The quiet news of the last few days was the leak/announcement of a $3 billion OpenAI acquisition of Windsurf. That’s not the largest private acquisition ever made — that honor goes to Google’s $30 billion acquisition of Wiz a few months prior — but man, it’s up there! $3B is the kind of exit startup founders dream about. Especially for a startup that’s been around for 2 years, with its current branding for about 5 months.
I assume most people don’t know what Windsurf is, which is fair because it has so few users that when you try to Google for that information you get data about the sport.
Maybe that’s unfair, supposedly the company has over a million users. But I’m always a bit skeptical of numbers like that. A person who uses Windsurf every day is obviously in a different category than one that installed the tool to play around for five minutes and quickly discarded it. This slipperiness has always been one of the benefits of working in the world of privately held companies. Anecdotally I know only one person who uses Windsurf, and I only kinda sorta know that person because he’s just a guy that I met at an SF house party.
If you aren’t familiar with Windsurf, you may know it by its previous name, Codeium. And if you aren’t familiar with Codeium, you may know its primary competition, a company called Cursor. And if you don’t know what Cursor is, a) you might know what GitHub Copilot is, and b) how did you find my blog?
All of these products are roughly in the same category of “AI tools for software engineers”. They all have basically the same form factor too — they integrate AI models directly into your coding workflow. And traditionally they operate on three levels of granularity:
-
Auto-complete. As you type the AI will suggest the rest of the line or function, which you can generally accept with a single button press.
-
Sidebar Q&A. The code window itself will have an integrated sidebar where you can ask models to modify a few files. You’ll get a diff, which you can then choose to apply or modify.
-
Agentic flows. The term “agent” is wildly underspecified, but I’m the AI coding space the term has generally come to mean “an AI model operates in a loop over an entire code base, often with a very vague or high level prompt has a starting point, and is equipped with tools to write, run, and otherwise analyze code and the computer system”.
Different companies aim to be best in class along different verticals. Some are better at the auto complete (Copilot), others at the agent flow (Claude Code). Some aim to be the best for non-technical people (Bolt or Replit), others for large enterprises (again, Copilot). Still, all of this “differentiation” ends up making a 1-2% difference in product. In fact, I can’t stress enough how much the UX and core functionality of these tools is essentially identical.
These all would have pejoratively been known as “GPT wrappers” just two years ago, because they do not actually compete on the model layer but rather allow users to choose and switch between any of the big LLM providers. To really emphasize how interchangeable these AI code assistants all are, I use an even lesser known tool called Avante, an entirely free and open source neovim plugin. It does the same things as all the other tools. I like it because I don’t have to leave vim.
But the similarities of these tools does not take away from how game changing they are. Once you get used to a form of AI powered coding, you cannot go back. The real issue with all of these products is that they are too easily verticalized. Anyone who wants to spin up a version of Windsurf with one slight change that targets a tiny market segment can do so fairly easily — again, I’m on Avante entirely because it supports vim. That means the addressable market for any of these companies may actually go down over time as more competitors and free alternatives enter the market, even as the number of “programmers” goes up. In point of fact, even though I love the ingenuity behind Cursor (which really spearheaded the current AI coding paradigm) I have openly said that their long term opportunities are slim. Even though Cursor had significant first mover advantage, they have no moat or stickiness. As with the rest of the AI market, switching cost remains extremely low, and there is simply no reason to use Cursor when you can use a free version or one with better enterprise support. Cursor isn’t even living on its own platform — it’s a fork of VSCode. I am personally convinced that their only long term exit opportunity is an acquisition by Microsoft, and even that seems less and less likely as Satya puts more resources into the already-VSCode-native Copilot as a real competitor.
All of which makes the $3B price tag for Windsurf seem eye wateringly high. Compared to Cursor, Windsurf has fewer users, has been around for less time, has less brand recognition, and has diminishing prospects for future growth. It’s not as tied to VSCode, which is a plus, I guess. But it all begs the question: why on earth is OpenAI paying so much?
This is especially strange in the context of OpenAI’s financial situation. Smart observers have caught on to Google’s inherent advantages in the space, something that I first publicly called out as early as December and more fully two weeks ago. OpenAI needs to shore up both its access to compute and its access to data in order to compete. But it’s once-sterling relationship with its previous patron, Microsoft, has frayed significantly. This has essentially forced the company to go to SoftBank (yes, that SoftBank) for additional capital.
It’s true that OpenAI managed to get $40bn committed, and it’s also true that this is the largest amount of capital ever raised by a privately held company. But they’re going against Google, one of the most valuable companies in the entire world, and extremely profitable to boot. That’s a hell of a war chest to compete with.
