AI and the Big Five – Stratechery by Ben Thompson

The story of 2022 was the emergence of AI, first with picture technology fashions, together with DALL-E, MidJourney, and the open supply Stable Diffusion, and then ChatGPT, the first text-generation mannequin to break through in a major way. It appears clear to me that this can be a new epoch in know-how.

To decide how that epoch may develop, although, it’s helpful to look again 26 years to considered one of the most well-known technique books of all time: Clayton Christensen’s The Innovator’s Dilemma, notably this passage on the completely different sorts of improvements:

Most new applied sciences foster improved product efficiency. I name these sustaining applied sciences. Some sustaining applied sciences will be discontinuous or radical in character, whereas others are of an incremental nature. What all sustaining applied sciences have in frequent is that they enhance the efficiency of established merchandise, alongside the dimensions of efficiency that mainstream prospects in main markets have traditionally valued. Most technological advances in a given trade are sustaining in character…

Disruptive applied sciences convey to a market a really completely different worth proposition than had been accessible beforehand. Generally, disruptive applied sciences underperform established merchandise in mainstream markets. But they produce other options that just a few fringe (and typically new) prospects worth. Products primarily based on disruptive applied sciences are usually cheaper, less complicated, smaller, and, often, extra handy to make use of.

It appears simple to look backwards and decide if an innovation was sustaining or disruptive by how incumbent firms fared after that innovation got here to market: if the innovation was sustaining, then incumbent firms grew to become stronger; if it was disruptive then presumably startups captured most of the worth.

Consider earlier tech epochs:

  • The PC was disruptive to just about all of the present incumbents; these comparatively cheap and low-powered gadgets didn’t have almost the functionality or the revenue margin of mini-computers, a lot much less mainframes. That’s why IBM was completely happy to outsource each the authentic PC’s chip and OS to Intel and Microsoft, respectively, in order that they may get a product out the door and fulfill their company prospects; PCs acquired quicker, although, and it was Intel and Microsoft that dominated as the market dwarfed all the things that got here earlier than.
  • The Internet was virtually solely new market innovation, and thus outlined by fully new firms that, to the extent they disrupted incumbents, did so in industries far faraway from know-how, notably these involving data (i.e. the media). This was the period of Google, Facebook, on-line marketplaces and e-commerce, and so forth. All of those functions ran on PCs powered by Windows and Intel.
  • Cloud computing is arguably a part of the Internet, however I believe it deserves its personal class. It was additionally extraordinarily disruptive: commodity x86 structure swept out devoted server {hardware}, and a complete host of SaaS startups peeled off options from incumbents to construct firms. What is notable is that the core infrastructure for cloud computing was primarily constructed by the winners of earlier epochs: Amazon, Microsoft, and Google. Microsoft is especially notable as a result of the firm additionally transitioned its conventional software program enterprise to a SaaS service, partially as a result of the firm had already transitioned mentioned software program enterprise to a subscription mannequin.
  • Mobile ended up being dominated by two incumbents: Apple and Google. That doesn’t imply it wasn’t disruptive, although: Apple’s new UI paradigm entailed not viewing the telephone as a small PC, a la Microsoft; Google’s new enterprise mannequin paradigm entailed not viewing telephones as a direct revenue middle for working system gross sales, however reasonably as a moat for their advertising business.

What is notable about this historical past is that the supposition I acknowledged above isn’t fairly proper; disruptive improvements do persistently come from new entrants in a market, however these new entrants aren’t essentially startups: a few of the largest winners in earlier tech epochs have been present firms leveraging their present enterprise to maneuver into a brand new area. At the identical time, the different tenets of Christensen’s concept maintain: Microsoft struggled with cellular as a result of it was disruptive, however SaaS was finally sustaining as a result of its enterprise mannequin was already aligned.

Given the success of present firms with new epochs, the most evident place to start out when desirous about the affect of AI is with the huge 5: Apple, Amazon, Facebook, Google, and Microsoft.


I already referenced considered one of the most well-known books about tech technique; considered one of the most well-known essays was Joel Spolsky’s Strategy Letter V, notably this well-known line:

Smart firms attempt to commoditize their merchandise’ enhances.

