1. As far as I’m aware, we don’t know the specs and efficiency of Google’s flash model serving search. It’s unclear then how close Google are or are not to the flash model frontier. They might already be the leader here. Or they might be subsidizing something inefficient.
2. What does OpenAI need to overcome?
In the last generation of competitive struggle, the following seemed to matter a lot: scale (for serving costs and network effects), hardware-software integration (use one to enter the other), cross-product data complementarities (search clicks or maps / phone data to monetize in ads), and product bundling (to envelope competitors).
Owning the user’s journey was key.
Can the current advantages enjoyed by frontier AI labs land up negating Google’s existing strengths in these areas above? I can imagine it weakening Google’s position vis-a-vie its end users, while reinforcing them for its producers. Why? Well Agentic AI consumes a ton of search, and a ton of maps, *but as an intermediary*. This is closer to producer (API) demand, perhaps. The consumer journey is being disintermediated. This must be a HUGE risk for Google. But for now manifests as higher API demand. And lower ads monetization. Just speculation.
3. Google’s ad network must be an enormous advantage. It is worth noting that OpenAI has had to start an ad network from scratch to avoid living inside Google’s ad network. How’s that working out for it?
The "winner takes all" mentality is almost being forced on the big AI companies given their CAPEX. It is not a given that either will "win" or be profitable. Open-source models are narrowing the performance gap, and their competitive advantage is foreseen by the US wanting to ban their use and importation. (Good luck with that.) There is already plenty of pushback on the siting of data centers for various reasons, and Google can read that public sentiment. Better to not be associated with those compute centers and sully what business they have.
It is also apparent that hyperscaling may not be necessary to achieve high performance, and certainly not appropriating the cultural artifacts for free to populate them for training. Just as electrification initially replaced the huge steam engines in factories with huge electric motors, it became obvious that the better way was to distribute electrical power to the point of use, and use small electric motors. Apple clearly wants to push appropriately-sized AI compute to edge devices, even as small as smartphones, which follows the electric power and motors history. If that is the future for AI, then Google can be the distributor of AI through apps and software running locally.
A further comment on distribution. I notice that Ollama is already set up to charge for model use that requires the cloud to support the model. This plays very much into Google's strength, as they already have Google Pay to muscle in on PayPal and Venmo. Google already has the infrastructure to handle advertising from vendors and buyers, so it should be relatively easy to offer the user options to select a model and pay for usage. (Far more useful than AI overviews in search.) Models will compete on price and performance, and will no doubt be rated by users. Like their success in dominating the online advertising market, Google takes a cut of the usage cost without incurring any risk of changing model preference. Users can inexpensively try different cloud-based models on the same platform, like Ollama, or offer free smaller models as they do with some versions of Gemini, or perhaps run locally. This avoids the user having to choose a platform that the "frontier AI Labs" want to try to become the winner. Selecting the model that works best for the user's application makes more sense, and will create an ecosystem of small domain-specific models and the information around those models, just like the book market.
The relationship between model distribution and payment systems is interesting. I have yet to dedicate time to reviewing the Stripe acquisition of OpenRouter beyond thinking at the time that it was an unexpected outcome for OpenRouter and that the Collison bros are very smart. Your comment helps me frame a better context for what will likely prove to be an important deal for the "intelligence payment rails" or whatever we come to call this ecosystem dynamic. Nice one.
“That sounds like someone who has argued that the bigger prize is being the layer that others’ AI runs on”. This is verbatim AWS’s strategy. Jassy uses the “picks and shovels” analogy for the AI gold rush.
Anthropic and openai ARE in diffusion mode. Their advances are primarily coming from superior harnesses, tooling and runtime modes, while raw intelligence plateaus and the economics of training become less and less viable with each generation. But their market is limited to coding, and they are all competing aggressively to own it. Whereas google's market is every phone and laptop user on earth, and they already own it and only need to keep it.