In that light, the decision to spend 3 out of 40 of those billions is even harder to rationalize. Even worse, it’s not yet clear OpenAI actually has $40bn to spend — so far they’ve only got $10bn actually lined up, with the rest being held by SoftBank contingent on OpenAI actually becoming a for-profit company by the end of the year.
It seems pretty obvious that Windsurf will not help OpenAI get more compute. Maybe Windsurf is providing OpenAI access to data? There’s certainly some possibility that this is the case — though it makes me wonder just how bad OpenAI’s relationship with Microsoft has gotten if they no longer have access to GitHub, which surely dwarfs any amount of code that Windsurf could provide.
The other possibility is that this is entirely a long term distribution play, akin to Facebook buying WhatsApp or Instagram. People criticized those deals for being overpriced too. OpenAI may think that Windsurf will be a crown jewel in how people access GPT models. There’s some sense in this — OpenAI has also announced a social media project, likely also an attempt at maintaining lines to unique data sources while providing more native ways to improve distribution and “commoditize their complement“.
But the issue that they will inevitably run into with Windsurf is that GPT just isn’t the best in class for programming. Everyone who’s using Windsurf is almost definitely using Claude or Gemini. Even though the “GPT wrapper” term was always meant as an insult, it is in practice a huge table stakes feature to be able to wrap around many different LLM providers. That flexibility is what allows a company like Windsurf to ride the machine learning wave, buoyed along by everyone else’s investments. Cursor really only took off when Claude suddenly got really good at programming, after all. If Windsurf ends up being tied exclusively to GPT, many of its users may leave the platform simply because it is now a worse platform. But if there isn’t any vendor lock in, we’re back to square one — what is the point?
Personally, I don’t get it. Maybe someone smarter than I am (or more connected than I am) can help me figure it out. But for now, I’m chalking this particular check size as a symptom of the AI market being way too hot right now.
The other quiet news of the last few days is the dawning realization of just how quiet it has been. The last two weeks saw the release of 3 new OpenAI models — o3, o4-mini, and GPT 4.1 — as well as the new Llama 4 model family from Meta and Grok-3 from Grok. And…nothing. It’s just crickets. In past months, a release calendar like this would have had headlines blazing. The hype train should be chugging at ridiculous speeds. But compared to what I’d expect, there’s nothing.
The reason is obvious: Google is still in the lead. Take a look at these two charts.
The former is the current state of the LMSYS chatbot arena; the latter maps chatbot arena performance against price. The over under? There’s no headlines because there’s nothing to write about. “OpenAI takes second place” no one cares!
It’s still too early to write about ‘general consensus’ since the OpenAI models were released only a few days ago. And to their credit, those models do top many of the LLM benchmarks. But so far, the reception has been extremely muted, with many saying something like:
Even though some of the new OpenAI models are quite powerful, they are simply too expensive and too slow for not enough extra juice.
I’ve already written extensively about Google’s Gemini 2.5 release, which quickly became the go-to model for just about everything. What I didn’t originally clock was just how much Google had shored up its model offerings all over the price/performance curve. Put bluntly: at every price point, the best model is a Google model.
That’s not all. I mentioned rumors that Google has disallowed new publications; that is now confirmed as of earlier this month:
Among the changes in the company’s publication policies is a six-month embargo before “strategic” papers related to generative AI are released. Researchers also often need to convince several staff members of the merits of publication, said two people with knowledge of the matter.
Google has also apparently started offering extremely generous non-compete deals to researchers, preferring to keep them on payroll doing nothing than have them go to competitors and leak secrets:
Google is making use of aggressive noncompete clauses and extended notice periods, contends former GoogDeepMinder Nando de Freitas in a recent post on X. In some cases, Google DeepMind’s employment contracts may lock an AI developer into doing nothing for as long as a year, notes Business Insider, to prevent its AI talent from moving to competing firms. That’s a long time away from working on the cutting edge in the rapidly developing world of AI.
And finally, Google has continued to release and improve its TPU offerings on GCP, giving them yet another method to profit off the back of the AI boom — even if they don’t win on the model, they can win by providing the underlying hardware.
With almost no fanfare, we all just woke up one day to a Google-dominated AI landscape. I have been critical of Sundar in the past, but I have to hand it to him — sometimes the showmanship really is just a distraction from executing a slow but precise strategy. Google is clearly now on a war footing. They are relentlessly poaching employees while trying to close up their shop as much as possible. The AI industry as a whole owes more to Google than any other organization. It’s unclear how many other players in that industry will survive when cut off from Google’s research. It’s also unclear how much this will last in the face of a continuing DOJ antitrust suit. More on that in a different article, though.