Spolsky wrote this line in the context of explaining why giant firms would put money into open supply software program:

Debugged code is NOT free, whether or not proprietary or open supply. Even in the event you don’t pay money {dollars} for it, it has alternative price, and it has time price. There is a finite quantity of volunteer programming expertise accessible for open supply work, and every open supply challenge competes with one another open supply challenge for the identical restricted programming useful resource, and solely the sexiest initiatives actually have extra volunteer builders than they will use. To summarize, I’m not very impressed by individuals who attempt to show wild financial issues about free-as-in-beer software program, as a result of they’re simply getting divide-by-zero errors so far as I’m involved.

Open supply isn’t exempt from the legal guidelines of gravity or economics. We noticed this with Eazel, ArsDigita, The Company Formerly Known as VA Linux and lots of different makes an attempt. But one thing remains to be occurring which only a few individuals in the open supply world actually perceive: lots of very giant public firms, with tasks to maximise shareholder worth, are investing some huge cash in supporting open supply software program, normally by paying giant groups of programmers to work on it. And that’s what the precept of enhances explains.

Once once more: demand for a product will increase when the value of its enhances decreases. In common, an organization’s strategic curiosity goes to be to get the value of their enhances as little as attainable. The lowest theoretically sustainable value can be the “commodity value” — the value that arises when you may have a bunch of opponents providing indistinguishable items. So, good firms attempt to commoditize their merchandise’ enhances. If you are able to do this, demand on your product will enhance and it is possible for you to to cost extra and make extra.

Apple invests in open supply applied sciences, most notably the the Darwin kernel for its working methods and the WebPackage browser engine; the latter suits Spolsky’s prescription as making certain that the net works nicely with Apple gadgets makes Apple’s gadgets extra priceless.

Apple’s efforts in AI, in the meantime, have been largely proprietary: conventional machine studying fashions are used for issues like suggestions and photograph identification and voice recognition, however nothing that strikes the needle for Apple’s enterprise in a serious means. Apple did, although, obtain an unbelievable reward from the open supply world: Stable Diffusion.

Stable Diffusion is outstanding not just because it’s open supply, but additionally as a result of the mannequin is surprisingly small: when it was launched it might already run on some shopper graphics playing cards; inside a matter of weeks it had been optimized to the level the place it could run on an iPhone.

Apple, to its immense credit score, has seized this chance, with this announcement from its machine studying group final month:

Today, we’re excited to launch optimizations to Core ML for Stable Diffusion in macOS 13.1 and iOS 16.2, together with code to get began with deploying to Apple Silicon gadgets…

One of the key questions for Stable Diffusion in any app is the place the mannequin is working. There are quite a few explanation why on-device deployment of Stable Diffusion in an app is preferable to a server-based method. First, the privateness of the finish person is protected as a result of any information the person offered as enter to the mannequin stays on the person’s gadget. Second, after preliminary obtain, customers don’t require an web connection to make use of the mannequin. Finally, regionally deploying this mannequin allows builders to scale back or eradicate their server-related prices…

Optimizing Core ML for Stable Diffusion and simplifying mannequin conversion makes it simpler for builders to include this know-how of their apps in a privacy-preserving and economically possible means, whereas getting the greatest efficiency on Apple Silicon. This launch includes a Python package deal for changing Stable Diffusion fashions from PyTorch to Core ML utilizing diffusers and coremltools, in addition to a Swift package deal to deploy the fashions.

It’s essential to notice that this announcement got here in two components: first, Apple optimized the Stable Diffusion mannequin itself (which it might do as a result of it was open supply); second, Apple up to date its working system, which due to Apple’s built-in mannequin, is already tuned to Apple’s personal chips.

Moreover, it appears secure to imagine that that is solely the starting: whereas Apple has been transport its so-called “Neural Engine” by itself chips for years now, that AI-specific {hardware} is tuned to Apple’s personal wants; it appears probably that future Apple chips, if not this yr than in all probability subsequent yr, can be tuned for Stable Diffusion as nicely. Stable Diffusion itself, in the meantime, could possibly be constructed into Apple’s working methods, with simply accessible APIs for any app developer.