It's like they're saying if intelligence will be commoditized by capable mid-tier models, then what comes next? Kurian builds the physical infra, Dean&co build the app infra, and Hassabis makes some next-gen bet he's not talking about.
You’re 100% right that history doesn’t travel in a straight line. This includes corporate history. Google’s moves show it wanting to be AT&T, but that’s not a long term plan.
Instead, I suspect Google wants to let the AI boys go into the market and build their own stuff, then they can buy whichever works best, rather than risking leadership by betting on one thing or the other. They’re going to be big enough to do that when the software boys are done.
They probably won't need to buy anything, all they need to do is fast-follow. Since non-competes are unlawful in California, all they have to do is poach a few key researchers from Frontier competitors, who will instantly spill the beans on the latest SOTA techniques.
Google’s advantage is that it doesn’t need to capture every dollar at the model layer. If AI demand expands compute, networking, storage, power, and cloud consumption by orders of magnitude, owning the infrastructure stack becomes an extremely valuable position. The Westinghouse analogy works precisely because widespread adoption can matter more than owning every application built on top.
@Tim O'Reilly Love the Westinghouse framing, it represents the right kind of history to pursue here. I'd humbly suggest it is only part of the story, one which becomes clearer if we widen the lens a little. Google isn't the only place this exact pattern is showing up. Ilya left OpenAI for SSI. LeCun left Meta for AMI Labs. Fei-Fei Li built World Labs. David Silver raised $1.1B for Ineffable within months of leaving DeepMind's RL team. In just eighteen months, three AI leaders left different major frontier labs. They are all making the same bet: the old way of doing things has hit its limit. They believe that what comes next requires a fresh start somewhere else.
The diffusion story and the frontier-loss story might both be true. They do not represent a single strategic choice. Instead, they capture two different challenges hitting Google and the larger AI industry at the same time. Moreover, there's a third piece worth adding: look at what's happening to the compute the incumbents keep, not just the talent that leaves. SpaceX announced $2.32 billion in AI compute revenue for the month. This comes from three deals with Anthropic, Google, and xAI. Interestingly, the infrastructure used was mostly sitting underutilized before these agreements. That's $27.8B annualized, more than SpaceX's entire 2025 revenue, at margins that could beat CoreWeave's 70%. Google's TPU/GCP story is the same shape. DeepSeek, an efficiency-focused company that proved massive compute advantages aren't bulletproof, is now investing $35 billion into a gigawatt data center. ByteDance is making similar moves, building its own chips and expanding its datacenter infrastructure. These giants are no longer just renting space; they are building the hardware to control their own future.
The point is clear at this stage of AI’s evolution: owning compute is where the profit lies, regardless of your company's location or mission. Everyone is building around the frontier struggle because compute is what drives the bottom line, not the theory of discovery. Public or near public capital demands this bottom line profit and monetization focus. Its driving the AI party bus, one no one can afford to stop. Private capital is more patient and risk bearing, so it is absorbing the new frontier risk. Your analysis and @Dylan Patel /SemiAnalysis complement each other. They highlight different aspects of the same larger story.
As a user, it already seems to me that Gemini displaces Google searches (and potentially, search revenues). Gemini seems pretty mature. How does Google benefit from advancing it much further? Perhaps as basic research, but it seems like it could well paint themselves into the corner.
My field is clean energy development. Google is head and shoulders above the other major companies pursuing clean energy. Their understanding, leadership, and execution is incredible. I marvel at how some markets get massive battery systems, when that is the right response (Minnesota); and others get comprehensive customer benefits (Missouri). A lower profile example is their work in Nebraska to replenish groundwater aquifers. They are zeroed in on what really matters (and not just from regulator or customer perspectives; the Nebraska example is physical).
So might they understand how to grow and excel in the cloud and AI? Count me as unsurprised.
