mmx1

I am not suggesting NT is dead on servers at all. I am positing it would be dead had it not been for owning the majority of desktops. Those use cases are primarily driven as an ancillary service to Windows desktop[1], and where they have wider applicability, like .NET and SQL Server, have been progressively unleashed from Windows. The realm of standalone server products were bulldozed by Linux; NT wouldn't have stood a chance either.

[1]: In fact, Active Directory was specifically targeted by EU antitrust lawsuit against Microsoft.

Obviously we'll never know, but I seriously doubt that parallel universe would've had a chance to materialize. Not the least due to "free as in beer" aspect of Linux whilst web/Apache was growing at the pace it did. All proprietary unices are basically dead. Sun was likely the sole company that had the best attitude to live alongside open source, but they also proved it wasn't a good enough business post bubble burst. NT and Darwin remain alive due to their desktop use, not server.

Realistically anyone who cared would be using something like Cygwin (and the original UNIX server market segment evaporated due to Linux and had zero interest in migrating to NT in that form--some did migrate due to application layer benefits like .NET but not for the same workloads.)

I don't think it is fair to brush it off under "same bucket; doesn't count." The syscalls are still different and there's quite a bit of nuance. I mean the lines you're drawing are out of superficial convenience and quite arbitrary. In fact, I'd argue macOS/Darwin/XNU are really Mach at their core (virtual memory subsystem, process management and IPC) and BSD syscalls are simply an emulated service on Mach, which is quite different from traditional UNIX. The fact that as a user you think of macOS much more similar to Linux is not really reflective of what happens under the hood. Likewise NT has very little to do with Win32 API in its fundamentals but Win2k feels the same to the user as WinME, but under your framing, you'd same-bucket those.

Indeed countless companies went under investing in tech, on various iterations of the exact things that later became enormously successful, e.g. iPhone. Success is only guaranteed in retrospect.

The point is precisely that being super fixated on Microsoft as a monopoly producing relatively trash-quality products is a static worldview. It is the case that Microsoft monopoly did not in any shape or form preclude a measly search startup and a beleaguered Cupertino company from surpassing it to the slightest (mind you, Windows is still the dominant desktop OS, which compared to what it meant in the 90s is immaterial; people rarely think about writing Windows apps anymore: they write web, iOS and Android apps.) This is the whole point of Thiel's thesis. The contemporary version of this world transition we are witnessing is from search to chatbot. Meanwhile Microsoft by pushing the "trash" products actually served the industry at large well in other ways: it provided a consistent standard base to widely deploy PCs and applications.

Aside: the "competition is for losers" tagline is not from Thiel himself: apparently the publisher came up with a punchy line, and I think that refers to his broader ideology. What I brought up is a very specific observation in his thesis that high margin monopolies are actually good from the perspective of society in a dynamic world where it does not merely translate rent-seeking long-term.

I’d say the point is more subtle than that: it is not mere efficiency that comes with such scale. Rather, such scale is necessary to enable certain technological investments and long-term risks and behavior. For example, if you were a non-monopoly company in a competitive environment you would not be able to steer Chrome and shape it the way it is. You’d be burdened by day to day business needs to keep the margin from eroding below zero at any given corner.

Under this theory it follows that if you actively kill any business with a monopoly market share like the EU, you’ll guarantee failure for the society in achieving certain technological breakthroughs. The only way to avoid it would be government sanctioned corporations (CCP style) or US style pre-Lina Khan relatively free market approach to scale.

This is a good take. Also hints at why Peter Thiel's thesis on tech monopolies in a dynamic world is correct; i.e. that they aren't necessarily bad. A lot of the value in big tech comes from scale and integration and is inherently not sensible to have a thousand mom and pop shops do it and keep the same value. To the extent that Google is doing anything nefarious as a monopoly with integration, any other company who is willing to value Chrome at any significant price is inevitably going to need a scaled product to tie this to and accomplish what Google did with Chrome. They cannot pull this off on their own even though the open-source Chromium exists, because Google is actually better than them at maintaining a browser ecosystem. I am certain they will do a worse job at owning Chrome, especially considering Google has shown a good faith effort in shipping a browser and keeping the engine open source.

Ironically, this whole saga is happening at the same time the "Google search business is under attack" is at its peak in the news media.

That's actually not true. GPLv2 by default means v2, not v3, unless you explicitly allow "or later."

Linux is actually the famed example of v2 but not v3.

Pretty sure the answer is yes. I have no direct knowledge of the matter for Gemini 2.5, but in general TPUs were widely used for training at Google. Even Apple used them to train their Apple Intelligence models. It’s not some esoteric thing to train on TPU; I would consider using GPU for that inside Google esoteric.

P.S. I found an on-the-record statement re Gemini 1.0 on TPU:

"We trained Gemini 1.0 at scale on our AI-optimized infrastructure using Google’s in-house designed Tensor Processing Units (TPUs) v4 and v5e. And we designed it to be our most reliable and scalable model to train, and our most efficient to serve."

What I've heard is that the extrapolation of compute needed so many additional CPU servers to keep running the existing workload types that it obviously justified dedicated hardware. Same for video encoding accelerators[1].

[1]: https://research.google/pubs/warehouse-scale-video-accelerat...

Google has done stuff primarily for negotiation purposes (e.g. POWER9 chips) but TPU ain't one. It's not a backup option or presumed "inferior solution" to NVIDIA. Their entire ecosystem is TPU-first.

I mean, it is obvious that you cannot sustain efficiency as you scale (Amdahl's law) but (1) $100M is not that crazy to be able to keep track of in your head, even for a single individual (I can imagine a successful real estate developer with a handful of ongoing projects and various other personal investments), and (2) in a high growth situation, it makes financial sense to sacrifice some economic gain for scale. In your original example, sure an investor would be better off, if they could actually find 10 good investments with zero cost, to spread their money, but very likely they'd be better off taking the big one and spend their energy raising more money.

Yes, IIRC (please correct me if I'm wrong), translate did utilize Seastar (TPU v1) which was integer only, so not easily useful for training.

You can’t be serious. Lots of businesses easily have that much just in cost of goods or marketing spend. $100M is not such a crazy amount especially considering the cost of hiring technical people.

Also note that the benchmark of “efficiency” should be a function of growth, not some absolute standard.