This raises the prospect of “ok” picture technology capabilities being successfully built-in to Apple’s gadgets, and thus accessible to any developer with out the must scale up a back-end infrastructure of the kind wanted by the viral hit Lensa. And, by extension, the winners on this world find yourself wanting so much like the winners in the App Store period: Apple wins as a result of its integration and chip benefit are put to make use of to ship differentiated apps, whereas small impartial app makers have the APIs and distribution channel to construct new companies.

The losers, on the different hand, can be centralized picture technology companies like Dall-E or MidJourney, and the cloud suppliers that undergird them (and, up to now, undergird the aforementioned Stable Diffusion apps like Lensa). Stable Diffusion on Apple gadgets gained’t take over the total market, to make certain — Dall-E and MidJourney are each “higher” than Stable Diffusion, at the very least in my estimation, and there may be in fact a giant world exterior of Apple gadgets, however built-in native capabilities will have an effect on the final addressable marketplace for each centralized companies and centralized compute.


Amazon, like Apple, makes use of machine studying throughout its functions; the direct shopper use instances for issues like picture and textual content technology, although, appear much less apparent. What is already essential is AWS, which sells entry to GPUs in the cloud.

Some of that is used for coaching, together with Stable Diffusion, which according to the founder and CEO of Stability AI Emad Mostaque used 256 Nvidia A100s for 150,000 hours for a market-rate price of $600,000 (which is surprisingly low!). The bigger use case, although, is inference, i.e. the precise utility of the mannequin to provide photographs (or textual content, in the case of ChatGPT). Every time you generate a picture in MidJourney, or an avatar in Lensa, inference is being run on a GPU in the cloud.

Amazon’s prospects on this area will rely upon quite a few elements. First, and most evident, is simply how helpful these merchandise find yourself being in the actual world. Beyond that, although, Apple’s progress in constructing native technology methods might have a big affect. Amazon, although, is a chip maker in its personal proper: whereas most of its efforts up to now have been centered on its Graviton CPUs, the firm might construct devoted {hardware} of its personal for fashions like Stable Diffusion and compete on value. Still, AWS is hedging its bets: the cloud service is a serious companion in terms of Nvidia’s choices as nicely.

The huge short-term query for Amazon can be in gauging demand: not having sufficient GPUs can be leaving cash on the desk; shopping for too many who sit idle, although, can be a serious price for a corporation attempting to restrict them. At the identical time, it wouldn’t be the worst error to make: considered one of the challenges with AI is the indisputable fact that inference prices cash; in different phrases, making one thing with AI has marginal prices.

This situation of marginal prices is, I believe, an under-appreciated problem when it comes to growing compelling AI merchandise. While cloud companies have at all times had prices, the discrete nature of AI technology could make it difficult to fund the kind of iteration needed to realize product-market match; I don’t suppose it’s an accident that ChatGPT, the largest breakout product to-date, was each free to finish customers and offered by an organization in OpenAI that each constructed its personal mannequin and has a sweetheart deal from Microsoft for compute capability. If AWS needed to promote GPUs for affordable that would spur extra use in the future.

That famous, these prices ought to come down over time: fashions will turn into extra environment friendly at the same time as chips turn into quicker and extra environment friendly in their very own proper, and there ought to be returns to scale for cloud companies as soon as there are adequate merchandise in the market maximizing utilization of their investments. Still, it’s an open query as to how a lot full stack integration will make a distinction, along with the aforementioned chance of working inference regionally.


I already detailed in Meta Myths why I believe that AI is an enormous alternative for Meta and price the large capital expenditures the firm is making:

Meta has large information facilities, however these information facilities are primarily about CPU compute, which is what is required to energy Meta’s companies. CPU compute can also be what was essential to drive Meta’s deterministic advert mannequin, and the algorithms it used to advocate content material out of your community.