Very much agree on diffusion. 💯 The economics of the frontier race never quite added up for me either. Billions for a benchmark lead that holds a few months? The winner-takes-all AGI case answers that, but only if you accept the premise. I am not sure I like the strategic vision behind it, though.
And Ding rightly points the other way: dominating the leading sector rarely decides who wins. And if diffusion is the deciding variable, there is a comparison here beyond Google vs. OpenAI. It's the US vs. China.
Because how is the West doing on diffusion? As we keep seeing across the big consultancy studies: still no material bottom-line impact. And where there is impact, it mostly comes from higher efficiency, optimizing existing processes. AI diffusion in the West is all about saving costs and much less about achieving real, new growth.
In China it runs the other way round: less bottom line, more top line. AI is already production infrastructure in manufacturing, logistics and commerce. Not pilots, growth engines. Take SHEIN's core model — LATR, large-scale automated test and re-order. Nobody hypes it as AI. But it is why they became the biggest online fashion destination on the globe in roughly a decade.
Which is what makes the Western blind spot dangerous. You cite Ding on Japan: leading the frontier through the 1980s in semiconductors and consumer electronics, and still losing the information revolution to the US - a country worse at making the chips but better at putting computers to work. Now swap in the US and China...
I wrote a longer analysis of this mechanism when the Citrini report came up earlier this year. If anyone is interested in this line of thinking and its implications for Western AI strategy:
Anthropic has pretty consistently chosen NOT to focus on a mass market diffusion strategy akin to Google search, choosing instead to work on the enterprise case and on certain high-value verticals. Open AI aspires to be EVERYTHING, but to your point it's not clear they're making enough diffusion progress quickly enough or will have the capital to drive all the way to that conclusion.
1. As far as I’m aware, we don’t know the specs and efficiency of Google’s flash model serving search. It’s unclear then how close Google are or are not to the flash model frontier. They might already be the leader here. Or they might be subsidizing something inefficient.
2. What does OpenAI need to overcome?
In the last generation of competitive struggle, the following seemed to matter a lot: scale (for serving costs and network effects), hardware-software integration (use one to enter the other), cross-product data complementarities (search clicks or maps / phone data to monetize in ads), and product bundling (to envelope competitors).
Owning the user’s journey was key.
Can the current advantages enjoyed by frontier AI labs land up negating Google’s existing strengths in these areas above? I can imagine it weakening Google’s position vis-a-vie its end users, while reinforcing them for its producers. Why? Well Agentic AI consumes a ton of search, and a ton of maps, *but as an intermediary*. This is closer to producer (API) demand, perhaps. The consumer journey is being disintermediated. This must be a HUGE risk for Google. But for now manifests as higher API demand. And lower ads monetization. Just speculation.
3. Google’s ad network must be an enormous advantage. It is worth noting that OpenAI has had to start an ad network from scratch to avoid living inside Google’s ad network. How’s that working out for it?
The "winner takes all" mentality is almost being forced on the big AI companies given their CAPEX. It is not a given that either will "win" or be profitable. Open-source models are narrowing the performance gap, and their competitive advantage is foreseen by the US wanting to ban their use and importation. (Good luck with that.) There is already plenty of pushback on the siting of data centers for various reasons, and Google can read that public sentiment. Better to not be associated with those compute centers and sully what business they have.
It is also apparent that hyperscaling may not be necessary to achieve high performance, and certainly not appropriating the cultural artifacts for free to populate them for training. Just as electrification initially replaced the huge steam engines in factories with huge electric motors, it became obvious that the better way was to distribute electrical power to the point of use, and use small electric motors. Apple clearly wants to push appropriately-sized AI compute to edge devices, even as small as smartphones, which follows the electric power and motors history. If that is the future for AI, then Google can be the distributor of AI through apps and software running locally.