One last thought. I’ve always been a staunch defender of capitalism and free markets, even though that’s historically been an unpopular opinion in my particular social circle. Watching the LLM market, I can’t help but feel extremely vindicated. Over the last 5 years, the cost per token has been driven down relentlessly even as model quality has skyrocketed. The brutal and bruising competition between the tech giants has left nothing but riches for the average consumer. There’s an alternative world where all of this is priced so high that only the wealthiest businesses can justify a “GPT license”, or where the government ends up keeping all the best AI technology for themselves. That world would objectively suck — not only would most people not be able to access the technology, there would also be significantly less interest in or ability to innovate. Just look at Google, which has finally risen like a beast from slumber to show the world what it means to innovate once more.
Since we’ve been talking about things that didn’t happen, I want to talk about one last notable absence: where the hell is Apple?
Something I’ve said repeatedly is that the LLM market has strong winner-take-all effects, and players in the market are heavily dependent on access to scientists, compute, and data. Apple is an extraordinarily wealthy company, so they have no problem getting access to scientists. But it seems like they have had a ton of issues on both of the latter two categories.
On the compute side, it seems like Apple sorta own goaled themselves? From the NYT:
The A.I. stumble was set in motion in early 2023. Mr. Giannandrea, who was overseeing the effort, sought approval from the company’s chief executive, Tim Cook, to buy more A.I. chips, known as graphics processing units, or GPUs, five people with knowledge of the request said. The chips, which can perform hundreds of computations at the same time, are critical to building the neural networks of A.I. systems, like chatbots, that can answer questions or write software code.
At the time, Apple’s data centers had about 50,000 GPUs that were more than five years old — far fewer than the hundreds of thousands of chips being bought at the time by A.I. leaders like Microsoft, Amazon, Google and Meta, these people said.
Mr. Cook approved a plan to double the team’s chip budget, but Apple’s finance chief, Luca Maestri, reduced the increase to less than half that, the people said. Mr. Maestri encouraged the team to make the chips they had more efficient.
The lack of GPUs meant the team developing A.I. systems had to negotiate for data center computing power from its providers like Google and Amazon, two of the people said. The leading chips made by Nvidia were in such demand that Apple used alternative chips made by Google for some of its A.I. development.
Well, that at least explains why they haven’t been putting out any decent models. Anecdotally, Apple obviously has data centers, but they aren’t a cloud provider like Google/Microsoft/Amazon, which at various points have powered DeepMind/OpenAI/Anthropic directly. So Apple is starting way behind on the whole chip thing. Maybe it makes some kind of strategic sense to try and double down on their own unique chip capacity — maybe try to do what Google has done with TPUs — but that’s really being extremely generous. The more obvious answer is the simple one: Apple cheaped out, and was penny wise pound foolish. As a result, the company that dominated the mobile wave is all but absent from the AI wave.
On the data side, Apple definitely own goaled themselves. In an environment of data hoarders and open disregard for information safety, Apple struck out as an ardent defender of user privacy. They made a brand out of it! They ran ads on it!
That all made sense a few years ago, when the data itself was more questionably useful and Apple had full control of the hardware stack. Apple was able to poke fun at Google while giving Meta a pretty serious black eye over their ads policy.
But now, data is almost literally fuel for deep learning models. Worse, there’s basically no way to avoid leaking data through the model! People have consistently been able to get models to directly reproduce training data! Google has more or less avoided using any training data because they can, they have the whole Internet already indexed. Meta, xAI, OpenAI, and Anthropic all train on public data — the former two from public posts on their social media platforms, and all four from extremely questionable flouting of copyright law.
Meanwhile, Apple is stuck with the same problem Google had when they got rid of their “Don’t be evil” motto.
“Don’t be evil” was a lot of things, and there were a lot of disagreeing interpretations about what it meant. One thing that no one disagreed about: it was hard to get rid of. Execs at Google ended up regretting the “Don’t be evil” motto, because no matter what Google did they would get raked over the coals for doing it. “I thought you said you wouldn’t be evil“, internet commenters would snidely say. They even got sued over it!
Apple is in a similar boat. Either they use the user data they have and risk serious brand damage that the rest of FAANG is sure to capitalize on, or they handicap themselves in the AI race. Which, really, is less of a race and more of an all out brawl, one in which Apple is fighting with both hands behind its back.
So far, they’ve taken the “handicap” approach. They’ve tried to pay their way out of the data access problem by straight up buying the copyright licenses for data they want to train on, but, come on, it’s just not anywhere near enough training data.
The worst case scenario for Apple is they decide to use user data late. In that setting, Apple incurs the brand risk while also being miles behind everyone else. That increasingly seems like what will happen, though, because I just can’t imagine Apple actually sitting the entire AI race out.
So yeah. All in all, a pretty quiet few weeks for AI.