The long-term resolution to ATT, although, is to construct probabilistic fashions that not solely determine who ought to be focused (which, to be honest, Meta was already utilizing machine studying for), but additionally understanding which advertisements transformed and which didn’t. These probabilistic fashions can be constructed by large fleets of GPUs, which, in the case of Nvidia’s A100 playing cards, price in the 5 figures; which will have been too dear in a world the place deterministic advertisements labored higher anyhow, however Meta isn’t in that world any longer, and it might be silly to not put money into higher concentrating on and measurement.

Moreover, the identical method can be important to Reels’ continued progress: it’s massively tougher to advocate content material from throughout the total community than solely from your pals and household, notably as a result of Meta plans to advocate not simply video but additionally media of every type, and intersperse it with content material you care about. Here too AI fashions can be the key, and the gear to construct these fashions prices some huge cash.

In the future, although, this funding ought to repay. First, there are the advantages to higher concentrating on and higher suggestions I simply described, which ought to restart income progress. Second, as soon as these AI information facilities are constructed out the price to take care of and improve them ought to be considerably lower than the preliminary price of constructing them the first time. Third, this large funding is one no different firm could make, apart from Google (and, not coincidentally, Google’s capital expenditures are set to rise as nicely).

That final level is maybe the most essential: ATT harm Meta greater than every other firm, as a result of it already had by far the largest and most finely-tuned advert enterprise, however in the future it ought to deepen Meta’s moat. This degree of funding merely isn’t viable for a corporation like Snap or Twitter or any of the different also-rans in digital promoting (even past the indisputable fact that Snap depends on cloud suppliers as a substitute of its personal information facilities); once you mix the indisputable fact that Meta’s advert concentrating on will probably begin to draw back from the area (exterior of Google), with the large enhance in stock that comes from Reels (which reduces costs), it will likely be a marvel why any advertiser would hassle going wherever else.

An essential think about making Meta’s AI work isn’t merely constructing the base mannequin but additionally tuning it to particular person customers on an ongoing foundation; that’s what will take such a lot of capability and it will likely be important for Meta to determine how to do that customization cost-effectively. Here, although, it helps that Meta’s providing will in all probability be more and more built-in: whereas the firm could have committed to Qualcomm for chips for its VR headsets, Meta continues to develop its personal server chips; the firm has additionally released tools to summary away Nvidia and AMD chips for its workloads, nevertheless it appears probably the firm is working by itself AI chips as nicely.

What can be fascinating to see is how issues like picture and textual content technology affect Meta in the future: Sam Lessin has posited that the end-game for algorithmic timelines is AI content material; I’ve made the identical argument when it comes to the Metaverse. In different phrases, whereas Meta is investing in AI to present customized suggestions, that concept, mixed with 2022’s breakthroughs, is customized content material, delivered by means of Meta’s channels.

For now it will likely be fascinating to see how Meta’s promoting instruments develop: the total strategy of each producing and A/B testing copy and photographs will be completed by AI, and no firm is best than Meta at making these kind of capabilities accessible at scale. Keep in thoughts that Meta’s promoting is primarily about the high of the funnel: the objective is to catch customers’ eyes for a services or products or app they didn’t know beforehand existed; because of this there can be lots of misses — the overwhelming majority of advertisements don’t convert — however that additionally means there may be lots of latitude for experimentation and iteration. This appears very nicely suited to AI: sure, technology could have marginal prices, however these marginal prices are drastically decrease than a human.


The Innovator’s Dilemma was printed in 1997; that was the year that Eastman Kodak’s inventory reached its highest value of $94.25, and for seemingly good purpose: Kodak, when it comes to know-how, was completely positioned. Not solely did the firm dominate the present know-how of movie, it had additionally invented the subsequent wave: the digital digital camera.

The downside got here all the way down to enterprise mannequin: Kodak made some huge cash with superb margins offering silver halide movie; digital cameras, on the different hand, had been digital, which suggests they didn’t want movie in any respect. Kodak’s administration was thus very incentivized to persuade themselves that digital cameras would solely ever be for amateurs, and solely once they grew to become drastically cheaper, which would definitely take a really very long time.

