Five US AI labs held every advantage and spent it on distraction, fear pricing and internal purges, while China closed most of the gap under chip sanctions.
Good article. Of the 5 only Google has a chance of coming out alive. While the open weight models will win out, so far the UX around them is pretty dire. The UX is the part you can build a house of brick around.
All the others are at least one hop away from the user is. Google and Apple are right there (with Microsoft hanging on to some degree). As the old adage goes, it's not what you've got, it's where you've got it. Apple seems to be skating to exactly where the puck will be in a few years. Google has all the pieces to copy. Everyone else is nowhere near.
So barring a totally new computing platform coming along, nothing much changes from today. It will be interesting to see what happens to Claude and ChatGPT once Siri 27 is released. Does the $20 a month market collapse at that point if new Siri ends up being good enough?
That said there is a gap for someone to reinvent the PC and the smartphone around AI. But outside of Apple and Google it's hard to see who that would be and what it would look like. Nvidia is probably the only other player with most of the pieces to pull it off.
Economically, yeah, I'm with you. Distribution plus their own silicon, that's the best hand of the five.
Product side though, no. And I can't really explain it. Benchmarks look fine, but every time I actually use Gemini for work it loses to Claude and Codex. Loses context in a big repo, makes up APIs, falls apart after a few turns. Something's just off and they've had years to fix it.
For the agentic coding market 100% agree. Claude is the tool to beat. Google with the exception of their integration with the Android tools aren’t as good. The Android integration is their moat.
For other segments, from what I’ve observed being close to the user is more important and also long as the results are good enough then there isn’t any need to take the next step to install a potentially better app. Google via Chrome and Android is right there.
But Google has a nasty habit of tripping over themselves so I guess we will see.
Re Musk's purchase of Cursor, you write: " It reads like a clean door into the market Grok’s own reputation had made harder to enter." Grok is now an option in Cursor. I know people who disabled Grok because Musk, and continue to use Cursor. Other now disfavour Cursor because Musk. So, don't be too sure.
I am unsure if Musk wants Grok to be profitable on its own. I could be wrong but I think Grok is kind there as a technology demonstrator and way to hold others accountable than to actually make money.
The reason I say this is that he has a history of doing this. He'll say one thing and actually deliver something that is about 45 degrees off but actually makes money.
Starlink and Falcon 9 being a great example. Starlink is able to put any ISP in the world out of business unless the ISP can offer gigabit speeds to the residential and SME market. Starlink makes money. Starlink also allowed SpaceX to develop its current launch cadence that means it has successfully launched more times than, AIUI, all other rocket makers combined. That has, in turn allowed SpaceX to be the cheapest launch for anything and , with starship, to get a simply enormous lead on anyone else.
Musk also understands manufacturing process. It is not impossible that Grok is the app that he needed to develop the expertise to build large datacenters. Colossus is probably the key - a data center built in well under a year from fencing the ground to power on. Everyone else trying to build 100 Megawatt plus data centers takes 2-3 years to do so. Grok was the excuse that drove the building of colossus and now Musk can sell AI compute in an extremely tight market and make bank.
Interesting take, and I hadn't looked at it from the compute side. One thing nagging me though: he doesn't need Grok for the datacenter play. You build and rent the capacity whether or not your own model runs on it. So Grok is a separate bet riding alongside Colossus, not the reason it got built, and that bet still has to answer for itself. If the compute is the business, a frontier model plus 60 billion for Cursor is a lot of motion for something the datacenter never required.
One thing to remember: Google is the only one with a global partner channel of resellers/distributors and has one of the top 3 marketplace with AWS and Microsoft. Distribution matters more than benchmarks. You win by having a seat at the table for every deal. Google can also loose 100s of billions on this and barely blink an eye.
How do you think Microsoft won this battle 30 years ago? By building a 1M members partner network.
Good article. Of the 5 only Google has a chance of coming out alive. While the open weight models will win out, so far the UX around them is pretty dire. The UX is the part you can build a house of brick around.
All the others are at least one hop away from the user is. Google and Apple are right there (with Microsoft hanging on to some degree). As the old adage goes, it's not what you've got, it's where you've got it. Apple seems to be skating to exactly where the puck will be in a few years. Google has all the pieces to copy. Everyone else is nowhere near.