A further comment on distribution. I notice that Ollama is already set up to charge for model use that requires the cloud to support the model. This plays very much into Google's strength, as they already have Google Pay to muscle in on PayPal and Venmo. Google already has the infrastructure to handle advertising from vendors and buyers, so it should be relatively easy to offer the user options to select a model and pay for usage. (Far more useful than AI overviews in search.) Models will compete on price and performance, and will no doubt be rated by users. Like their success in dominating the online advertising market, Google takes a cut of the usage cost without incurring any risk of changing model preference. Users can inexpensively try different cloud-based models on the same platform, like Ollama, or offer free smaller models as they do with some versions of Gemini, or perhaps run locally. This avoids the user having to choose a platform that the "frontier AI Labs" want to try to become the winner. Selecting the model that works best for the user's application makes more sense, and will create an ecosystem of small domain-specific models and the information around those models, just like the book market.
The relationship between model distribution and payment systems is interesting. I have yet to dedicate time to reviewing the Stripe acquisition of OpenRouter beyond thinking at the time that it was an unexpected outcome for OpenRouter and that the Collison bros are very smart. Your comment helps me frame a better context for what will likely prove to be an important deal for the "intelligence payment rails" or whatever we come to call this ecosystem dynamic. Nice one.
“That sounds like someone who has argued that the bigger prize is being the layer that others’ AI runs on”. This is verbatim AWS’s strategy. Jassy uses the “picks and shovels” analogy for the AI gold rush.
Anthropic and openai ARE in diffusion mode. Their advances are primarily coming from superior harnesses, tooling and runtime modes, while raw intelligence plateaus and the economics of training become less and less viable with each generation. But their market is limited to coding, and they are all competing aggressively to own it. Whereas google's market is every phone and laptop user on earth, and they already own it and only need to keep it.
It's like they're saying if intelligence will be commoditized by capable mid-tier models, then what comes next? Kurian builds the physical infra, Dean&co build the app infra, and Hassabis makes some next-gen bet he's not talking about.
You’re 100% right that history doesn’t travel in a straight line. This includes corporate history. Google’s moves show it wanting to be AT&T, but that’s not a long term plan.
Instead, I suspect Google wants to let the AI boys go into the market and build their own stuff, then they can buy whichever works best, rather than risking leadership by betting on one thing or the other. They’re going to be big enough to do that when the software boys are done.
They probably won't need to buy anything, all they need to do is fast-follow. Since non-competes are unlawful in California, all they have to do is poach a few key researchers from Frontier competitors, who will instantly spill the beans on the latest SOTA techniques.
Google’s advantage is that it doesn’t need to capture every dollar at the model layer. If AI demand expands compute, networking, storage, power, and cloud consumption by orders of magnitude, owning the infrastructure stack becomes an extremely valuable position. The Westinghouse analogy works precisely because widespread adoption can matter more than owning every application built on top.
@Tim O'Reilly Love the Westinghouse framing, it represents the right kind of history to pursue here. I'd humbly suggest it is only part of the story, one which becomes clearer if we widen the lens a little. Google isn't the only place this exact pattern is showing up. Ilya left OpenAI for SSI. LeCun left Meta for AMI Labs. Fei-Fei Li built World Labs. David Silver raised $1.1B for Ineffable within months of leaving DeepMind's RL team. In just eighteen months, three AI leaders left different major frontier labs. They are all making the same bet: the old way of doing things has hit its limit. They believe that what comes next requires a fresh start somewhere else.
The diffusion story and the frontier-loss story might both be true. They do not represent a single strategic choice. Instead, they capture two different challenges hitting Google and the larger AI industry at the same time. Moreover, there's a third piece worth adding: look at what's happening to the compute the incumbents keep, not just the talent that leaves. SpaceX announced $2.32 billion in AI compute revenue for the month. This comes from three deals with Anthropic, Google, and xAI. Interestingly, the infrastructure used was mostly sitting underutilized before these agreements. That's $27.8B annualized, more than SpaceX's entire 2025 revenue, at margins that could beat CoreWeave's 70%. Google's TPU/GCP story is the same shape. DeepSeek, an efficiency-focused company that proved massive compute advantages aren't bulletproof, is now investing $35 billion into a gigawatt data center. ByteDance is making similar moves, building its own chips and expanding its datacenter infrastructure. These giants are no longer just renting space; they are building the hardware to control their own future.