In truth, Kodak’s administration was proper: it took over 25 years from the time of the digital digital camera’s invention for digital digital camera gross sales to surpass movie digital camera gross sales; it took longer nonetheless for digital cameras for use in skilled functions. Kodak made some huge cash in the meantime, and paid out billions of {dollars} in dividends. And, whereas the firm went bankrupt in 2012, that was as a result of customers had entry to higher merchandise: first digital cameras, and finally, telephones with cameras in-built.

The concept that this can be a completely happy ending is, to make certain, a contrarian view: most view Kodak as a failure, as a result of we anticipate firms to stay ceaselessly. In this view Kodak is a cautionary story of how an progressive firm can enable its enterprise mannequin to steer it to its eventual doom, even when mentioned doom was the results of customers getting one thing higher.

And thus we arrive at Google and AI. Google invented the transformer, the key know-how undergirding the newest AI fashions. Google is rumored to have a dialog chat product that’s far superior to ChatGPT. Google claims that its picture technology capabilities are higher than Dall-E or anybody else on the market. And but, these claims are simply that: claims, as a result of there aren’t any precise merchandise on the market.

This isn’t a shock: Google has lengthy been a pacesetter in utilizing machine studying to make its search and different consumer-facing merchandise higher (and has provided that know-how as a service by means of Google Cloud). Search, although, has at all times relied on people as the final arbiter: Google will present hyperlinks, however it’s the person that decides which one is the appropriate one by clicking on it. This prolonged to advertisements: Google’s providing was revolutionary as a result of as a substitute of charging advertisers for impressions — the worth of which was very troublesome to determine, notably 20 years in the past — it charged for clicks; the very individuals the advertisers had been attempting to achieve would resolve if their advert was ok.

I wrote about the conundrum this offered for Google’s enterprise in a world of AI seven years in the past in Google and the Limits of Strategy:

In yesterday’s keynote, Google CEO Sundar Pichai, after a recounting of tech historical past that emphasised the PC-Web-Mobile epochs I described in late 2014, declared that we’re transferring from a mobile-first world to an AI-first one; that was the context for the introduction of the Google Assistant.

It was a yr previous to the aforementioned iOS 6 that Apple first launched the concept of an assistant in the guise of Siri; for the first time you would (theoretically) compute by voice. It didn’t work very nicely at first (arguably it nonetheless doesn’t), however the implications for computing typically and Google particularly had been profound: voice interplay each expanded the place computing could possibly be completed, from conditions by which you would commit your eyes and arms to your gadget to successfully in all places, even because it constrained what you would do. An assistant must be much more proactive than, for instance, a search outcomes web page; it’s not sufficient to current attainable solutions: reasonably, an assistant wants to present the proper reply.

This is a welcome shift for Google the know-how; from the starting the search engine has included an “I’m Feeling Lucky” button, so assured was Google founder Larry Page that the search engine might ship you the actual consequence you wished, and whereas yesterday’s Google Assistant demos had been canned, the outcomes, notably when it got here to contextual consciousness, had been much more spectacular than the different assistants on the market. More broadly, few dispute that Google is a transparent chief in terms of the synthetic intelligence and machine studying that underlie their assistant.

A enterprise, although, is about greater than know-how, and Google has two vital shortcomings in terms of assistants specifically. First, as I defined after this yr’s Google I/O, the firm has a go-to-market gap: assistants are solely helpful if they’re accessible, which in the case of tons of of thousands and thousands of iOS customers means downloading and utilizing a separate app (or constructing the kind of expertise that, like Facebook, customers will willingly spend in depth quantities of time in).

Secondly, although, Google has a business-model downside: the “I’m Feeling Lucky Button” assured that the search in query wouldn’t make Google any cash. After all, if a person doesn’t have to select from search outcomes, mentioned person additionally doesn’t have the alternative to click on an advert, thus selecting the winner of the competitors Google created between its advertisers for person consideration. Google Assistant has the very same downside: the place do the advertisements go?

That Article assumed that Google Assistant was going for use to distinguish Google telephones as an unique providing; that ended up being mistaken, however the underlying evaluation stays legitimate. Over the previous seven years Google’s major enterprise mannequin innovation has been to cram ever extra advertisements into Search, a very efficient tactic on cellular. And, to be honest, the kind of searches the place Google makes the most cash — journey, insurance coverage, and so forth. — might not be well-suited for chat interfaces anyhow.