So barring a totally new computing platform coming along, nothing much changes from today. It will be interesting to see what happens to Claude and ChatGPT once Siri 27 is released. Does the $20 a month market collapse at that point if new Siri ends up being good enough?
That said there is a gap for someone to reinvent the PC and the smartphone around AI. But outside of Apple and Google it's hard to see who that would be and what it would look like. Nvidia is probably the only other player with most of the pieces to pull it off.
Economically, yeah, I'm with you. Distribution plus their own silicon, that's the best hand of the five.
Product side though, no. And I can't really explain it. Benchmarks look fine, but every time I actually use Gemini for work it loses to Claude and Codex. Loses context in a big repo, makes up APIs, falls apart after a few turns. Something's just off and they've had years to fix it.
For the agentic coding market 100% agree. Claude is the tool to beat. Google with the exception of their integration with the Android tools aren’t as good. The Android integration is their moat.
For other segments, from what I’ve observed being close to the user is more important and also long as the results are good enough then there isn’t any need to take the next step to install a potentially better app. Google via Chrome and Android is right there.
But Google has a nasty habit of tripping over themselves so I guess we will see.
Re Musk's purchase of Cursor, you write: " It reads like a clean door into the market Grok’s own reputation had made harder to enter." Grok is now an option in Cursor. I know people who disabled Grok because Musk, and continue to use Cursor. Other now disfavour Cursor because Musk. So, don't be too sure.
Great article. Just one quibble
I am unsure if Musk wants Grok to be profitable on its own. I could be wrong but I think Grok is kind there as a technology demonstrator and way to hold others accountable than to actually make money.
The reason I say this is that he has a history of doing this. He'll say one thing and actually deliver something that is about 45 degrees off but actually makes money.
Starlink and Falcon 9 being a great example. Starlink is able to put any ISP in the world out of business unless the ISP can offer gigabit speeds to the residential and SME market. Starlink makes money. Starlink also allowed SpaceX to develop its current launch cadence that means it has successfully launched more times than, AIUI, all other rocket makers combined. That has, in turn allowed SpaceX to be the cheapest launch for anything and , with starship, to get a simply enormous lead on anyone else.
Musk also understands manufacturing process. It is not impossible that Grok is the app that he needed to develop the expertise to build large datacenters. Colossus is probably the key - a data center built in well under a year from fencing the ground to power on. Everyone else trying to build 100 Megawatt plus data centers takes 2-3 years to do so. Grok was the excuse that drove the building of colossus and now Musk can sell AI compute in an extremely tight market and make bank.
Interesting take, and I hadn't looked at it from the compute side. One thing nagging me though: he doesn't need Grok for the datacenter play. You build and rent the capacity whether or not your own model runs on it. So Grok is a separate bet riding alongside Colossus, not the reason it got built, and that bet still has to answer for itself. If the compute is the business, a frontier model plus 60 billion for Cursor is a lot of motion for something the datacenter never required.
I generally assume that Musk has multiple angles on anything. Grok serves to validate the Data center, but it has other uses too.
For example Grok serves to also keep OpenAI and Anthropic mostly honest and pressured. Especially if he releases the previous model as open source which he said he was going to do. No it isn’t as good as Kimi. But it sets a floor and grok is probably more trustable for people than Chinese models. Quite possibly another point is to make S©am altman fail, considering the bad blood between Musk and Altman. I mean he may be the only person who can afford to screw Altman over good and proper.
I really enjoyed your article, been reading you for months :-) God bless you, your nation, and your comrades in arms. Slava Ukraini!
Appreciate you sticking around all these months, and thank you for the kind words. They land. Glory to Ukraine!
Excellent article!
One thing to remember: Google is the only one with a global partner channel of resellers/distributors and has one of the top 3 marketplace with AWS and Microsoft. Distribution matters more than benchmarks. You win by having a seat at the table for every deal. Google can also loose 100s of billions on this and barely blink an eye.
How do you think Microsoft won this battle 30 years ago? By building a 1M members partner network.