The point is clear at this stage of AI’s evolution: owning compute is where the profit lies, regardless of your company's location or mission. Everyone is building around the frontier struggle because compute is what drives the bottom line, not the theory of discovery. Public or near public capital demands this bottom line profit and monetization focus. Its driving the AI party bus, one no one can afford to stop. Private capital is more patient and risk bearing, so it is absorbing the new frontier risk. Your analysis and @Dylan Patel /SemiAnalysis complement each other. They highlight different aspects of the same larger story.
Fallout from the civil war. Google is creating a new retention vehicle.
As a user, it already seems to me that Gemini displaces Google searches (and potentially, search revenues). Gemini seems pretty mature. How does Google benefit from advancing it much further? Perhaps as basic research, but it seems like it could well paint themselves into the corner.
My field is clean energy development. Google is head and shoulders above the other major companies pursuing clean energy. Their understanding, leadership, and execution is incredible. I marvel at how some markets get massive battery systems, when that is the right response (Minnesota); and others get comprehensive customer benefits (Missouri). A lower profile example is their work in Nebraska to replenish groundwater aquifers. They are zeroed in on what really matters (and not just from regulator or customer perspectives; the Nebraska example is physical).
So might they understand how to grow and excel in the cloud and AI? Count me as unsurprised.
The quality and volume of interest & opinions brings back memories of late 90s excitement.
Or maybe the can see the coming implosion more clearly than others. He who runs away lives to fight another day.
Very much agree on diffusion. 💯 The economics of the frontier race never quite added up for me either. Billions for a benchmark lead that holds a few months? The winner-takes-all AGI case answers that, but only if you accept the premise. I am not sure I like the strategic vision behind it, though.
And Ding rightly points the other way: dominating the leading sector rarely decides who wins. And if diffusion is the deciding variable, there is a comparison here beyond Google vs. OpenAI. It's the US vs. China.
Because how is the West doing on diffusion? As we keep seeing across the big consultancy studies: still no material bottom-line impact. And where there is impact, it mostly comes from higher efficiency, optimizing existing processes. AI diffusion in the West is all about saving costs and much less about achieving real, new growth.
In China it runs the other way round: less bottom line, more top line. AI is already production infrastructure in manufacturing, logistics and commerce. Not pilots, growth engines. Take SHEIN's core model — LATR, large-scale automated test and re-order. Nobody hypes it as AI. But it is why they became the biggest online fashion destination on the globe in roughly a decade.
Which is what makes the Western blind spot dangerous. You cite Ding on Japan: leading the frontier through the 1980s in semiconductors and consumer electronics, and still losing the information revolution to the US - a country worse at making the chips but better at putting computers to work. Now swap in the US and China...
I wrote a longer analysis of this mechanism when the Citrini report came up earlier this year. If anyone is interested in this line of thinking and its implications for Western AI strategy:
https://www.ognibeni.de/content/ending-the-innovation-theater-cycle-why-the-real-ai-danger-isnt-what-you-think
infrastructure diffusion always wins over pure frontier speculation.... physical scale and compute economics dictating the endgame
Loved this analogy!
Btw if you’re into the Westinghouse and Edison story you may enjoy our show on the history: https://www.stepchange.show/grid
Anthropic has pretty consistently chosen NOT to focus on a mass market diffusion strategy akin to Google search, choosing instead to work on the enterprise case and on certain high-value verticals. Open AI aspires to be EVERYTHING, but to your point it's not clear they're making enough diffusion progress quickly enough or will have the capital to drive all the way to that conclusion.
Microsoft's bedfellow. Open AI sounds like Windows, Office, Network, and every other Microsoft product over the past 30 years.