That, although, ought solely enhance the concern for Google’s administration that generative AI could, in the particular context of search, signify a disruptive innovation as a substitute of a sustaining one. Disruptive innovation is, at the very least in the starting, not so good as what already exists; that’s why it’s simply dismissed by managers who can keep away from desirous about the enterprise mannequin challenges by (appropriately!) telling themselves that their present product is best. The downside, in fact, is that the disruptive product will get higher, at the same time as the incumbent’s product turns into ever extra bloated and exhausting to make use of — and that actually sounds so much like Google Search’s present trajectory.

I’m not calling the high for Google; I did that previously and was hilariously wrong. Being mistaken, although, is as a rule a matter of timing: sure, Google has its cloud and YouTube’s dominance solely appears to be rising, however the define of Search’s peak appears clear even when it throws off money and income for years.


Microsoft, in the meantime, appears the greatest positioned of all. Like AWS it has a cloud service that sells GPU; it is usually the unique cloud supplier for OpenAI. Yes, that’s incredibly expensive, however provided that OpenAI seems to have the inside observe to being the AI epoch’s addition to this listing of high tech firms, that implies that Microsoft is investing in the infrastructure of that epoch.

Bing, in the meantime, is like the Mac on the eve of the iPhone: sure it contributes a good bit of income, however a fraction of the dominant participant, and a comparatively immaterial quantity in the context of Microsoft as a complete. If incorporating ChatGPT-like results into Bing dangers the enterprise mannequin for the alternative to realize large market share, that could be a guess nicely price making.

The latest report from The Information, in the meantime, is that GPT is finally coming to Microsoft’s productiveness apps. The trick can be to mimic the success of AI-coding software GitHub Copilot (which is constructed on GPT), which discovered how one can be a assist as a substitute of a nuisance (i.e. don’t be Clippy!).

What is essential is that including on new performance — maybe for a payment — suits completely with Microsoft’s subscription enterprise mannequin. It is notable that the firm as soon as considered a poster little one for victims of disruption will, in the full recounting, not simply be born of disruption, however be well-placed to achieve better heights due to it.

There is a lot extra to jot down about AI’s potential affect, however this Article is already loads lengthy. OpenAI is clearly the most fascinating from a brand new firm perspective: it’s attainable that OpenAI turns into the platform on which all different AI firms are constructed, which might finally imply the financial worth of AI exterior of OpenAI could also be pretty modest; that is additionally the bull case for Google, as they might be the most well-placed to be the Microsoft to OpenAI’s AWS.

There is one other chance the place open supply fashions proliferate in the textual content technology area along with picture technology. In this world AI turns into a commodity: that is in all probability the most impactful end result for the world however, paradoxically, the most muted when it comes to financial affect for particular person firms (I believe the largest alternatives can be in industries the place accuracy is important: incumbents will subsequently underinvest in AI, a la Kodak under-investing in digital, forgetting that know-how will get higher).

Indeed, the largest winners could also be Nvidia and TSMC. Nvidia’s funding in the CUDA ecosystem means the firm doesn’t merely have the greatest AI chips, however the greatest AI ecosystem, and the firm is investing in scaling that ecosystem up. That, although, has and will proceed to spur competitors, notably when it comes to inner chip efforts like Google’s TPU; everybody, although, will make their chips at TSMC, at the very least for the foreseeable future.

The largest affect of all although, although, might be off our radar fully. Just earlier than the break Nat Friedman told me in a Stratechery Interview about Riffusion, which makes use of Stable Diffusion to generate music from textual content through visible sonograms, which makes me marvel what else is feasible when photographs are actually a commodity. Right now text is the universal interface, as a result of textual content has been the basis of data switch since the invention of writing; people, although, are visible creatures, and the availability of AI for each the creation and interpretation of photographs might essentially remodel what it means to convey data in methods which are unimaginable to foretell.

For now, our predictions have to be way more time-constrained, and modest. This could also be the starting of the AI epoch, however even in tech, epochs take a decade or longer to rework all the things round them.

Related Posts