Financial Times would report that Anthropic is trying to go public as early as 2026 and that Microsoft and NVIDIA's money would "form part of a funding round expected to value the group between $300bn and $350bn."

For some reason, Anthropic is hailed as some sort of "efficient" competitor to OpenAI, at least based on what both The Information and Wall Street Journal have said, yet it appears to be raising and burning just as much as OpenAI. Why did a company that's allegedly “reducing costs” have to raise $13 billion in September 2025 after raising $3.5 billion in March 2025, and after raising $4 billion in November 2024? Am I really meant to read stories about Anthropic hitting break even in 2028 with a straight face? Especially as other stories say Anthropic will be cash flow positive “as soon as 2027.”

And if this company is so efficient and so good with money, why does it need another $15 billion, likely only a few months after it raised $13 billion? Though I doubt the $15 billion round closes this year, if it does, it would mean that Anthropic would have raised $31.5 billion in 2025 — which is, assuming the remaining $22.5 billion comes from SoftBank, not far from the $40.8 billion OpenAI would have raised this year.

In the event that SoftBank doesn't fund that money in 2025, Anthropic will have raised a little under $2 billion less ($16.5 billion) than OpenAI ($18.3 billion, consisting of $10 billion in June split between $7.5 billion from SoftBank and $2.5 billion from other investors, and an $8.3 billion round in August) this year.

I think it's likely that Anthropic is just as disastrous a business as OpenAI, and I'm genuinely surprised that nobody has done the simple maths here, though at this point I think we're in the era of "not thinking too hard because when you do so everything feels crazy.”

Which is why I'm about to think harder than ever!

I feel like I'm asked multiple times a day both how and when the bubble will burst, and the truth is that it could be weeks or months or another year, because so little of this is based on actual, real stuff. While our markets are supported by NVIDIA's eternal growth engine, said growth engine isn't supported by revenues or real growth or really much of anything beyond vibes. As a result, it's hard to say exactly what the catalyst might be, or indeed what the bubble bursting might look like.

Today, I'm going to sit down and give you the scenarios — the systemic shocks — that would potentially start the unravelling of this era, as well as explain what a bubble bursting might actually look like, both for private and public companies.

This is the spiritual successor to August's AI Bubble 2027, except I'm going to have a little more fun and write out a few scenarios that range from likely to possible, and try and give you an enjoyable romp through the potential apocalypses waiting for us in 2026.

专题：openai_strategy

AI行业泡沫破裂迹象：竞争加剧与需求下滑